Guidance on the Use of Models and
Other Analyses for Demonstrating
Attainment of Air Quality Goals for
Ozone, PM2 5, and Regional Haze

-------
                                EPA -454/B-07-002
                                April 2007
Guidance on the Use of Models and Other
Analyses for Demonstrating Attainment of
 Air Quality Goals for Ozone, PM2 5, and
               Regional Haze
        U.S. Environmental Protection Agency
      Office of Air Quality Planning and Standards
           Air Quality Analysis Division
            Air Quality Modeling Group
       Research Triangle Park, North Carolina

-------
                           ACKNOWLEDGMENTS

      We would like to acknowledge contributions from members of an external review group
and U.S. EPA Regional Office modeling staffs in providing detailed comments and suggestions
regarding the final version of this guidance.  In particular, we would like to thank staff members
of the Lake Michigan Air Directors Consortium (LADCO), Carolina Environmental Program
(CEP), South Coast Air Quality Management District (SCAQMD), California Air Resources
Board (CARB), Texas Commission on Environmental Quality (TCEQ), North Carolina Division
of Air Quality (NCDAQ), New York State Department of Environmental Conservation
(NYSDEC), Georgia Environmental Protection Divsion (GAEPD) and Computer Sciences
Corporation (CSC) for testing our ideas for a modeled attainment test and sharing the results
with us.

      We would also like to acknowledge the contributions and accomplishments of Ned
Meyer. Ned wrote the original drafts of the ozone and PM2 5 modeling guidance documents. He
also developed the relative attainment tests and put his vision on paper.  The final version of this
guidance is shaped by Ned's words and thoughts.
                                         in

-------
ACKNOWLEDGMENTS	iii

FOREWORD	ix

1.0    Introduction	1
      1.1   What Is The Purpose Of This Document?	2
      1.2    Does The Guidance In This Document Apply To Me?	2
      1.3   How Does The Perceived Nature Of Ozone, PM, and Regional Haze Affect
            My Attainment Demonstration?	3
            1.3.1  Ozone, PM, and Regional Haze 	3
            1.3.2  How is Measuring and Modeling Particulate Matter Different (and
                  Often More Complicated) than Ozone?  	5
      1.4   What Topics Are Covered In This Guidance? 	11

Part I.      How Do I Use Results Of Models And Other Analyses To Help
            Demonstrate Attainment?	14

2.0    What Is A Modeled Attainment Demonstration?—An Overview	15
      2.1   What Is The Recommended Modeled Attainment Test for Ozone and PM2 5?-
            An Overview	15
      2.2   What Is The Recommended Modeling Assessment For Regional Haze?- An
            Overview	16
      2.3   What Does  A Recommended Supplemental Analysis/Weight Of Evidence
            Determination Consist Of? —An Overview
             	17
      2.4   Why Should A Model Be Used In A "Relative" Sense And Why May
            Corroboratory Analyses Be Used In A Weight Of Evidence Determination?18

3.0    What Is The Recommended Modeled Attainment Test?	20
      3.1   Calculating site-specific baseline concentrations	21
      3.2   Identifying surface grid cells near a monitoring site	24
      3.3   Choosing model predictions to calculate a relative response factor (RRF)j
            near a monitor	26
      3.4   Estimating  design values at unmonitored locations: what is an unmonitored
            area analysis and why is it needed?  	29
            3.4.1  Why does the unmonitored area analysis need to use both ambient
                  data and model output?	29
            3.4.2  Implementation of Gradient adjusted Spatial Fields	30
            3.4.3  Using the Results of the Unmonitored Area Analysis 	32
      3.5   Which base year emissions inventory should be projected to the future for
            the purpose of calculating RRFs?  	33
      3.6   Choosing a year to project future emissions	34
                                       IV

-------
4.0    Ozone Attainment Test	37
      4.1    Limiting modeled 8-hour daily maxima chosen to calculate RRF	37
      4.2    How Do I Apply The Recommended 8-Hour Ozone Modeled Attainment Test?
             	40

5.0    What Are The Modeled Attainment Tests For The Two PM2 5 NAAQS? 	43
      5.1    What Is The Recommended Modeled Attainment Test For The Annual
                        NAAQS?	43
            5.1.1  What Ambient PM2 5 Data is Used in the Attainment Test? 	44
            5.1.2  FRM Monitors that Don't Have Speciation Data	45
            5.1.3   New Species Calculations and Definitions 	46
            5.1.4  Recommended Treatment of Species Data 	47
      5.2    What Is The Recommended Modeled Attainment Test For The 24-Hour NAAQS?
             	56
      5.3    Special Considerations for Primary PM2 5	63
            5.3.1  Analysis of Primary PM25 Impacts at Monitors	64
      6.0    What Is The Recommended Modeling Analysis for Regional Haze?	75
      6.1    Regional Haze Rule Background	75
      6.2    Uniform Rate of Progress	76
      6.3    How are Regional Haze Trends Measured?  	77
      6.4    How Do I Apply A Model To Calculate Changes in Visibility ?	83
      6.5    How Do I Select Appropriate Inputs For The Uniform Rate of Progress Analysis?
             	93

7.0    How Can Additional Analyses Can Be Used to Support the Attainment
      Demonstration? 	98
      7.1    What Types of Additional Analyses Should Be Completed as Part of the
            Attainment Demonstration?	99
      7.2    What Is Entailed In A Weight Of Evidence Determination? 	105
      7.3    What Role Should Weight Of Evidence Play In Visibility-Related Analyses?
             	107

8.0    What Additional Analyses and Data Collection Can Be Completed to Assess
      Progress towards Attainment	110

9.0 What Documentation Do I Need To Support My Attainment Demonstration?
       	113

-------
Part II.     How Should I Apply Air Quality Models To Produce Results
            Needed To Help Demonstrate Attainment?  	120

10.0   How Do I Apply Air Quality Models?-- An Overview	121

11.0   How Do I Get Started?- A "Conceptual Description"	126
      11.1   What Is A "Conceptual Description"?	126
            11.1.1 8-Hour Ozone NAAQS  	126
            11.1.2 Annual PM25 NAAQS	128
            11.1.3 Example Regional Haze Application	131

12.0   What Does A Modeling/Analysis Protocol Do, And What Does Developing One
      Entail?	133
      12.1   What Is The Protocol's Function?	133
      12.2   What Subjects Should Be Addressed In The Protocol?	133

13.0   What Should I Consider In Choosing An Air Quality Model?	135
      13.1   Types of Air Quality Models	135
      13.2   What Prerequisites Should An Air Quality Model Meet To Qualify For Use
            In An Attainment Demonstration?  	136
      13.3   What Factors Affect My Choice of A Model For A Specific Application?   137
      13.4   What Are Some Examples Of Air Quality Models Which May Be Considered?
             	138
      14.0  How are the Meteorological Time Periods (Episodes) Selected? 	140
      14.1  What Are The Most Important Criteria For Choosing Episodes?	141
            14.1.1 8-Hour Ozone NAAQS  	142
            14.1.2 Recommendations for the 24-Hour PM2 5 NAAQS	147
            14.1.3 Annual PM25	148
            14.1.4 Regional Haze	149
      14.2  What Additional, Secondary Criteria May Be Useful For Selecting Time
            Periods  to Model?  	150

15.0   What Should Be Considered When Selecting The  Size And Horizontal/Vertical
      Resolution Of The Modeling Domain?  	152
      15.1  How is the Size of the Modeling Domain Chosen?	152
      15.2  How are the Initial and Boundary Conditions Specified?	153
      15.3  What Horizontal Grid Cell Size Is Necessary?	155
            15.3.1 Horizontal Resolution Recommendations for Ozone	157
            15.3.2 Horizontal Resolution Recommendations for the PM2 5 NAAQS .. 157
            15.3.3 Horizontal Resolution Recommendations for Regional Haze	158
      15.4  How Should the Vertical Layers Be Selected?  	158
                                       VI

-------
16.0  How are the Meteorological Inputs Prepared for Air Quality Modeling?	160
      16.1  What Issues are Involved in the Generation and Acquisition of
            Meteorological Modeling Data?	160
      16.2 How Should the Meteorological Modeling Analysis be Configured?	161
      16.3 How Should the Performance of the Meteorological Modeling Be Evaluatedl63

17.0  How Are the Emission Inputs Developed? 	168
      17.1  What Emission Inventory Data are Needed to Support Air Quality ModelsOSS
      17.2  Can The National Emission Inventory Be Used As a Starting Point? 	172
      17.3  What Other Data are Needed to Support Emissions Modeling?	172
      17.4  How Are Inventory Data Converted Into Air Quality Model Input?	174
      17.5  Are there Other Emissions Modeling Issues?	177
      17.6  How Are Emissions Estimated for Future Years?	179
            17.6.1 Identify and prioritize key sectors and regions for projections .... 180
            17.6.2 Collect available data and models for each sector needing projections
                    	180
            17.6.3 Evaluate and refine data for projections in modeling region 	187
            17.6.4 Create future-year inventories and air quality model inputs	188
      17.7  Emissions Preparation for a Local Area Analysis 	188

18.0  What are the Procedures for Evaluating Model Performance and What is the Role
      of Diagnostic Analyses?  	190
      18.1  How Do I Assess Model Performance And Make Use of Operational and
            Diagnostic Evaluations?  	190
      18.2  What Pollutants Are of Interest for Model Performance?  	191
      18.3  What Are the PM2 5 Measurement versus Model Concentration Issues?  . 192
      18.4  How Should the Operational Evaluation of Performance Be Completed?  197
            18.4.1 Ozone Operational Evaluation	198
            18.4.2 PM/RH Operational Evaluation	202
      18.5  How Do I Assess Model Performance and What Types of Analyses Can be
            Done to Evaluate the Accuracy of the Model Response: Diagnostic
            Evaluations?	208
            18.5.1 Observational models  	209
            18.5.2 Probing Tools	210
            18.5.3 Receptor Models (Multi-variate Source Apportionment)	213
            18.5.4 Retrospective Analyses 	214
            18.5.5 Alternative Base Cases (Sensitivity  tests)  	215
      18.6  How Should the Results of the Model Evaluation be Assessed (Interpreted)?
              	216

REFERENCES	218

Glossary  	239

                                        vii

-------
APPENDIX A	240

Appendix B: Summary of recent model performance evaluations conducted by other
      modeling groups	243
                                      Vlll

-------
                                    FOREWORD
       The purpose of this document is to provide guidance to EPA Regional, State, and Tribal
air quality management authorities and the general public, on how to prepare 8-hour ozone and
PM25 attainment demonstrations and regional haze uniform rate of progress analyses using air
quality models and other relevant technical analyses.  This guidance is designed to implement
national policy on these issues. This document does not substitute for any Clean Air Act
provision or EPA regulation, nor is it a regulation itself. Thus, it does not impose binding,
enforceable requirements on any party, nor does it assure that EPA will approve all instances of
its application. The guidance may not apply to a particular situation, depending upon the
circumstances. The EPA and State decision makers retain the discretion to adopt approaches on
a case-by-case basis that differ from this guidance where appropriate. Any decisions by EPA
regarding a particular State Implementation Plan (SIP) demonstration will only be made based
on the statute and regulations, and will  only be made following notice and opportunity for public
review and comment.  Therefore, interested parties will be able to raise questions and objections
about the contents of this guidance and the appropriateness of its application for any particular
situation.

       This guidance is a living document and may be revised periodically. Updates, revisions,
and additional documentation will be provided at http://www.epa.gov/ttn/scram/. Any mention
of trade names or commercial products in this document is not intended to constitute
endorsement or recommendation for use. Users are cautioned not to regard statements
recommending the use of certain procedures  or defaults as either precluding other procedures or
information, or providing guarantees that using these procedures or defaults will result in actions
that are fully approvable. As noted above, EPA cannot assure that actions based upon this
guidance will be fully approvable in all instances, and all final actions will only be taken
following notice and opportunity for public comment.  The EPA welcomes public comments on
this document and will consider those comments in any future revisions of this guidance
document, providing such approaches comply with all applicable statutory and regulatory
requirements.
                                           IX

-------
1.0    Introduction

       This document describes how to estimate if an emissions control strategy will lead to
attainment of annual and 24-hour national ambient air quality standards (NAAQS) for particles
smaller than 2.5 |im in diameter (PM25) and the 8-hour NAAQS for ozone.  We also describe
how to use modeled and monitored data to estimate the visibility improvement in Class I areas
(e.g., national parks, wilderness areas) as part of a uniform rate of progress analysis1.

       The document describes how to apply air quality models to generate the predictions later
used to evaluate attainment and/or uniform rate of progress assessments.  Modeling to show
attainment of the NAAQS primarily applies to nonattainment areas2 for which modeling is
needed, or desired. Modeling to assess uniform rate of progress for regional haze applies to all
States3

       The guidance consists of two major parts. Part I addresses the question, "how should I
use the results of models and other analyses to help demonstrate attainment and/or assess
uniform rate of progress?" We begin by describing a modeled attainment test for the 8-hour
ozone NAAQS, the annual PM2 5 NAAQS, and the 24-hour PM2 5 NAAQS. We also recommend
a modeled test to assess uniform rate of progress goals to reduce regional haze. We explain
what is meant by a modeled attainment demonstration, a modeled attainment test, and a weight
of evidence determination. We also identify additional data which, if available, can enhance the
credibility of model results. Part I concludes by identifying what documentation States/Tribes
should include as part of an attainment demonstration.

       Part II of the guidance describes how to apply air quality models.  The recommended
procedure for applying a model has nine steps.  The results of this process are then used to apply
the modeled attainment test to support an attainment demonstration, as described in Part I of the
guidance.
       Modeling is not used to determine reasonable progress goals and does not determine
whether reasonable progress has been met.  Modeling is one part of the reasonable progress
analysis. The modeling results are used to determine if future year visibility at Class I areas are
estimated to be on a (glide)path towards reaching natural background. This is called a uniform
rate of progress analysis or "glidepath" analysis. The uniform rate of progress analysis is
described in more detail in section 6.

       2While this guidance document is primarily directed at modeling applications in
nonattainment areas, it may also be useful as a guide for modeling in maintenance areas or to
support other rules or sections of the Clean Air Act.

       3The Regional Haze rule assumes that all 50 States either contain a Class I area or impact
visibility within a Class I area.  Therefore, a regional haze State  Implementation Plan (SIP) is
required for all States (40 CFR 50.308(b)).

-------
       1. Develop a conceptual description of the problem to be addressed.
       2. Develop a modeling/analysis protocol.
       3. Select an appropriate model to support the demonstration.
       4. Select appropriate meteorological time periods to model.
       5. Choose an appropriate area to model with appropriate horizontal/vertical resolution
       and establish the initial and boundary conditions that are suitable for the application.
       6. Generate meteorological inputs to the air quality model.
       7. Generate emissions inputs to the air quality model.
       8. Run the air quality model with basecase emissions and evaluate the performance.
       Perform diagnostic tests to improve the model, as necessary.
       9. Perform future year modeling (including  additional control strategies, if necessary) and
       apply the attainment test.

       Model applications require a substantial effort. States/Tribes should work closely with
the appropriate U.S. EPA Regional  Office(s) in executing each step.  This will increase the
likelihood of approval of the demonstration at the end of the process.

       1.1    What Is The Purpose Of This Document?

       This document has two purposes.  The first  is to explain how to interpret whether results
of modeling and other analyses support a conclusion that attainment of the ozone and/or PM25
NAAQS, and/or uniform rate of progress for regional haze will occur by the appropriate date for
an area. The second purpose is to describe how to apply an air quality model to produce results
needed to support an attainment demonstration or a uniform rate of progress analysis.

       The guidance herein should be viewed as recommendations rather than requirements.
Although this guidance attempts to  address issues that may arise in attainment demonstrations,
situations which we have failed to anticipate may occur.  These should  be resolved on a case by
case basis in concert with the appropriate U.S. EPA Regional Office.

       1.2     Does  The Guidance In This Document Apply To Me?

       This guidance applies to all  locations required to submit a State Implementation Plan
(SIP), or Tribal  Implementation Plan (TIP) revision with an attainment demonstration designed
to achieve attainment of the ozone or PM25 NAAQS. The guidance also applies to SIPs
developed to address regional haze  rule requirements. Areas required to submit an attainment
demonstration and/or a reasonable progress demonstration are encouraged to follow the
procedures described in this document. Details on when a State is  required to submit a modeled
attainment or reasonable progress demonstration can be found in the regional haze rule4, the 8-
       440CFR 50.308

-------
hour ozone implementation rule5 and the PM2 5 implementation rule6.

       Implementation plan revisions for ozone and PM2 5 are due three years from the effective
designation date after an area is designated "nonattainment" (e.g., June 15, 2007 for areas whose
effective designation dates are June 15, 2004).  Regional haze SIPs (for the first planning period)
are due no later than December 17, 2007 (40 CFR 50.308(b)) Attainment and uniform rate of
progress analyses supporting these revisions should be completed in time to allow sufficient time
to complete the rulemaking process by the SIP due date.
       1.3    How Does The Perceived Nature Of Ozone, PM, and Regional Haze Affect
             My Attainment Demonstration?

       Guidance for performing attainment demonstrations needs to be consistent with the
perceived nature of the pollutant. In this section, we identify several premises regarding ozone,
PM2 5, and regional haze.  We then describe how the guidance accommodates each.

             1.3.1  Ozone, PM, and Regional Haze

Premise 1. There is uncertainty accompanying model predictions. "Uncertainty" is the
notion that model  estimates will not perfectly predict observed air quality at any given location,
neither at the present time nor in the future. Uncertainty arises for a variety of reasons, for
example, limitations in the model's formulation which may be due to an incomplete
representation in the model of physiochemical processes and/or meteorological and other input
data base limitations, and uncertainty in forecasting future levels of emissions.  States/Tribes
should recognize these limitations when preparing their modeled attainment demonstrations.

       We recommend several  qualitative means for recognizing model limitations and resulting
uncertainties when preparing an attainment demonstration. First, we recommend using models
in a relative sense in concert with observed air quality data (i.e., taking the ratio of future to
present predicted air quality and multiplying it times an "ambient" design value)7. As described
later, we believe this approach should reduce some of the uncertainty attendant with using
absolute model predictions alone. Second, we recommend that a modeling analysis be preceded
by analyses of available air quality, meteorological, and emissions data to gain a qualitative
understanding of an area's nonattainment problem.  Such a description should be used to help
guide a model application and may provide a reality check on the model's predictions.  Third, we
recommend that States/Tribes use several model outputs, as well as other supporting analyses, to
provide corroborative evidence concerning the adequacy of a proposed strategy for meeting the
       540 CFR 50.908

       640 CFR 50.1007
       7  Ambient design values are based on observations made at monitor locations.

                                           3

-------
NAAQS. Modeling results and other supporting analyses can be weighed to determine whether
or not the resulting evidence suggests a proposed control strategy is adequate to meet the
NAAQS. Finally, we identify several activities/analyses which States/Tribes could undertake, if
they so choose, to apply models and corroborative approaches in subsequent reviews and
analyses of a control strategy,  such as mid-course reviews.  These subsequent reviews are useful
for determining whether a SIP is achieving progress as expected.

Premise 2. For many areas,  nested regional/urban model applications will be needed to
support the attainment demonstration.  Available air quality data suggest ozone and PM2 5
concentrations approach levels specified in the NAAQS throughout much of the eastern U.S. and
in large parts of California (U.S. EPA 2004a, U.S.  EPA 2004b).  A number of analyses
(U.S.EPA, 1998b and U.S. EPA, 2005a and 2005b) show that regional ozone and PM transport
can impact areas several hundred miles or more downwind. The regional extent of ozone and PM
transport patterns and distances in some areas will likely necessitate nested regional model
applications.

       This guidance identifies several modeling systems* with nesting capabilities to resolve
meteorological parameters, emissions, chemistry, and transport.  We believe it is not beneficial
to identify any modeling system as the preferred, or "guideline model" for ozone, PM2 59 or
regional haze modeling. States/Tribes may use any appropriate modeling system provided that
the requirements of 40 CFR 51.112 are met.  In this guidance, we provide certain criteria to
assist States/Tribes in justifying the use of such modeling systems.  These criteria apply equally
to U.S.EPA models and alternative air quality model(s). The guidance also provides
recommendations for developing meteorological, air quality and  emissions inputs used in nested
regional model applications, and makes suggestions for quality assuring inputs and evaluating
performance of emissions, meteorological and air quality models.

Premise 3. Resource intensive approaches may often be needed to support an adequate
attainment demonstration. This follows from the regional nature of ozone and PM25
concentrations in the Eastern U.S.  While we believe that existing and future regional reductions
in NOx and SO2 emissions will reduce ozone and PM2 5 over much of the eastern U.S., elevated
regional ozone and PM2 5 concentrations will continue to affect local strategies needed to attain
the NAAQS in the remaining nonattainment areas.
       8A modeling system includes a chemical model, an emissions model and a meteorological
model. Terms, such as this one, which are introduced using italics are defined more fully in a
glossary at the back of this guidance. "Modeling system" and "air quality model" are used
interchangeably. "Air quality model" means "modeling system" in this guidance.

       9In this context, we are not referring to dispersion modeling of primary PM25 sources
subject to New Source Review (NSR) or Prevention of Significant Deterioration (PSD) analyses.
We are only referring to PM25 analyses for SIP attainment demonstrations.

-------
       This guidance recommends using regional modeling domains. Regional modeling
applications require coordination, quality assurance and management of data bases covering
large areas of the country. Resources used to run recommended models for generating
meteorological and emissions inputs  and the air quality model itself can be substantial.
States/Tribes facing the need to develop an attainment demonstration requiring resource
intensive techniques may wish to consider pooling resources in some manner.  Examples might
include delegating responsibilities for certain parts of the analyses to a single State/Tribe which
can "specialize" in that kind of analysis. Another example might be formation of a regional
technical center to perform analyses as directed by its client group of States/Tribes (e.g., multi-
state and tribal organizations such as the Regional Planning Organizations (RPO) 10, LADCO,
and the Ozone Transport Commission (OTC)).

Premise 4. High concentrations of ozone, PM2 5 and regional haze often have a common
origin. Ozone formation and formation of secondary particulates result from several common
reactions and reactants. Secondary particulates are a major part of PM25  Often similar sources
contribute precursors to both ozone and PM2 5.  In some regions of the U.S., high regional ozone
and secondary particulates are observed under common types of meteorological conditions.
Reducing PM2 5 is the principal controllable means for improving regional haze.  Reducing PM2 5
precursors can also lead to reductions in ozone. Models intended to address secondary
particulate matter problems need also to be capable of simulating ozone formation and transport.
U.S. EPA policy is to encourage "integration" of programs to reduce ozone, PM25 and regional
haze to ensure they do not work  at cross purposes and to foster maximum total air quality benefit
for lower costs.

              1.3.2  How is Measuring and Modeling Particulate Matter Different (and
                    Often More Complicated) than Ozone?

       Before modeling the expected PM2 5 benefits from emissions controls, it is important to
understand the unique and complicated aspects of measuring and modeling particulate matter.
For many reasons, PM2 5 and regional haze modeling presents many more difficulties compared
to ozone modeling. In this section, we  identify attributes of PM25 that are applicable to most
attainment or uniform rate of progress demonstrations. There are of course exceptions. As we
discuss in Section 11.0, States need to develop a conceptual description of the PM25 or regional
haze problem in each of their areas subject to a demonstration. If a substantially different picture
emerges from the general one presented in this section, modeling/analysis procedures which
differ from some of those we believe are generally applicable may be warranted.

Premise PM1. Particulate matter is a mixture. Unlike a compound (e.g., ozone) or an
element (e.g., lead), a mixture has components which (a) can behave independently of one
another (e.g., primary vs. secondary components) or (b) are related to one another in a complex
       10EPA provides funding to five regional planning organizations to address regional haze
and related issues,  http://www.epa.gov/air/visibilitv/regional.html

-------
way (e.g., different secondary components).  Thus, if one only considers PM25 as a single entity,
rather than as the sum of its major components, there is a greater risk of choosing an ineffective
control strategy.  This follows, because models may  not perform equally well in predicting major
components of PM25. Nevertheless, balancing errors could (erroneously) indicate good model
performance predicting PM25 If a control strategy focused on reducing a component of PM25
which was overestimated by the model, the subsequently observed impact on PM2 5 could be less
than expected.

       Characteristics of PM25 as a mixture and the  possibility that models perform unevenly in
predicting the major components of the mixture have two important implications for our
guidance. First, the modeling should divide PM25 into a half a dozen or so major components
and note the effects of a strategy on each.  The effect on PM2 5 should be estimated as a sum of
the effects on individual components.  Second, to reduce the effects  of uneven performance and
possible major bias in predicting absolute concentrations of one or more components, models are
best used in a "relative" sense in concert with measured PM25 and estimated composition of the
measured PM2 5 derived from speciated measurements. That is, responses predicted by models
should be applied to observed component concentrations derived from PM2 5 measurements and
composition of PM25 estimated from measurements of ambient species.

Premise PM2. In most parts of the country, "Secondary" PM is a more important part of
PM2 5 than it is of PM10.  Size-differentiated ambient particulate data suggest that mass of
particulate matter follows a bimodal distribution, with one peak (fine mode} reflecting particles
with aerodynamic diameters ~ 0.1-1.0 micrometers arising from nucleation and accumulation
phenomena, and a second (coarse mode) occurring with aerodynamic diameters in the range of
1.0-20 micrometers. As shown in Figure 1.1, derived from Wilson and Suh, (1997), mass of fine
particulate matter (i.e., PM25) attributable to coarse mode particulate matter < 2.5 micrometers is
relatively small.  Mass attributable to fine mode particulate matter with aerodynamic diameters >
2.5 micrometers (i.e., coarse particulate matter) is also relatively small.

       Origins of coarse and fine mode particulate matter are usually quite different. The former
results mostly from physical types of activities (e.g.,  crushing, grinding, resuspension due to
motions, etc.). Nearly all these activities result in particulate matter, with little subsequent
chemical change. We call such emissions of particulate matter "primary emissions", because
they are measured  in more or less the same form in which they are emitted.  In contrast, origins
of fine mode particulate matter are more diverse.  For example, some fine mode particulate
matter is directly emitted to the atmosphere as a result of combustion. Such emissions occur
either directly as  particles or as a result of condensation which occurs very shortly after the
emissions occur.  These are primary emissions, because what is measured in the ambient air is
essentially unchanged (chemically) from what is released. However, many fine mode particles
are the result of physicochemical reactions which occur in the atmosphere among gaseous
precursors or  through absorption or adsorption onto previously existing aerosols.  Such particles
constitute "secondary" particulate matter, because they undergo transformations in the
atmosphere causing the chemical and/or physical nature of what is measured to be different from
what is emitted.

-------
       Because of the size distribution of ambient particulate matter, most measured PM25 is
likely to be fine mode particulate matter.  As a result, it is dominated to a much larger extent
than PM10 by "secondary" particulate matter and primary particulate emissions arising from
combustion. Some of the physicochemical processes leading to secondary particulate matter
formation may take hours or days, as do some of the removal processes. Thus, many of the
sources of measured secondary particulate matter may not be local emitted sources. This implies
that modeling to support attainment demonstrations for PM2 5 (and, as we will discuss later,
regional haze-related applications) will need to cover a very large domain, and will need to
include chemical/physical mechanisms important in formation/removal of secondary particulate
matter. Because several of the processes  are slow and first require thorough mixing with the
environment, spatially detailed treatment  near sources of precursors may not be necessary.
Individual treatment of precursor emissions from relatively nearby large sources of primary
PM2 5 may be needed on a case by case basis.

-------
                                       PM
                                            10
                        PM
                             2.5
 &
  2.5 |im aerodynamic diameter.
Further, certain components of PM25 are more efficient at scattering or absorbing light than
others. Many of the most efficient are secondary particulate species. For example, sulfates
(secondary), nitrates (secondary) and organic (secondary and primary) components scatter light
more efficiently than do primary particles composed of soil material. Light extinction is also
exacerbated by high relative humidity.  Water vapor combines with hygroscopic particulate
matter (e.g., sulfates and nitrates) to greatly increase the light scattering efficiency of these
species. Previously, we noted that secondary particulate matter is likely to comprise an

-------
important fraction of measured PM25. Secondary participate matter will be even more important
as a cause of regional haze.  This follows from the greater efficiency with which these already
important components of PM2 5 scatter light.  This importance can be enhanced further by high
relative humidity, which is especially relevant in the Eastern U.S.

       The discussion in the preceding paragraph suggests that modeling to assess uniform rate
of progress for regional haze will need to address secondary particulate matter.  This, in turn,
means that large modeling domains will be necessary.  Class I areas are generally likely to be far
removed from most sources  of precursors for secondary parti culate matter.  Additionally, the
measure of visibility (deciviews) with which we are most concerned, addresses maximum range
of visibility. This measure reflects an effect which is integrated over a relatively large distance.
Thus, the need for a large domain, the prevalence of secondary particles, the relative remoteness
of Class I areas from most sources of precursors and the visibility measure of greatest interest
suggest that modeling related to regional haze may be done without a fine degree of spatial
resolution.

Premise PM4. Sampling anomalies associated with the Federal Reference Method for PM2 5
need to be considered as part of the model attainment demonstration. ANAAQS needs to
be related to observed health effects.  In order to establish this link most clearly,  the U.S. EPA
has adopted a Federal Reference Method (FRM) for measuring PM2 5 similar to procedures used
in epidemiological studies in which adverse health effects are associated with exposure to PM2 5
(40 CFR Part 50, Appendix L). Since the FRM is included in the definition of the NAAQS, the
modeled attainment test we recommend uses the measured concentrations of FRM PM25 as the
"ground truth".  The FRM sampling protocol is known to be a precise measurement, but it may
not accurately measure PM2 5 in the air. There are many known positive and negative sampling
artifacts associated with the  measurement and analysis protocol.  Among them are positive
artifacts associated with organics and water retained on the Teflon filter and negative artifacts
associated with volatile nitrates and organics (Frank, 2006).

The attainment test recommends using speciated ambient data to estimate PM2 5 components.
These measurements are from the Speciation and Trends Network (STN) for urban sites and the
Interagency Monitoring of PROtected Visual Environments (IMPROVE) network for Class I
areas. The Speciation measurements use different protocols than FRM measurements.
Therefore, the species measured by STN and IMPROVE are often not directly comparable to the
FRM PM2 5.  The attainment test recommends a  set of default procedures to adjust the speciation
measurements to make them more comparable to FRM measurements.

Premise PM5. Spatial gradients for primary  particulate matter may be more pronounced
than those for secondary particulate matter or those for ozone.  As previously noted,
secondary particulate matter and ozone result from an interaction of meteorology and chemistry
taking several hours to days. Because of the time scales and mixing required, sharp spatial
gradients in concentrations of these pollutants are unlikely.  In contrast, primary  particulate
matter is emitted in the form it appears at monitoring sites.  It is likely that concentrations of
primary particulate matter are greatest near major source areas  of primary particulate matter.

-------
       The preceding implies that it may be necessary to estimate concentrations of primary
participate matter using models with finer spatial resolution than is necessary for secondary
particles. Further, there may be several sources or concentrations of sources of primary
particulate matter within an area designated as "nonattainment" for PM2 5. The guidance will
need to address how to evaluate model performance and how to estimate whether attainment of
the NAAQS is likely in such locations.

Premise PM6. Seasonal differences are likely in emissions of PM2 5 and its precursors, as
well as in meteorological conditions affecting source/receptor relationships.  Emissions from
several potentially important sources of PM25, such as residential wood burning, wildfires,
agricultural burning, prescribed burning and biogenic sources have distinctive seasonal patterns.
Further, meteorological factors which may affect PM2 5 or regional haze,  such as relative
humidity, sunlight intensity, mixing heights, precipitation and temperature, have marked
seasonal differences in many parts of the United States.  The annual NAAQS for PM2 5 and the
regional haze rule address a composite of conditions measured over many days. To understand
how such composites respond to changes in emissions, it will be necessary to model a variety  of
days with varying emissions and meteorological conditions. This implies that States will need to
develop base emissions estimates for a full year (or multiple years) or a representative portion of
days may need to be modeled from each season.

Modeling for the 24-hour PM2 5 NAAQS is further complicated by the fact that violations of the
24-hour standard may occur during different times of the year and may be caused by a variety of
different PM2 5 species and precursors.

Premise PM7. Causes of PM2 5 concentrations which violate NAAQS may be diverse.
Modeling approaches needed to address primary vs. secondary particulate matter differ in their
requirements.  Earlier, we noted differing requirements for size and resolution of the modeling
domain. Another difference is the need to consider atmospheric chemistry in the modeling. It is
essential to have some understanding of the nature of an area's PM25 or visibility problem before
modeling begins.  Otherwise, a State runs the risk of selecting inappropriate analysis tools as
well as selecting a strategy which will prove to be ineffective at reducing its problem.  Thus, a
State needs to perform data analysis before using air quality models. This analysis should be
used to develop a conceptual description of the problem at hand. The conceptual description
may then be used to select a broad strategy (e.g., do I focus on reducing primary or secondary
particulate matter or both?)  as well as to help implement a modeling protocol to best address the
nature of the problem and the qualitative strategy which  has been tentatively selected to address
it. The guidance needs to provide States with flexibility in choosing model(s) to address specific
problems.

Premise PM8. Ability of models to predict future concentrations of the components of
PM25 is limited by a variety of factors. Our ability to characterize emissions on a day to day
basis or on a source-specific basis is limited.  Fully characterizing meteorological conditions on
any given day is also problematic. Further, for regulatory models to be tractable, they must
characterize chemical and physical processes by simplifying them in some reasonable manner.

                                           10

-------
In some cases, most notably for secondary organic participate matter, the extent to which current
simplifications are reasonable is uncertain.  These limitations (and others) add to the uncertainty
of the model's ability to accurately predict concentrations of PM25 and its components at a given
time and location.

       The preceding paragraph has several implications for using models to demonstrate future
attainment of a NAAQS for PM25 or assessing uniform rate of progress for regional haze.  It
suggests that we should focus on composite responses of the model averaged over several days
to help circumvent the problem of not knowing all of the details on an individual day.  This
composite response then needs to be related to the form of the air quality goal in some manner.
Limitations in available models and their underlying data bases also suggest that the guidance
should recognize a need for performing other, corroboratory analyses to confirm conclusions
reached with a model.

       1.4    What Topics  Are Covered In This Guidance?

       This guidance addresses two broad topics: Part I, "How do I use results of models and
other analyses to help demonstrate attainment?", and Part II, "How should I apply air quality
models to produce results needed to help demonstrate attainment?". Part I is divided into 8
sections (i.e., Sections 2-9). Part II consists of 9 sections (Sections 10-18).

       Part I ( "How do I use results of models and other analyses to help demonstrate
attainment?") begins in Section 2 with an overview of the procedure for using modeling results
to help demonstrate attainment of the 8-hour ozone NAAQS and the annual and 24-hour
NAAQS for PM2 5.  Section 2 also summarizes the uniform rate of progress modeling analysis.

       Section 3 describes the recommended modeled attainment tests.

       Section 4 describes the attainment test for ozone in more detail.

       Section 5 describes the attainment tests for each of the two NAAQS for PM2 5.in more
detail.
       Section 6 outlines a recommended analysis to assess uniform rate of progress at Class I
areas.

       Section 7 describes how supporting analyses should be performed to  complement the
attainment test and the uniform rate of progress analysis, as well as how it should be used in a
weight of evidence determination.

       Section 8 identifies several data gathering activities and analyses which States/Tribes
could undertake to enhance the credibility of the modeling and corroborative analyses to support
subsequent reviews on progress toward attainment and/or visibility improvements.

       Section 9 identifies the documentation necessary to adequately  describe  the analyses

                                           11

-------
used to demonstrate attainment of the ozone NAAQS.
       Part II  ("How should I apply air quality models to produce results needed to help
demonstrate attainment?") begins in Section 10 with an overview of the topics to be covered.

       Section  11 identifies a series of meteorological, emissions and air quality data analyses
which should be undertaken to develop a qualitative description of an area's nonattainment
problem prior to a model application. As we describe, this qualitative description should be used
to guide the subsequent model application.

       Section  12 describes the purpose, function, and contents of a modeling protocol.

       Section  13 addresses what criteria should be considered in choosing a model to support
the ozone and PM attainment demonstrations and uniform rate of progress analyses. Several
guidelines are identified for accepting the use of a model for this purpose.

       Section  14 provides guidance for selecting suitable time periods (or episodes) to model
for an attainment demonstration or uniform rate of progress analysis.  Topics include a
discussion of the form of the NAAQS and its resulting implications for episode selection.

       Section  15 identifies factors which should be considered in choosing; a model domain,
the horizontal and vertical resolution, and the initial/boundary conditions for an air quality
modeling application.

       Section  16 addresses how to develop and evaluate meteorological inputs for use in a
modeling exercise supporting an attainment demonstration or uniform rate of progress analysis..

       Section  17 discusses how to develop appropriate emissions estimates for use in the
selected air quality model.

       Section  18 outlines the structure of model  performance evaluations and discusses the use
of diagnostic analyses.

       The guidance concludes with references and a glossary of important terms which may be
new to some readers.
                                           12

-------
This Page is intentionally left blank
                13

-------
Part I.   How Do I Use Results Of Models And Other
        Analyses To Help Demonstrate Attainment?
                         14

-------
2.0    What Is A Modeled Attainment Demonstration?--An Overview

       A modeled attainment demonstration consists of (a) analyses which estimate whether
selected emissions reductions will result in ambient concentrations that meet the NAAQS and (b)
an identified set of control measures which will result in the required emissions reductions. As
noted in Section 1, this guidance focuses on the first component of an attainment demonstration,
that is, completion and interpretation of analyses to estimate the amount of emission reduction
needed to attain the NAAQS. Emission reduction strategies should be simulated by reducing
emissions from specific source categories rather than through broad "across-the-board"
reductions from all sources.

       States/Tribes should estimate the amount of emission reduction needed to demonstrate
attainment of the NAAQS using the modeled attainment test. We recommend a similar analysis
for assessing progress toward reducing regional haze. In addition, a State/Tribe should consider
a broader set of model results, as well as perform a set of other corroboratory analyses to further
support whether a proposed emission reduction will lead to attainment of the NAAQS or uniform
rate of progress.

       2.1    What Is The Recommended Modeled Attainment Test for Ozone and PM2 5?-
             An Overview

       A modeled attainment test is an exercise in which an air quality model is used to simulate
current and future air quality. If future estimates of ozone and/or PM25 concentrations are less
than the NAAQS, then this element of the attainment test is satisfied11.  Our recommended test is
one in which model estimates are used in a "relative" rather than "absolute" sense. That is, we
take the ratio of the model's future to current (baseline) predictions at monitors.  We call these
ratios, relative response factors. Future ozone and/or PM2 5 concentrations are estimated at
existing monitoring sites by multiplying a modeled relative response  factor at locations "near"
each monitor by the observation-based, monitor-specific, "baseline" design value.  The resulting
predicted "future concentrations" are compared to NAAQS.

       The test for ozone is based on the calculation of a single mean ozone RRF for each
monitor.  The PM2 5 attainment test is more complicated and reflects the fact that PM2 5 is a
mixture.  In the test,  ambient PM25 is divided into major components. These are

       - mass associated with sulfates
       - mass associated with nitrates
       -mass associated with ammonium
       - mass associated with organic carbon
       nAs detailed in Section 7, additional corroborative analyses are still needed to
supplement the modeled attainment test, even when predicted ozone and/ or PM2 5 concentrations
are less than the NAAQS.

                                              15

-------
       - mass associated with elemental carbon
       -mass associated with particle bound water
       - mass associated with "other"primary inorganic particulate matter
       -and passively collected mass

       A separate RRF is calculated for each of the PM25 components (except passive mass).
We call each  of these site-specific ratios, component-specific RRFs.  Future PM25 design values
are estimated at existing monitoring sites by multiplying modeled relative response factors
"near" each monitor times the observed "component specific design value".  This latter quantity
is estimated using measured site-specific design values for PM2 5 in concert with available
measured composition data. Future site-specific PM2 5 design values at a site are estimated by
adding the future year values of the seven PM25 components. If all future site-specific PM25
design values are < the concentration specified in the NAAQS, the test is passed.

       2.2 What Is The Recommended Modeling Assessment For Regional Haze?- An
Overview

       The recommended modeled test for assessing uniform rate of progress in reducing
regional haze is similar conceptually to the recommended tests for the two NAAQS for PM25.
Models are used to develop relative response factors for each of 6 components of particulate
matter between a base period (2000-2004) and a future 5-year period which will be reviewed in
2018.  Components used for regional haze-related applications differ slightly from those  used for
NAAQS-related applications.  They are:

       - mass associated with sulfates;
       - mass associated with nitrates;
       - mass associated with organic carbon;
       - mass associated with elemental carbon;
       - mass associated with fine soil (i.e., crustal material);
       - mass associated with coarse particulate matter (i.e., PM10 - PM25).

       Current speciated measurements in a Class I area are used in an empirically derived
equation to estimate light extinction for each day with measurements. Days are ranked
according to their resulting light extinction (measured in deciviews). This ranking is used to
identify the 20% of days with worst and 20% of days with best visibility during each year in the
base period.  The 20% worst and best days  are examined to estimate appropriate observed
concentrations for the components of PM on "best" and "worst" days.

       Observed component concentrations are multiplied by the corresponding relative
response factors to estimate future concentrations for each component on "best" and "worst"
days. Future  component concentrations are then inserted into the equation relating light
extinction to concentrations of particulate matter.  The resulting estimates for future light
extinction on "best" and "worst" days are compared with observations made during the base
period to see  assess the future year visibility improvement.

                                           16

-------
       2.3    What Does A Recommended Supplemental Analysis/Weight Of Evidence
             Determination Consist Of? —An Overview

       As we describe in more detail in Section 7, States/Tribes should always perform
complementary analyses of air quality, emissions and meteorological data, and consider
modeling outputs other than the results of the attainment test.  Such analyses are instrumental in
guiding the conduct of an air quality modeling application.  Sometimes, the results of
corroboratory analyses may be used in a weight of evidence determination to show that
attainment is likely despite modeled results which may be inconclusive. The further the
attainment test is from being passed, the more compelling contrary evidence produced by
corroboratory analyses must be to draw a conclusion differing from that implied by the modeled
attainment test results. If a conclusion differs from the outcome of the modeled test, then the
need for subsequent review (several years hence) with more complete data bases is increased.  If
the test is failed by a wide margin (e.g., future design values outside the recommended range at
an individual site or multiple sites/locations), it is far less likely that the more qualitative
arguments made in a weight of evidence determination can be sufficiently convincing to
conclude that the NAAQS will be attained.  Table 2.1 contains guidelines for assessing when
corroboratory analyses and/or weight of evidence determinations may be appropriate.
Table 2.1 Guidelines For Weight of Evidence Determinations
Results of Modeled Attainment Test
Ozone
Future Design
Value < 82 ppb,
all monitor sites
Future Design
Value 82 - 87
ppb, at one or
more sites/grid
cells
Future Design
Value > 88 ppb,
at one or more
sites/grid cells
Annual PM2 5
Future Design
Value <14.5 ug/m3,
all monitor sites
Future Design
Value 14.5-15.5
ug/m3, at one or
more sites/grid
cells
Future Design
Value > 15.5
ug/m3,
at one or more
sites/grid cells
24-Hour PM2 5
Future Design
Value <62 ug/m3,
all monitor sites
Future Design
Value 62-67 ug/m3,
at one or more
sites/grid cells
Future Design
Value > 68 ug/m3,
at one or more
sites/grid cells
Supplemental Analyses

Basic supplemental analyses
should be completed to confirm
the outcome of the modeled
attainment test
A weight of evidence
demonstration should be
conducted to determine if
aggregate supplemental
analyses support the modeled
attainment test
More qualitative results are
less likely to support a
conclusion differing from the
outcome of the modeled
attainment test12.
       12 Regional modeling completed by EPA indicates that, on average, considerable
amounts of precursor control (e.g., 20-25 percent or more) will be needed to lower projected
                                           17

-------
       In a weight of evidence (WOE) determination, States/Tribes should review results from
several diverse types of air quality analyses, including results from the modeled attainment test.
As a first step, States/Tribes should note whether or not the results from each of these analyses
support a conclusion that the proposed strategy will meet the air quality goal. Secondly,
States/Tribes should weigh each type of analysis according to its credibility, as well as its ability
to address the question being posed (i.e., is the strategy adequate for meeting the NAAQS by a
defined deadline?).  The conclusions derived in the two preceding steps are combined to make an
overall assessment of whether meeting the air quality goal is likely. This last step is a qualitative
one.  If it is concluded that a strategy is inadequate to demonstrate attainment, a new strategy is
selected for review, and the process is repeated. States/Tribes should provide a written rationale
documenting how and why the conclusion is reached regarding the adequacy of the final selected
strategy.  Results obtained with air quality models are an essential part of a weight of evidence
determination and should  ordinarily be very influential in deciding whether the NAAQS will be
met.

       2.4    Why Should A Model Be Used In A "Relative" Sense And Why May
             Corroboratory Analyses Be Used In A Weight Of Evidence Determination?

       The procedure we  recommend for estimating needed emission reductions differs from
that in past guidance (U.S. EPA, 1996a) for ozone and PM in two major respects. First, we
recommend a modeled attainment test in which model predictions are used in a relative rather
than  absolute sense.  Second, the role of the weight of evidence determination, when used, has
been expanded. That is, these results can now be used as a rationale for concluding that a control
strategy will meet the NAAQS, even though the modeled attainment test alone may not be
conclusive. There are several  reasons why we believe these changes are appropriate.

1. Starting with an  observed concentration as the base value reduces problems in
interpreting model results. If a model under (or over) predicts an observed daily maximum
concentration, the appropriate  target prediction is not as clear as might be desired. For example,
if an 8-hour daily maximum ozone concentration of 120 ppb were observed and a model
predicted 100 ppb on that  day, should  the target for the day still be 84 ppb? In the relative
attainment test, observed data  is used to define the target concentration. This has the effect of
anchoring the future concentrations to a "real" ambient value. Although good model
performance remains a prerequisite for use of a model in an attainment demonstration, problems
posed by less than ideal model performance on individual days are reduced by the new
procedure.

2. The form of the 8-hour ozone and PM2 5 NAAQS necessitates such an attainment test.
The NAAQS for ozone and PM2 5 requires ambient data to be averaged over three consecutive
years. This presents  difficulties using the resource-intensive Eulerian models we believe are
ozone design values by 3 ppb or PM2 5 annual design values by 0.5 ug/m3 or 24-hour PM2 5
design values by 3 ug/m3.
                                           18

-------
necessary to capture spatially differing, complex non-linearities between ambient ozone and
precursor emissions.  That is, it is difficult to tell whether or not a modeled exceedance obtained
on one or more days selected from a limited sample of days is consistent with meeting the
NAAQS. To do so would require modeling several years and, perhaps, many strategies.  This
problem is reduced by using the monitored design value, as an inherent part of the modeled
attainment test.

3. PM2 5 consists of a diverse mix of primary and secondary components. This raises a
concern about a model's potential inability to correctly predict values for each component which
are proportional to the observed mix of components. Failure to predict major measured
components of PM25  in the correct proportion increases the possibility of choosing ineffective
control strategies on the basis of incorrect model predictions. This possibility is reduced if the
model responses are instead applied to components of PM25 which are derived from
measurements.

4. Model results and projections will continue to have associated uncertainty. The procedure
we recommend recognizes this by including modeling plus other analyses to determine whether
all available evidence supports a conclusion that a proposed emission reduction plan will suffice
to meet the NAAQS.  For applications in which the modeled attainment test is not passed, a
weight of evidence analysis may be used to support a determination that attainment will be
achieved, despite the  results of the modeled attainment test.  The weight of evidence
determination includes several modeling results which are more difficult to relate to the form of
the 8-hour ozone and 24-hour PM2 5 NAAQS. These results address relative changes in the
frequency and intensity of high modeled ozone or PM25 concentrations on the sample of days
selected for modeling. If corroboratory  analyses produce strong evidence that a control strategy
is unlikely to meet the NAAQS, then the strategy may be inadequate, even if the modeled
attainment test is passed.

5. Focusing the modeled attainment test only at monitoring sites could result in control
targets which are too low if the monitoring network is limited or poorly designed.  We
recommend a test which includes a review of the strategy's impact at locations without monitors.
This exercise provides a supplemental test to determine whether there is a need for further action
despite passing the modeled attainment test at all monitoring sites. While this test may indicate
potential unmonitored violations, ultimately, the best way to account for a limited or poorly
designed monitoring network is to use the model results, or other available analyses, to help
determine locations where additional monitors should be sited.
                                           19

-------
3.0    What Is The Recommended Modeled Attainment Test?

       In Section 2, we provided an overview of the recommended modeled attainment test. In
this section we provide more details on the calculation of baseline design values, the definition
of "nearby grid cells", and an "unmonitored area analysis" which provides estimates of future
year values in unmonitored areas. Sections 4,  5, and 6 provide more specific details and
examples for attainment tests for ozone and PM25, and the uniform rate of progress assessment
for regional haze respectively.

       We begin by establishing the basic equation on which all of the tests are built upon.
Equation (3.1) describes the recommended modeled attainment test, applied near monitoring site
I.

       (DVF)I = (RRF)I(DVB)I                                                   (3.1)

   where

    (DVB)j = the baseline concentration monitored at site I, units in ppb or ug/m313;

    (RRF)j = the relative response factor, calculated near site I, unitless
          The relative response factor is the ratio of the future concentration predicted near a
          monitor (averaged over multiple days) to the baseline  concentration predicted near
          the monitor (averaged over the same days), and

   (DVF)j = the estimated future design value for the time attainment is required,  ppb or ug/m3.

       Equation (3.1) looks simple enough. However,  several issues must be resolved before
applying it.

(1) How is a "site-specific" baseline design value ((DVB)j) calculated?

(2) In calculating the (RRF^ , what do we mean by "near"  site I?
(3) Several surface grid cells may be "near" the monitor, which one(s) of these should be used to
calculate the
(4) How do you calculate future design values in unmonitored areas?

(5) Which base year emissions inventory should be projected to the future for the purpose of
calculating RRFs?
       13The units for ozone is ppb and the units for PM2 5 is ug/m3.

                                          20

-------
(6) Which future year should emissions be projected to in order to assess attainment using the
modeled attainment test?

   3.1    Calculating site-specific baseline concentrations.

       The modeled attainment test is linked to the form of the ozone and PM25 NAAQS
through the use of monitored design values. Official design values are calculated in different
ways for 8-hour ozone, annual average PM2 5, and 24-hour average PM2 5. The following is a
brief description of the calculations:

Ozone

The 8-hour ozone design value is calculated as the 3 year average of the fourth highest
monitored daily 8-hour maximum value at each monitoring site.  The standard is met if the
design value is <0.8 ppm (in practice, this translates to < 85 ppb). The detailed description of the
methodology can be found in 40 CFR Part 50.10, and Appendix I to Part 50.

Annual Average PM2 5

The annual average PM2 5 standard is met if, over a consecutive three year period, the average
arithmetic mean concentration of PM25 is less than or equal to 15.0 |ig/m3.  The annual
arithmetic mean at a monitoring  site is calculated by averaging the  four quarterly arithmetic
mean concentrations observed during a calendar year at the site.  The annual mean concentration
averaged over three years must be < 15.0 |ig/m3 at all monitoring sites. The details can be found
in 40CFR Part 50, Appendix N

In some cases, two nearby sites can be spatially averaged to determine the annual average design
value for both sites. For details on spatial averaging requirements see 40 CFR Part 58, Appendix
D, Section 2.8.

24-Hour Average PM2 5

The 24-hr NAAQS for PM2 5 is met if the 98th percentile 24-hour average concentration of
particulate matter with aerodynamic diameter < 2.5 micrometers, averaged over three
consecutive years, is < 65 |ig/m3.14 The test applies at all monitoring sites—spatial averaging is
not allowed for the 24-hour NAAQS.  The "98th percentile" concentration (and, thus, the design
value concentration) depends on the number of days on which PM2 5 is monitored during a year.
See 40CFR Part 50, Appendix N for details.  Attainment is reached if the design value at the site
is < 65 |ag/m3.
       14See 40CFR Part 50, Appendix N for a discussion of how data completeness affects the
determination of attainment status.

                                           21

-------
How is a "site-specific" baseline design value ((DVB)j) calculated?

       The baseline measured concentrations at each monitoring site is the anchor point for
future year projected concentrations.  The baseline design values are projected to the future using
RRFs.  In practice, the choice of the baseline design value can be critical to the determination of
the estimated future year design values. Therefore, careful consideration should be given to the
calculation of baseline values.  The baseline design values should have the following attributes:

1) Should be consistent with the form of the applicable NAAQS.
2) Should be easy to calculate.
3) Should represent the baseline inventory year.
4) Should take into account the year-to-year variability of meteorology.
5) Should take into account the year-to-year variability of emissions.

Several possible methodologies to calculate baseline design values are:

1) The designation design value period (i.e. 2001-2003).
2) The design value period that straddles the baseline inventory year (e.g., the 2001-2003 design
value period for a 2002 baseline inventory year).
3) The highest (of the three) design value periods which include the baseline inventory year (e.g.
the 2000-2002, 2001-2003,  and 2002-2004 design value periods for a 2002 baseline inventory
year).
4) The average (of the three) design value periods which straddle the baseline inventory year.

       For the modeled attainment tests we recommend using the average of the three design
value periods which include the baseline inventory year  (choice number 4 from above). Based
on the attributes listed above, the average of the three design value periods best represents the
baseline concentrations, while taking into account the variability of the meteorology and
emissions (over a five year period).

       The three design values that are averaged in the calculation cover a five year period, but
the average design value is not a straight five year average.  It is, in effect, a weighted average of
the annual averages. For example, given a baseline inventory year of 2002, the years used to
calculate the average design value range from 2000-2004. In the average of the 2000-2002,
2001-2003, and 2002-2004  periods, 2002 is "weighted" three times, 2001 and 2003 are weighted
twice, and 2000 and 2004 are weighted once. This has the desired effect of weighting the
projected ozone or PM25 values towards the middle year of the five year period, which is the
emissions year (2002 in this example). The average design value methodology is weighted
towards the inventory year (which is the middle year) and also takes into account the emissions
and meteorological variability that occurs over the full five year period (although the emissions
and meteorology from the other years are weighted less than the middle year of the 5 year
                                           22

-------
period)15.  Because of this, the average weighted design value is thought to be more
representative of the baseline emissions and meteorology period than other methodologies such
as choosing the highest single design value period.

       Additionally, the average design value will be more stable (less year to year variability)
than any single  design value period. An analysis of recent ambient design value data at 471
ozone monitors, over the period 1993-2004, shows that the median standard deviation of design
values was 3.3 ppb whereas the standard deviation of the 5 year weighted average design values
was 2.4 ppb (Timin, 2005a). Also, moving from the period ending in 2003 to the period ending
in 2004, the median change in the ozone design values was 4.0 ppb. The median change in the 5
year weighted average ozone design values was only 0.8 ppb (there was not a long enough data
record to do the same analyses for PM2 5).  These  analyses show that the average design values
are clearly more stable and will therefore provide  a "best estimate" baseline year design value
(DVBj) for use in future year model projections.

       The recommended averaging technique assumes that at least five complete years of
ambient data is  available at each monitor.  In  some cases there will less than five years of
available data (especially at relatively new monitoring sites). In this case we recommend that
data from the monitor is used if there is  at least three consecutive years of data.  If there are three
years of data then the baseline design value will be based on a single design value.  If there are
four years of data then the baseline design value will be based on an average of two design value
periods. If a site has less than three years of data,  then the site should not ordinarily be used in
the attainment test.

Calculating site-specific "baseline" design values16 (DVB) to use in the attainment test

Example  3.1

Given:

(1) The baseline inventory year is 2002 (i.e., 2002 emissions are being modeled).
       15The emissions trends and meteorological data during the five year period should be
evaluated to determine if any large emissions changes and/or "extreme" meteorological
conditions have occurred during the period.  It is especially important to consider the emissions
and meteorology during the middle year of the period. When choosing a base emissions year
and calculating baseline design values, extreme conditions that may lead to either abnormally
high or abnormally low concentrations should be discussed with the appropriate EPA Regional
Office.  See section 14 for more information on selecting modeling time periods.

       16The "baseline design value" is an average of several design values and thus is
technically not a design value.  The guidance continues to refer to the average design values as
"design values" even though they are based on averages of observed design values.

                                           23

-------
(2) For purposes of illustration, suppose the area contains only three ozone monitors.

Find: The appropriate site-specific baseline design values to use in the modeled attainment test.

Solution: Since the inventory reflects 2002, we need to examine monitored design values for
overlapping 3-year periods that include 2002. The three design values are then averaged for
each site. These are the values for site-specific baseline design values (DVB) in the modeled
attainment test. The procedure is shown in Table 3.1.

Table 3.1 Example Illustrating Calculation Of Baseline Design Values
Monitor
1
2
3
2000-2002
Design Value,
ppb
88
86
88
2001-2003
Design
Value, ppb
87
84
86
2002-2004
Design
Value, ppb
90
91
85
Baseline Design Value
(DVB) Used In The
Modeled Attainment
Test, ppb
88.317
87.0
86.3
       3.2
Identifying surface grid cells near a monitoring site.
       There are four reasons why we believe it is appropriate, in the modeled attainment test,
to consider cells "near" a monitor rather than just the cell containing the monitor.  First, one
consequence of a control strategy may be "migration" of a predicted peak. If a State were to
confine its attention only to the cell containing a monitor, it might underestimate the RRF (i.e.,
overestimate the effects of a control strategy).  Second,  monitor siting guidelines recommend
spatial scales (for ozone and PM2 5) which are generally urban or neighborhood scale. In many
cases, the representative spatial scale of a monitor is larger than a single grid cell.  Third, we
believe that uncertainty in the formulation of the model and the model inputs is consistent with
       17 The average design value should carry one significant figure to the right of the decimal
point for ozone and 24-hour PM2 5 and 2 significant figures for annual PM2 5. There are several
calculations in the modeled attainment test which carry the tenths digit or hundreths digit. We
have found that rounding and/or truncating concentrations and RRFs can lead to an overestimate
or underestimate of the impact of emissions controls. In some cases, a few tenths of a ppb
change (or a few tenths of a percent reduction) can be meaningful. Rounding or truncating can
make the change appear to be equal to a full ppb  (or 1%) or equal to zero change.  For 8-hour
ozone, it is recommended to round to the tenths digit until the last step in the calculation when
the final future design value is truncated. For annual PM2 5 it is recommended to round  to the
hundreths digit until the last step when the final future design value is rounded to the nearest
tenths place. For daily PM2 5 it is recommended to round to the tenths digit until the last step
when the final future design value is rounded to the nearest whole number.
                                           24

-------
recognizing some leeway in the precision of the predicted location of daily average PM25 and
daily maximum ozone concentrations.  Finally, standard practice in defining a gridded modeling
domain is to start in the southwest corner of the domain, and determine grid cell location from
there. Considering several cells "near" a monitor rather than the single cell containing the
monitor diminishes the likelihood of inappropriate results which may occur from the geometry of
the superimposed grid system.

       Earlier ozone modeling guidance (U.S. EPA,1996a) has identified 15 km as being "near"
a site. This is also consistent with the broad range of intended representativeness for urban scale
ozone monitors identified in 40CFR Part 58, Appendix D.  As noted in section 1.3, secondary
particulate matter is likely to constitute an important fraction of PM25.  Further, a major purpose
of urban monitoring performed to determine attainment of the NAAQS for PM25 is to estimate
likely exposure to PM2 5 concentrations over 24-hour periods. For PM2 5, sites having
"neighborhood" or "urban" scales of representativeness are generally most suitable for
estimating exposure or compliance with the NAAQS.  Because of the rather long sampling time
(i.e., 24 hours) inherent in the PM25 NAAQS, locations within the lower range of "urban" scale
(4-50 km) would seem consistent with a definition of "near" a site.  Therefore, we recommend
following the same ground rules used for defining "near" as we originally defined for ozone
modeling (U.S.EPA, 1999a).

       For ease in computation, States/Tribes may assume that a monitor is at the center of the
cell in which it is located and that this cell is at the center of an array of "nearby" cells.  The
number of cells considered "nearby" (i.e., within about a 15 km radius of) a monitor is a function
of the size of the grid cells used in the modeling. Table 3.2 provides a set of default
recommendations for defining "nearby" cells for grid systems having cells of various sizes.
Thus, if one were using a grid with 4 km grid cells, "nearby" is defined by a 7 x 7 array of cells,
with the monitor located in the  center cell.

       The use of an array of grid cells near a monitor may have a large impact on the RRFs in
"oxidant limited" areas (areas where NOx decreases may lead to ozone or secondary PM
increases) and in areas with large gradients in primary PM2 5. The array methodology could lead
to unrealistically small or large RRFs, depending on the specific case. Care should be taken in
identifying an appropriate array size for these areas. States/Tribes may consider the presence of
topographic features, demonstrated mesoscale flow patterns (e.g., land/sea, land/lake interfaces),
the density  of the monitoring network, the density of emissions (especially primary PM)18, and/or
other factors to deviate from our default definitions for the array of "nearby" grid cells, provided
the justification for doing so is  documented.
       18It may be inappropriate to use an array as large as 7 X 7 in areas with large gradients of
primary PM25 concentrations. Averaging of multiple grid cells may mask changes in local
concentrations between the base year and future years.

                                           25

-------
Table 3.2.  Default Recommendations For Nearby Grid Cells Used To Calculate RRF's
Size of Individual Cell, km
4-519
>5-8
>8-15
>15
Size of the Array of Nearby
Cells, unitless
7x7
5x5
3x3
1x1
       3.3    Choosing model predictions to calculate a relative response factor (RRF)j
             near a monitor.

8-Hour Ozone NAAQS

       Given that a model application produces a time series of estimated 1-hour ozone
concentrations (which can be used to calculate running 8-hour averages), what values should be
chosen from within the time series?  We recommend choosing predicted 8-hour daily maximum
concentrations from each modeled day (excluding "ramp-up" days) for consideration in the
modeled attainment test. The 8-hour daily maxima should be used, because they are closest to
the form of concentration  specified in the NAAQS.

       The second decision that needs to be made is, "which one(s) of the 8-hour daily maxima
predicted in cells near a monitor should we use to calculate the RRF?" We recommend choosing
the nearby grid cell with the highest predicted 8-hour daily maximum concentration with
baseline emissions for each day considered in the test, and the grid cell with the highest
predicted 8-hour daily maximum concentration with the future emissions for each day in the test.
Note that, on any given day, the grid cell chosen with the future emissions need not be the same
as the one chosen with baseline emissions.

       We believe selecting the maximum (i.e., peak) 8-hour daily maxima on each day for
subsequently calculating the relative response factor (RRF) is preferable for several reasons.
First, it is likely to reflect  any phenomenon which causes peak concentrations within a plume to
migrate as a result of implementing controls. Second, it is likely to take better advantage of data
produced by a finely resolved modeling analysis.

       The relative response factor (RRF) used in the modeled attainment test is computed by
taking the ratio of the mean of the 8-hour daily maximum predictions in the future to the mean
of the 8-hour daily maximum predictions with baseline emissions, over all relevant days.  The
       19The appropriate size of the array for horizontal grid cells < 4km should be discussed
with the appropriate U.S EPA Regional Office.

                                          26

-------
procedure is illustrated in Example 3.2.

Example 3.2

Given: (1) Four primary days have been simulated using baseline and future emissions.

(2) The horizontal dimensions for each surface grid cell are 12 km x 12 km.

(3) In each of the 9 grid cells "near" a monitor site I, the maximum daily predicted future
concentrations are 87.2, 82.4, 77.5, and 81.1 ppb.

(4) In each of the 9 grid cells "near" a monitor site I, the maximum daily predicted baseline 8-
hour daily maximum ozone concentrations are 98.3, 100.2, 91.6, and 90.7 ppb.

Find: The site-specific relative response factor for monitoring site I,  (RRF)j

Solution:

(1) For each day and for both baseline  and future emissions, identify the 8-hour daily maximum
concentration predicted near the monitor. Since the grid cells are 12 km, a 3 x 3 array of cells is
considered "nearby" (see Table 3.2).

(2) Compute the mean 8-hour daily maximum concentration for (a) future and (b) baseline
emissions.

Using the information from above,

       (a) (Mean 8-hr daily max.)&ture = (87.2 + 82.4 + 77.5 + 81.1)/4 = 82.1 ppb
       and

       (b) (Mean 8-hr daily max.)baseline = (98.3 + 100.2 + 91.6 + 90.7)74 = 95.2 ppb

(3) The relative response factor for site I is

      ! = (mean 8-hr daily max.)&ture/(mean 8-hr daily max.)baseline

         82.1/95.2 = 0.862
                                           27

-------
           Figure 3.1.  Choosing Ozone Predictions To Estimate RRF's
                  (a) Predictions With Baseline Emissions
        Day 1
      Day 2
Day3
Day 4
97.2
97.1
97.2
95.5
95.2
98.3
96.2
89.1
97.6
87.8
90.9
88.5
90.1
91.6
89.4
89.9
88.7
90.2
85.9
87.9
86.9
87.9
90.5
87.3
88.9
90.7
88.4
        98.3
      100.2
91.6
  90.7
       Mean Baseline Ozone Concentration = (98.3 + 100.2 + 91.6 + 90.7) 14 = 95.2 ppb
        Day 1
(b) Predictions With Future Emissions

      Day 2               Day 3
                    Day 4
82.2
82.4
81.4
80.8
79.9
77.8
81.2
80.5
80.1
72.1
74.6
76.9
76.1
77.5
77.4
75.5
74.3
75.6
75.4
80.8
80.4
78.8
79.5
76.9
79.8
80.9
81.1
        87.2
      82.4
  77.5
  81.1
      Mean Future Ozone Concentration = (87.2 + 82.4 + 77.5 + 81.1) 14 = 82.1 ppb
PM2 5 NAAQS

       For PM2 5 applications, we recommend choosing predicted 24-hour average
concentrations from each modeled day (excluding "ramp-up" days) for consideration in the
modeled attainment test20. The 24-hour average values should be used, because they are the
basis for design value calculations.

       The second decision that needs to be made is, "which of the 24-hour average values
predicted in cells near a monitor should we use to calculate the RRF?" For the annual PM2.,5
NAAQS, we recommend taking the spatially averaged value of the nearby predictions (mean
value of the grid cell array). Each component-specific relative response factor (RRF)j used in the
modeled attainment test is computed by taking the ratio of the mean of the spatially averaged
daily predictions in the future to the mean of the spatially averaged daily predictions with current
       20This includes modeled days for which there is no corresponding monitored data.

                                           28

-------
emissions.  The procedure is the same as the ozone example presented above, except the spatial
mean is used instead of the highest concentration near the monitor.

       For the 24-hour PM2 5 NAAQS and for annual PM2 5 NAAQS sites which exhibit strong
spatial  concentration gradients of primary PM25, we recommend calculating an RRF based on
the single value in the grid cell which contains the monitor. Particularly in areas with strong
primary PM concentration gradients, the use of a single grid cell is more likely to accurately
capture the local change in PM2 5 components.

       3.4     Estimating design values at unmonitored locations: what is an unmonitored
              area analysis and why is it needed?

       An additional review is necessary, particularly in  nonattainment areas where the ozone or
PM2 5 monitoring network just meets or minimally exceeds the size of the network required to
report data to Air Quality System (AQS). This review is  intended to ensure that a control
strategy leads to reductions in ozone or PM2 5 at other locations which could have baseline (and
future) design values exceeding the NAAQS were a monitor deployed there.  The test is called
an "unmonitored area analysis".  The purpose of the analysis is to use a combination of model
output  and ambient data to identify areas that might exceed the NAAQS if monitors were located
there.

       The unmonitored area analysis should identify areas where future year design values are
predicted to be greater than the NAAQS.  The unmonitored area analysis for a particular
nonattainment area is intended to address potential problems within or near that nonattainment
area. The analysis should include, at a minimum, all nonattainment counties and counties
surrounding the nonattainment area (located within the State).  In large States, it is possible that
unmonitored area violations may appear in counties far upwind or downwind of the local area of
interest. In those cases, the distance to the nonattainment area and ability of the modeling to
represent far downwind areas should be evaluated on a case by case basis.  In order to examine
unmonitored areas in all portions of the domain, it is recommended to use interpolated spatial
fields of ambient data combined with gridded modeled outputs.
             3.4.1  Why does the unmonitored area analysis need to use both ambient
                    data and model output?

       Ambient data can be interpolated to provide a set of spatial fields. The spatial fields will
provide an indication of concentrations in monitored and unmonitored areas.  But a simple
interpolation of the ambient data cannot identify unmonitored areas with higher concentrations
than those measured at monitors.  The interpolated concentration between monitors will
generally be the same or lower than the measured concentration at the monitors (assuming that
more sophisticated statistical techniques are not used, such as adding a nugget effect or a trend
surface). The interpolation technique does not account for emissions or chemistry information
that may be needed to identify potential unmonitored violations.

                                           29

-------
       The gridded model output (absolute) concentrations can also be used to examine
unmonitored area concentrations. The model provides an hourly concentration for every grid
cell. The concentrations can be analyzed to determine unmonitored areas where the model
predicts high values.  But the absolute predictions from the model may not be entirely accurate.
The model output is only as accurate as the emissions and meteorological input.  But unlike the
interpolated ambient data, the model output explicitly accounts for emissions, chemistry, and
meteorology over the entire domain.

       Both the interpolated ambient data  and the model outputs have major weaknesses. But
they also both have strengths.  We can take advantage of the strengths of each dataset by
combining the two  types of data.  The interpolated spatial fields of ambient data  provide a strong
basis for estimating accurate pollutant concentrations at monitors and near monitors. Given that
information, the model outputs can be used to adjust the interpolated spatial fields (either up or
down) so that more accurate estimates can be derived in the unmonitored areas21. The best way
to use the model to adjust the spatial fields is to use modeled gradients.  It is preferable to
assume that the model is predicting  areas of generally high  or low ozone or PM,  as compared to
assuming that the absolute predictions from the model  are correct. For example,  in areas where
the model predicts relatively high ozone or PM concentrations, the spatial fields  can be adjusted
upward. In areas where the model predicts relatively low ozone or PM concentrations, the
spatial fields can be adjusted downward.   In this way, it may be possible to predict unmonitored
areas that may have high concentrations. At the same time, concentrations in rural areas, (which
may be overly influenced by high monitored concentrations near urban areas), may be adjusted
downward. It may  also be possible  to predict high PM2 5 concentrations in unmonitored areas
which contain large sources of primary PM emissions.  The combination of interpolated spatial
fields and modeled output will be referred  to as "gradient adjusted spatial fields"

             3.4.2  Implementation of Gradient adjusted Spatial Fields

       Gradient adjusted spatial fields are  first created for the base year. Future year estimates
can then be created by applying gridded RRFs to the gradient adjusted spatial fields. The basic
steps are as follows:

1) Interpolate base year ambient data to create a set of spatial fields.
2) Adjust the spatial fields using gridded model output gradients (base year values).
3) Apply gridded model  RRFs to the gradient adjusted spatial fields.
4) Determine if any unmonitored areas are predicted to exceed the NAAQS in the future.

       EPA has developed a software package called "Modeled Attainment Test Software"
(MATS) which will spatially interpolate data, adjust the spatial fields based on model output
gradients and multiply the fields by  model  calculated RRFs (steps 1-3 above) (ABT, 2007).
       21The accuracy of interpolated fields can be tested by removing sets of monitors to see
how well the interpolation scheme estimates known concentrations at ambient monitoring sites.

                                           30

-------
States will be able to use the EPA-provided software or are free to develop alternative techniques
that may be appropriate for their areas or situations.

Stepl

       The first step in the analysis is to interpolate ambient data. Ideally, design values should
be interpolated. The same 5 year weighted average design values that are used in the monitor
based model attainment test can be used in the development of ambient spatial fields. Care
should be taken so that the interpolated fields are not unduly influenced by monitoring sites that
do not have complete data.  Since the design values can vary significantly from year to year, it is
important to use a consistent set of data. In some cases, it may be preferable to interpolate
individual years of data or individual design values, and then average those up to get the 5 year
weighted average.

       We are not recommending a single interpolation technique. EPA has provided example
analyses in the past using the Kriging interpolation technique (U.S.EPA, 2004c). Alternatively,
EPA's BenMAP software, which was used to create interpolated fields for the Clean Air
Interstate Rule (CAIR) uses the Voronoi Neighbor Averaging (VNA) technique (Abt, 2003)22.

       For ozone analyses, a single spatial field of ozone values is needed. But for the PM25
NAAQS, we recommend creating interpolated spatial fields for PM2 5 (FRM) and for each
component of PM2 5.  For the annual PM2 5 NAAQS, a set of quarterly average spatial fields can
be created. The 4 quarters are averaged to get an annual average set of fields.

       For the 24-hour PM2 5 standard, a spatial field can be created using the high end of the
distribution of 24-hour PM2 5 concentrations.  This is best represented by interpolating the
measured high values from each quarter.  To be consistent with the attainment test, we
recommend interpolating the PM2 5 concentration in each quarter which is equal to or less than
the 98th percentile value for the year. For the PM2 5 component species, we recommend
interpolating the high PM2 5 days in each quarter. This can be based on the highest monitored
days in each quarter.

       Interpolated data should be evaluated to quantify the errors associated with the
interpolation technique.  Additional information on evaluation of spatial fields can be found in
(U.S. EPA, 2004c).

Step 2

       The second step in the process involves the use of gridded model output to adjust the
spatial fields. The BenMAP software contains an example of this technique called eVNA.  It
uses seasonal average model output data to adjust interpolated spatial fields (MATS also uses the
       2MATS also uses a version of VNA.

-------
same technique).  The eVNA technique has been used in health benefits assessments (U.S. EPA,
2005a). A specific metric is needed to determine the model predicted gradient in concentrations
for each of the NAAQS. For ozone, a logical metric is the 4th highest ozone prediction in each
grid  cell23. For the annual PM2 5 NAAQS, the model predicted quarterly mean concentrations for
PM2 5 and PM2 5 species can be used to adjust the ambient spatial fields.  For the 24-hour PM2 5
NAAQS, the gradient adjusted fields can be derived from the high end of the distribution of daily
averages in each quarter. This could be for all days with 24-hour predicted (or measured) PM2 5
> 65 ug/m3 (or some relatively high value) or simply the top 10% or 25% of all (PM25) days in
each quarter.

StepS

       The next step is to create future year fields by multiplying the base year gradient adjusted
spatial fields by model derived gridded RRFs.  The RRFs for the unmonitored area analysis are
calculated in the same way as the monitored based attainment test (except that the grid cell array
is not used in the spatial fields based analysis). The future year concentrations are equal to the
base year concentration times the RRF in each grid cell.  The future year gradient adjusted
spatial fields are then analyzed to determine if any grid cells are predicted to remain above the
NAAQS.

       For ozone, a single spatial field is multiplied by a single set of model derived RRFs.  For
PM2 5, the RRFs for each of the species, for each quarter, are multiplied by the spatial fields for
each species, for each quarter.

              3.4.3   Using the Results of the Unmonitored Area Analysis

       It should be stressed that due to the lack of measured data, the examination of ozone and
PM2 5 concentrations as part of the unmonitored area analysis is more uncertain than the monitor
based attainment test.  As a result, the unmonitored area analysis should be treated as a separate
test from the monitor based attainment test. The results from the unmonitored area analysis
should, at a minimum, be included as a supplemental analysis. While it is expected that
additional emissions controls are needed to eliminate predicted violations of the monitor based
test,  the same requirements may not be appropriate in unmonitored areas.

       The cause of predicted violations of the unmonitored area analysis may not be obvious.
While careful analysis of the emissions in the area may reveal likely sources of an ozone
violation, elimination of the violation may not be simple or straightforward. Alternatively,
predicted violations of the unmonitored area analysis for PM2 5 may, in some cases be simple to
diagnose. They may occur in areas with high emissions of primary PM25. Elimination of the
       23The metric should approximate the measured design values at monitoring sites.
Depending on the days modeled, other metrics, such as the 2nd or 3rd highest predicted ozone
values or an average of several days may be a better proxy for the design value.

                                           32

-------
predicted violation may be possible by placing controls on one or more sources.  Additional
dispersion modeling may be helpful in diagnosing the source(s) of high primary PM25 in a
particular grid cell. The results of a more refined dispersion modeling analysis may help identify
emissions controls or locations where additional monitoring may be appropriate.

       It is recommended that predicted violations of the unmonitored area analysis are carefully
scrutinized to determine whether they are likely to exist in the ambient air or whether they may
be caused by an error or uncertainty in the modeling system.  At a minimum, it may be
appropriate to commit to additional deployment of ambient monitors in areas where the
unmonitored area analysis predicts future violations24. This monitoring would allow a better
assessment in the future  of whether the NAAQS is being met at currently unmonitored locations.

       Violations of the unmonitored area analysis should be handled on a case by case basis.
As such, additional analyses and/or tracking requirements may be needed depending on the
nature of the problem and the uncertainty associated with the potential violation(s).
       3.5    Which base year emissions inventory should be projected to the future for
              the purpose of calculating RRFs?

       The modeled attainment test adjusts observed concentrations during a baseline period
(e.g., 2000-2004) to a future period (e.g., 2009) using model-derived "relative response factors".
It is important that emissions used in the attainment test correspond with the period reflected by
the chosen baseline design value period (e.g., 2000-2004). Deviations from this constraint will
diminish the credibility of the relative response factors. Therefore, it is important to choose an
appropriate baseline emissions year.  There are potentially two different base year emissions
inventories.  One is the base case inventory which represents the emissions for the meteorology
that is being modeled. These are the emissions that are used for model performance evaluations.
For example, if a State is modeling a base year of 1998 (or episodes from 1998), "base case"
emissions and meteorology would be for 1998. As described in Section 18, it is essential to use
base case emissions together with meteorology occurring in the modeled episode(s) in order to
evaluate model performance.

       Once the model has been shown to perform adequately, it is no longer necessary to model
the base case emissions. It now becomes important to model emissions corresponding to the
period with a recent observed  design value.  The second potential base year inventory
corresponds to the middle year of the baseline average design value (e.g 2002 for a 2000-2004
average design value).  This is called the baseline inventory.  The baseline emissions inventory is
the inventory that is ultimately projected to a future year.
       24It would also be appropriate to commit to additional emissions controls in lieu of
additional monitoring in unmonitored areas.

                                           33

-------
       In section 17 we recommend using 2002 as the baseline inventory year for the current
round of ozone, PM25, and Regional Haze SIPs.  If States/Tribes model the full year of 2002 (or
episodes from 2002)  then the base case and baseline inventory years will be the same25.  But if
States/Tribes model other years or episodes or full seasons from other years, then the base case
inventories should be projected (or "backcasted") to 2002 to provide a common starting point for
future year projections.

       Alternatively, the baseline emissions year could be earlier or later than 2002, but it
should be a relatively recent year (preferably within the 5 year design value window).  In order
to gain confidence in the model results, the emissions projection period should be as short as
possible. For example, projecting emissions from 2002 to 2009 (with a 2000-2004 baseline
average design value) should be less uncertain than projecting emissions from 1995 to 2009
(with a 1993-1997 baseline average design value). Use of an older baseline average design value
period is discouraged.

       It is desirable to model meteorological time periods occurring during the period reflected
by the baseline design value (e.g., 2000-2004). However, modeling time periods need not be
selected from the period corresponding to the baseline design value, provided they are
representative of meteorological conditions which commonly occur when exceedances of the
ozone or PM25 standard occur (or represent typical annual concentrations of PM25) . The idea is
to use selected representative time periods to capture the sensitivity of predicted ozone or PM25
to changes  in emissions during commonly occurring conditions. There are at least three  reasons
why using time periods outside the period with the baseline design value may be acceptable: (1)
availability of air quality and meteorological data from an intensive field study, (2) the desire to
use meteorological data which may be "more representative" of typical conditions compared to
the baseline design value period and (3) availability of a past modeling analysis in which the
model performed well.

       Under the regional haze rule, the period for establishing baseline visibility is defined as
2000-2004. Therefore, emissions and meteorological time periods should be chosen from the
base time period. Modeling the same year (or time period) for PM2 5 NAAQS applications and
uniform rate of progress analyses is a logical way to conserve  resources.

       3.6    Choosing a year to project future emissions.

       States/Tribes should project future emissions to the attainment year or time period, based
on the area's classification.  The "Final Rule to Implement the 8-Hour Ozone National Ambient
Air Quality Standard, Phase 1" provides a schedule for implementing emission reductions
       25 The year may be the same, but the emissions may still differ. The base case inventory
may include day specific information (e.g. wildfires, CEM data) that is not appropriate for using
in future year projections. Therefore the baseline inventory may need to replace the day specific
emissions with average or "typical" emissions (for certain types of sources).

                                           34

-------
needed to ensure attainment by the area's attainment date (40 CFR 51.908).  Specifically, it
states that emission reductions needed for attainment must be implemented by the beginning of
the ozone season immediately preceding the area's attainment date. The PM25 implementation
rule contains similar provisions.  It states that emissions reductions should be in place by the
beginning of the year preceding the attainment date (40 CFR 51.1007).

       The logic for requiring emissions reductions by the year (or season) immediately
preceding the attainment year follows from language in the Clean Air Act. The Clean Air Act
(172(a)(2)(C)) identifies a process by which attainment is expected to occur. Shortly after the
attainment deadline, EPA is to evaluate whether areas have attained.  If the area is clean for the
last three years, then the area has met the attainment deadline.  But if the area was above the
standard for earlier years/seasons but has clean data for the single year/season immediately
preceding the attainment date (and if the area has met planning requirements), the area is
eligible for a 1-year extension of the attainment deadline. If the air is clean again the following
year, the area is eligible for a  second  1-year extension of the attainment deadline.  The
expectation is then that the third year's data, in combination with data from years 1 and  2, will
show attainment (This is commonly referred to as the "clean data policy".) Therefore, emissions
in the year preceding the attainment year should be at a level that is consistent with attainment.
It also follows that the year preceding the attainment year should be modeled for attainment
planning purposes.

        As an example, based on the language in the Clean Air Act and the ozone and PM2 5
rules, areas with an attainment date of no later than June 15th 2010 (for moderate ozone
nonattainment areas) or April 5th, 2010 (for PM2 5 nonattainment areas)26, need to have emission
reductions implemented no later than the beginning of 2009 (for PM2 5) or  the beginning of the
2009 ozone season (for 8-hour ozone). Therefore, modeling the year before the attainment
deadline (2009 in this case) is generally  appropriate for both ozone and PM2 5 attainment
demonstrations. However, attainment dates are expressed as "no later than"  three, five, six, or
nine years after designation and nonattainment areas are required to attain  as expeditiously as
practicable.  Therefore, additional considerations are necessary before a modeling year  can be
established.

       Using moderate ozone nonattainment areas as an example, these areas have an attainment
date of no later than June 15, 2010, or as expeditiously as practicable.  States/Tribes are required
to conduct a Reasonably Available Control Measures (RACM) analysis to determine if they can
       26One key difference between ozone and PM2 5 planning is that all PM2 5 nonattainment areas
have a single classification and a single set of attainment deadlines.  The PM25 nonattainment areas are
required to attain the standard within 5 years of designation or as expeditiously as practicable.  They can
also request up to a 5 year extension of the attainment deadline. The procedures for justifying an
extension of the attainment date are contained in the PM implementation rule (40 CFR 50.1004). In this
guidance document, we will assume that areas will model a future year based on attaining within the first
5 year period. Areas that request and are granted attainment date extensions should consult with their
Regional Office regarding future year modeling analyses and requirements

                                            35

-------
advance their attainment date by at least a year  . A RACM analysis is also required for PM2
                                                                                   2.5
nonattainment areas28. Since areas are required to attain as expeditiously as practicable and
perform a RACM analysis, results of the analysis may indicate attainment can be achieved
earlier, (e.g., 2008). In this case, the timing of implementation of control measures should be
used to determine the appropriate projection year. For example, if emission reductions
(sufficient to show attainment) are implemented no later than the beginning of the 2008 ozone
season, then the attainment year and the future projection year should be no later than 2008. In
all cases, the selection of the future year(s) to model should be discussed with the appropriate
EPA Regional Office as part of the modeling protocol development process.

       For regional haze assessments, the first review of progress will occur in 2018.  This will
likely be based on monitored data from the 2013-2017 period. Therefore, a logical future period
to model is the middle of the five year review period (2015).  But the uniform rate of progress
level can be calculated for any year in the future. Therefore, it is appropriate to model any future
year between 2015 and 2018.  Unless there are reasons to the contrary, any of the years between
2015 and 2018 can be considered to be representative of the regional haze future planning
period.  The calculation of the uniform rate of progress should be consistent with the emissions
projection period. For example, if a future year of 2015 is modeled, the  uniform rate of progress
calculation should be based on the number of years between the base period and 2015.
Similarly, if a future year of 2018 is modeled, then the uniform rate of progress should be based
on the number of years between the base period and 2018.
       2740CFR51.912(d)

       2840CFR51.1010(b)
                                           36

-------
4.0 Ozone Attainment Test

       Section 3 contained the details of the modeled attainment test that are common to both
ozone and PM2 5. This section describes several additional details related to the 8-hour ozone
attainment test.  The section concludes with an example calculation.

       4.1   Limiting modeled 8-hour daily maxima chosen to calculate RRF.

       On any given modeled day, meteorological conditions may not be similar to those
leading to high concentrations (i.e., values near the site-specific design value) at a particular
monitor.  If ozone predicted near a monitor on a particular day is much less than the design
value, the model predictions for that day could be unresponsive to controls (e.g., the location
could be upwind from most of the emissions in the nonattainment area on that day). Using
equation (3.1) could then lead to an erroneously high projection of the future design value.

       In order to examine this issue, we analyzed modeled baseline and future emissions for 30
episode days during 1995 using a grid with 12 km x 12 km cells and 9 vertical layers29. We
examined modeled RRF's computed near each of 299 monitoring sites in the eastern half of the
United States (Timin, 2005b). The study examined the day to day variability of (daily) RRFs at
each site.  One purpose of the study was to assess the extent to which a relative response factor
(RRF)  is dependent on the magnitude of modeled current 8-hour daily maxima.

       Figure 4.1 shows an example of the raw data from the analysis for all of the monitoring
sites in the Baltimore region. The plot shows the daily RRFs vs. the base case daily maximum
modeled concentrations for all days (above 60 ppb in this case).  In this example, it can be seen
that the model tends to respond more to emissions reductions (lower RRFs) at higher predicted
ozone concentrations.  There appears to be a general pattern in the model results such that the
model  predicts less benefit from  emissions reductions at lower concentrations.  The greater
model  response at higher concentrations is likely due to more "controllable ozone" at higher
concentrations.  On days with high ozone concentrations, there is a relatively high percentage of
locally and/or regionally generated ozone compared to days with low basecase concentrations.
Days with low concentrations likely have a high percentage of ozone due to background and
boundary conditions (background ozone is uncontrollable, although boundary conditions may be
controllable, depending on the size of the domain.) Since we are generally interested in the
model  response on high ozone days, these results tend to suggest that the RRF calculation should
be limited to days when the model predicts high ozone concentrations.
       29See http://www.epa.gov/cair/pdfs/fmaltech02.pdf for documentation of the base case
and future year modeling.

                                           37

-------
1 n1; -, 	
•\ _
u.
o;
g;_
k.
"o n QC; _
Co
u- :
c
o
'•ff
o
3 n q
ฃ
1
- 085
K
0 8
•
+
+ * •
• • .
" X
A
-iปl -
W^ *• ป
^^>Air •
"^C^r*^^"^ + •* •
^r^^^^r-
* ป • ^
•
X
•


• 240030014
• 240030019
240051007
240053001
X 240130001
• 240251001
+ 240259001
. 245100053
	 Power (24003001 4)
Power (240051 007)
	 Power (24003001 9)

60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 135 140
2001 Base Model Value (ppb)
Figure 4.1 - Daily relative response factors as a function of daily maximum base modeled
concentrations for monitors in the Baltimore nonattainment area.

       The analysis examined daily RRFs, but in practice, the RRFs are not calculated on a daily
 basis30. A mean RRF is calculated based on the mean base case concentration (across model
 days) divided by the mean future case concentration (across the same model days). As such, we
 also calculated mean RRFs using various minimum concentration thresholds.  The minimum
 concentration thresholds examined ranged from 70-85 ppb in 5 ppb increments31. The
       30Although it is recommended to calculate mean RRFs, examination of daily RRFs can
 provide important information. Information gleaned from daily RRFs can be used to examine
 model response under different meteorological regimes. Additionally, days that are particularly
 unresponsive to emissions controls can indicate problems in the model inputs and/or formulation
 and may also provide a focus for the development of control strategies. For example,  controlling
 specific upwind sources may provide benefits on particular  days that appear to be otherwise
 unresponsive.

       31It was clear from the plots that a minimum threshold value of less than 70 ppb would
 not be appropriate.  An upper threshold of 85 ppb was examined because it is equal to the
 NAAQS.  We are generally concerned about the model response on days that exceed the

                                           38

-------
monitoring sites were screened to eliminate sites that had a limited number of "high" modeled
ozone days. All sites that had less than 10 days with an 8-hour average maximum modeled
concentration of 85 ppb were dropped from the analysis.  This left 206 sites in the analysis32.

       Table 4.1 shows that the mean RRF (averaged across all sites) is sensitive to the
minimum ozone threshold.  As the threshold is raised from 70 ppb to 85 ppb, the RRFs become
smaller (larger percent decrease in ozone). On average, the model predicts a 0.2% greater ozone
reduction for every 5 ppb increase in the minimum threshold. Additionally, the variability of the
daily RRFs (as measured by the standard deviation of the daily RRFs) is reduced as the threshold
is increased. This is an important finding because lower variability in day to day RRFs indicates
lower uncertainty in the mean RRFs, and hence the attainment test results.
Minimum Threshold
70 ppb
75 ppb
80 ppb
85 ppb
Mean RRF
0.879
0.876
0.874
0.872
Mean Standard
Deviation
0.030
0.029
0.028
0.026
Table 4.1- Mean RRFs and standard deviations as a function of various minimum
thresholds.

       As a result of the sensitivity in model response at concentrations below the NAAQS,
combined with the increased variability of the RRFs at lower concentrations, we recommend
using a minimum concentration threshold of 85 ppb33. This will result in RRFs that adequately
represent high ozone days and also provide for more robust RRFs and future design values. For
a more detailed description of selecting modeling days for ozone, see Section 14.1.1.
NAAQS. But at the same time, it is necessary to have a sufficient number of days in the mean
RRF calculation to develop a robust estimate (this is addressed in more detail in section 14.1.1).
       32'
        The same set of sites was used for each of the threshold concentrations (70-85 ppb).

       33The analysis suggests a threshold of greater than 85 ppb may be appropriate, but 85 ppb
was chosen as the upper end of the threshold because it is equal to the NAAQS. For areas with
very high baseline design values (>110 ppb) it may be necessary to use a higher threshold.
                                          39

-------
Example 4.1

Example 4.1 illustrates how to apply the minimum concentration threshold.

Given: The same simulations as performed in Example 3.2 yield low predictions near site I with
baseline emissions on day 3, such that the 8-hour daily maximum ozone concentration predicted
for that day is 65.0 ppb (rather than the 91.6 ppb shown in Example 3.2).

Find: The relative response factor near site I ((RRF)j).

Solution:  (1) Calculate the mean 8-hour daily maximum ozone concentration obtained near site
I for baseline and future emissions. Exclude results for day 3 from the calculations.  From
Example 3.2,

(a) (mean 8-hr daily max)&ture = (87.2 + 82.4 + 81.1)73 = 83.6 ppb

(b) (mean 8-hr daily max)baseline = (98.3 +  100.2 + 90.7)73 = 96.4 ppb.

(2) Compute the relative response factor by taking the ratio of future/baseline.

(RRF)! = 83.6/96.4 = 0.867

       4.2     How Do I Apply The Recommended 8-Hour Ozone Modeled Attainment
              Test?

       States/Tribes should apply the modeled attainment test at all monitors within the
nonattainment area plus other nearby counties within the State34. Inputs described in Section 3.1
are applied in Equation (3.1) to estimate a future design value at all monitor sites and grid cells
for which the modeled attainment test is applicable.  When determining compliance with the 8-
hour ozone NAAQS, the standard is  met if, over three consecutive years, the average 4th highest
8-hour daily maximum ozone concentration observed at each monitor is < 0.08 ppm (i.e., < 84
ppb using rounding conventions)35. Thus, if all resulting predicted future design values (DVF)
are < 84 ppb, the test is passed.  The modeled attainment test is applied using 3 steps.
       34States are responsible for submitting SIPs for all areas of their State. As such, the
monitored and unmonitored area attainment tests should be applied both within and outside of
the nonattainment area. In the modeling protocol, States should identify the appropriate areas to
apply the respective tests.  States should work with their EPA Regional Office to help determine
the appropriate areas.

       3540CFR Part 50.10, Appendix I, paragraph 2.3

                                           40

-------
Step 1.  Compute baseline design values.  Compute site-specific baseline design values
(DVBs) from observed data by using the average of the design value periods which include
the baseline inventory year.

       This is illustrated in Table 3.1 for specific sites.  The values in the right hand column of
Table 3.1 are site-specific baseline design values.

Step 2.  Estimate relative response factors.  Use air quality modeling results to estimate a
relative response factor for each grid cell near a monitoring site.

       This step begins by computing the mean 8-hour daily maximum ozone concentrations for
future and baseline emissions. The relative  response factor for site I is given by Equation 4.1.

       = (mean 8-hr daily max)&ture/ (mean  8-hr daily max)baseline                         (4.1)

       Using Equation (4.1), the relative response factor is calculated as shown in example 3.2.
Note that the RRF is calculated to three significant figures to the right of the decimal place. The
last significant figure is obtained by rounding, with values of "5" or more rounded upward.  For
the illustration shown in Table 4.2, we have assumed that the same four days described
previously in Example 4.1 have been simulated. Note that on day 3, model baseline 8-hour daily
maximum ozone concentration was < 85 ppb.  As discussed in Section 4.1, predictions for this
day are not included in calculating the mean values shown in the last row of the table.  We have
also assumed that the monitored baseline design value (DVB) at site I is 102.0 ppb.

Step 3. Calculate future design values for all monitoring sites in the nonattainment area.
Multiply the observed baseline design values obtained in Step 1 times the relative response
factors obtained in Step 2.

       In Table 4.2, we see (column (2)) that the baseline observed design value at monitor site I
is 102.0 ppb.  Using Equation (3.1), the predicted future design value for monitor site I is,

       (DVF)! = (102.0 ppb) (0.867) = 88.434 ppb = 88 ppb

Note that the final future design value is truncated36 and in this example, the modeled attainment
test is not passed at monitor site I.
       36This effectively defines attainment in the modeled test as <= 84.999 ppb and
nonattainment as >= 85.0 ppb.

                                           41

-------
Table 4.2 Example Calculation of a Site-Specific Future Design Value (DVF)T
Day
1
2
3
4
Mean
Calculated
baseline design
value, (DVB),,
(ppb)




102.0
Baseline 8-hr daily
max. concentration
at monitor (ppb)
98.3
100.2
65.0
90.7
96.4
Future predicted
8-hr daily max.
concentration at
monitor (ppb)
87.2
82.4
Not Considered
81.1
83.6
Relative
response
factor(RRF),
-
-
-
-
0.867
(i.e., 83.6/96.4)
Future design
value, (DVF)j,
(ppb)
-
-
-
-
88.434=
88 ppb
                                       42

-------
5.0    What Are The Modeled Attainment Tests For The Two PM2 5 NAAQS?

       In this section, we begin by illustrating the monitored tests for the annual and 24-hour
NAAQS. We next identify how to derive needed inputs from available modeled data for the two
modeled attainment tests and highlight some implications that result from the recommended
tests.  We conclude the section by describing a primary PM2 5 analysis which may be used to
better capture the contributions of primary PM2 5 to nonattainment and the benefits of controlling
primary PM2 5 in areas with relatively large contributions from primary sources.
       5.1    What Is The Recommended Modeled Attainment Test For The Annual
             NAAQS?


       As described in section 2, the monitored attainment test for PM2 5 utilizes both PM2 5 and
individual PM2 5 component species. We therefore call the attainment test for PM2 5, the
Speciated Modeled Attainment Test (SMAT).  A separate RRF is calculated for each PM2 5
species. In order to perform the recommended modeled attainment test, States should divide
observed mass concentrations of PM25 into 7 components (plus passive mass):

       - mass associated with sulfates
       - mass associated with nitrates
       -mass associated with ammonium
       - mass associated with organic carbon
       - mass associated with elemental carbon
       -mass associated with particle bound water
       - mass associated with "other"primary inorganic particulate matter
       -and passively collected mass

       To apply the test, States must first have run an air quality model at least twice-to
simulate current emissions and to simulate future year emissions.  We recommend a modeled
attainment test which has 4 basic steps.  Additional details are provided later in this section.

Step 1. Compute observed quarterly mean PM2 5 and quarterly mean composition for each
monitor.

Derive current quarterly mean concentrations37 for each of the major components of PM25  This
is done by multiplying the monitored quarterly mean concentration of Federal Reference Method
       "Concentrations should be calculated based on calendar quarters for two reasons.  First,
the NAAQS is calculated for a calendar year, so the quarters need to fit evenly within a year.
Second, the monitored data used to calculate design values is averaged on a calendar quarter
basis before calculating annual averages.

                                          43

-------
(FRM) (EPA, 1997'a) derived PM25 by the monitored fractional composition of PM25 species for
each quarter, (e.g., 20% sulfate x 15.0 ug/m3 PM25 =3.0 ug/m3 sulfate).

Step 2.  Using model results, derive component-specific relative response factors (RRF) at
each monitor for each quarter.

For each quarter, apply an air quality model to estimate current and future concentrations for
each of the components of PM25. Take the ratio of future to current predictions for each
component. The result is a component-specific relative response factor (RRF).

 The relative response factor for component j at a site i is given by the following expression:

       (RRF)ij = ([Cj, projectedMCj, current]);

       where Cj current is the quarterly mean concentration predicted at or near the monitoring site
       with emissions characteristic of the period used to calculate the baseline design value for
       annual PM2 5

       Cj, projected ls the future year quarterly mean concentration predicted at or near the
       monitoring site.

 (e.g., given model predicted base year sulfate of 10.0 ug/m3 and future year concentration of 8.0
ug/m3' then RRF for sulfate is 0.800).

Step 3.  Apply the component specific RRFs to observed air quality to obtain projected
quarterly species estimates.

For each quarter, multiply the current quarterly mean component concentration (step 1) times the
component-specific RRF  obtained in step 2.  This leads to an estimated future quarterly mean
concentration for each component, (e.g., 3.0 ug/m3 sulfate x 0.800 = future sulfate of 2.4 ug/m3).
Step 4. Calculate a future year annual average PM2 5 estimate.

Sum the quarterly mean components to get quarterly mean PM2 5 values. Then average the
quarterly mean PM2 5 concentrations to get a future year annual average PM2 5 estimate for each
FRM site.

Compare the projected average annual arithmetic mean PM25 concentration obtained in Step 4
with 15.0 i-ig/m3.  If all values are < 15.0 i-ig/m3, the test is passed.

              5.1.1   What Ambient PM25 Data is Used in the Attainment Test?

       PM2 5 data collected  at FRM sites is used for nonattainment designations. Therefore we

                                           44

-------
recommend using FRM data as the base data for projecting future PM2 5 concentrations. As can
be seen from the list of steps, the modeled attainment test is critically dependent on the
availability of species component mass at FRM sites. This raises several issues.  First, the FRM
filter measurements and PM2 5 speciation measurements do not always measure the same mass
(Frank, 2006). There are numerous known issues with positive and negative sampling artifacts.
And second, the majority of FRM sites do not have co-located STN38 speciation samplers39.
Both of these issues are addressed below.
              5.1.2  FRM Monitors that Don't Have Speciation Data

       There are approximately 1200 FRM sites and only -200 urban speciation monitoring
sites.  This makes it difficult to apply the attainment test at the majority of FRM sites.  Species
concentration data and/or species fractions are needed in order to apply SMAT. There are
several possible ways to estimate species concentrations at FRM monitors that lack speciation
data.  Among them are:

1) Use concurrent data from a nearby speciation monitor to estimate species concentrations
and/or fractions at one or more FRM sites.

2) Use representative speciation data (from a different time period) collected in an area to
estimates species data at FRM sites.

3) Use interpolation techniques to create spatial fields using ambient speciation data.

4) Use interpolation techniques to create spatial fields using ambient speciation data and gridded
model outputs to gradient adjust the species concentrations.

       In general, we recommend using spatial interpolation techniques to estimate species
concentrations at FRM sites that do not have speciation data (numbers 3 and 4 above). But in
some  cases, interpolating data from nearby sites may not be appropriate, may  not be feasible, or
simply may not be necessary.  For the CAIR analysis, a relatively simple interpolation technique
was used to estimate species concentrations at all FRM sites in the country. The analysis used a
Voronoi Neighbor Averaging (VNA) technique.  Other interpolations have been done using
Kriging and other more complicated  methodologies. States are encouraged to explore
techniques that are most appropriate for their area and situation.  EPA's MATS software
contains routines to  interpolate data and model output using the VNA technique.
       38References to STN monitors in this document always refer to the overall speciation
network which includes both trends sites and SLAMS sites.

       39There are ~1100 FRM measurement sites and -200-250 urban speciation sites (trends
and SLAMS).

                                           45

-------
       For areas which contain one or more speciation sites, and where FRM sites exhibit very
little spatial gradients, it may be appropriate to assume that the speciation site is representative of
the entire area. Areas which exhibit strong spatial gradients in PM2 5 concentrations should
strongly consider using more sophisticated techniques to estimate species concentrations.
Combining ambient data with model output concentrations (number 4 above) may help adjust
concentrations in areas with strong gradients, but limited speciation data.  This technique has
great promise but has not been used extensively for PM2 5 analyses. The technique is further
limited by uncertainties in the representativeness  of the model outputs and emissions inventories.
             5.1.3   New Species Calculations and Definitions

       Recent data analyses (Frank, 2006) have noted that the FRM monitors do not measure the
same components and do not retain all of the PM2 5 that is measured by routine speciation
samplers and therefore cannot be directly compared to speciation measurements from the
Speciation Trends Network (STN)40.  By design, the FRM mass measurement does not retain all
ammonium nitrate and other semi-volatile materials (negative sampling artifacts) and includes
particle bound water associated with sulfates, nitrates and other hygroscopic species (positive
sampling artifacts). This results in concentrations (and percent contributions to PM2 5 mass)
which may be different than the ambient levels of some PM2 5 chemical constituents.

       The FRM data is used to determine compliance with the NAAQS. Therefore, it is critical
to estimate the species composition as measured on the FRM filters. In addition, for the
purposes of predicting changes in PM2 5 components, constructed PM2 5 mass should match the
composition of mass retained by the FRM. As such, we are recommending changes to the
calculation and definition of PM25  species compared to the  (original 2001) draft PM25 modeling
guidance41.

       We recommend a SMAT technique which uses an FRM mass construction methodology
which results in reduced nitrates (relative to the amount measured by routine speciation
networks), higher mass associated with sulfates and nitrates (reflecting water included in
gravimetric FRM measurements) and a measure of organic  carbonaceous mass which is  derived
from the difference between measured PM2 5 and its non-organic carbon components (Frank,
       40The information in this section applies to the most common samplers in the STN
network. Some networks use alternative special purpose samplers to collect both PM2 5 and
PM2 5 speciation data.  The characteristics of the sampler and the analytical procedures used to
produce chemical speciation data should be considered in determining which, if any adjustments
are appropriate to make the data useful for comparison to FRM data.

       41The draft PM2 5 modeling guidance recommended a mass reconstruction scheme that
was identical to the IMPROVE technique.  We are still recommending use of IMPROVE
assumptions for regional haze modeling (see section 6).

                                          46

-------
2006). This characterization of PM25 mass also reflects "other" primary inorganic PM25 and
other minor constituents. Frank (2006) terms this approach "sulfate, adjusted nitrate, derived
water, inferred carbonaceous material balance approach (SANDWICH)". The resulting
characterization provides a complete mass balance. It does not have any unknown mass which is
sometimes presented as the difference between measured PM2 5 mass and the characterized
chemical components derived from routine speciation measurements. The recommended SMAT
characterizations should yield more accurate assessments of future PM25 concentrations as
measured by FRM monitors.

              5.1.4  Recommended Treatment of Species Data

      As noted above, special treatment of species component mass is recommended in the
SMAT methodology.  The goal is to reconstruct the measured species components so that they
add up to the measured FRM mass.  This concept can generally be represented by the following
equation:

RCFMFRM = [Ammoniated Sulfate Mass] + [Retained Nitrate Mass] + [Retained Carbonaceous
Mass] + [Other Primary PM2 5 ] + [Other Components]                         [5.1]

      In the above characterization, RCFM equals reconstructed fine mass and all of the listed
chemical components reflect those retained during sampling and equilibration on the FRM's
Teflon filter. Sulfate and nitrate mass include associated ammonium which may be different than
assumed ammonium sulfate and ammonium nitrate compounds.  Also, sulfates, nitrates and
carbonaceous mass includes particle bound water associated with these hygroscopic aerosols. In
this characterization, other primary PM2 5 mass is intended to be a more general term that
includes fine soil, and oxides that result from other PM emissions.  The following section
describes the recommended treatment for each of the components.

Retained Nitrate Mass

      The first step in the procedure for identifying FRM mass components is to estimate the
retained nitrate mass on the FRM filters.   The FRM does not capture all  of the semi-volatile
components of the ambient air, such as ammonium nitrate. The retained amount of nitrate ion,
however, can be accurately estimated by  a simple thermodynamic model that involves 24-hr
ambient nitrate speciation concentrations (as measured by a standard speciation sampler using a
nylon filter preceded by a HNO3 denuder) together with hourly ambient  temperature and
humidity. Atmospheric nitrates are higher during the cooler months. Retention on the FRM is
also higher during the cooler months and essentially all the nitrates are lost during the summer.
The retention does not appear to depend on ambient NH3 or HNO3. More NO3 is  retained  at
low temperatures and high humidity which varies by sampling location and time of year.

      Because nitrate retention varies by site and season, an ammonium nitrate equilibrium
model can be used to predict the amount  of nitrates retained on the FRM Teflon filter.  As used
by Hering (Hering,  1999; Zhang, 1992),
                                          47

-------
                                               24
                                                     '/2
              delta NO3 (ug/m3)= 745.7/TR* 1/24* S (K; 2 )                  [5.2]
where delta NO3 is the amount of volatilized nitrate mass, TR is the reference temperature for the
sampled air volume in degrees Kelvin and Ki is the dissociation constant for ammonium nitrate
evaluated at the ambient temperature for hour i.  The nitrate loss can be predicted for each day,
based on hourly temperature and relative humidity data. The SMAT analysis for CAIR used
National Weather Service temperature and relative humidity data for the closest station.  Other
sources of meteorological data may also be appropriate. Further details on the nitrate loss
calculations can be found in (Frank, 2006).

Ammonium Associated with Sulfates and Retained Nitrates

       To determine the mass associated with nitrates, we first assume retained nitrate is
probably all ammonium nitrate. Thus the ammonium associated with nitrates can be derived
directly from the measured or predicted NO3FRM as

                    NH4N03  =0.29*N03FRM                              [5.3]

       The difference between total FRM NH4 (amount associated with nitrates and sulfates),
termed NH4FRM, and the measured NH4, termed NH4STN, is needed to determine the ammoniated
form of sulfates as described by equation 5.1.  Recent measurement study by Collett (Collett,
2004) shows that NH4 may not be completely retained during collection on nylon filters
preceded by a nitric acid denuder. During sampling conditions associated with nitrate
volatilization, ammonium nitrate dissociates but the HNO3 downstream of the denuder is
recaptured on the basic nylon media and the result is reported as particle nitrate.  On the other
hand, the NH4+ volatilizes to gaseous NH3 and apparently passes thru the filter. At several FRM
study sites, the STN NH4 which is adjusted for evaporated NH4NO3 tends to more closely
correspond to the measured NH4 from the FRM Teflon filter. However, for other sites, the
measured STN NH4 appears to agree with FRM NH4.

       In the CAIR analysis, 50% of the ammonium associated with volatilized nitrate was
assumed to also volatilize. But more recent information suggests that using measured
ammonium (assuming none is volatilized) may be a better assumption. We recommend using
measured ammonium, but further analysis of this issue is warranted.

       There are several ways to estimate ammonium mass for use in SMAT.  The most direct
way is to use measured ammonium concentrations from the STN network (IMPROVE does not
measure ammonium ion).

       A second, more indirect method is to calculate the ammonium associated with sulfate and
the degree of neutralization of sulfate (DON), and then use the resulting information to calculate
ammonium mass. Due to uncertainties associated with the ammonium measurements and the
lack of ammonium measurements in rural areas, this indirect method was used for the CAIR

                                          48

-------
analysis. The ambient data is such that all of the ammonium data is from urban sites (STN), but
the sulfate and nitrate data is from both urban (STN) and rural (IMPROVE) sites. This leads to
an overestimation of ammonium concentration in rural areas when ammonium is directly
interpolated. Therefore, in the CAIR analysis, calculated DON, SO4 and NO3FRM were
interpolated to get quarterly average concentrations at each FRM site. The interpolated species
concentrations were then used to calculate NH4FRM using the following equation:

NH4FRM = DON * S04 + 0.29*NO3FRM                                             [5.4]

       The indirect calculation of ammonium mass from interpolated fields tends to smooth out
the gradients in mass.  This was deemed to be beneficial, due to the uncertainty in the
measurements.

Particle Bound Water

       Because ammoniated sulfate and ammonium nitrate are hygroscopic, the retained sulfate
and nitrate mass will include water. Particle bound water (PBW) can be estimated using an
equilibrium model. The CAIR  analysis used the Aerosol Inorganic Model (AIM) (Clegg, 1998)
to calculate PBW. PBW was derived from quarterly average FRM concentrations of sulfate,
ammonium, and nitrate as describe above. Estimated hydronium ion, H+, needed to achieve
ionic balance was derived from the latter values.  The  model enables the distribution of water
and ions to be calculated between liquid, solid and vapor phases for specific temperature and
relative humidity conditions. Typical FRM filter equilibration conditions of 35% RH and 22 deg
C (295 deg K) temperature were used.

       Application of AIM  at the specified FRM filter equilibration conditions show that PBW
is much more dependent on  sulfate concentration compared to nitrate and that the relationship
varies somewhat by season.  There is proportionally less estimated PBW water for wintertime
aerosol which has higher NO3 and lower  SO4.  The PBW concentrations are also sensitive to the
degree of neutralization of the sulfate particles (determined by the relative concentration of
NH4FRM).

       For computational convenience, a polynomial regression equation was fit to the
calculated water mass from AIM and the three input values that fed into AIM (sulfate, nitrate
and ammonium).  A polynomial equation can then be used in all SMAT analyses to estimate
water.

       We recommend calculating PBW as a component of PM2 5 mass. The AIM model (or
other equilibrium models) can be used, or a regression equation can be developed to simplify the
process.

Other Primary PM2 5

        The terms "crustal", "fine soil", "major metal  oxides", "inorganic particulates", and

                                          49

-------
"other PM" are sometimes used interchangeably.  For PM2 5 NAAQS calculations we will refer
to this material as "other primary PM2 5"(OPP). For regional haze calculations, we will continue
to refer to this material as fine soil.

      For the purposes of estimating other primary PM2 5 for SMAT, all measured non-carbon
mass that is not organic in nature (not associated with sulfate and/or nitrate) should be counted.
As with the other PM2 5 components measured on the FRM filter, there is uncertainty associated
with this estimate. The "crustal" or "fine soil" definition from IMPROVE can be used to
estimate other primary PM2 5 42 or an alternative formula can be defined which better estimates
the urban nature of the FRM measurements. The IMPROVE definition of "fine soil" accounts
for the typical crustal components (and attached mass) that would be expected in remote Class I
areas. Fine soil is represented as five elements (Al, Si, Fe, Ti, and Ca) with coefficients to
represent various oxides and material which may  be attached to or associated with the major
elements.  In urban areas, inorganic PM which is  not elemental carbon, or associated with sulfate
or nitrate, may come from many sources such as re-suspended dust or industrial sources (stack or
fugitives). It is generally "crustal" in nature (dominated by silicon), but urban PM is more likely
to contain heavy  metals and industrial components.

      Although the composition of inorganic PM may differ between urban and rural (remote)
areas, we recommend using an equation similar to the IMPROVE fine soil equation to estimate
other primary PM2 5 for the PM2 5 attainment test.  The recommended equation is suggested by
(Frank, 2006) and uses only four elements. The equation removes aluminum (and accounts for
associated mass by increasing the coefficient for Si), due to the fact that aluminum is often
missing from the speciation measurements.  This  allows for more complete data.

The recommended equation is as follows:

Other primary PM25 mass = 3.73 x [Si] + 1.63 x [Ca] + 2.42 x [Fe]  + 1.94 x [Ti]

Blank mass

      The other quantifiable components of PM25 mass include passively  collected mass,
represented by a field blank concentration of 0.3-0.5ug/m3 (U.S. EPA, 2002a).  This appears to
constitute a contamination of the filter resulting from handling or contact with the FRM cassette.
This value is deemed to be an important constituent of PM2 5 mass (it is assumed to not be
dependent on pollutant emissions). We recommend using a default nominal blank mass value of
0.5 ug/m3.  This value can be modified based on local FRM blank mass measurements. The
blank mass is assumed to remain constant through time (RRF=1.0).
       42IMPROVE estimates fine soil as:  2.2 x [Al] + 2.49 x [Si] + 1.63 x [Ca] + 2.42 x [Fe]
+ 1.94 x [Ti]

                                          50

-------
Calculation of Carbonaceous Mass

       Organic carbon mass is typically estimated from blank corrected speciation data, where
organic carbonaceous mass is first estimated by multiplying the organic carbon concentrations
by 1.4 or alternative factors to account for the oxygen, hydrogen and other elements associated
with ambient carbon particles.

       There are many uncertainties in estimating carbonaceous mass from carbon
measurements (Turpin, 2001; Chow, 2004). Uncertainties include differences in carbon
measurement protocol between urban and rural monitoring locations; a relatively "bumpy"
surface of urban carbon concentrations as derived from urban and rural organic carbon
measurements; and lack of carbon measurements at all FRM locations. We therefore
recommend an alternative approach to estimate the organic carbon contribution to PM2 5 mass.

       The recommended SMAT approach estimates organic carbon by mass balance. Precisely
measured FRM PM25 mass (U.S. EPA, 2003a) is compared to the sum of its non-organic carbon
components. The latter  are sulfates, ammonium, nitrates, estimated particle bound water,
elemental carbon, estimated other primary PM2 5 material plus 0.5 ug/m3 blank mass as
discussed earlier.

       This approach estimates retained organic carbon FRM mass  and explicitly accounts for
the following important and difficult to estimate carbon mass properties: (1) regional and urban-
rural differences in the mix of carbonaceous aerosols, i.e. the amount of oxygen, hydrogen, etc;
(2) retained water associated with hygroscopic carbon compounds (Saxena, 1996; Yua, 2004);
(3) volatile carbonaceous material measured by speciation samplers, but not retained in FRM
mass; and (4) uncertainties associated with blank corrections of measured organic carbon.

Organic Carbon Mass by mass balance (OCMmb) is defined as ,

   OCMmb = PM25 - {[S04] + [N03FRM] + [NH4FRM] + [water] + [EC] + [OPP] + [0.5]}   [5.5]

In this expression, all of the above components represent the mass retained on FRM Teflon
filters.

       This approach completely accounts for FRM mass43 and OCMmb is often greater than the
       43The OC by mass balance technique assumes that all other mass is accounted for and
therefore all remaining mass is OCM.  This may not always be a good assumption. The results
of the technique should be carefully evaluated to ensure that OC mass is not overestimated (and
therefore other mass components are underestimated). This may be a problem in areas that do
not have nearby speciation data and have relatively large concentrations of primary PM25.  The
OC by mass balance technique may inadvertently apportion mass to organic carbon which may
actually be EC or "other" primary PM2 5 mass. All available ambient data and modeling data

                                          51

-------
amount that would be derived directly from speciation measurements. Because of uncertainties
in speciation measurements and their estimates from interpolated surfaces, setting a lower limit
(floor) may be necessary so that the OCMrnb is not unreasonably low. In the CAIR analysis, the
floor was set so that OCMmb could not be more than 30% lower than measured OCM (OC*1.4).
We used interpolated measured values of OCM to calculate the floor. The lower limit was equal
to interpolated OC * 1.4 * 0.7. If the OCMmb concentration was less than the lower limit, it was
set equal to the lower limit. The recommended approach is to simply set the OC floor equal to
measured OC.

       There may also be situations where an OCM "ceiling" is needed. In remote urban areas
with relatively high FRM concentrations that may be surrounded by rural background
concentrations, the OC by mass balance technique may apportion 95% or more of the PM25 mass
to OCM. If this is not a reasonable assumption, then a ceiling may be needed to cap the OCM as
a percentage of PM25 mass. Based on measured data (FRM sites with co-located speciation
data), it appears that on a quarterly average basis, OCM is rarely more than 80% of total PM25
mass. This may be a reasonable  default ceiling, but a lower value (or in rare circumstances a
higher value) may be more appropriate in many regions of the country.

Summary of PM0 < Composition Calculations

       The terms of equation 5.6 reflect the final estimated composition of the particles
measured by the FRM (for each quarter).  Quarterly average FRM mass is equal to the sum of
the  seven species plus blank mass.
 PM25FRM = { [OCMmb] + [EC] + [SO4] + [NO3FRM] + [NH^J + [water] + [OPP] + [0.5] }
[5.6]

The recommended order to generate the data is as follows:

       1) Calculate adjusted nitrate using hourly meteorology and 24-hour average nitrate
       measurements.

       2) Calculate quarterly averages for adjusted nitrate, sulfate, elemental carbon, ammonium
       (or degree of sulfate neutralization (DON)), OPP mass , and measured OCM.44

       3) Quarterly average ammonium is calculated from the adjusted nitrate, sulfate, and DON
       values (if measured ammonium is not used directly).

       4) Calculated ammonium, sulfate, and nitrate values are input into the polynomial water
       equation to derive particle bound water concentrations.
should be used to evaluate the species apportionment results.

      44The measured OCM is only used to calculate the "floor" for OCMmb

                                          52

-------
       5) Carbon mass by difference (OMCmb) is calculated from the PM2 5 mass, adjusted
       nitrate, ammonium, sulfate, water, elemental carbon, other primary PM2 5, and blank mass
       values. The sum of the 7 species plus blank mass is equal to the FRM mass.

We illustrate application of the recommended test in example 5.1.

Example 5.1

Given: (1) Area "C" has 2 monitoring sites.

(2) Monitored FRM air quality data show the following average quarterly mean PM2 5
concentrations based on a 5 year weighted average of observations from 2000-2004 at each site
(values are in |ig/m3).

Table 5.1
Site
1
2
Quarter 1
17.21 ug/m3
15.39
Quarter 2
16.34
17.98
Quarter 3
20.30
18.23
Quarter 4
14.76
13.76
(3) The area has 1 urban speciation site which is co-located with FRM site 2.  The speciation
data is only available for 2003 and 2004. Therefore, the species data for the two years has been
interpolated to derive estimated species concentrations at each site.  The species concentrations
(derived from 2003 and 2004 data) are matched up with FRM data for 2003 and 2004 to derive
representative species fractions.
Average monitored air quality for the 3rd quarter of 2003 and 2004 at site 1 is as follows
                                                                               45.
 Table 5.2
FRM
Mass
(ug/m3)

22.57
Blank
Mass
(ug/m3)

0.5
Non-blank
Mass
(ug/m3)

22.07
Sulfate
(ug/m3)


8.51
Nitrate
(ug/m3)


1.11
Organic
Carbon
Mass
(ug/m3)
5.21
Elemental
Carbon
(ug/m3)

0.91
Water
(ug/m3)


2.31
Ammonium
(ug/m3)


3.31
OPP
(ug/m3)


0.71
The blank mass is subtracted before species fractions are calculated because the blank mass is
held constant at 0.5 ug/m3 throughout the analysis. In the example, the measured FRM mass for
quarter 3 is 22.57 ug/m3.  The non-blank FRM mass is 22.07 ug/m3. The mass of the seven
       45The species concentrations can be derived from co-located speciation data, nearby
speciation data or from interpolated spatial fields. In this example, FRM mass is not
interpolated.
                                           53

-------
species add up to the non-blank mass.
(4) Species fractions are calculated for each quarter for each species.  In the example below, a
fraction of non-blank mass is calculated for each of the seven species.  Blank mass remains fixed
at 0.5 ug/m3.

Table 5.3
FRM
Mass
(ug/m3)
22.5
Blank
Mass
(ug/m3)
0.5
Non-blank
Mass
(ug/m3)
22.0
%
Sulfate

38.6
%
Nitrate

5.0
%
Organic
aerosol
23.6
%
Elemental
Carbon
4.1
%
Water

10.5
%
Ammonium

15.0
%
OPP

3.2
The percentages in table 5.3 above are the relative composition for the 3rd quarter of 2003 and
2004. It is assumed that these species fractions are representative of the 2000-2004 time
period46.

(5) The weighted quarterly average FRM design values are used as the baseline FRM value for
each monitoring site (2000-2004). The species fractions from the 2003/2004 speciation data
were used to estimate the species concentrations for the baseline year FRM PM2 5 data. The
percentage compositions for 2003/2004 are applied to the quarterly weighted average design
values as shown in table 5.4. In the example below, the weighted average design value for the
3rd quarter for the site from  table 5.1 is 20.30 ug/m3.  This leads to the following concentrations
of PM25 species:

Table 5.4 Calculation of the "current" species  concentrations
Weighted
Avg.
FRM
Mass
(ug/m3)
20.30
Blank
Mass
(ug/m3)


0.5
Non-blank
Mass
(ug/m3)


19.8
Sulfate
(ug/m3)



7.64
Nitrate
(ug/m3)



0.99
Organic
aerosol
(ug/m3)


4.67
Elemental
Carbon
(ug/m3)


0.81
Water
(ug/m3)



2.08
Ammonium
(ug/m3)



2.97
OPP
(ug/m3)



0.63
       This procedure is repeated for each PM2 5 site and quarter to complete the calculation of
current (baseline) ambient concentrations used as the basis for future estimates of PM2 5 mass
and its components.
       46If there are less than 5 years of speciation data, then an assumption of
representativeness needs to be made. It is generally assumed that the inter-annual variability of
the species fractions is small compared to the variability of species concentrations. In the near
future, 5 full years of concurrent FRM and speciation data will be available.
                                           54

-------
(6) Modeled results show the following relative response factors (RRF) in predicted mass of 5
components of PM25for the 3rd quarter:

Table 5.5
RRF
Sulfate
0.876
RRF
Nitrate
0.943
RRF
Organic
aerosol
0.971
RRF
Elemental
Carbon
0.932
RRF
OPP
1.042
(7) The quarterly mean RRFs from table 5.5 are multiplied by the weighted quarterly average
species concentrations from table 5.4 to derive future year concentrations.

From the example above, the future year 3rd quarter concentrations are:

SulfateFuture = 7.64 * 0.876 = 6.69 ug/m3
NitrateFuture= 0.99 * 0.943 = 0.93 ug/m3
Organic carbon massFuture= 4.67 * 0.971  = 4.53 ug/m3
Elemental CarbonFuture= 0.81 * 0.932 = 0.75 ug/m3
OPPFuture= 0.63 *  1.042 = 0.66 ug/m3

(8) The future year concentrations derived in step 5 are used to calculate the future year
concentration of ammonium (if the direct ammonium RRF is not used) and particle bound water.

       The future year ammonium concentrations are calculated from the sulfate, nitrate, and
(current year) DON values.  Assuming that the DON is unchanged from the current year47, the
ammonium is calculated using the following formula:
NH4fi,,,rp = DON * S04
    Tfuture
                      foture
0.29*NO3
                                     futures
In the example above, assuming the base year DON is 0.336,
       47In the CAIR analysis, the DON was assumed to stay constant through time due to the
uncertainty in the ammonium measurements. The water calculation is sensitive to the ammonium (and
therefore the DON value) concentrations. Keeping the DON constant allows for the future year
ammonium and water values to be solely a function of the change in sulfate and nitrate concentrations.
Otherwise, the water concentration can go up when the sulfate and nitrate concentrations go down. This
may occur if sulfate becomes more neutralized in the future. It is a somewhat illogical outcome (although
scientifically possible) and is highly dependent on an uncertain measurement (ammonium).  Therefore,
use of a constant DON creates a more stable set of calculations. If the measured and modeled ammonium
concentrations are believed to be accurate and respond in a reasonable way to emissions controls, then it
would be more scientifically credible to use the model predicted change in ammonium.  Otherwise, it is a
reasonable assumption to keep the DON constant over time.
                                            55

-------
AmmoniumFuture= 0.336 * 6.69 + 0.29 * 0.93 = 2.52 ug/m3

       The NH4&ture, SO4foture and NO3future concentrations can then be input into an equilibrium
model (AIM or another alternative model) or through a polynomial equation to predict future
year particle bound water concentration.

(9) The future species concentrations at each FRM site are then summed over the seven species
plus blank mass to estimate the future quarterly average PM2 5 concentration.

In the example above, the total PM2 5Future=
6.69 + 0.93 + 4.53 + 0.75 + 0.66 + 2.52 + 1.73 + 0.5 = 18.31 ug/m3
(assuming that the future year water concentration is equal to 1.73 ug/m3)

(10) The same calculations are completed for the other 3 quarters to get a future year PM2 5
concentration for each quarter.  The 4 quarters are then averaged to get a final future year annual
average PM2 5 concentration for each FRM site.

(11) The future year annual average concentration is compared to  15.0 ug/m348.
       5.2    What Is The Recommended Modeled Attainment Test For The 24-Hour
             NAAQS?

       Our recommended modeled attainment test for the 24-hour NAAQS for PM2 5 is similar
to the previously described test for the annual NAAQS in that it uses model predictions in a
relative sense to reduce site-specific observations (averaged over 5 years).  In the test, we are
interested in reducing the current design values at each site to < 65 |ig/m3 49'50.
       48In the final step, the future year concentration should be rounded to the tenths digit. A
(rounded) value of 15.0 ug/m3 meets the NAAQS. A value of 15.1 ug/m3 or greater violates the
NAAQS.

       49PM2 5 nonattainment areas are not specifically designated for the annual or 24-hour
standard. Therefore, the attainment demonstration must show that the area will meet both forms
of the standard. Areas that only violate the annual standard and whose design value is well
below the 24-hr standard may not need to apply the 24-hour modeled attainment test.
Attainment test requirements should be discussed with the appropriate EPA Regional Office.

       50The PM25 24-hour standard was lowered to 35 ug/m3 in 2006 (71 FR 61224) (40 CFR
50.13). Planning requirements for areas that violated the old standard are unchanged. We
expect the attainment test for the new standard to be the same or similar to the procedures in this
version of the modeling guidance.  As part of the implementation process for the revised
standard, the procedures in the guidance will be re-evaluated. If necessary, we will release a
revised modeling guidance document that addresses  any changes to the 24-hour test for the new

                                           56

-------
       Ideally, the modeled attainment test should reflect model results obtained for days in
each season having observed PM2 5 concentrations slightly below as well as above the design
value.  This may seem strange at first.  The underlying reasons are that PM25 consists of a
mixture of pollutants.  Composition of the mixture could vary substantially from season to
season. Second, there could be a substantial amount of uncertainty associated with predictions
on any single day. Thus, our test is most likely to be reliable when relative response factors
reflect composite responses from many days.  Therefore, we recommend modeling as many days
as feasible where observed PM2 5 is greater than 65 |ig/m3.   Alternatively, the test can focus on
the high end of the distribution of days in each quarter, (e.g. the top 25% of PM25 days)
assuming that many of the high days violate the NAAQS (or are representative of days that
violate the NAAQS). As with the annual NAAQS (and for the same reasons), the preferred
approach is to develop relative response factors which are season (i.e., quarter) specific.

       We have noted that it is desirable to base our conclusions on a composite response of the
model(s) over many days. However, there is not likely to be many days with observed
concentrations greater than 65 |ig/m3. If this results in a sample size of smaller than about 5 days
per quarter then the analysis should focus on the high end of the distribution of PM25 days in
each quarter (e.g. the top 25% of PM25 days51).

       The 24-hour attainment test should be based on the same 5 year weighted average
methodology that was used for the annual  standard, with some slight modifications.  The 24-hour
design values are calculated from the 98th percentile value for each year. We recommend
applying the attainment test on an annual basis and then averaging the annual average results up
to 3 year averages and then averaging again to get a 5 year weighted average.

       PM25 is a complex mixture of components that have significant seasonal variations. If the
observed 98th percentile value  came from the 1st quarter, it is possible that an observed
concentration from a different quarter, which may have a slightly lower concentration, could
become the 98th percentile value in the future. It is important to demonstrate that the standard
will be met for all seasons.  Therefore we recommend applying the test by quarter for each
year52.
standard.

       51The top 25% of days may seem like too many days per quarter to use for a 98th
percentile based standard, but for sites with a 1 in 6 day sampling schedule meeting minimal
completeness criteria (assuming 11 samples per quarter), the top 25% of days is only 3 days per
quarter. For most sites, the top 25% of monitored days per quarter will represent between 3 and
8 days.

       52In some areas it may not be necessary to model and evaluate the NAAQS for all
quarters. For example, if PM25 concentrations only exceed the NAAQS in the  1st and 4th
quarters, and concentrations in the 2nd and 3rd quarters are very low, then it may not be necessary

                                           57

-------
       Similar to the annual PM2 5 attainment test, we recommend interpolation techniques for
FRM monitors that do not have co-located speciation data. Because the 24-hour standard is a
98th percentile based value, the component composition on high concentration days may be
highly variable from day to day and from site to site. Therefore, while interpolation techniques
may need to be used, we strongly recommend collecting speciation data at all FRM sites that
violate the 24-hour NAAQS. A precise estimate of the PM2 5 components at violating sites will
help reduce uncertainty in the future year concentration estimates.

We recommend a modeled attainment test for the 24-hour PM2 5 NAAQS with 6 steps.

Step 1. Compute observed 98th percentile values for each year and next highest
concentrations for each quarter.

 The first step in projecting the daily design value is to identify the maximum daily average
PM2 5 concentration in each quarter that is less than or equal to the annual 98th percentile value
over the entire year. This results in data for each year (for five years) for each site which
contains one quarter with the 98th percentile value and three quarters with the maximum values
from each quarter which are less than or equal to the 98th percentile value53.

Step 2. Identify the species composition for each quarter for each monitor.

These quarterly PM2 5 concentrations are then separated into their component species by
multiplying the quarterly maximum daily concentration at each site by the estimated fractional
composition of PM25 species, by quarter and year, based on observed species fractions for the
days > 65 ug/m3 or from the days in the high end of the distribution of PM25 concentrations
(using the same methodology as used in the annual average calculations54) (e.g., 30% nitrate  x
70.0 ug/m3 PM2 5 =21.0 ug/m3 nitrate).
to model the full year. But for areas that have monitored violations (or high values that are close
to the NAAQS) in all 4 seasons, the entire year should be evaluated.

       53The test should be performed for each monitoring site that meets the data completeness
criteria for calculating a 98th percentile value under the 24-hr NAAQS. The highest
concentration from the other 3 quarters (less than or equal to the 98th percentile) is taken from
any values that are measured at the site. It is necessary to examine the concentrations in other
quarters which are less than the design value because PM species may have different responses
to emissions controls in different parts of the year.  Basecase concentrations that are less than the
design value concentrations may end up being higher than the design value concentration in the
future (after RRFs are applied).  There may not always be data available for all four quarters.

       54The ambient speciation data for the 24-hour attainment test should use the same
adjustments that were recommended for the  annual test.  The adjusted data should be carefully
examined.  Adjustments to daily concentrations of speciation data (used for the 24-hr test) may
in some cases be large compared to adjusted quarterly average data.

                                           58

-------
Step 3.  Using model results, derive component-specific relative response factors (RRF) at
each monitor for each quarter.

For each quarter, apply an air quality model to estimate current and future concentrations for
each of the components of PM25. Take the ratio of future to current predictions for each
component on the modeled days with PM2 5 > 65 ug/3 or the high end of the distribution  of
modeled days. The result is a component-specific relative response factor (RRF). (e.g., given
model predicted base year nitrate of 20.0 ug/m3 and future year concentration of 15.0 ug/m3' then
RRF for nitrate is 0.750).

 The relative response factor for component j at a site i is given by the following expression:

       (RRF)ij = ([Cj, projectedMCj, current]);

       where Cjj current is the mean concentration (for the high modeled days for each quarter)
       predicted at or near the monitoring site with emissions characteristic of the period used to
       calculate the baseline design value for 24-hour PM2 5

       Cj, projected is the future year mean concentration (for the high modeled days for each
       quarter) predicted at or near the monitoring site with emissions characteristic of the
       future year period.

Step 4.  Apply  the component specific RRFs to observed air quality to obtain projected
quarterly species estimates.

For each quarter, multiply the quarterly component concentration (step 2) times the component-
specific RRF obtained in step 3. This leads to an estimated future quarterly concentration for
each component, (e.g., 21.00 ug/m3 nitrate x 0.750 = future nitrate of 15.75 ug/m3).
Step 5. Sum the quarterly components to get quarterly "potential" 98th percentile PM2 5
values.

Sum the quarterly components to get quarterly "potential" 98th percentile PM2 5 values.  Repeat
this procedure for each quarter and for each of the 5 years of ambient data. The highest daily
value (from the 4 quarterly values) for each year at each monitor is considered to be the
estimated 98th percentile value for that year.

Step 6.  Calculate future year 5 year weighted average 24-hour design values and compare
to the NAAQS.

The estimated 98th percentile values for each of the 5 years are averaged over 3 year intervals
to create design values and then the 3 design values are averaged to create the 5 year weighted
average for each monitor. Compare the projected 24-hour design values to 65 ug/m3.


                                           59

-------
The recommended test is illustrated in example 5.2. In the interest of brevity, some of the steps
are only illustrated for quarter 1.  The actual calculations would also use analogous data for the
other three quarters.

Example 5.2

Given: (1) Area "D" has 1 monitoring site. The 98th percentile values for 5 years of observed
       data are as follows (year 3 is being modeled for the attainment test):

Year 1-80 ug/m3
Year 2- 68 ug/m3
Year 3-71 ug/m3
Year4-58ug/m3
Year 5- 74 ug/m3

Step 1- Compute observed 98th percentile values for each year and next highest
concentrations for each quarter.

The maximum quarterly values from the 5 years which are equal to or less than the 98th
percentile concentration are as follows:

Quarter 1
Quarter 2
Quarter 3
Quarter 4
Year 1
80.0 ug/m3
67.6
41.7
52.8
Year 2
65.0
45.5
47.6
68.4
YearS
57.0
46.4
32.5
71.3
Year 4
58.0
45.3
39.4
49.5
YearS
74.0
52.2
45.3
64.2
                                          60

-------
Step 2- Identify the species composition for each quarter for each monitor.

The mean species composition55 for the observed high PM2 5 days in quarter 1 is:

Nitrate- 30%
Sulfate- 10%
Organic carbon- 23%
Elemental carbon- 5%
Other primary PM2 5- 5%
Ammonium- 15%
Water- 12%

This leads to the following baseline concentrations for quarter 1 (ug/m3):

Year 1
Year 2
YearS
Year 4
YearS
PM25
79.5
64.5
56.5
57.5
73.5
Nitrate
23.85
19.35
16.95
17.25
22.05
Sulfate
7.95
6.45
5.65
5.75
7.35
OCM
18.29
14.84
13.00
13.23
16.91
EC
3.98
3.23
2.83
2.88
3.68
OPP
3.98
3.23
2.83
2.88
3.68
Ammonium
11.93
9.68
8.48
8.63
11.03
Water
9.54
7.74
6.78
6.90
8.82
Step 3- Using model results, derive component-specific relative response factors (RRF) at
each monitor for each quarter.

The following are RRFs for the 1st quarter derived from the modeled baseline and future year
concentrations for the high PM2 5 days.

RRF
Nitrate
0.942
Sulfate
0.843
OCM
0.917
EC
0.930
OPP
1.040
       55The percentages are derived from several steps in the processing. The nitrate, sulfate,
elemental carbon, and other primary PM2.5 are derived from measured or interpolated data.  The
ammonium is calculated from interpolated sulfate, nitrate, and DON (or directly interpolated
ammonium).  Water is calculated from ammonium, nitrate, and sulfate using a model or
equation. The OMC is calculated by mass balance.
                                          61

-------
Step 4- Apply the component specific RRFs to observed air quality to obtain projected
quarterly species estimates.

The quarter 1 concentrations for each year are multiplied by the component RRFs.  Ammonium
concentrations are calculated from the future year sulfate, nitrate, and (baseline) DON value.
Water concentrations are calculated from the future year nitrate, sulfate, and ammonium
concentrations, using an empirical equation.

Yearl
Year 2
Year3
Year 4
YearS
Nitrate
22.47
18.23
15.97
16.25
20.77
Sulfate
6.70
5.44
4.76
4.85
6.20
OCM
16.77
13.60
11.92
12.13
15.50
EC
3.70
3.01
2.63
2.68
3.43
OPP
4.14
3.36
2.94
3.00
3.83
Ammonium
10.41
8.45
7.40
7.53
9.62
Water
8.41
6.83
5.98
6.09
7.78
Step 5. Sum the quarterly components to get quarterly "98th percentile" PM2 5 values.

The year 1 PM25 concentration is equal to the sum of the 7 components plus 0.5 ug/m3 blank
mass:
Year 1= 22.47 + 6.70 + 16.77 + 3.70 + 4.14 + 10.41 + 8.41 + 0.5 = 72.6

Filling in the quarter 1 values for the other 4 years, plus the other 3 quarters for each year; the
full matrix of quarterly values is as follows:

Quarter 1
Quarter 2
Quarter 3
Quarter 4
1999
72.6 ug/m3
57.5
38.8
46.6
2000
58.9
38.7
44.3
60.5
2001
51.6
39.4
30.2
63.0
2002
52.5
38.5
36.6
43.6
2003
67.1
44.4
42.1
57.1
Step 6.  Calculate future year 5 year weighted average 24-hour design values and compare
to the NAAQS.

In the example, the highest quarterly value for each year in step 5 represents the future year
                                           62

-------
annual 98th percentile. Therefore, the 3 year average design values are:

DV1 = (72.6 + 60.5 + 63.0)/3 = 65.4
DV2 = (60.5 + 63.0 + 52.5)73 = 58.7
DV3 = (63.0 + 52.5 + 67.1)73 = 60.9
And finally, the 5 year weighted average is equal to the average of the 3 future year design
values:

5 year weighted average future DV= (65.4 + 58.7 + 60.9)73 = 61.7 = 62 ug/m3
       5.3     Special Considerations for Primary PM2 5

       Primary particulate matter does not undergo chemical transformation between
its being emitted and its arrival at a receptor location.  Thus, a relatively lengthy travel time from
source to receptor (to enable thorough mixing and chemistry to proceed) is not needed for high
concentrations of primary particulate matter to occur.  Therefore, unlike secondary particulate
matter, we would often expect concentrations of primary particulate matter to increase the closer
one gets to its source(s) of emissions.  Therefore, if a designated nonattainment area contains a
few (as opposed to many which are spread out) concentrated sources of primary particulate
matter, we would expect there to be some substantial spatial gradients in the primary portion of
the organic carbon component and in the inorganic particulate matter (OPP) and elemental
carbon (EC) components of ambient PM2 5 (these are often called "hotspots").  Substantial
gradients are most likely to be a potential problem in addressing whether a proposed control
strategy is sufficient to attain the 24-hour NAAQS for PM2 5. This follows, because orientation
of a source's plume varies. For many hours, one might expect to find no impact from such a
source at a given location.  Occasions with no likely impact tend to balance those large impacts
occurring over some periods.

       The PM25 NAAQS rulemaking (U.S. EPA 1997a) as well as ambient monitoring rule
language56 indicates that high PM2 5 concentrations occuring in "unique" population oriented
areas should not be compared to the annual NAAQS and should only be compared to the 24-hour
NAAQS (especially when caused by a single source). Furthermore, high PM25 concentrations
measured in non-population oriented areas should not be compared to either of the PM2 5
NAAQS. PM25 monitor siting guidance (U.S. EPA, 1997b) defines the appropriate scales of
influence for the PM2 5 monitoring network.

       High primary PM concentrations can occur at (or near) existing monitors or in
unmonitored areas. The modeled attainment test is primarily a monitor based test. As such, the
       5640 CFR part 58 subpart D 58.30

                                           63

-------
focus of the attainment test is whether attainment can be reached at existing monitors.  To
address the issue of PM25 concentrations in unmonitored areas, we have recommended an
"unmonitored area analysis" (see section 3.4).  The unmonitored area analysis is intended to be
the primary means for identifying high PM2 5 concentrations outside of traditionally monitored
locations.  The spatial resolution of the modeling that is the underlying basis of the unmonitored
area analysis will determine how well it addresses primary PM hotspot issues.  The finer the
resolution of the grid model, the more likely that primary PM hotspots will be recognized.
Based on the monitoring guidance, we believe that an unmonitored area analysis conducted at
12km or finer resolution is sufficient to address unmonitored PM2 5 for the annual NAAQS.
Conducting the unmonitored analysis at 4km or finer resolution will provide an even more
detailed analysis of the spatial gradients of primary PM25, especially when evaluating violations
of the 24-hr. NAAQS.

              5.3.1   Analysis of Primary PM25 Impacts at Monitors

       Most PM2 5 monitors are located  at "urban scale" sites. These sites tend to represent
relatively large spatial scales and do not have large gradients compared to other monitors in the
area.  Some sites are classified as "neighborhood" scale and may be influenced by more local
sources. For reasons stated above, local influences creating spatial gradients are likely to consist
mostly of primary PM2 5  (OC,  EC, and other primary particulates). These sources may be local
point sources, or they may be nearby roads or other mobile or area sources.

       It may be necessary to  evaluate the impact of local primary PM sources for contributions
to both the 24-hour and annual NAAQS. As stated earlier, it may not be appropriate to compare
population oriented FRM sites that are dominated by point sources, to the annual NAAQS.  But
there are numerous cases where the impact from local sources is not dominant, but a sizable
contributor to total PM25 (-10-30% of total annual average PM25). In these cases, a more
refined analysis of the contribution of local primary PM25 sources to PM25 at the monitor(s) will
help explain the causes of nonattainment at the monitor and may lead to the more efficient ways
to attain the standard by  controlling emissions from local sources.

       There are several modeling tools that can be used to evaluate contributions of local PM25
sources. A grid model can be  run at very high resolution (down to 1 or 2 km) or a Gaussian
dispersion can be used. Grid based models usually simulate chemical transformation and
complex meteorological  conditions, while dispersion models are generally more simplistic;
being limited to a local-scale and using Gaussian approximations with little or no chemistry.
Therefore, while dispersion models may not be an appropriate tool for determining secondary
PM2 5 concentrations, they work well for use in determining local primary PM2 5 impacts in a
small area. The model(s) and  model  setup should be evaluated to determine the most appropriate
tools for a specific situation.

       Regardless of which type  of models are used to evaluate changes in primary PM25 at
monitors, we recommend that  the model results be used in a relative manner.  This is consistent
with the already described attainment test. If a grid model is used at very high resolution, the

                                           64

-------
attainment test as described in Section 5.1 should be followed. If a Gaussian dispersion model is
used then the application of the attainment test will vary slightly. The test will need to combine
results from the photochemical grid based modeled attainment test and the results from local-
scale dispersion modeling. If such an approach is taken, the suggested variations on the
attainment test are discussed in the following section.
              5.3.2  Analysis of Primary PM2 5 Impacts at Monitors using a Gaussian
                    Dispersion Model

       To apply a dispersion model in an attainment test, it becomes important to determine the
local component of primary PM2 5 at the monitor and the sources that are contributing to that
component.  There is no single, simple method for quantifying this contribution, but detailed
analysis of ambient data and advanced data analysis techniques, such as receptor modeling, may
help quantify the contribution. The simplest method for identifying the local component of
PM2 5 is to examine local ambient PM2 5  concentrations. For this analysis, it is important to
identify the local contributions from as small an area as possible.  This will make it easier to
identify contributing sources.  It is most appropriate to examine monitored concentration
differences between urban monitors (with the highest concentrations) and more suburban
measurements.  This is likely to be representative of the excess contribution from a relatively
local area. "Urban excess" calculations  which pair an urban monitor with a rural background
monitor (U.S. EPA, 2004b) are more likely to indicate "local" contributions that may be more
representative of an entire metropolitan area. In most cases, the local component will include
contributions from more than one source.

       Sources identified as contributing to the monitor will be modeled with a dispersion model
(or alternatively, a very high resolution grid model) .  It is common practice to run dispersion
models for limited numbers of point sources in relatively small areas (out to a distance of-50
km). Dispersion models can also be run with all sources of primary PM in a limited area
(including mobile, non-road, and area sources).

       When applying the model to evaluate changes in primary PM2 5, one should determine the
PM2 5 species that make up the local primary particulate matter57.  We recommend that the
individual components of primary PM25 are tracked separately in the dispersion model. This will
allow more thorough evaluation of the dispersion modeling results (comparing model output to
speciated ambient data) and may aid in the development of primary PM25 control strategies.

       The majority of the primary PM will consist of the primary portion of the organic carbon
       "The PM2 5 emissions should be broken out into individual species using the best
information available. Typically, SCC specific speciation profiles are used to "speciate" the PM
into individual components. Local, source specific information should be used whenever
possible.

                                           65

-------
component, elemental carbon (EC), and the general category of "other primary particulate
matter" (OPP). In some cases, directly emitted sulfate (and in rare cases nitrate) may also be a
significant component of the local primary PM25.

       The dispersion modeling results should be evaluated to ensure adequate model
performance. Similar to grid modeling, the dispersion model results should be compared to
ambient data to ensure that the model is working well. Although section 18 of this guidance is
geared towards evaluating grid models, many of the same statistical calculations can be made for
primary PM2 5 and PM2 5 components predicted by a Gaussian dispersion model.  Since
secondary PM25 is often a large component of total PM2 5 concentrations, it may be difficult to
separate the primary and secondary components of ambient PM2 5. EC and OPP should be
considered to be  primary PM2 5. Much of the rest of the  primary PM2 5 concentration will be
primary OC.

       As part of the analysis, an estimated concentration of primary OC is needed. There are
several methods available for estimating the primary vs. secondary portion of ambient OC.
Among these are the EC tracer method and receptor modeling.  The EC tracer method is the
most common method used to estimate secondary and primary OC concentrations (Turpin,
1995), (Strader, 1999) (Cabada, 2004), (Chu, 2005), (Saylor, 2006) .  The method uses
measurements of OC and EC and calculated OC to EC ratios to identify periods when OC is
likely to be mostly primary.  This information  is then used to calculate the secondary
contribution to OC .  Receptor models such as CMB and PMF have also been used to estimate
secondary organic concentrations (Na, 2004), (Yuan,  2006).

       The following sections discuss a suggested methodology for examining decreases or
increases in PM concentrations  due to local primary sources for the Annual and 24-hour PM2 5
NAAQS. Because each nonattainment area has unique emissions sources and source-receptor
relationships, States  should work closely with their EPA  Regional Office in developing local
area analysis applications.

Methodology for the Annual PM2 5 NAAQS

The suggested  steps  outlined below assume that the attainment test has been performed using the
photochemical modeling and that the annual DV has been computed as discussed in Section 5.1.

(1) Identify local primary PM2 5 sources that are thought to be are contributing to the monitor
and causing non-attainment of the annual NAAQS.

(2) By quarter, identify at the monitor, primary versus secondary components of
PM25.

(3) Identify the percentage of total primary PM25 at the monitor due to the local sources
identified in (1).  Convert the percentage to an estimated  concentration in ug/m3
                                          66

-------
(4) Model with a dispersion model the current and future concentrations of these
local sources using actual emissions.
       -Place receptors at and near the monitor location(s)58.
       -The receptor spacing should be fine enough to get an accurate picture of the
       concentration at the monitor.
       -Model the base year and future annual average
       -For both base and future year use the same 1 year of met data - the
       same year as used for the photochemical modeling (if available, use more than 1 year of
       met data).

(5) From the modeling in (4), calculate quarterly RRFs59 for total
primary PM2 5 modeled for the selected sources (using concentration values
at receptors at monitors).

(6) Multiply the primary  PM2 5 identified in (3) by the RRFs from (5).

(7) Subtract the result in (6) from the primary PM2 5 identified in (3)

(8) For each quarter, subtract the amount in (7) from the future year quarterly DV provided by
the photochemical modeling.

(9) Average the four quarters to get an annual average DV.
Methodology for the 24-Hour NAAQS

The suggested steps outlined below assume that the attainment test has been performed using the
photochemical modeling and that the quarterly DV's have been computed as discussed in
Section 5.2.  As with the annual design value, monitoring data for 5 years are used as the basis
for the project!on.

(1) Determine the local primary PM2 5 sources that are thought to be contributing to the monitor
for each of the 4 quarters60.
       58Additional receptors should be placed near monitors to examine predicted concentration
gradients. Receptors should only be located in areas that are appropriate for placing FRM
monitors for comparison to the annual NAAQS (see EPA, 1997b).

       59RRFs should be calculated as a ratio of the base and future mean concentrations (a ratio
of the means, not a mean of the daily ratios).

       60The selection of these sources should be representative of the 5-year period used to
calculate the 5 year weighted average design values. If large emissions changes have occurred
over the 5 year period then it may be necessary to determine the local contribution of identified

                                           67

-------
(2) For each of the four representative quarters, determine the primary versus secondary
components of PM2 5 at the monitor on the 98th percentile day and all high days.

(3) Identify percentage of total primary PM25 at monitor due to the local sources
identified in (1). Convert the percentage to an estimated concentration in ug/m3.

(4) Model with a dispersion model the current and future concentrations of these
local sources using actual emissions that are representative of emissions on high PM25 days61.
       -Place receptors at and near the monitor location(s)62.
       -The receptor spacing should be  fine enough to get an accurate picture of the
       concentration at the monitor.
       -Model the base and future year.
       -For both base and future year use the same 1 year of met data used for the
       photochemical modeling (if available, use more than  1 year of met data).
       -Calculate the 24-hour average for each day. Average the high days63 within each quarter
       to determine a modeled high concentration for each quarter in the base and future year.

(5) From the modeling in (4), calculate quarterly RRFs64 for total primary PM2 5 modeled for the
selected sources.

(6) Multiply the primary PM25 identified in (3) by the RRFs from (5).
sources for each year. The identification of the contribution to the monitor should be most
influenced by the modeling year (the middle year of the 5 year period) since it is the year being
modeled and it has the strongest weighting in the design value calculations.

       61Because the test is relative, in most cases, actual emissions should be used. The actual
emissions should be representative of emissions on high PM25 days (days that exceed the
NAAQS).  Since the absolute predicted concentrations are not used directly, allowable emissions
may overestimate the changes in concentrations due to the identified sources.  This should be
evaluated  on a case-by-case basis. Allowable emissions may be appropriate if there are sources
that often  emit above their typical levels on high PM2 5 days.

       62Additional receptors should be placed near monitors to examine predicted concentration
gradients.  Receptors should only be located in areas that are appropriate for placing FRM
monitors for comparison to the 24-hour NAAQS (see EPA, 1997b).

       63High days may be all days > 65 ug/m3 or the high end of the distribution of modeled
days (e.g top 25% days)

       64RRFs should be calculated as a ratio of the base and future mean concentrations (a ratio
of the means, not a mean of the daily ratios).

                                           68

-------
(7) Subtract the result in (6) from the primary PM2 5 identified in (3)

(8) For each of the 5 years used to calculate the 5 year weighted average design value, subtract
the result from (7) from each quarter for each year.  The "new" annual 98th percentile value is the
highest quarterly concentration in each year.

(9) Average the annual 98th percentile concentrations to get 3 design values and then average the
3 DV to get the final future year 5 year weighted average DV.

Annual NAAQS Example:

For quarter 4, a monitor has a 5 year weighted mean concentration of 17.0 ug/m3.

(1) It is believed that several steel mills in the area are contributing to the local
component of PM25 at the monitor

(2) Based on comparisons to other monitors in the area, it is believed that the local PM2 5
component at the monitor for quarter 4 is 2 ug/m3.

(3) Using receptor modeling, it is determined that, on average, the steel mills are  contributing a
local PM25 component in quarter 4 of approximately 1 ug/m3.

(4) Modeling with a dispersion model shows the following contribution at the monitor from the
identified local sources (steel mills):
Base: 2.0 ug/m3
Future: 1.5 ug/m3

(5) Quarter 4 RRFJocal = 1.5/2.0 = 0.750.

(6) 1 ug/m3 * 0.750 = 0.75 ug/m3.

(7) 0.75 ug/m3 - 1 ug/m3 = -0.25 ug/m3 (total Q4 reduction from identified local sources is 0.25
ug/m3)

(8) Quarter 4 New DV =17.0  - 0.25 ug/m3 = 16.75 ug/m3

(9) Steps 1-8 are repeated for each quarter with similar results.  The annual DV = (Ql New DV +
Q2 New DV + Q3 New DV + Q4 New DV)/4
24 Hour NAAQS Example:

       The baseline design values are calculated from the 5 year base period centered around the
modeling year.  The dispersion model is run with emissions and meteorology for year 3 (the

                                          69

-------
center year of the 5 year period). The maximum quarterly values from the 5 years which are
equal to or less than the 98th percentile concentration are as follows (from example 5.2):

Quarter 1
Quarter 2
Quarter 3
Quarter 4
Year 1
80.0 ug/m3
67.6
41.7
52.8
Year 2
65.0
45.5
47.6
68.4
YearS
57.0
46.4
32.5
71.3
Year 4
58.0
45.3
39.4
49.5
YearS
74.0
52.2
45.3
64.2
The future year PM2 5 concentrations from the photochemical modeling described in Section 5.2,
Step 5 are as follows:

Quarter 1
Quarter 2
Quarter 3
Quarter 4
Year 1
72.6 ug/m3
57.5
38.8
46.6
Year 2
58.9
38.7
44.3
60.5
YearS
51.6
39.4
30.2
63.0
Year 4
52.5
38.5
36.6
43.6
YearS
67.1
44.4
42.1
57.1
(1) It is believed that several steel mills in the area are contributing to the local
component of PM25 at the monitor

(2) Based on comparison to other monitors in the area, it is believed that the primary local PM2 5
component at the monitor (on high PM days) is the following for each of the quarters in 2001:
Quarter 1:15 ug/m3
Quarter 2: 13 ug/m3
Quarter 3: 8 ug/m3
Quarter 4: 20 ug/m3
It is believed that this is representative of the other four years within the 5-year period.

(3) Using receptor modeling, it is determined that the steel mills are contributing a local
PM2 5 component (on high PM days) of the following for each of the quarters in 2001:
Quarter 1: 8 ug/m3
Quarter 2: 4ug/m3
Quarter 3: 4 ug/m3
Quarter 4: 5 ug/m3
                                           70

-------
It is believed that this is representative of the other four years within the 5-year period.

(4) Modeling with a dispersion model shows the following contribution at the monitor from the
identified local sources (steel mills) for total primary PM2 5 in each of the quarters in 2001:
Quarter 1: Base: 11.7 ug/m3  Future: 8.3 ug/m3
Quarter 2: Base: 6.3 ug/m3  Future: 4.4 ug/m3
Quarter 3: Base: 5.3 ug/m3  Future: 3.1 ug/m3
Quarter 4: Base: 8.3 ug/m3  Future: 6.2 ug/m3
For each quarter, we computed the 24-hr average for each day and computed the mean of the
high days for each quarter to get the base and future year concentrations.

(5) Based on the results from the modeling in (4), the RRF values are:
Quarter 1: 8.3/11.7 =  0.709
Quarter 2: 4.4/6.3 = 0.698
Quarters: 3.1/5.3 = 0.585
Quarter 4: 6.2/8.3 = 0.747

(6) Multiplying the primary PM2 5 identified in (3) by the RRFs from (5).
Quarter 1: 8 ug/m3 *  0.709 = 5.7 ug/m3
Quarter 2: 4 ug/m3 *  0.698 = 2.8 ug/m3
Quarter 3: 4 ug/m3 *  0.585 = 2.3 ug/m3
Quarter 4: 5 ug/m3 *  0.747 = 3.7 ug/m3

(7) Subtracting the result in (6) from the primary PM25 identified in (3)
Quarter 1: 5.7 ug/m3  - 8 ug/m3 = -2.3 ug/m3
Quarter 2: 2.8 ug/m3  - 4 ug/m3 = -1.2 ug/m3
Quarter 3: 2.3 ug/m3  - 4 ug/m3 = -1.7 ug/m3
Quarter 4: 3.7 ug/m3  - 5 ug/m3 = -1.3 ug/m3

(S)Subtracting the result from (7) from each quarter for each of the five years:

Year 1:
Ql:  72.6 ug/m3 - 2.3 ug/m3 = 70.3 ug/m3
Q2:  57.5- 1.2 = 56.3
Q3:  38.8- 1.7 = 37.1
Q4:  46.6- 1.3 = 45.3

Year 2:
Ql:  58.9-2.3 =56.6
Q2:  38.7- 1.2 = 37.5
Q3:  44.3- 1.7 = 42.6
Q4:  60.5- 1.3 = 59.2
                                          71

-------
Year 3:
Ql: 51.6-2.3 = 49.3
Q2: 39.4- 1.2 = 38.2
Q3: 30.2- 1.7 = 28.5
Q4: 63.0- 1.3 = 61.7
Year 4:
Ql: 52.5-2.3 = 50.2
Q2: 38.5- 1.2 = 37.3
Q3: 36.6- 1.7 = 34.9
Q4: 43.6- 1.3 = 42.3

Year 5:
Ql: 67.1 -2.3 = 64.8
Q2: 44.4- 1.2 = 43.2
Q3: 42.1 - 1.7 = 40.4
Q4: 57.1 - 1.3 = 55.8

The "new" 98th percentile values are:

Year 1:70.3
Year 2: 59.2
Year 3: 61.7
Year 4: 50.2
Year 5: 64.8

(9) Averaging the 3 design values:
DV1 = (70.3 + 59.2 + 61.7)/3 = 63.7
DV2 = (59.2 + 61.7 + 50.2)73 = 57.0
DV3 = (61.7 + 50.2 + 64.8)73 = 59.9

Averaging the 3 DV to get the final future year 5 year weighted average DV:
(63.7 + 57.0 + 59.9)73 = 60.2 = 60 ug/m3

       In the above examples, emissions reductions from steel mills were modeled with a
dispersion model, which provided additional reductions to the future year 5 year weighted
average annual and 24-hour design values. Changes in design values derived from
photochemical modeling should be evaluated to determine the magnitude (on concentrations) of
emissions  changes from the flagged  steel mills. The change in concentration in the
photochemical model from emissions controls on the steel mills may need to be subtracted from
the final design value to ensure that there is no double counting of emission reductions.
                                          72

-------
Double Counting

       The methodology discussed above may result in double counting of local emissions
sources and reductions. The change in emissions from local sources is counted once in the
photochemical modeling and again in the dispersion modeling. There are several ways to
account for this issue.

1) Run the photochemical model without the primary PM2 5 emissions from the flagged sources
in both the base case and future case(s).

2) Run the grid model with a source tagging approach. The flagged sources can be tagged so
that their contribution in the grid model can be explicitly tracked (Douglas, 2006) (Environ,
2006a), (Environ, 2006b).

3) Employ a methodology to combine the results of the grid and dispersion models so that
double counting is eliminated or minimized.  Several techniques have been developed to do this
(Isakov,2007).

4) Assume that double counting is small enough to be ignored.
       The grid resolution of the photochemical model and the number of flagged sources in the
dispersion model will determine the magnitude of the potential double counting. The higher the
resolution of the grid model, the more double counting will be a problem (because the local
sources will be more finely resolved). And the more sources that are flagged, the more double
counting will be a problem.  Therefore, the nature of the issue should be evaluated on a case by
case basis.  The simplest way to evaluate the magnitude of the problem is to run the
photochemical model with and without the flagged sources from the dispersion model. This will
indicate the maximum impact the sources can have on the photochemical modeling results. A
very small impact means that double counting is not a problem.  A large impact indicates that
double counting needs to be explicitly accounted for.

Quality Assurance

       As with any modeling exercise, it is important to quality assure the model inputs and
outputs. In particular, we recommend a focused quality assurance check on emissions from
sources flagged for a dispersion modeling analysis.  Prior to applying a model, States should
review available information to ensure that there are no major discrepancies between modeled
estimates and nearby monitored data for particulate matter. The emission factors, activity data,
and speciation profiles of the PM25 emissions should also be analyzed for accuracy. If a
speciation monitor exists, the speciated data from the monitor can be compared to the speciation
profiles of the flagged sources. Receptor models can also be used as a QA tool.  Discrepancies
between receptor modeling results (which are based on analyzing ambient data) and speciated

                                           73

-------
emissions may indicate a problem with the magnitude and/or the speciation of the emissions
sources.  If discrepancies are found, those implementing the modeling protocol (see section 12)
should consult with the appropriate U.S. EPA regional office to reach agreement on what, if any,
adjustments should be made to the emissions estimates for the identified sources.

       It is also important to quality assure the model outputs.  Modeling each species of
primary PM25 separately within the dispersion model should aid in this analysis, especially if the
selected control strategies applied in the future year do not control each primary species to the
same degree. If a speciation monitor exists, the speciated data from the monitor may also help
quality assure the dispersion model results.  However, the relative application of the dispersion
model results will help to reduce the impact of possible over- or under-estimations by the
dispersion model due to emissions, meteorology, or general selection of other model input
parameters.
                                           74

-------
6.0    What Is The Recommended Modeling Analysis for Regional Haze?

       In this section, we describe the recommended modeled test to assess visibility
improvement.  The visibility analysis has many similarities to the attainment tests described in
Section 5. The recommended visibility analysis and the attainment tests both use monitored data
to define current air quality. They divide PM25 into major species, and component-specific
relative response factors are multiplied by current monitored values to estimate future
concentrations.

       Section 6.1  provides background information on the requirements of the Regional Haze
rule and defines the components of a reasonable progress analysis.  The rest of the section
describes the recommended modeling analysis which supports the regional haze rule.

       6.1     Regional Haze Rule Background

       Section 169 A of the Clean Air Act states "Congress hereby declares as a national goal
the prevention of any future, and the remedying of any existing, impairment of visibility in
mandatory class I Federal areas which impairment results from manmade air pollution." Section
169B calls for EPA to "carry out the Administrator's regulatory responsibilities under [section
169A], including criteria for measuring 'reasonable progress' toward the national goal."

       In response to these mandates, EPA promulgated the regional haze rule (RHR) on July 1,
1999.65  Under section 51.308(d)(l) of this rule, States must "establish goals (expressed in
deciviews) that provide for reasonable progress towards achieving natural visibility conditions
for each Class I area within a State".  These reasonable progress goals must provide for an
improvement in visibility for the most impaired days over the period of the implementation plan
and ensure no degradation in visibility for the least impaired days over the same period.66
Reasonable Progress Goals (RPGs), measured in deciviews, are interim goals that represent
incremental visibility improvement over time toward the goal of natural background conditions
and are developed in consultation with other affected States and Federal Land Managers.67

       In determining what would constitute reasonable progress, section 169A(g) of the Clean
Air Act requires States to consider the following four factors:

       The costs of compliance;
       The time necessary for compliance;
       The energy and non-air quality environmental impacts of compliance; and
       65 64 FR 35714 (codified at 40 CFR 51.300-309).
       66 40CFR51.308(d)(l).

       67 40 CFR 51.308(d)(l)(iv) and 51.308(1).
                                           75

-------
•      The remaining useful life of existing sources that contribute to visibility impairment.68

States must demonstrate in their SIPs how these factors are taken into consideration in selecting
the RPG for each Class I area in the State.  More details on the setting of RPGs and the
appropriate use of the four factors is contained in "Guidance for Setting Reasonable Progress
Goals Under the Regional Haze Program" (U.S. EPA, 2007).

       The regional haze rule also establishes an additional analytical requirement in the process
of establishing RPG. States are required to determine the rate of improvement in visibility
needed to reach natural conditions by 2064, and to set each RPG taking this "glidepath" into
account.69  The glidepath is the amount of visibility improvement needed in each review period
to stay on a linear path towards visibility improvement to natural conditions by 2064. The
glidepath represents a linear or uniform rate of progress. Therefore, the amount of visibility
improvement needed to stay on the glidepath is also referred to as "uniform rate of progress".
This document uses the terms "glidepath" and "uniform rate of progress" interchangeably.

       This document focuses on the modeling analysis required to assess future visibility
improvement relative to the glidepath or uniform rate of progress (for each Class  I area). As
mentioned previously, this document does not address the process of setting and evaluating
reasonable progress goals.  The actual determination of reasonable progress is a multi-step
process which looks at the uniform rate of progress, the predicted rate of progress, and the four
factor test.  In the end, the reasonable progress goal will be determined by the states for each
Class I area by considering the factors and the modeling analysis.  The results of the  modeling
analyses will be one of several components in the determination of reasonable progress.

       6.2    Uniform Rate of Progress

       Regional haze is calculated by estimating light scattering and absorption by components
of PM2 5. Baseline conditions represent the visibility conditions which exist for the best and
worst days at the time the regional haze program is established for each Class I area.   Once
established, the baseline represents the starting point from which the uniform rate of progress
will be measured. The RHR also requires an estimate of "natural conditions" for each Class I
area which represents the visibility conditions that would exist in the absence of man-made
impairment.

       As explained in the RHR, the baseline for each Class I area is the average visibility (in
deciviews) for the 20% most impaired days, or "worst days", and for the 20% least impaired
days, or "best days," for the years 2000 through 2004™.
       68 42 U.S.C. ง 7491(g)(l); 40 CFR 51.308(d)(l)(i)(A).
       69 40 CFR 51.308(d)(l)(i)(B).
       7064FR at 35730.
                                           76

-------
       States are asked to calculate the reduction in deciviews (on the worst visibility days) by
an amount which is proportional to the number of years between the base period and the first
review period vs. the number of years between the base period and 2064. This establishes the
glidepath or uniform rate of progress. For example, if a Class I area's mean worst visibility
during the 2000-2004 base period were 35 deciviews and natural background were estimated to
be 8 deciviews., the glidepath for 2018 would be to reduce visibility impairment on the worst
days by 6.3 deciviews (i.e., [35 -8][(2018-2004)/(2064-2004)]). More detailed examples of this
calculation are contained in (U.S. EPA, 2003b).

       The goal of the regional haze program is to return to natural conditions by 2064, and
States are required to demonstrate, by the end of the first planning period (2018), reasonable
progress toward meeting that goal.  The U.S. EPA has developed guidance on how to track
visibility for the regional haze rule (U.S. EPA, 2003b).  In Section 6.3, we describe how the
ambient trends in regional haze will be determined. The modeling assessment for uniform  rate
of progress needs to provide a good predictor for what the trend in observed air quality will be
in the future.  This is most likely to happen if we reproduce the calculations which will be
applied to the ambient observations as closely as practical. In Section 6.4, we present and
recommend a modeling analysis which closely replicates the ambient data analysis. Section 6.5
addresses how to develop key inputs for the recommended modeling assessment of uniform rate
of progress.

       6.3     How are Regional Haze Trends Measured?

       Regional haze is measured by an extinction coefficient (bext) which represents light
attenuation resulting from scattering and absorption of light from ambient parti culate matter plus
scattering of light due to gas molecules in the air (i.e., Rayleigh scattering).  Although bext can be
estimated by several different methodologies, the regional haze rule requires that it be estimated
using measured ambient parti culate matter. This follows since, for a given set of meteorological
conditions, visibility can be improved by reducing concentrations of ambient particulate matter.
Thus, deriving bext in this manner provides a direct link between regional haze and related
pollutant concentrations. Equation (6.1) may be applied in each Class I  area to estimate bext
(Sisler, 1996, IMPROVE, 2000). The equation reflects empirical relationships derived between
measured mass of particulate matter components and transmissometer measurements of bext at
monitoring sites in Class I areas within the IMPROVE network.
4
-------
f(rh),f'(rh) are relative humidity adjustment factors applied to the light scattering efficiency (to
be described in greater detail shortly), dimensionless,

SO4 is the mass associated with sulfates, ug/m3,

NO3 is the mass associated with nitrates, ug/m3,

OC is the mass associated with organic carbon, ug/m3,

EC is the mass associated with elemental carbon, ug/m3,

Fine Soil is inorganic primary particulate matter (excluding primary sulfate and nitrate particles)
associated with soil components with aerodynamic diameter < 2.5 um, i-ig/m3,

CM is coarse particulate matter with aerodynamic diameter > 2.5 um, but < 10 um, ug/m3,

brayieigh l$ light-scattering attributable to  Rayleigh scattering, Mm"1 (i.e., inverse "mega-meters"),
assumed to be 10 Mm"1.

bext is the estimated extinction coefficient, Mm"1.


New IMPROVE Algorithm

       The IMPROVE program has recently revised the IMPROVE algorithm (IMPROVE,
2006); (Hand, 2006) . The new algorithm is intended to reduce biases in light extinction
estimates compared to the old algorithm. The new algorithm is somewhat more complicated
than the old algorithm as follows:

bext * 2.2 x fs(RH) x [Small Sulfate] + 4.8 x fL(RH) x [Large Sulfate]

       + 2.4 x fs(RH) x [Small Nitrate] + 5.1 x fL(RH) x [Large Nitrate]

       + 2.8 x {Small Organic Mass] + 6.1 x [Large Organic Mass]
       + 10 x [Elemental Carbon]

       + 1 x [Fine Soil]

       + 1.7 x fss(RH)  x [Sea Salt]

       + 0.6 x [Coarse Mass]
       + Rayleigh Scattering (site specific)

       + 0.33 x [NO2 (ppb)]                                          [6.2]
                                           78

-------
       The total sulfate, nitrate and organic carbon compound concentrations are each split into
two fractions, representing small and large size distributions of those components. The organic
mass concentration used in the new algorithm is 1.8 times the organic carbon mass
concentration, changed from 1.4 times carbon mass concentration used for input into the old
IMPROVE algorithm. New terms have been added for sea salt (important for coastal locations)
and for absorption by NO2 (only used where NO2 data are available). Site-specific Rayleigh
scattering is calculated for the elevation and annual average temperature of each IMPROVE
monitoring site.

       The apportionment of the total concentration of sulfate compounds into the
concentrations of the small and large size fractions is accomplished using the following
equations.

[Large Sulfate] = [Total Sulfate]/20 ug/m3 x [Total Sulfate], for [Total Sulfate] < 20 ug/m3

[Large Sulfate] = [Total Sulfate], for [Total Sulfate] > 20 ug/m3

[Small Sulfate] = [Total Sulfate] -  [Large Sulfate]
       The same equations are used to apportion total nitrate and total organic mass
concentrations into the small and large size fractions.

       Sea salt is calculated as 1.8 x [Chloride] or 1.8 x [Chlorine] if the chloride measurement
is below detection limits, missing or invalid.  The algorithm also uses three new water growth
adjustment terms. They are for use with the small size distribution and the large size distribution
sulfate and nitrate compounds and for sea salt (fs(RH), fL(RH) and fss(RH) respectively).

       Testing of the new algorithm reveals that, in most regions of the country, the new
equation gives somewhat higher light extinction, especially on the worst visibility days. But
compared to the  old equation, the relative percentage of extinction attributed to the various PM
components is nearly unchanged. Therefore, the revised equation is not expected to make a
large difference in the rate of visibility changes resulting from potential control strategy
simulations. However, the new equation could lead to differences in the relative change in light
extinction predicted by the models. Either of the IMPROVE equations are acceptable for use in
reasonable progress analyses, although States should provide documentation concerning the
choice of equation, especially in situations where large differences  are noted between the new
and old71.
       71Both algorithms can be independently used to estimate changes in extinction between a
base period and a future year. But the same algorithm needs to be used in each calculation
(don't use the old algorithm to calculate base extinction and the new algorithm to calculate
future extinction or vice versa).  Also, the same algorithm should be used to calculate the

                                           79

-------
Deciview Haze Index

       The regional haze rule stipulates that reasonable progress is to be measured in terms of
changes in a deciview haze index or simply "deciviews"(dv). Deciviews are defined as the
natural logarithm of the ratio of the extinction coefficient to Rayleigh scattering (Pitchford and
Malm, 1994)

       Deciview = 10 ln(bext/10)                                                   (6.3)

       Where the units of bext and light scattering due to Rayleigh scattering72 (i.e., the "10" in
       the denominator of the logarithmic expression) are both expressed in Mm"1.

A change in deciviews, which is how we track progress is given by Equation (6.4).  A 1
deciview change is equivalent to -10% change in bext.

       Adv = dvfuture - dvbaseline                                              (6.4)

A negative number indicates a reduction in deciviews, which is an improvement in visibility.

Estimating Mass Associated With Components Of Particulate Matter.

       Regional haze calculations for the modeled uniform rate of progress analysis utilizes
IMPROVE measurements at Class I areas. All Class I areas have on-site speciated PM2 5
measurements or have been assigned a representative IMPROVE monitoring site with
measurements73. Therefore, it is generally not necessary to interpolate measured PM2 5 species
data to a site. The existing IMPROVE database of PM25 measurements should be adequate to
provide data for all Class I areas74.

       FRM data is not used in the regional haze analysis.  Therefore, it is not necessary for the
sum of species components to equal gravimetric measurements obtained with a Federal
glidepath (using ambient data) and the modeled visibility changes.

       72Even though the "new" IMPROVE equation uses a site specific Rayleigh Scattering
value, the denominator in the deciview equation should always be 10.  In this way, the deciview
calculation will be consistent across all Class I areas.  Under very clean conditions, the deciview
value can be negative, but this should not be considered a problem.
       73See U.S. EPA 2003b, Appendix A, Table A-2.

       74There may be a few IMPROVE sites that don't have enough complete data to provide
baseline condition information for the 2000-2004 period. It may be necessary to substitute data
from other monitoring networks or to interpolate data to the site(s).

                                           80

-------
Reference or equivalent method for measuring PM2 5.  Therefore, for regional haze calculations,
it is not necessary to use the speciated data adjustment procedures introduced in Section 3 (the
"SANDWICH" adjustments).

       Sisler (1996) has developed a set of default assumptions75 for mass associated with each
of the components of particulate matter for Equation (6.1). These are presented in Table 6.1.
The components in Table 6.1 are similar to those used in the modeled attainment demonstrations
for the PM2 5 NAAQS.  Notice however, that sulfate and nitrate mass are assumed to be fully
neutralized (therefore, ammonium is not needed as a separate component); organic carbon is
assumed to be equal to 1.4 times measured OC mass (1.8 times OC in the new IMPROVE
equation); there is no water component; and there is a term for coarse particulate matter.
   Table 6.1.  Default Assumptions Used To Derive Aerosol Species Concentrations For
                           Estimating Extinction Coefficients
(1)
Species
Sulfate
Nitrate
EC
OC
Fine Soil
PM25
CM
(coarse mass)
(2)
Formula
4.125 (measured sulfur) or 1.375
(measured SO4)
1.29 (measured nitrate)
high + low temperature EC
1.4 (organic carbon) or 1.8* OC
for the new IMPROVE equation
2.2(A1) + 2.49(Si) + 1.63(Ca) +
2.42(Fe)+ 1.94(Ti)
measured gravimetrically
(PM10) - (PM2.5)
(3)
Assumptions
All elemental sulfur is from
sulfate, & all sulfate is from
ammonium sulfate
Denuder efficiency is ~100% &
all nitrate is from ammonium
nitrate
All high temperature carbon is
elemental
Average organic molecule is 70%
carbon
(Soil K)=0.6(Fe), FeO & Fe2O3
are equally abundant, a factor of
1.16 is used for MgO, Na2O, H2O
&C03
Represents dry ambient fine
aerosol mass for continental sites
Consists only of insoluble soil
particles
       75The default assumptions in table 6.1 are applicable to the "old" IMPROVE equation.
The new IMPROVE equation assumes that OCM= 1.8 x [OC] and sea salt = 1.8 x [chloride].
                                          81

-------
       Normalizing trends in regional haze- f(rh) factors.  It is clear from equations 6.1 and
6.2 that relative humidity can have a substantial effect on estimated extinction coefficients, as
well as on the relative importance of changes in different components of particulate matter can
have on trends in regional haze. Because of the importance of relative humidity as a determinant
of regional haze, it is necessary to normalize any apparent trend in the estimated extinction
coefficient for differences in relative humidity.  This enables us to assess whether an emissions
control strategy will reduce regional haze, without confounding effects of different relative
humidity during the base and future periods.

       There are two obvious potential ways to normalize trends in visibility for changes in
relative humidity.  The first is to assume that the same day to day changes in relative humidity
which were observed in the base period calculations will occur in future years. Thus, one would
use the relative humidity observations made on a specific day together with measured
components of parti culate matter on that day to compute the day specific visibility extinction
coefficient on that day. Subject to the uncertainties posed by the empirically derived coefficients
in Equation (6.1), this approach is the most likely to identify the 20% best and worst visibility
days during the base period at a Class I site. However, the approach could lead to misleading
conclusions if humidity observations were missing for some days or if the humidity observations
are atypical in some way. Further, if a State or regional planning organization wanted to perform
visibility calculations in a number of Class I areas, they would need to obtain hourly relative
humidity data for each area.

       The second approach to normalize trends in the extinction  coefficient is to review relative
humidity data over a long period of record to derive climatological estimates for relative
humidity adjustment factors. These climatological estimates would then be used in Equations
6.1 and/or 6.2 to estimate visibility  extinction coefficients.   These estimates are more likely to
reflect "typical" relative humidity at different times of year and, thus, expected future visibility
extinction.

       The need to use  climatological f(rh) data for tracking progress of measured data is
obvious.  The measured relative humidity on the 20% worst days in a baseline period will be
different than the measured relative humidity in a future period. Therefore the only way to
normalize for relative humidity is to use a single climatological  value. In the context of
modeling, there is a choice. We can use climatological data or we can use the modeled (or
measured) relative humidity data from the air quality model. Since the meteorological data is
held constant when predicting future year air quality, the modeled relative humidity values
would also be constant.  This would be one way to normalize the data.

       Throughout this section we make every attempt to be consistent with the "Guidance for
Tracking Progress Under the Regional Haze Program"(hereafter referred to as the "Tracking
Guidance") (U.S. EPA, 2003b). It will be easier to interpret the modeling results and compare
them to future measured values, if the data are calculated in the  same manner. We therefore
recommend using climatological f(rh) values to calculate baseline and future visibility.
Appendix A (Table A-l) from the Tracking Guidance displays the relationship between relative

                                           82

-------
humidity and f(rh) values.  The f(rh) values were calculated from data reported by Tang (1996).
The f(rh) values are 1.00 up to 36% relative humidity and grow to a maximum value of 7.40 at a
relative humidity of 95%76.  The Tracking Guidance contains monthly climatological average
f(rh) values for each Class I area. These values were calculated using 10 years of meteorological
data.

       For calculations with the new IMPROVE equation, three separate f(rh) curves are
needed. These can be found in (Hand, 2006). The monthly f(rh) values for both the old and new
IMPROVE equations have been incorporated into precalculated  daily extinction and deciview
data that can be found on the Visibility Information Exchange Web System (VIEWS) website (
http://vista.cira.colostate.edu/views/).  The use  of precalculated  f(rh) and visibility data will
make it more simple for States and RPOs to calculate visibility changes and will also provide for
more consistency.

       6.4 How Do I Apply A Model To Calculate Changes in Visibility ?

       The purpose of a modeling assessment of uniform rate of progress is to determine if a
proposed control strategy will result in uniform  rate  of progress being met when measured
concentrations of particulate matter are used to estimate visibility impairment at some future
date. The analysis we recommend has 6 steps.

       1. For each Class I area, rank visibility (in deciviews) on each day with observed
       speciated PM2 5 data plus PM10 data for each  of the 5 years comprising the base period.

       2. For each of the 5 years comprising the base period, calculate  the mean deciviews for
       the 20% of days with worst and 20% of days with best visibility. For each Class I area,
       calculate the 5 year mean deciviews for worst and best days from the five year-specific
       values.

       3. Use an air quality model to simulate base period emissions and future emissions. Use
       the resulting information to develop relative response factors for each component of
       particulate matter identified on the right hand side of Equations (6.1) and/or (6.2).

       4. Multiply the relative response factors  times the measured species concentration data
       during the base period (for the measured 20% best and worst days).  This results in daily
       future year species concentrations data.

       5.  Using the results in Step 4 and the IMPROVE algorithm (equation 6.1 or 6.2),
       calculate the future daily extinction coefficients for the 20% best and 20% worst
       visibility days in each of the five base years.
       76For the regional haze calculations, the f(rh) factors have been capped at the f(rh) value
associated with 95% relative humidity. Relative humidity value above 95% use the same value.

                                              83

-------
       6. Calculate daily deciview values (from total daily extinction) and then compute the
       future average mean deciviews for the worst and best days for each year.  Then average
       the 5 years together to get the final future mean deciview value for the worst and best
       days.

       We describe each of these steps more fully below. The methodology follows the same
basic procedures outlined in the "Tracking guidance." Further details (for steps 1 and 2) can be
found in that document. We conclude this subsection with an  example illustrating the
recommended modeled uniform rate of progress analysis.

       Step 1. Using monitored data, rank baseline visibility for each day with PM10, PM2 5
and speciated PM2 5  measurements within a Class I area

       Ranking should be performed  separately for each of the 5 years comprising the base
period77. The deciview (dv) should serve as the basis for ranking. Day-specific observations for
mass associated with  SO4, NO3, OC, EC,  soil, and CM, as defined in Table 6.1, should be used to
calculate bext for each day.  The appropriate month- and area-specific climatological relative
humidity adjustment factor(s) (f(rh)} should be used.  Total bext for all components should be
converted to deciviews for each day to get a daily deciview value.

       Step 2. Calculate the average baseline deciviews for the 20% of days with worst and
the 20% days with best visibility.

       For each of the 5 years in the base period, order all days considered in Step  1 from worst
(highest deciview value) to best (lowest deciview value) visibility. For each year, note the 20%
of days with worst and the 20% of days with best visibility. Calculate the arithmetic mean
deciview value for the identified 20% worst- and best visibility days in each year.  Average the
resulting 5 yearly mean deciview values reflecting worst visibility.  This represents the value
subject to improvement (i.e., reduction) to meet the glidepath for regional haze. Average the 5
yearly  mean deciview values reflecting mean visibility on the days with best visibility.

       Step 3. Estimate relative response factors (RRF) for each component of PM2 5 and
for CM.

       This  should be done by using emissions during the base period in  air quality model
simulations performed for a large number of days78. Air quality corresponding with future
       77Pre-calculated and ranked extinction and deciview calculations for all Class I areas for
the base period 2000-2004 (using both the "old" and "new" IMPROVE algorithms) are available
on the VIEWS website at http://vista.cira.colostate.edu/views/. This data can be used to satisfy
steps 1 and 2.

       78How many and which days to simulate is discussed in Section 6.3 and in Section 14.1.

                                           84

-------
emissions (reflecting effects of growth and controls) should be simulated for the same days.
Take the (temporal) arithmetic mean concentration for each PM2 5 component (and coarse mass)
computed near the Class I monitoring  site with future emissions and divide this by the
corresponding arithmetic mean concentration for each component obtained with current
emissions.  The resulting quotients are the component-specific RRF's. A separate set of RRF
values are calculated for the "worst" and "best" visibility days identified in step 2. The RRFs
are calculated using the identified 20% best and 20% worst monitored days at each Class I area.
This will likely be a different set of days at each monitor.

       Step 4. Using the RRFs and ambient data, calculate future year daily concentration
data for the best and worst days.

       Multiply the relative response factors derived in Step 3 times measured daily
concentration data for each component of PM2 5 and CM to get future daily estimates of species
concentrations for PM2 5 components and CM on "worst visibility" and "best visibility"days.
These multiplications produce future concentration estimates for SO4, NO3,  OC, EC,  Soil and
CM for each of the previously selected "worst" and "best" visibility days. This calculation is
performed for each best and worst day for the five year period using an RRF for each PM
component (a separate set of RRFs for the best days and the worst days)

       Step 5. Use the information developed in Step 4 to compute future year daily bext
values for the best and worst days.

       Use the future year concentration data calculated in step 4 to calculate future year daily
bext values for each PM component for each of the  best and worst days for the five year period.
This is accomplished by applying either the old or new IMPROVE visibility algorithm
(equations 6.1 and 6.2).

       Step 6. Use the daily total bext values from step 5 to calculate future mean deciview
values for the best and worst days.

       The total daily bext for each day is converted to deciviews.  This gives a future year daily
deciview value for each of the best and worst days.

       Next, compute the arithmetic mean future deciview value for the "worst" and "best"
visibility days for each year.  This leads to 5 future estimated mean deciview values for the
"worst" and 5 future estimated mean deciview values for the "best" visibility days. Compute the
arithmetic mean of the 5 mean values for deciviews on the "worst" days, and the arithmetic  mean
of the 5 mean deciview values estimated for the "best" visibility days.

       The resulting 5 year average mean values for deciviews on "worst" and "best" visibility
days can be compared to the baseline mean deciview values calculated in step  2.  If the resulting
change in deciviews is a negative number (future - base), this represents an improvement in
visibility. Compare the change in visibility to the previously  calculated uniform rate  of progress

                                           85

-------
value for each Class I area.

Example 6.2

       We use example 6.2 to illustrate the modeled uniform rate of progress assessment. For
ease of presentation, we assume there are only 10 speciated samples for PM in the first of 5 years
comprising the base period. Since sampling will occur every third day, we anticipate a usual
sample size of about 120 days per year. We go through the calculations for the first base year
and then furnish information regarding mean deciviews for the other four base years to illustrate
subsequent steps in the test. The example uses the old IMPROVE algorithm.  The procedure is
the same for the new IMPROVE algorithm.

Given:

Ten days have measured components of PM in a Class  I area during the first year of a 5-year
base period.  Table 6.2 below shows the measurements (in |ig/m3) for each of the 10 days. Table
6.2 also shows the date of each measurement and the corresponding climatological relative
humidity adjustment factor (made up for this example)  for the appropriate month and area.

Table 6.2
Day
1
2
3
4
5
6
7
8
9
10
Date
2/15
3/15
4/15
5/15
6/15
7/15
8/15
9/15
9/30
10/30
f(rh)
1.7
2.9
3.5
3.7
4.3
4.6
4.4
4.1
4.1
2.7
SO4
Gig/m3)
4.53
5.12
5.67
6.59
7.47
8.33
9.22
8.05
6.84
5.32
NO3
(Hg/m3)
2.23
1.78
1.25
1.12
0.93
0.79
0.89
0.97
1.15
1.41
OC
(|ig/m3)
3.37
2.92
2.44
2.86
3.49
3.50
3.24
3.12
3.21
3.26
EC
(Hg/m3)
0.89
0.78
0.54
0.50
0.61
0.69
0.67
0.73
0.85
0.76
Soil
(Hg/m3)
0.32
0.44
0.43
0.66
0.98
1.08
0.89
0.71
0.42
0.64
CM
(Hg/m3)
7.33
9.17
10.25
9.80
10.99
11.38
10.78
8.25
8.82
8.11
The following uniform rate of progress has been calculated: (a) mean visibility on the 20% of
days with worst visibility should be improved by 2.0 deciviews; (b) mean visibility on the 20%
of days with best visibility should not deteriorate.
                                          86

-------
Find: Is the control strategy simulated in this model analysis sufficient to meet the uniform rate
of progress in this Class I area?

Solution:

Step 1. Using monitored data, rank baseline visibility for each day with PM10, PM2 5 and
speciated PM2 5 measurements within a Class I area.

       First, estimate the extinction coefficient for each day with the needed PM measurements.
This is done using the information in table 6.2 with Equation (6.1). For day 1 in year 1, the
current extinction coefficient is:
bex, = (3)(1.7)[4.53] + (3)(1.7)[2.23]
(10)[0.89]
                                                                              (0.6)[7.33]
+ 10
       bext = 71.6Mm-1

Then convert extinction into deciviews:

       dv = 10 *ln(71.6/10) = 19.7 dv

Current extinction coefficients and deciviews for the remaining 9 days with monitored data in
year 1 are calculated in a similar manner.  The days are then ranked.  The day with the highest
deciviews (i.e., worst visibility) is given a rank of "1".  The results of these calculations are
displayed in the table 6.3 below. Based on these rankings, days 6 and 7 comprise the 20% of
days with worst visibility.  Days 1 and 10 comprise the 20% of days with best visibility

Table 6.3
Day
1
2
3
4
5
6
7
8
9
10
Date
2/15
3/15
4/15
5/15
6/15
7/15
8/15
9/15
9/30
10/30
f(rh)
1.7
2.9
3.5
3.7
4.3
4.6
4.4
4.1
4.1
2.7
S04
(Mm1)
23.1
44.5
59.5
73.1
96.4
115.0
121.7
99.0
84.1
43.1
NO3
(Mm1)
11.4
15.5
13.1
12.4
12.0
10.9
11.7
11.9
14.1
11.4
oc
(Mm1)
13.5
11.7
9.8
11.4
14.0
14.0
13.0
12.5
12.8
13.0
EC
(Mm1)
8.9
7.8
5.4
5.0
6.1
6.9
6.7
7.3
8.5
7.6
Soil
(Mm1)
0.3
0.4
0.4
0.7
1.0
1.1
0.9
0.7
0.4
0.6
CM
(Mm1)
4.4
5.5
6.2
5.9
6.6
6.8
6.5
5.0
5.3
4.9
bext
baseline
(Mm1)
77.6
95.5
104.4
118.6
146.0
164.7
170.5
146.4
135.3
90.7
Deci
views
19.7
22.6
23.5
24.7
26.8
28.0
28.4
26.8
26.1
22.0
Rank
10
8
7
6
4
2
1
3
5
9
                                           87

-------
Step 2.  Calculate the average baseline deciviews for the 20% of days with worst visibility
and the 20% days with best visibility.

       For year 1, mean worst visibility = (28.0 + 28.4) 12 = 28.2 dv, and

mean best visibility = (19.7 + 22.0) / 2 = 20.9 dv

Mean worst and best visibility for years 2-5 is provided in table 6.4.  Table 6.4 below
summarizes mean worst and best visibility for each of the 5 years in the base period.

Table 6.4
Year
1
2
3
4
5
Mean dvcurrent,
Worst Visibility Days
28.2
29.1
30.2
27.8
28.5
Mean dvcurrent,
Best Visibility Days
20.9
20.4
18.7
19.1
19.5
The "average mean baseline worst and best visibility" is obtained by taking the arithmetic
average of the mean worst and best visibility for the 5 years.  Thus, the average mean worst
visibility is given by

       dvbaseline = (28.2 + 29.1 + 30.2 + 27.8 + 28.5) / 5 = 28.8 dv

The average mean best visibility is

       dvbaseline = (20.9 + 20.4 + 18.7 + 19.1 + 19.5) / 5 = 19.7 dv

Step 3.  Apply a model to develop component specific RRF's for SO4, NO3, OC, EC, Soil
and for coarse mode particulate matter (CM).

       Tables 6.5 and 6.6 show the procedure for calculating component-specific relative
response factors using an air quality model. Component-specific relative response factors are
computed as described in Section 4.3. The example shows the calculation of RRFs for the worst
days in the year modeled. The same calculation is repeated for the best days.
                                           88

-------
Table 6.5 Base year modeled species concentrations on worst days (in ug/m3)
Modeled
Output
Dayl
Day 2
Day 3
Day 4
Day 5
Mean Base
Concentration
S04
2.12
8.33
9.33
6.43
10.53
7.35
NO3
6.54
2.11
1.23
1.67
0.57
2.42
OC
3.56
4.67
6.32
4.56
4.12
4.65
EC
0.87
1.23
1.45
0.54
0.69
0.96
Soil
1.23
0.34
0.87
1.07
1.54
1.01
CM
5.43
7.32
8.21
4.67
10.32
7.19
Table 6.6 Future year modeled species concentrations on worst days (in ug/m3
Modeled
Output
Dayl
Day 2
Day 3
Day 4
Day5
Mean Future
Concentration
S04
2.01
7.56
6.54
5.32
7.23
5.73
NO3
4.23
1.89
1.11
1.45
0.43
1.82
OC
3.45
4.55
5.99
4.23
3.99
4.44
EC
0.77
1.01
1.09
0.35
0.45
0.73
Soil
1.21
0.32
1.02
1.1
1.51
1.03

CM
5.75
7.54
8.71
5.02
10.47
7.50
Worst days
RRF
RRF (mean
future/mean
base)
S04

0.780


NO3

0.752


OC

0.956


EC

0.768


Soil

1.022


CM

1.043


Step 4.  Multiply the relative response factors times the measured daily species
concentration data during the base period to compute future daily species concentrations.

       In year 1, we previously identified days 6 and 7 as those included in the 20% of days
with worst visibility (i.e., see Step 1).  Similarly, days 1 and 10 are the 20% of days with best
visibility. In this step, we need to estimate future concentrations for components of PM2 5 and
for CM for these two sets of days. This is done using information shown in tables presented  in
Steps 1 and 3  as well as the best days RRFs given in table 6.7 below:
                                          89

-------
Table 6.7
Best days RRF
RRF (mean
future/mean
base)
S04
0.870


NO3
0.823


OC
0.976


EC
0.784


Soil
1.112


CM
1.058


Worst Days

Day 6: [SO4]&ture = (RRF)S04 [SO4]baseline = (0.780) [8.33] = 6.50
       [N03]&ture = (0.752) [0.79] = 0.59
       [OC]&ture = (0.956) [3.50] = 3.35
       [EC]&ture = (0.768) [0.69] = 0.53
       (Soil)&ture = (1.022) [1.08] = 1.10
       [CM]&ture = (1.043) [11.38] = 11.87
Day 7: [SO4]&ture = (0.780) [9.22] = 7.19
       [N03]&ture = (0.752) [0.89] = 0.67
       [OC]&ture = (0.956) [3.24] = 3.10
       [EC]&ture = (0.768) [0.67] = 0.51
       (Soil)&ture = (1.022) [0.89]  = 0.91
       [CM]&ture = (1.043) [10.78] = 11.24
Best Days

Day 1:  [SO4]&ture = (0.870) [4.53] = 3.94
       [N03]&ture = (0.823) [2.23] = 1.84
       [OC]&ture = (0.976) [3.37] = 3.29
       [EC]&ture = (0.784) [0.89] = 0.70
       (Soil)&ture = (1.112) [0.32] = 0.36
       [CM]&ture = (1.058) [7.33] = 7.76
Day 10:   [SO4]&ture = (0.870) [5.32] = 4.63
       [N03]&ture = (0.823) [1.41] = 1.16
       [OC]&ture = (0.976) [3.26] = 3.18
       [EC]&ture = (0.784) [0.76] = 0.60
       (Soil)&ture = (1.112) [0.64] = 0.71
       [CM]&ture = (1.058) [8.11] = 8.58
                                             90

-------
Similar calculations (using the same model derived component specific RRFs) are performed for
each of the "worst" and "best" days in each of the other 4 years in the base period79.

Step 5. Using the results in Step 4, calculate the future year extinction coefficients for the
20% best and 20% worst visibility days in each of the five base years.

       Using future PM components obtained in Step 4, we can estimate future daily total bext.

For year 1

Worst Days
Day 6: bext = (3)(4.6)[6.50] + (3)(4.6)[0.59] + (4)[3.35] + (10)[0.53] + 1.10 + (0.6)[11.87] + 10
= 134.8 Mm'1

Day 7: bext = (3)(4.4)[7.19] + (3)(4.4)[0.67] + (4)[3.10] + (10)[0.51] + 0.91 + (0.6)[11.24] + 10 =
139.0 Mm'1

Best Days

Day 1: bext = (3)(1.7)[3.94] + (3)(1.7)[1.84] + (4)[3.29] + (10)[0.70] + 0.36 + (0.6)[7.76] + 10 =
64.6 Mm'1

Day 10: bext = (3)(2.7)[4.63] + (3)(2.7)[1.16]  + (4)[3.18] + (10)[0.60] + 0.71 + (0.6)[8.58] + 10 =
8 1.4 Mm'1

Step 6. Use the daily total bext values from step 5 to calculate future mean deciview values
for the best and worst days.

       Calculate daily deciview values (from total daily extinction) and then compute the future
average mean deciviews for the worst and best days for each year.  Then average the 5 years
together to get the final future mean deciview value for the worst and best days.

For Year 1

Worst days

Day 6: 134.8 Mm4 = 10* ln(134.8/10) = 26.0 dv

Day 7: 139.0 Mm'1 = 10* ln(139.0/10) = 26.3 dv
       79The same mean RRFs (for each component) are applied to the concentrations on all of
the worst days for each of the 5 years of data. The mean RRFs are derived from the modeled
days.  A separate set of RRFs are derived (in the same manner) for the best days.

                                           91

-------
Future mean visibility on worst days = (26.0 + 26.3) 12 = 26.2 dv

Best Days

Day 1: 64.6 Mm'1 = 10* ln(64.6/10) = 18.7 dv

Day 10:  81.4 Mm'1 = 10* ln(81.4/10) = 21.0 dv

Future mean visibility on best days = (18.7 + 21.0) 12 = 19.8 dv

Similar calculations are performed for previously selected "worst" and "best" days in each of
years 2-5. To illustrate the uniform rate of progress assessment, assume these other calculations
yield the following estimates for future mean dv on worst and best visibility days.

Table 6.7
Year
1
2
3
4
5
Future Mean dv On Worst
Visibility Days
26.2
27.3
27.6
25.9
26.1
Future Mean dv On Best
Visibility Days
19.8
19.3
17.6
18.3
18.6
Using results in table 6.7, we see that the estimated future average mean visibility for the 20%
days with worst visibility is

       dvfuture = (26.2 + 27.3 + 27.6 + 25.9 + 26.1) / 5 = 26.6

The estimated future average mean extinction coefficient for the 20% days with best visibility is

       dvfuture = (19.8 + 19.3 + 17.6 + 18.3 + 18.6) / 5 = 18.7

The results generated in step 6 are then used to estimate the difference in deciviews for days with
worst visibility and then the difference in deciviews for days with best visibility.

For the 20% days with worst visibility,

       dv difference = 26.6 - 28.8  = - 2.2 deciviews
                                            92

-------
For the 20% days with best visibility,

       dv difference = 18.7 - 19.7 = -1.0 deciviews

       The future year strategy leads to an improvement in visibility of 2.2 deciviews. Thus, the
visibility improvement for the 20% days with the worst visibility is predicted to exceed the
uniform rate of progress (-2.2 dv vs. -2.0  dv). Also visibility is not predicted to deteriorate on
the best days but, in fact, is estimated to improve (by 1.0 deciviews)

       This information is used in the process of determining reasonable progress. The
remaining steps in the process are beyond the modeling guidance. The modeling can only be
used to determine the predicted improvement in visibility (and whether the improvement is on,
above, or below the glidepath).  It cannot determine the reasonable progress goals and it cannot
determine whether reasonable progress is met.
       6.5 How Do I Select Appropriate Inputs For The Uniform Rate of Progress
Analysis?

       In Section 6.4, we described the recommended modeled uniform rate of progress
analysis. An important part of the analysis requires using component-specific relative response
factors (RRF's), obtained with models, to estimate future concentrations of these components
and, subsequently, future visibility.  In this subsection, we address more details concerning the
calculation of RRFs.  Second, there are several assumptions inherent in the recommended
modeled uniform rate of progress analysis. We identify these in this subsection and comment on
their underlying rationale.  More specifically, we address eight issues:

- how to estimate base period air quality in a Class I area without monitored data;
- how to handle a base year without data or with a small sample size;
- how to consider a day with missing data for one or several components of PM;
- use of the same days to estimate changes in visibility for the worst days and a different set of
same days to estimate changes in visibility for the best days;
- which predictions to use to derive relative response factors;
- how many  and what kind of days to use to develop RRF values
- use of relative response factors which are specific for days with poor vs. days with good
visibility
-alternative RRF calculations

       Estimating baseline "worst" and "best" visibility in a Class I area without monitors.
There are 156 Class I areas and there are 110 IMPROVE sites in or near Class I areas.
Therefore, approximately 46 Class I areas do not have co-located IMPROVE monitors. EPA's
"Tracking Guidance" recommends using IMPROVE data from a nearby site to represent the
visibility at each Class I area that does not have ambient data. Table A-2 in Appendix A (of the
                                           93

-------
tracking guidance) lists the recommended monthly f(rh) values80 for each Class I area as well as
the representative site for each Class area.  The representative IMPROVE site data will be used
to track regional haze progress for the Class I areas.  Therefore, it follows that visibility
improvement should be predicted at the monitor, not at the actual Class I area. For the purposes
of deriving ambient data for modeling, we recommend following the same representative site
assignments contained in the tracking guidance. In this way, the 20% worst and best days can be
derived for each Class I area from the network of 110 IMPROVE sites81. Similarly, the
modeling results should be extracted for the location of the representative monitor, not the actual
location of the Class I area82.

       Considering a base year with little or no monitored particulate matter or missing
data. The Tracking Guidance recommends calculating baseline visibility values for sites with at
least 3 out of 5 complete years of data.  It further contains recommendations for determining if a
year has complete data. In general, a site should have 50% data completeness in all quarters and
meet a 75% completeness criteria for the full year. There should be no periods with more than
30 consecutive days without data. The guidance assumes that all IMPROVE sites will have at
least 3 complete years  of data in the base 2000-2004 period.  The guidance also contains
procedures for filling in missing data as part of the calculation of the 20% best and worst days.
We recommend that these same procedures are followed for calculations in the modeled test.

       There are several data completeness issues that may cause problems within the modeling
analysis.  First, a site may have less than 3 years of complete data.  This will be a problem for the
purpose of modeling and for tracking progress. States should work with their EPA Regional
Office and  Federal Land Managers to determine how to estimate baseline visibility for these
area(s).

       Another issue that is a more specific problem for modeling occurs when data is missing
during the meteorological time period that was modeled. It is likely that most States will only be
modeling a single year (or less).  Therefore it is possible that some or all of the ambient data at
one or more Class I areas is missing during that year.  Without ambient data, it is impossible to
identify the 20% best and worst days (used to calculate modeled RRFs).
       80The f(rh) values in the Tracking Guidance were developed for the "old" IMPROVE
algorithm and should not be used to calculate visibility using the "new" IMPROVE algorithm.

       81Bering Sea Wilderness (Alaska) is the only Class I area that has no IMPROVE monitor
and no representative IMPROVE site. On-site IMPROVE or representative IMPROVE data can
be found for the other 155 sites.

       82It may be informative to compare model response at the representative IMPROVE site
versus the actual location of the Class I area. Large differences in the model response may
indicate that the "representative" site may not adequately represent visibility at the Class I area.

                                           94

-------
       Again, if this occurs, States should work with their Regional Office and FLM to
determine the best way to calculate visibility on the best and worst days. Potential options are to
use data from another nearby IMPROVE site, use nearby data from a different ambient data
network, or interpolate ambient data to the site. Another alternative is to estimate the 20% best
and worst days from the model outputs83.

       Using a constant sample of days to estimate baseline and future visibility. For a
typical Class I area, there will be about 120 days per year having measurements needed to
estimate (dv)baseline with Equation (6.1). Thus, there should be about 24 "worst" and 24 "best"
visibility days for each of the 5 years in the base period. It is conceivable that the identity of
these "worst" and "best" days could change if emissions were altered to reflect net effects of
controls and growth. The recommended test described in Section 6.4 assumes that the identity of
the "worst" and "best" days remains unchanged. This is done primarily to avoid having to
perform iterative analyses to identify future worst and best visibility days and to  keep the test
relatively simple and more readily understood.  This assumption could cause improvement in
visibility to be overestimated for the "worst" days and could also cause the test to overestimate
the difficulty in preventing deterioration of visibility on the  "best" days.  However, for the
reasons described below, we do not believe the effects of this assumption are substantial.

       It is unlikely that there would be any wholesale change in the identity of "worst" or
"best" days with future vs. current emissions. Analyses performed by Meyer, et  al. (1997) have
shown that the predicted ranked severity of high ozone days is largely unaffected by simulated
controls and growth (i.e., highest days tend to remain the highest days after the effects of growth
and controls are simulated). There is no reason to expect a different outcome for other
secondary pollutants. If there are differences, we would expect these to occur near the
borderline between  the "worst" days and more moderate days.

       Because the uniform rate of progress analysis relies on mean visibility values on 20 or
more "worst" visibility days and most of these days are unlikely to change, we would expect
little difference in the outcome of the analysis. Further, because of the shape of the distribution
of extinction coefficients, the mean of the worst days is more heavily influenced  by extreme days
rather than those on the borderline between "worst" and more moderate light extinction.   There
could be differences in some "best" visibility days corresponding with pre- and post-control
emissions. However, because the differences in concentrations of particulate matter on such
       83This is only recommended if the model performance for all species is good (at sites with
data). While it is expected that the air quality models will do an adequate job of predicting the
PM2 5 components which are responsible for the largest visibility degradation, the models may
not perform as well in estimating total visibility impairment. Accurately estimating the 20% best
and worst days depends on accurately estimating all of the PM components. In particular,
overprediction of certain components (e.g nitrates in the winter) could lead to identification of
worst or best days in the wrong season and/or for the wrong reasons (e.g. wrong mix of species
on the best or worst days).

                                           95

-------
days are likely to be relatively low, differences in the computed mean visibility for "best" days
are likely to be small.  Further, any resulting difference in the progress analysis for "best" days is
likely to be protective of the environment. If our recommended procedure leads to suspected
problems in the outcome of a test, a State may perform a more rigorous version of the analysis
(in which the identity of pre-control and post-control days changes) as part of a weight of
evidence determination.

       Selecting predictions to use in deriving RRF. Relative response factors should be
developed for each Class I area.  When a Class I area contains a monitoring site, the RRF
estimates should be derived using predictions which are made "near" that site. "Near" is defined
in Section 3.3. For each day, daily average surface predictions of each component of PM made
near a monitor should be estimated. These nearby estimates should then be spatially averaged to
estimate a spatially representative daily concentration. Spatially representative daily
concentrations obtained for each modeled day with monitored data should then be averaged.
This final average should be used to compute the RRF. Thus, component-specific RRF values
for a Class I area with a monitor are the ratio of the temporally averaged spatial mean of nearby
concentrations predicted with future emissions to that predicted with baseline emissions.  The
recommended procedure is illustrated in section  3.3 for a grid whose cells are 12 km on a side.
Similar to the PM2 5 NAAQS attainment test, nearby grid cells should be averaged. Note that for
cells larger than 15 km on a side, no spatial averaging is necessary—States should just use the
prediction in the cell containing the monitor.

       Selecting days to derive RRF values.  RRF values should be estimated by taking the
ratio of future predictions averaged over several days to current predictions averaged over the
same several days. It may often happen that a regional planning organization or a group of
States decides to model effects of a strategy for numerous Class I areas simultaneously. As we
note in Section 14, this may make it advisable to simulate (at least) a full year so that relative
response factor (RRF) values for each Class I area is based on a substantial number of observed
"best" and "worst" days. For the "worst" days in the chosen year, the RRF for a component of
PM should be estimated as the ratio of its arithmetic mean predicted value on the 20% worst
days with future emissions to that with baseline emissions. Thus, the RRF should reflect values
averaged over ~ 24 "worst" days in that year. The same procedure is followed to derive RRFs
over the -24 "best" days in the year.

       If it is not feasible to model an entire year, or if only a small number of Class I areas is to
be considered, a State should examine when worst visibility is observed to occur.  Choose a
sample of days from each quarter in which an incident of "worst" visibility occurs and calculate
a RRF estimate for each component of PM.  The appropriate RRF value would be applied to
monitored "worst visibility" days.  There is not, as yet, a good basis for suggesting a minimum
number of days to choose for this purpose.  However, information presented in Section 14.1.4
suggests that this number should be > -10 days.

       Since meteorological  conditions and/or emissions may be markedly different on "best"
visibility vs. "worst" visibility days, we recommend calculation of a separate set of RRF values

                                           96

-------
for "best" visibility days.  As with "worst" days, the preferred approach is to model an entire
year and select an RRF value for concentrations averaged over the 20% "best" visibility days for
each Class I area. If this is not feasible or only a limited number of Class I areas are to be
considered in an analysis, States may review when "best" visibility days are observed to occur in
the base period.  Model > ~ 10 days with observed "best" visibility and average predicted
baseline and future concentrations for each PM component. The RRF values are the ratios of the
future to baseline modeled averages. The appropriate RRF values should then be used in concert
with each observed "best" day to estimate future concentrations for each component on each
identified "best" day.

       Alternative RRF calculations. The default glidepath analysis is relatively simple in that
a single mean RRF is calculated for each PM component (separate RRFs on worst and best
days).  A series of tests with more complicated methods has shown that 1) the difference
between various versions of the test are usually small and 2) each of the alternative tests has
limitations in its applicability (Environ, 2005).  Possible variations include the use of day
specific RRFs, season (or quarter) specific RRFs, or climatological based RRFs. In some cases,
these more complicated techniques may provide different answers, but sometimes not. There are
specific limitations noted with each of these alternatives.  We have chosen to keep the single
mean RRF test as the default recommendation.  States are encouraged to explore other methods
for estimating RRFs if it is felt that the default recommendation is too simplistic to accurately
capture the change in future visibility at any particular Class I area. The SIP demonstration
should use the most appropriate method of projecting future concentrations for the characteristics
of each Class I area84.  Alternative methods should be discussed in the modeling protocol and
discussed with the appropriate EPA Regional Office and FLMs.
       84In particular, issues may arise when dealing with significant visibility contributions
from fires, coarse mass and fine soil (mostly wind-blown dust), and international transport (and
possibly other issues). Each of these issues should be addressed in the modeling protocol and
solutions should be discussed with the appropriate EPA Regional Office(s) on a case by case
basis.

                                           97

-------
7.0 How Can Additional Analyses Can Be Used to Support the
Attainment Demonstration?

       By definition, models are simplistic approximations of complex phenomena.  The
modeling analyses used to assess whether various emission reduction measures will bring an
individual area into attainment for the NAAQS contain many elements that are uncertain (e.g.,
emission projections, meteorological inputs, model response). These uncertain aspects of the
analyses can sometimes prevent definitive assessments of future attainment status. The
confidence in the accuracy of the quantitative results from a modeled attainment test should be a
function of the degree to which the uncertainties in the analysis were minimized.  In general, by
following the recommendations contained within this guidance document, EPA expects that the
attainment demonstrations will mitigate the uncertainty as much as is possible given the current
state of modeling inputs, procedures, and  science. However, while Eulerian air quality models
represent the best tools for integrating emissions and meteorological information with
atmospheric chemistry and no single additional analysis can match the expected reliability of
these models' results, EPA believes that all attainment demonstrations will be strengthened by
additional  analyses that can supplement the modeling analysis to enhance the assessment of
whether the planned emissions reductions will result in attainment.

       Supplemental evidence should accompany all model attainment demonstrations.
Generally, those modeling analyses that show that attainment will be reached in the future with
some margin of safety (e.g., estimated concentrations below 82 ppb for ozone, 14.5 ug/m3 for
annual PM2 5, and 62 ug/m3 for 24-hour PM2 5) will need more limited supporting material. For
other attainment cases, in which the projected future design value is closer to the NAAQS, more
rigorous supporting analyses  should be completed. As noted earlier (see section 2.2), there may
be some areas that can expect to achieve timely attainment despite failing the model attainment
test, and vice versa. This section of the guidance will discuss some specific additional analyses
that can be used to supplement the model  projections.  Of particular interest are analyses that
help  determine whether the model likely overpredicts, underpredicts, or accurately predicts the
air quality improvement projected to occur by the attainment date.  States should review these
supplemental analyses, in combination with the modeling analysis, in a "weight of evidence"
assessment of whether each area is likely to achieve timely attainment. Additional examples of
possible weight of evidence determinations are provided in existing EPA guidance (U.S. EPA,
1996a). Again, it should be noted that no single supplemental analysis can serve as an adequate
substitute for the air quality model; however, in aggregate, supplemental analyses may provide
information which may indicate a different outcome than the modeled test.

       In Section 7.1, we identify several broad types of analyses which can be used to
corroborate one another as part of a set of supplemental analyses. States should utilize each of
these types of analysis to ensure that conclusions regarding adequacy of a proposed strategy are
based on a variety of independent analyses.  In Section 7.2, we more specifically address the
submittal of a weight of evidence demonstration and recommend a framework for compiling and
submitting a weight  of evidence demonstration.
                                          98

-------
       The goals for regional haze between the base period and 2018 relate to trends in light
extinction rather than to some absolute value, as is the case for the NAAQS.  Thus, supplemental
analyses differ in some respects for visibility-related applications. Section 7.3 describes how to
use supplemental analyses in regional haze analyses.
       7.1    What Types of Additional Analyses Should Be Completed as Part of the
Attainment Demonstration?

       There are three basic types of analyses that are recommended to supplement the primary
modeling analysis. They are:

1) Additional modeling

2) Analyses of trends in ambient air quality and emissions

3) Observational models and diagnostic analyses

These broad groups of supplemental analyses are discussed in more detail below:

Modeling:  The relative attainment tests described in sections 3, 4, and 5 are the primary
modeling tools used in an attainment demonstration. The application of a photochemical grid
model, developed on a regional or local scale, is the best tool available to judge the impacts of
changes in future year emissions. In addition to this "primary" modeling analysis, there are
various other models, applications, and tools that can be used to supplement the results of the
modeled attainment test. These include, but are not limited to:

•      Use of available regional or national scale modeling applications that are suitable85 for
       the local area,

•      Use of other appropriate local modeling attainment demonstrations that include the
       nonattainment area of interest,

•      Use of photochemical source apportionment and/or process analysis modeling tools to
       help explain why attainment is (or is not) demonstrated,
       85 The resolution, emissions, meteorology, and other model inputs should be evaluated for
applicability to the local nonattainment area. Additionally, model performance of the regional
modeling for the local nonattainment area should be examined before determining whether the
regional model results are suitable for use in the local attainment demonstration.

                                           99

-------
•      Use of multiple air quality models / model input data sets (e.g., multiple meteorological
       data sets, alternative chemical mechanisms or emissions inventories, etc.).  For results to
       be most relevant to the way we recommend models be applied in attainment
       demonstrations, it is preferable that such procedures focus on the sensitivity of estimated
       relative response factors (RRF) and resulting projected design values to the variations in
       inputs or model formulations.

•      Use of dispersion models to address primary PM2 5 contributions to monitors.  In areas
       with large spatial gradients of primary PM25, dispersion models are best suited to
       characterizing the change in primary PM2 5 in the future. Areas that are relying on local
       primary PM controls to reach attainment should submit a local area analysis as part of the
       primary attainment demonstration. In other areas, a local area analysis may be useful as a
       supplemental analysis.

•      Application of the attainment test with alternative procedures compared to the default
       recommendations in Sections 3,  4, and 5 of this guidance.  Any alternate approaches
       should be accompanied with a technical justification as to why the approach is
       appropriate for the area in question and should be discussed with the appropriate EPA
       regional office.

       As discussed in Section 2, EPA has determined that the best approach to using models to
demonstrate attainment of the NAAQS is to use a model in a relative mode. However, for some
model applications there may be strong evidence from the performance evaluation that the
model is able to reproduce detailed observed data bases with relatively little error or bias.
Particularly for cases such as these, some types of "absolute" modeling results may be used to
assess general progress towards attainment from the baseline inventory to the projected future
inventory86. There are several metrics that can be considered as part of this type of additional
analysis:

•      Percent change in total amount of ozone or PM2 5 >= NAAQS87 within the nonattainment
       area
•      Percent change in number of grid cells >= NAAQS within the nonattainment area
•      Percent change in grid cell-hours (days) >= NAAQS within the nonattainment area
•      Percent change in maximum modeled 8-hour ozone within the nonattainment area
       86 Care should be taken in interpreting absolute metrics if the model evaluation shows a
large underprediction or overprediction of ozone or PM2 5 concentrations. An underprediction of
observed concentrations will make it artificially easy to show progress towards absolute
attainment levels and an overprediction will make it artificially difficult to show progress
towards attainment.

       87For each of these metrics, the appropriate comparison to the level of the NAAQS is 85
ppb for 8-hour ozone; 65 ug/m3 for 24-hour PM25; and 15 ug/m3 for annual PM25.

                                           100

-------
       While these metrics can be used to estimate the magnitude, frequency, and relative
amount of ozone or PM25 reductions from any given future emissions scenario, there are no
threshold quantities of these metrics that can directly translate to an attainment determination.
Generally, a large reduction in the frequency, magnitude, and relative amount of 8-hour ozone
nonattainment (i.e., >= 85 ppb) or PM2 5 nonattainment (24-hour and/or annual) is consistent
with a conclusion that a proposed strategy would meet the NAAQS.  In the context of a weight
of evidence determination, these metrics could be used to suggest that a particular location may
be "stiff or relatively unresponsive to emissions controls, while the rest of the modeling
domain/nonattainment area is projected to experience widespread reductions. If a sound
technical argument can be made for why atypically high RRFs at any particular location are not
reasonable, then these types of supplemental modeling metrics would suggest that attainment is
more likely to be achieved than the modeling analysis alone would indicate.

       As discussed in section 3.4, an unmonitored area analysis may provide evidence that the
area may not achieve timely attainment, even if modeling suggests that attainment will occur at
all monitoring locations. In such cases, assessment of metrics concerning the frequency,
magnitude, and relative amount of nonattainment may help supplement the information from the
unmonitored area analysis. If application of the unmonitored area test indicates that most (if not
all) of the unmonitored areas will be in attainment, then that information would be evidence that
future attainment may be likely.

       Uncertainty estimates associated with the spatial interpolation technique can also be
considered when reviewing and interpreting the results of an unmonitored area analysis.  When
making a decision on whether attainment is likely to occur, areas with very high uncertainty
estimates for interpolated design values should be given less weight than areas with low
uncertainty estimates88.

       The overall modeling analyses can also be evaluated to determine how appropriate the
modeling systems are for making regulatory decisions.  Roth (2005) has proposed an "idealized
evaluation framework" for judging the quality of modeling applications.  The paper lists a series
of twenty questions which can be used to judge the overall model application. These questions
provide an objective way to compare the quality, and identify deficiencies in modeling
applications.  The absence of major deficiencies may provide a strong basis for acceptance of
model results.

Trends in Ambient Air Quality and Emissions: Generally, air quality models are regarded as
the  most appropriate tools for assessing the expected impacts of a change in  emissions.
However, it may also be possible to extrapolate future trends in ozone or PM2 5 based on
measured historical trends of air quality and emissions. There are several elements to this
analysis that are difficult to quantify.  First, in most cases, the ambient data trends are best
        8This is particularly true for PM2 5 in areas with little or no ambient speciation data.

                                           101

-------
assessed by normalizing to account for year-to-year meteorological variations. Second, one must
have an accurate accounting of the year-to-year changes in actual emissions (NOx, VOC, and/or
SO2 and NH3) for the given area and any surrounding areas whose emissions may impact local
concentrations. Third, one must have a solid conceptual model of how ozone or PM25 is formed
in the local area (e.g., NOx-limited, ammonia limited, transport-influenced, etc.).  Assuming all
of these prerequisites can be met, then it may be possible to develop a curve that relates past
emissions changes to historical and current air quality.  Once the relationship between
past/present emissions and air quality is established, this "response  factor" can be applied to the
expected emissions reductions from a particular control strategy.

       If available, meteorologically adjusted  ozone  and PM2 5 concentrations can be used to
establish air quality trends.  There are several techniques that have been used to examine the
influence of meteorology on air quality. Among them are (a) statistical modeling  (U.S. EPA,
2005c ); (b) filtering techniques (Rao and Zurbenko,  1995, Flaum, et al.. 1996, Milanchus, et al..
1998, Hogrefe, et al.. 2000), (c) using a probability distribution of meteorological  severity based
on climatological data (Cox and Chu, 1993, 1996), (d) using CART analysis to identify
meteorological classes and selecting days from each year so that the underlying frequency of the
identified meteorological classes remains the same (Stoeckenius, 1990, Deuel and Douglas,
1996). Most of this work has  examined the relationship between ozone and meteorology. Only
recently have analyses examined the relationship between meteorology and PM2 5. Additionally,
compared to PM2 5, the established relationship between ozone and  meteorological variables is
stronger (higher r-square values). In the case of PM25, the relationship between concentration
and meteorology is complicated by the fact that PM2 5 components experience high
concentrations at different times of the year and for different reasons.  This makes it more
difficult to meteorologically adjust PM2 5 concentrations.

       If a meteorologically adjusted trend in ozone or PM25 can be estimated, then the
information can be used to establish a link between emissions and air quality trends.  This is not
always straightforward due to the multitude of emissions precursors that may lead to high ozone
and PM25 concentrations. A careful analysis of (meteorologically adjusted) air quality trends
and emissions trends of each of the ozone and  PM precursors (as well as primary PM) is needed
to fully establish relationships. Detailed emissions information as well as a solid understanding
of the conceptual model of ozone or PM2 5 formation  is needed.  If a trend can be established
based on past emissions changes and air quality changes, then future year predicted emissions
levels can be used to extrapolate future air quality.

       A simpler (and more uncertain) way to qualitatively assess progress toward attainment is
to examine recently observed air quality and emissions trends.  Downward trends  in observed air
quality and in emissions (past and projected) are consistent with progress towards attainment.
Strength of the evidence produced by emissions and air quality trends is increased if an extensive
monitoring network exists and if there is a good correlation between past emissions reductions
and current trends in ozone or PM2 5.  EPA recently prepared a report that analyzed statistically
significant trends in ozone (U.S. EPA, 2004a and  U.S. EPA, 2005c) and ozone precursor
                                           102

-------
emissions as well as a report which examined trends in PM2 5 concentrations and precursors
(U.S. EPA, 2004b).

       Weight given to trend analyses depends on several factors. Analyses that use more air
quality data and apply a greater variety of trend parameters provide more credible results. More
weight can be placed on the results if the procedure used to normalize the trend for
meteorological differences explains much of the variability attributable to these differences.  In
addition, trend analysis is more believable if the extrapolation does not extend very far into the
future. Finally, trend analysis is most credible if the contemplated strategy is similar to a past
strategy (e.g., both strategies focus on reducing sulfates for PM or NOx for ozone). For
example, if a past strategy focused on reducing sulfates, but a future one envisions controlling
OC, there is no guarantee that ambient OC will respond similarly to changes in past emissions.

Observational Models In some cases ambient data can be used to corroborate the effects of a
control strategy (e.g., Blanchard et al, 1999; Croes et al, 2003; Koerber and Kenski, 2005).
Observational models take advantage of monitored data to draw conclusions about the relative
importance of different types of emissions and precursors as factors contributing to observed
PM2 5 and ozone, as well as inferences which might be drawn about the effectiveness of various
strategies to reduce concentrations. Observational models can be used to examine days which
have not been modeled with an air quality model, as well  as days which have been modeled.
The resulting information may be useful for drawing conclusions about the general
representativeness of the responses simulated with the air quality model for a limited sample of
days. However, their ability to estimate how much control is needed is limited. Thus,
observational approaches are suitable to corroborate results from more quantitative techniques,
like air quality models. Additionally, observational models are limited to analyses based on
current or past ambient conditions. A change in the relative mix of precursor emissions in the
future (compared to current conditions) may lead to  a different set of future control strategies
compared to what the observational model may indicate based on current data.  There are at least
two types of observational models: source apportionment (i.e., "receptor") models and indicator
species approaches.

Receptor models  A large body of literature describes the theory and use of receptor models to
identify and/or apportion sources which may be contributing to monitored air quality. Seigneur,
(1997), summarized in Seigneur, (1999), provides a review of receptor models, contains a more
complete description of the major  approaches, summarizes findings obtained in a number of
applications, and provides an extensive list of references.  Receptor models are most useful for
identifying contributions of various source categories to observed primary components of
particulate matter,  although, to a lesser degree, they can also be used to infer contributions from
ozone and secondary PM precursors.

       There are two major types  of receptor models. The first type is the chemical mass
balance model (CMB). A description and user's guide is  available for the CMB model (U.S.
EPA, 2004d). This model assumes that the user already has a good idea of what source
categories potentially contribute to observations at a monitoring site.  Speciated emissions

                                           103

-------
profiles for all source categories are then compared with speciated air quality measurements at a
monitor on a given day. A combination of source category contributions which minimizes
observed differences between calculated and observed concentrations for a set of PM species
(i.e., "fitting elements") is derived. One key assumption of the CMB approach is that there is no
substantial chemical transformation of particulate matter between the point of emissions and the
site of the ambient measurements.  Thus, the technique is limited when a large portion of
measured PM2 5 is secondary particulate matter.

       The second type of receptor models are multi-variate statistical models. Unlike with
CMB, a priori assumptions about contributing source categories are not needed. Another
distinction from the CMB approach is that multi-variate models look at day to day variations in
speciated observations rather than focusing on individual days.  The object of this is to identify
sets of chemical species which track one another well.  Statistical methods like cluster analysis
or principal component analysis are used to identify groups of days or groups of chemical
species which are associated with one another.  Common multi-variate statistical models include
positive matrix factorization (PMF) (Paatero and Tapper, 1994), and UNMIX (Maykut, 2003;
Poirot et. al, 2001) These models may be useful when confirming whether a strategy is reducing
the right sorts of sources.

Indicator Species Indicator species approaches are based on the predicted sensitivity of a
secondary pollutant's concentration to changes in different precursors for that pollutant. It is
possible to identify ratios of certain species which are good indicators of whether a secondary
pollutant is sensitive to reductions in precursor A or precursor B. Measurement of "indicator
species" is a potentially useful means for assessing which precursor category (e.g., VOC or
NOx) limits further production of ozone or secondary PM25 at a monitor's location at various
times of day and under various sets of meteorological conditions.  Several indicator ratios  have
been developed to examine the sensitivity of ozone to changes in NOx and VOC, and the
sensitivity of particulate nitrate to changes in NOX, VOC, and ammonia.

       Sillman (1998, 2002) and Blanchard, (1997, 1999, 2000, 2001) identify several sets of
indicator species which can be compared to suggest whether ozone is limited by availability of
VOC or NOx. Comparisons are done by looking at ratios of these species. States/Tribes should
consult the Sillman (1998, 2002) and Blanchard, (1997, 1999, 2000, 2001) references for further
details on measurement requirements and interpretation of observed indicator ratios (also see
Section 18.5.1 for more details).

       Ansari and Pandis (1998) have developed an indicator ratio of species and applied  it to
several combinations of secondary particulate matter present under different environmental
conditions. They use this ratio to predict how mass of particulate matter will respond to
reductions in sulfate, nitrate and ammonia.  Blanchard, et al. (2000) have also examined how
indicator species might be used to assess whether particulate nitrate concentrations are limited by
NOx, VOC, or by ammonia emissions using mechanisms which incorporate reactions dealing
with secondary particulate matter. If a model accurately predicts observed ratios of indicator
                                           104

-------
species, then one can conclude with additional confidence that the predicted change in ozone or
PM may be accurate.

       The strength of the evidence produced by observational models is increased if an
extensive monitoring network exists and at least some of the monitors in the network are capable
of measuring pollutants to the degree of sensitivity required by the methods. Evidence produced
by observational models is more compelling if several techniques are used which complement
one another and produce results for which plausible physical/chemical explanations can be
developed. Indications of a strong quality assurance analysis of collected data and
measurements that are made by a well trained staff also lend credence to the results.
       7.2    What Is Entailed In A Weight Of Evidence Determination?

       As discussed in Section 2, augmenting a modeled attainment test with supplemental
analyses may yield a conclusion differing from that indicated by the modeled attainment test
results alone. Past modeling analyses have shown that future design value uncertainties of 2-4
ppb for ozone89, can result from use of alternate, yet equally appropriate, emissions inputs,
chemical mechanisms, and meteorological inputs (Jones, 2005; Sistla, 2004). Because of this
uncertainty, EPA believes that weight of evidence determinations can be used in some cases to
demonstrate attainment conclusions that differ from the conclusions of the model attainment test.

       As part of their recommendations to transform the SIP process into one that is more
performance-oriented, the Clean Air Act Advisory Committee (CAAAC) recommended
increased use of weight of evidence within State/Local attainment demonstrations (AQMWG,
2005). One of the workgroup's recommendations to EPA was that "EPA, in conjunction with
affected stakeholders, should modify its guidance to promote weight-of-evidence (WOE)
demonstrations for both planning and implementation efforts. In particular,  these demonstrations
should reduce reliance on modeling data as the centerpiece  for SIP/TIP planning, and should
increase use of monitoring data and analyses of monitoring data, especially for tracking progress.
Enhanced tracking and ambient monitoring data is a better use of available resources than
intensive local modeling."

       A weight of evidence (WOE) determination examines results from a diverse set of
analyses, including the outcome of the primary attainment test, and attempts to summarize the
results into an aggregate conclusion with respect to whether a chosen set of  control strategies
will result in an area attaining the NAAQS by the appropriate year.  The supplemental analyses
discussed above are intended to be part of a WOE determination, although the level of detail
       89Few studies have been done to examine similar uncertainties for PM2 5.  Based on recent
modeling analyses, a similar range of+- 2-4% of the NAAQS seems appropriate for PM25.  That
translates to roughly 0.3-0.6 ug/m3 for the annual PM2 5 standard. Consequently, the
recommended weight of evidence range for PM2 5 is nominally +- 0.5 ug/m3.

                                          105

-------
required in a WOE submittal will vary as a function of many elements of the model application
(e.g., model performance, degree of residual nonattainment in the modeled attainment test,
amount of uncertainty in the model and its inputs, etc.). Each weight of evidence determination
will be subject to area-specific conditions and data availability.  Area-specific factors may also
affect the types of analyses which are feasible for a nonattainment area, as well as the
significance of each. Thus, decisions concerning which analyses to perform and how much
credence to give each need to be made on a case by case basis by those implementing the
modeling/analysis protocol. States/Tribes are encouraged to consult with their EPA Regional
office in advance of initiating supplemental analyses to determine which additional analyses may
be most appropriate for their particular area.

       The most useful supplemental analyses are those providing the best evidence as to how
much air quality improvement can be expected as compared to the improvement projected by the
air quality modeling analysis.  Each analysis is weighed qualitatively, depending on: 1) the
capacity of the analysis to  address the adequacy of a strategy and 2) the technical credibility of
the analysis. If the overall weight of evidence produced by the combination of the primary
modeling analysis and the  various supplemental analyses supports the attainment hypothesis,
then attainment of the NAAQS is demonstrated with the proposed strategy.  The end product of a
weight of evidence determination is a document which describes analyses performed, data bases
used, key assumptions and outcomes of each  analysis, and why a State/Tribe believes that the
evidence, viewed as a whole, supports a conclusion that the area will, or will not, attain the
NAAQS despite a model-predicted DVF concluding otherwise.  In conclusion, the basic criteria
required for an attainment  demonstration based on weight of evidence are as follows:

       1) A fully-evaluated, high-quality modeling analysis that projects future values that are
close to the NAAQS.

       2) Multiple supplemental analyses in each of the three various categories discussed above
(modeling, air quality/emissions trends analyses, observational models).

       3) A qualitative weighting for each  separate analysis based on its ability to quantitatively
assess the ability of the proposed control measures to yield attainment.

       4) A description of each of the individual supplemental analyses and their results.
Analyses that utilize well-established analytical procedures and are grounded with sufficient data
should be weighted accordingly higher.

       5) Where applicable, a written description as to why the full set of evidence leads to a
conclusive determination regarding the future attainment status of the area that differs from the
results of the modeled attainment test alone.
                                           106

-------
       7.3    What Role Should Weight Of Evidence Play In Visibility-Related Analyses?

       We believe a weight of evidence analysis is an appropriate option for States to use when
examining visibility trends. Unlike the NAAQS, progress goals address trends in air quality
rather than some absolute level of air quality. Thus, the focus of weight of evidence analyses
differs from those performed in NAAQS attainment demonstrations. In this subsection, we note
some potential supplemental analyses that may be used to examine visibility trends.  We then
identify several refinements to our recommended modeled test for uniform rate of progress
analyses if a State or regional planning organization believes these are warranted for a weight of
evidence determination.  We conclude by noting some potential concerns about the ability of
models to address days with very good visibility.

       Additional air quality modeling.  Sensitivity tests can be performed to see if
conclusions about trends in "worst" and "best" visibility are robust. One example of such an
analysis is applying a model with and without a more finely resolved nested grid near one or
more Class I areas.  A purpose of this would be to see whether conclusions are affected by the
degree of detail in which nearby sources are considered  A second example of an analysis would
be to consider alternative future emissions and/or differing growth rate assumptions.  This may
be a particular concern for regional haze analyses because the emissions projection period is
generally longer than for most ozone and PM2 5 attainment demonstrations. Uncertainty in
emissions and growth rates become more important as the projection period is lengthened.

       If trends in visibility for "worst" and/or "best" days are similar using sensitivity tests,
alternative models and/or alternative modeling approaches, this finding supports conclusions
reached in the uniform rate of progress analysis.

       Review of trends.  A review of trends generally involves a comparison, sometimes
qualitative, between past trends in reconstructed visibility and estimated changes in emissions
(e.g., early '90's to mid 'OO's). The IMPROVE dataset has a longer record than the
corresponding urban FRM and speciation sites. Some IMPROVE sites have data going back to
1989.  This information could be used to confirm that more control measures on previously
reduced emissions of a component or its precursors is likely to be useful. It may  also be used to
see whether certain PM components are becoming increasingly important sources of light
extinction.

       Observational models. Receptor models are potentially useful for flagging potential
importance of local sources, if any, in influencing measurements made in a Class I area. This
could lead to a refined treatment of a local source, either through a more finely resolved nested
grid model application or plume-in-grid modeling.  Trajectory models may also be useful for
identifying the types of meteorological conditions most often corresponding to observed "worst"
and "best" visibility in various Class I areas. This, in turn, may enable States to draw inferences
about the orientation of areas containing sources most likely to influence visibility in a Class I
area on days with "poor" and "good" visibility. Grid model based techniques such as tagging,

                                           107

-------
DDM, or source apportionment may also be useful in identifying areas and emissions sources
most responsible for visibility impairment on the worst or best days.

       Refinements to the recommended uniform rate of progress analysis. If a strategy for
meeting the glidepath appears generally successful, but the glidepath is not met in a limited
number of Class I areas, States may consider refining the recommended uniform rate of progress
analysis in some manner.  Refinements are best made if they are based on local
observations/analyses which suggest that some of the underlying assumptions in the
recommended assessment may not be applicable. We list some potential refinements which
could be considered. The list is intended to illustrate types of additional analyses which could be
performed.

       - Use an alternative light extinction equation such as the revised IMPROVE equation or
       an area-specific version.

       -Available speciated data and other information may be reviewed to see whether the
       outcome of the test is being influenced by including one or more days with extraordinary
       events (e.g., a nearby major forest fire lasting a number of days or transported dust
       events).  If convincing arguments can be made that the event is a "natural" one,
       excluding these days from the calculations should be discussed with the appropriate U.S.
       EPA regional office.

       -Daily component specific RRFs can be examined to determine if one or more days are
       responding in a different way compared to the majority of the best or  worst days.
       Determining why days may be more or less responsive to emissions controls may lead to
       conclusions regarding the suitability of particular days to be represented in the mean
       response. It may be appropriate, in some cases, to re-calculate mean RRFs with suspect
       days removed.

       - Re-rank future estimated light extinction (i.e., bext values) for all days with current
       measurements and recompute mean future "best" and "worst" visibility (i.e., do not
       assume that the identity of baseline "best" and "worst" days remains the same).
       Concerns about modeling days with "best" visibility. In some parts of the United
States, concentrations of the components of particulate matter used in visibility calculations may
be within a i-ig/m3 or two of background levels on days with "best" visibility. Measurements and
model estimates may be subject to more relative uncertainty (i.e., lower signal to noise ratio) on
days where observed concentrations of parti culate matter are very low (and light extinction is
also low).  Utility of weight of evidence determinations is heightened in such cases.  If a State
has reason to believe that an atmospheric simulation model's  ability to estimate concentrations
of components of parti culate matter is limited on such days, performance tests described in
Section 18 should be applied to the extent feasible for the specific Class I area in question.  Next,
a State should see whether a model's inability to accurately predict one or more individual

                                          108

-------
components of particulate matter has a substantial effect on the extinction coefficient calculated
with Equations (6.1) or (6.2).  If it does, and diagnostic tests (also described in Section 18) are
unable to resolve a performance problem, a State may need to address the goal for "best"
visibility days in the particular Class I area(s) without using results from a grid model.
                                            109

-------
8.0  What Additional Analyses and Data Collection Can Be Completed to
Assess Progress towards Attainment

       The purpose of an attainment demonstration is to provide a best estimate as to whether
the control measures included in a State Implementation Plan will result in attainment of the
NAAQS by a specific date in the future.  In most cases, it will be desirable to periodically track
the air quality improvements resulting from the SIP to ensure that the plan is going to result in
attainment by the appropriate dates. One possible tracking approach is a mid-course review
(MCR). Also, the regional haze rule requires periodic tracking of progress towards visibility
goals (every 5 years). In this section, we identify measurements and activities which will
provide better support for mid course reviews, future modeling exercises and other supplemental
analyses designed to determine  the progress toward attainment of the NAAQS or reasonable
progress goals.  Improved data bases will increase the reliability of reviews and enable
identification of reasons for attainment or non-attainment of the NAAQS.

       Deploying additional air quality monitors. One type of additional monitoring which
should be considered has already been mentioned in Section 3. Additional ozone and/or PM2 5
monitors should be deployed in unmonitored locations where future design values are predicted
to exceed the NAAQS via the unmonitored area test. This would allow a better future
assessment of whether the NAAQS is being met at unmonitored locations. Also, additional PM
speciation monitors should be deployed in areas that are expected to continue to exceed the
NAAQS in the future (particularly if there is not a speciation monitor at the highest FRM sites in
a nonattainment area).  Additional ambient data can provide an increased understanding of the
nature of the problem, which can help lead to the most cost  efficient solution.

       Collecting Measurements for Indicator Species Measurement of "indicator species" is
a potentially useful means for assessing which precursor category limits further production of
ozone or secondary PM25 at a monitor's location at various  times of day and under various sets
of meteorological conditions. Many of the gas phase species needed to calculate indicator ratios
are not routinely measured.  Some of the species needed to calculate ozone indicator ratios are
NOy and HNO3. Some of the species needed to calculate the gas ratio and excess ammonia (for
nitrate analyses) include NH3 and HNO3.  Thus, realizing the potential of the "indicator species
method" as a tool for model  performance evaluation and for diagnosing why observed
concentrations do or do not meet previous expectations may depend on deploying additional
monitors and/or measurements.

       Making measurements aloft. Almost all measured ambient air quality and
meteorological data are collected within 20 meters  of the earth's surface. However, the
modeling domain extends many kilometers above the surface. Further, during certain times of
day (e.g., at night) surface measurements are not always representative of air quality or
meteorological conditions aloft. Measurements aloft can be made by aircraft (usually during
special studies) or on tall buildings or towers.  Concentrations aloft can have marked effects
when they are mixed with ground-level emissions during daytime.  Thus, the weight given to
                                          110

-------
modeling results can be increased if good agreement is shown with air quality measurements
aloft.  The most important of these measurements are ozone, NOy, NO, NO2, as well as several
relatively stable species like CO and selected VOC species.  Measurements of SO2 may also be
helpful for identifying presence of plumes from large combustion sources.

       Measurements of altitude, temperature, water vapor, winds and pressure are also useful.
Continuous wind measurements, made aloft in several locations, are especially important.  They
provide additional data to "nudge" meteorological model fields, but more importantly also allow
for construction of more detailed conceptual models  of local ozone formation (Stehr, 2004).
For example, measurements of aloft winds and temperatures from lower atmosphere radar
profilers can detect low level jets and be used to infer mixing depths. This information can be
used to evaluate meteorological and air quality model outputs.  This provides greater assurance
that the air quality model correctly reflects the configuration of sources contributing to ozone
formation.

       Special Studies. Over the last 20 years, many States have embarked upon short term
special studies to examine both ozone and PM issues. Data collected from special studies can be
used to make improvements in the conceptual model of ozone and/or PM formation and to
improve the inputs to models.  The results of current modeling can be examined to determine the
largest sources of uncertainty.  This information can  be used to design special studies in an effort
to collect data that might allow future improvements in emissions, meteorological,  and air
quality modeling. Examples of special studies include the Central California Ozone Study
(CCOS) (http://www.arb.ca.gov/airwavs/CCOS/CCOS.htmX California Regional Particulate Air
Quality Study (CRPAQS) (http://www.arb.ca.gov/airways/crpaqs/publications.htm). Southern
California Ozone Study (SCOS) (http://www.arb.ca.gov/research/scos/scos.htm). Texas Air
Quality Study (TexAQS) and TexAQS II
(http://www.tceq.state.tx.us/implementation/air/airmod/texaqs-files/TexAQS_II.html), Northern
Front Range Air Quality Study (NFRAQS) (http://www.nfraqs.colostate.edu/) , and various
NARSTO sponsored field studies (http://www.narsto.org/section.src?SID=9 ).

       Extending measurements year round. Many  States do not collect ozone and VOC data
(and possibly other species) outside of the officially designated ozone season. Because PM25 is
a year round issue (especially for the annual NAAQS),  PM precursor data is needed both during
and outside of the ozone season to adequately evaluate secondary PM25 performance and
sensitivity to emissions controls. This is particularly relevant in areas with high wintertime
nitrate concentrations. Collection of ozone, VOC, and precursor data on a year round basis may
help improve performance evaluations and in particular, diagnostic analyses.

       Collecting locally applicable speciated emissions data. While the U.S. EPA maintains
a library of default VOC and PM emissions species profiles  (U.S. EPA,  1993); (U.S. EPA,
2006a), some of these may be dated or may not properly reflect local sources. Use of speciated
emissions data is a critical input to air quality models. For example, the accurate representation
of the VOC speciation of current and future gasoline emissions may have an important impact on
future ozone concentrations. Efforts to improve speciation profiles for local sources (especially

                                          111

-------
for large sources of primary PM25) should enhance credibility of the modeling as well as several
of the procedures recommended for use in supplemental analyses and the weight of evidence
determinations.
                                          112

-------
9.0  What Documentation Do I Need To Support My Attainment
      Demonstration?

      States/Tribes should follow the guidance on reporting requirements for attainment
demonstrations provided in U.S. EPA (1994b). The first seven subjects in Table 9.1 are similar
to those in the 1994 guidance.  The 1994 guidance envisions an air quality model as the sole
means for demonstrating attainment. However, the current guidance (i.e., this document)
identifies supplemental analyses as well as a possible weight of evidence determination as a
means for corroborating/refuting the modeled attainment test in an attainment demonstration.  In
addition, feedback received  since the earlier guidance has emphasized the need for technical
review of procedures used to identify a sufficient control strategy.  Thus, we have added two
additional subject areas which should be included in the documentation accompanying an
attainment demonstration. These are a description of the supplemental analyses and/or weight of
evidence  determination, and identification of reviews to which analyses used in the attainment
demonstration have been subject.  In the end, the documentation submitted by the States/Tribes
as part of their attainment demonstration should contain a summary section which addresses the
issues shown in Table  9.1.

Table 9.1    Recommended Documentation for Modeled Attainment Demonstrations and
             Regional Haze Analyses
        Subject Area
       Purpose of
    Documentation
       Issues Included
  Conceptual Description
Characterization
(qualitative and
quantitative) of the
area's nonattainment
problem; used to guide
the development of the
modeling analysis.
Emissions and air quality
assessment;

Measurements used;

Analyses performed;

Processes, conditions, and
influences for ozone, PM,
and/or regional haze
formation.
                                        113

-------
      Subject Area
       Purpose of
    Documentation
       Issues Included
Modeling/Analysis
Protocol
Communicate scope of
the analysis and
document stakeholder
involvement.
Names of organizations
participating in preparing and
implementing the protocol;

Types of analyses performed;
Steps followed in each type of
analyses;

Rationale for choice of the
modeling system  and model
configurations.
Emissions Preparations
and Results
Assurance of valid,
consistent emissions data
base.  Appropriate
procedures are used to
derive emission
estimates needed for air
quality modeling.
Data base used and quality
assurance methods applied;

Data processing used to
convert data base to model-
compatible inputs;

Deviations from existing
guidance and underlying
rationale;

VOC, NOx, SO2, NH3, PM2 5,
PM10, and CO (as
appropriate) total emissions
by State/County and for major
source categories.
                                      114

-------
       Subject Area
       Purpose of
    Documentation
       Issues Included
Air Quality/Meteorology
Preparations and Results
Assurance that
representative air
quality and
meteorological inputs
are used in analyses
Description of data base and
procedures used to derive and
quality assure inputs for
modeling;

Departures from guidance and
their underlying rationale.

Performance of meteorological
model used to generate
meteorological inputs to the
air quality model.
Performance Evaluation
for Air Quality Model (and
Other Analyses)
Show decision makers
and the public how well
the model (or other
analyses) reproduced
observations on the days
selected for analysis for
each nonattainment area
and appropriate sub-
regions.
Summary of observational
data base available for
comparison;

Identification of performance
tests used and their results
(including diagnostic
analyses);

Ability to reproduce observed
temporal and spatial patterns;

Overall assessment of what the
performance evaluation
implies.
                                      115

-------
       Subject Area
       Purpose of
     Documentation
        Issues Included
Description of the Strategy
Demonstrating Attainment
Provide the EPA and the
public an overview of the
plan selected in the
attainment demonstration.
Qualitative description of the
attainment strategy;

Reductions in VOC, NOx, SO2,
NH3, and/or CO emissions from
each major source category for
each State/county/Tribal land
from current (identify) emission
levels;

Clean Air Act mandated
reductions and other reductions;

Show predicted relative response
factors for ozone and/or each
component of PM25 and regional
haze(as applicable)

Show predicted future design
values for the selected control
scenario and identify any
location(s) which fails the
unmonitored area analysis
described in Section 3;

Identification of authority for
implementing emission
reductions in the attainment
strategy.

Evidence that emissions remain
at or below projected levels
throughout the 3-year period
used to determine future
attainment for PM2 5 and ozone
and/or 5-year period for uniform
rate of progress assessments.
                                          116

-------
       Subject Area
       Purpose of
    Documentation
       Issues Included
Data Access
Enable the EPA or other
interested parties to
replicate model
performance and
attainment simulation
results, as well as results
obtained with other
analyses.
Assurance that data files are
archived and that provision
has been made to maintain
them;

Technical procedures for
accessing input and output
files;

Identify computer on which
files were generated and can
be read, as well as software
necessary to process model
outputs;

Identification of contact
person, means for
downloading files and
administrative procedures
which need to be satisfied to
access the files.
                                       117

-------
       Subject Area
       Purpose of
    Documentation
       Issues Included
Supplemental
Analyses/Weight of
Evidence Determination
Assure the EPA and the
public that the strategy
meets applicable
attainment tests and is
likely to produce
attainment of the
NAAQS and/or uniform
rate of progress by the
required time.
Description of the modeled
test and observational data
base used;

Identification of air quality
model used;

Identification of other analyses
performed;

Outcome of each analysis,
including the modeled
attainment test;

Assessment of the credibility
associated with each type of
analysis in this application;

Narrative describing process
used to conclude the overall
weight of available evidence
supports a hypothesis that the
selected strategy is adequate to
attain the NAAQS.
Review Procedures Used
Provide assurance to the
EPA and the public that
analyses performed in
the attainment
demonstration reflect
sound practice
Scope of technical review
performed by those
implementing the protocol;

Assurance that methods used
for analysis were peer
reviewed by outside experts;

Conclusions reached in the
reviews and the response.
                                      118

-------
This page is intentionally left blank
                119

-------
Part II.  How Should I Apply Air Quality Models To
        Produce Results Needed To Help Demonstrate
        Attainment?
                     120

-------
10.0  How Do I Apply Air Quality Models?-- An Overview

       In Part I of this guidance, we described how to estimate whether a proposed control
strategy will lead to attainment of the NAAQS or meet the uniform rate of progress goals within
a required time frame. We noted that air quality models play a major role in making this
determination. We assumed that modeling had been completed, and discussed how to use the
information produced. We now focus on how to apply models to generate the information used
in the modeled attainment demonstration. The procedure we recommend consists of nine steps:

    1. Formulate a conceptual description of an area's nonattainment problem;
    2. Develop a modeling/analysis protocol;
    3. Select an appropriate air quality model to use;
    4. Select appropriate meteorological episodes to model;
    5. Choose a modeling domain with appropriate horizontal and vertical resolution and
    establish the initial and boundary  conditions to be used;
    6. Generate meteorological and air quality inputs to the air quality model;
    7. Generate emissions inputs to the air quality model;
    8. Evaluate performance of the air quality model and perform diagnostic tests, as necessary.
    9. Perform future year modeling (including additional control strategies,  if necessary) and
    apply the attainment test

       In this section, we briefly describe each of these steps to better illustrate how they are
inter-related.  Because many of these steps require considerable effort to execute, States/Tribes
should keep the appropriate U.S. EPA Regional Office(s) informed as they proceed. This will
increase the likelihood of having an approvable attainment demonstration when the work is
completed. The steps outlined in this section are described in greater depth in Sections 11-18.

1. Formulate a conceptual description of an area's nonattainment problem.  A State/Tribe
needs to have an understanding of the  nature of an area's nonattainment problem before it can
proceed with a modeled attainment demonstration.  For example, it would be difficult to identify
appropriate stakeholders and develop a modeling protocol without knowing whether resolution
of the problem may require close coordination and cooperation with other nearby States.

       The State/Tribe containing the designated nonattainment area or the RPO containing the
Class I area is expected to initially characterize the problem. This characterization provides a
starting point for addressing steps needed to generate required information by those
implementing the protocol. Several examples of issues addressed in the initial description of a
problem follow.  Is it a regional or local problem? Are factors outside of the  nonattainment area
likely to affect what needs to be done locally? Are monitoring sites observing violations located
in areas where meteorology is complex or where there are large emission gradients? How has
observed air quality responded to past efforts to reduce precursor emissions?  Are there ambient
measurements suggesting which precursors and sources are important to further reduce ozone,
PM2 5, and/or regional haze? What information might be needed from potential stakeholders?
As many of the  preceding questions imply, an initial conceptual description may be based

                                          121

-------
largely on a review of ambient air quality data.  Sometimes, methods described in Sections 7 and
8 (e.g., trend analysis, observational models) may be used. Other times, these types of analyses
may be deferred until after a team is in place to develop and implement steps following a
modeling/analysis protocol. The initial conceptual picture may be based on less resource-
intensive analyses of available data.

2. Develop a modeling/analysis protocol.  A protocol describes how modeling will be
performed to support a particular attainment demonstration or uniform rate of progress analysis.
The content of the protocol and identification of participating stakeholders are influenced by the
previously developed conceptual description of the problem. The protocol outlines methods and
procedures which will be used to perform the subsequent steps needed to generate the modeling
results and then apply the modeled attainment test and unmonitored area analysis as well as
other corroborating analyses.  This is accomplished by: a) identifying those responsible for
implementing the modeling, b) outlining the specific steps needed to complete the attainment
demonstration, c) identifying those who will review each step as it occurs, and d) identifying
procedures to be used to consider input/suggestions from those potentially affected  by the
outcome (i.e., "stakeholders"). In short, the protocol defines the "game plan" and the "rules of
the game".

3. Select an appropriate model for use.   This step includes reviewing non-proprietary, grid-
based photochemical models to select the model that is most appropriate for the application in
terms of (a) state-of-the science  algorithms  to represent the chemical and physical processes
associated with ozone and/or PM formation, transport, and removal, (b) peer review, (c) model
performance in prior applications, and (d) ease of use.  Identifying the air quality model to be
used is an early step in the  process,  since it  may affect how emissions and meteorological
information are input to the model.  It could also affect size of the area modeled and choice of
the horizontal/vertical resolution considered. For PM2 5 nonattainment applications where large
spatial gradients of primary PM25 exist, a dispersion model may be needed in addition to a
photochemical model.

4. Select appropriate meteorological time periods to model.  Like the preceding step, this
step requires review of available air quality and meteorological data. It also requires a thorough
understanding of the form of the ozone and PM2 5 NAAQS, the regional haze goals, and the
modeled attainment tests described in Sections 3-6.  Finally, it requires a review of
meteorological conditions which represent annual PM2 5 concentrations, both good and bad
visibility days, and time  periods  which have been observed to accompany monitored
exceedances of the 8-hour ozone or 24-hour PM2 5 NAAQS. The object of these reviews is to
select periods which a) include days with observed concentrations exceeding site-specific design
values (8-hour ozone or 24-hour PM25 NAAQS), b) select a representative mix of days for each
quarter in applications dealing with the annual  NAAQS and/or c) select days which are
representative of those corresponding with good and poor visibility.

       Due to increased computer speeds, it is now prudent to recommend modeling relatively
long time periods. For 8-hour ozone and 24-hour PM2 5,  at a minimum, modeling episodes which

                                           122

-------
cover full synoptic cycles is desirable.  Depending on the area and the time of year, a synoptic
cycle may be anywhere from 5-15 days. Modeling even longer time periods of up to a full
season may simplify the episode selection process and provide a rich database with which to
apply the modeled attainment test. For annual average PM and regional haze applications, we
recommend modeling (at least) a full year.

5. Choose a modeling domain with appropriate horizontal and vertical resolution and
establish the initial and boundary conditions.  Nested grid models will typically be used to
support the modeled attainment test.  In order to provide reasonable boundary conditions for the
local nonattainment area, in many cases it is important to model a large regional domain with
relatively coarse resolution, and a  smaller sub-regional domain with relatively fine horizontal
resolution. Meteorological and air quality data corresponding to the time periods that will be
modeled need to be reviewed prior to choosing size of the area modeled.  Appropriate domain
size is influenced by the air quality goal being addressed, whether the model is being applied to
address ozone and/or primary and/or secondary particulate matter, and the choice of days
modeled.

       The presence  of topographical features or mesoscale meteorological features (e.g.,
land/sea breeze) near or in the nonattainment area of principal interest are factors to consider in
choosing size of individual grid cells and the number of required vertical layers for that portion
of the modeling grid. Another factor affecting the choice of grid cell size is the available spatial
detail in the emissions data used as input to an emissions model. Finally, factors which cannot
be ignored in choosing size of a domain and its grid cells include the feasibility of managing
large data bases and the resources  needed to estimate meteorological inputs and air quality in
many grid cells.

6. Generate meteorological inputs to the air quality simulation model.  Prognostic
meteorological models will ordinarily be used to generate the meteorological inputs used in the
attainment demonstration modeling.  The application of meteorological models and the choice of
model grid resolution in the preceding  step are closely related. Meteorological conditions near
the area which is the focus of an attainment demonstration may dictate the required spatial
resolution.  On the other hand, cost and data management difficulties increase greatly for finely
resolved grids.  Thus, those implementing the protocol will likely be faced with a tradeoff
between cost/feasibility of running air quality and meteorological models and resolution at which
it might be most desirable to treat dispersion of nearby emissions.

7. Generate emissions inputs to the air quality simulation model.  Emissions are the central
focus in a modeled attainment demonstration because they are the only input which is altered
between the present and future  case scenarios and represent the model input to which control
strategies are applied. Emissions inputs to an air quality model are generated using an emissions
model. Applying such a model is as complicated as the air quality model itself, and demands at
least as much attention.  In current emissions models, emissions from some of the major source
categories are affected by meteorological conditions. This requires an interface between
meteorological inputs and emissions.  The development of emissions data must also take into

                                           123

-------
account the horizontal/vertical model resolution of the model configuration and the size of the
area to be modeled. In short, treatment of emissions is a central and complex one which, itself,
involves several steps. These include deriving emission inventories, quality assuring results,
applying results in an emission model(s), and (again) quality assuring results. Emission inputs
may be needed for number of scenarios including; (1) a base case corresponding to that of the
selected meteorological time periods, (2) a baseline corresponding to that represented by the
baseline monitored design value, (3) a future base case when attainment of the NAAQS or
regional haze progress needs to be demonstrated, and (4) control scenarios in which additional
emissions controls are applied to emissions in the future base case.

8. Evaluate performance of the air quality simulation model and perform diagnostic tests.
The credibility of a modeled attainment test and other model outputs is affected by how well the
model replicates observed historical air quality. Evaluating model performance and conducting
diagnostic tests depend on the prior definition of the modeling exercise and specification of
model inputs.  Hence, this is generally the last step prior to using the model to support an
attainment demonstration or glidepath assessment.  In the past, the performance evaluation for
ozone has relied almost exclusively on numerical tests comparing predicted and observed ozone,
or visual inspection of predictions and observations.  These are still important tools. However,
photochemical grid models have many inputs, and it is possible to get similar predicted ozone
and/or PM25 concentrations with different combinations of inputs. There is no guarantee that
ozone and/or secondary PM2 5 will respond the same way to controls with these different
combinations of inputs.  Thus, we place greater emphasis on additional kinds of tests than was
the case in past guidance. These include use of precursor observations, indicator species, and
corroborative analyses with observational models. Diagnostic tests are separate simulations
which are performed to determine the sensitivity of a model's predictions to various inputs to the
model. This can be done for a variety of purposes, including selection of effective control
strategies, prioritizing inputs needing greatest quality assurance and assessing uncertainty
associated with model predictions.  In performing such tests, States/Tribes should remember how
model results are used in the modeled attainment tests recommended in Sections 3-6. Model
results are used in a relative rather than absolute sense.  Thus, diagnostic tests should be used to
consider how relative, as well as absolute predictions, are affected by changes to model inputs.

9. Perform future year modeling (including additional control strategies, if necessary) and
apply the attainment test.  The base case model runs for performance evaluations should
contain emissions inventories on a highly resolved basis which can best simulate the ozone
and/or PM2 5 concentrations that were measured.  For some sources, it may not be appropriate to
project day specific emissions to the future because they may not be representative of typical
base case conditions. This is commonly the case for wildfire and continuous emissions monitor
(CEM) based utility emissions.  If needed, a separate baseline model run should be completed for
the purpose of establishing a base to compare against future year model outputs (for calculating
relative response factors).

       The next step is to run the future year base case model run.  The inventory should
contain all known emissions controls expected to be in place in the future year, as well  as

                                           124

-------
projected growth of emissions to the future.  The attainment test should be performed using the
future base case and the (current year) baseline. If attainment cannot be shown, then model runs
which contain additional control measures are needed. Multiple future year control strategy runs
may need to be completed until the attainment test is passed.
                                           125

-------
11.0 How Do I Get Started?- A "Conceptual Description"

       A State/Tribe should start developing information to support a modeled attainment
demonstration by assembling and reviewing available air quality, emissions and meteorological
data. Baseline design values should be calculated at each monitoring site, as described in
Section 3. For PM applications, speciated data should be reviewed to get a sense of what
component(s) might be contributing most significantly to nonattainment or light extinction. If
past modeling has been performed, the emission scenarios examined and air quality predictions
may also be useful.  Readily available information should be used by a State/Tribe to develop an
initial conceptual description of the nonattainment or regional haze problem in the area which is
the focus of a modeled demonstration.  A conceptual description is instrumental for identifying
potential stakeholders and for developing a modeling/analysis protocol.  It may also influence a
State's choice of air quality model, modeling domain, grid cell size, priorities for quality
assuring and refining emissions estimates, and the choice of initial diagnostic tests to identify
potentially effective control strategies.  In general, a conceptual description is useful for helping
a State/Tribe identify priorities and allocate resources in performing a modeled demonstration.

       In this Section, we identify key parts of a conceptual description. We then present
examples of analyses which could be used to describe each of these parts. We note that initial
analyses may be complemented later by additional efforts performed by those implementing the
protocol.

       11.1  What Is A "Conceptual Description"?

       A "conceptual description" is a qualitative way of characterizing the nature of an area's
nonattainment or regional haze problem. It is best described by identifying key components of a
description. Examples are listed below.  There are 3 different examples. One each for ozone,
PM2 5, and regional haze.  The examples are not comprehensive.  There could be other features of
an area's problem which are important in particular cases. For purposes of illustration later in
the discussion, we have answered each of the questions  posed below. Our responses appear in
parentheses.
              11.1.1 8-Hour Ozone NAAQS

    1. Is the nonattainment problem primarily a local one, or are regional factors important?

    (Surface measurements suggest transport of ozone close to 84 ppb is likely. There
    are some other nonattainment areas not too far distant.)

    2. Are ozone and/or precursor concentrations aloft also high?

    (There are no such measurements.)
                                            126

-------
3. Do violations of the NAAQS occur at several monitoring sites throughout the
nonattainment area, or are they confined to one or a small number of sites in proximity to
one another?

(Violations occur at a limited number of sites, located throughout the area.)

4. Do observed 8-hour daily maximum ozone concentrations exceed 84 ppb frequently or
just on a few occasions?

(This varies among the monitors from 4 times up to 12 times per year.)

5. When 8-hour daily maxima in excess of 84 ppb occur, is there an accompanying
characteristic spatial pattern, or is there a variety of spatial patterns?

(A variety of patterns is seen.)

6. Do monitored violations occur at locations subject to mesoscale wind patterns (e.g., at a
coastline) which may differ from the general wind flow?

(No.)

7. Have there been any recent major changes in emissions of VOC or NOx in or near the
nonattainment area? If so, what changes have occurred?

(Yes, several local measures [include a list] believed to result in major reductions in VOC
[quantify in tons per summer day] have been implemented in the last five years.
Additionally, the area has seen large regional NOx reductions from the NOx SIP call.)

8. Are there discernible trends in design values or other air quality indicators which have
accompanied a change in emissions?

(Yes, design values have decreased by about 10% at four sites over the past [x] years.
Smaller or no reductions are seen at three other sites.)

9. Is there any apparent spatial pattern to the trends in design values?

(No.)

10. Have ambient precursor concentrations or measured VOC species profiles changed?

(There are no measurements.)

11. What past modeling has been performed and what do the results suggest?
                                        127

-------
    (A regional modeling analysis has been performed. Two emission scenarios were modeled:
    current emissions and a substantial reduction in NOx emissions throughout the regional
    domain. Reduced NOx emissions led to substantial predicted reductions in 8-hour daily
    maximum ozone in most locations, but changes near the most populated area in the
    nonattainment area in question were small or nonexistent.)

    12. Are there any distinctive meteorological measurements at the surface or aloft which
    appear to coincide with occasions with 8-hour daily maxima greater than 84 ppb?

    (Other than routine soundings taken twice per day, there are no measurements aloft. There
    is no obvious correspondence with meteorological measurements other than daily maximum
    temperatures are always > 85 F on these days.)

      Using responses to the preceding questions in this example, it is possible to construct an
initial conceptual description of the nonattainment area's ozone problem.  First, responses to
questions 1 and 11 suggest there is a significant regional component to the area's nonattainment
problem. Second, responses to questions 3, 4, 7, 8, and 11 indicate there is an important local
component to the area's nonattainment problem. The responses to questions 4, 5 and 12 indicate
that high ozone concentrations may be observed under several sets of meteorological conditions.
The responses to questions 7, 8, and 11 suggest that ozone in and near  the nonattainment area
may be responsive to both VOC and NOx controls and that the extent of this response may vary
spatially. The response to question 6 suggests that it may be appropriate to develop a strategy
using a model with 12 km grid cells.

      The preceding conceptual  description implies that the  State/Tribe containing the
nonattainment area in this example will need to involve stakeholders from other, nearby
States/Tribes to develop and implement a modeling/analysis protocol.  It also suggests that a
nested regional modeling analysis will be needed to address the problem.  Further,  it may be
necessary to model at least several distinctive types of episodes and additional analyses will be
needed to select episodes. Finally, sensitivity (i.e.,  diagnostic) tests, or other modeling probing
tools, will be needed to assess the effects  of reducing VOC and NOx emissions separately and at
the same time.

             11.1.2 Annual PM25 NAAQS

    1. Is the nonattainment  problem primarily a local one, or are regional factors important?

           (Surface measurements suggest that only design values in or immediately downwind
    of the city violate the NAAQS. However, other nearby design values come close to the
    concentration specified in the NAAQS)

    2. What is the relative importance of measured primary and secondary components of PM25
    measured at sites violating the NAAQS?
                                            128

-------
       (Secondary components (i.e., SO4, NO3, OC) constitute about 80% of the measured
mass of PM25. There are higher concentrations of primary PM25 in the core urban area
compared to the suburbs and more rural areas.)

3. What are the most prevalent components of measured PM25?

       (The most important components in ranked order are mass associated with SO4, OC
and other primary particulate matter).

4. Does the measured mix of PM components appear to roughly agree with mix of emission
categories surrounding the monitoring sites?

       (No. Relative importance of measured other primary PM25 (OPP) appears less than
what might be inferred from the inventory).

5. Do there appear to be any areas with large gradients of primary PM2 5 in monitored or
unmonitored areas?

       (Cannot really tell for sources of primary PM2 5 material until we resolve the
preceding inventory/monitoring discrepancy. There are  no other  obvious major sources of
primary particulate matter).

6. Is there any indication of what precursor might be limiting formation of secondary
particulate matter?

       (No indicator species analyses have been performed. Past analyses performed for
ozone-related SIP revisions suggest that ozone in this area may be limited by availability of
VOC).

7. Do monitored violations occur at locations subject to mesoscale wind patterns (e.g., at a
coastline) which may differ from the general wind flow?

       (No.)

8. Have there been any recent major changes in emissions of PM  or its precursors  in or near
the nonattainment area? What?

       (Yes, measures believed to result in major reductions in VOC and NOx have been
implemented in the last 5 years. Reductions in power plant NOx  have resulted from the
NOx SIP call and SO2 emissions reductions have resulted from the national program to
reduce acid deposition.)

9. Are there discernible trends in design values or other air quality indicators which have
accompanied a change in emissions?

                                        129

-------
           (The trend appears to be downward, but the most recent air quality data has been
    higher.  Overall, the period of record is insufficiently long to tell).

    10. Is there any apparent spatial pattern to the trends in design values?

           (No.)

    11. What past modeling has been performed and what do the results suggest?

           (A regional modeling analysis has been performed for ozone and PM2 5. Two
    emission scenarios were modeled:  current emissions and a substantial reduction in NOx and
    SO2 emissions throughout a regional domain. Reduced NOx emissions led to substantial
    predicted reductions in 8-hour daily maximum ozone in most locations. Modeled SO2
    reductions from the CAIR rule had a strong impact on sulfate concentrations)

    12. Are there any distinctive meteorological measurements at the surface or aloft which
    appear to coincide with occasions with PM2 5 concentrations in excess of 15.0  |ig/m3?

           (Other than routine soundings taken twice per day, there are no measurements aloft.
    There is no obvious correspondence with meteorological measurements other than daily
    maximum temperatures are often > 85F on days with the highest PM2 5 observations.)

    13. Do periods with high measured parti culate matter or components of parti culate matter
    appear to track each other or any other measured pollutant?

           (There appears to be some correspondence between measured high concentrations of
    SO4 and ozone).

       Using responses to the preceding questions in this example, it is possible to construct an
initial conceptual description of the nonattainment area's PM25 problem.  First, responses to
questions 1, 2 and 3  suggest there is a significant regional component to the area's  nonattainment
problem. Second, responses to questions 1 and 3 indicate there is a local component to the
problem..  The responses to questions 11,12 and 13 suggest that there may be a link between
reducing ozone and reducing particulate matter.  Thus, it may be appropriate to assess effects of
previously committed to strategies to reduce ozone and national PM control measures before
simulating additional control measures. The responses to questions 4 and 5 suggest that it is
premature to determine whether a "local area analysis" will be needed.  The response to  question
7 suggests that it may not be necessary to model with very small grid cells, at least for the
secondary components of PM25.

       The preceding conceptual description implies that the State containing the nonattainment
area in this example will need to involve stakeholders from other, nearby  States to  develop and
implement a modeling/analysis protocol.  It also suggests that a nested regional modeling
analysis will be needed to address the problem.

                                          130

-------
              11.1.3 Example Regional Haze Application

    1. What components of participate matter appear to have high concentrations on days with
    poor visibility?

           (Mass associated with SO4 and coarse paniculate matter (CM) seem to have the
    highest concentrations on most such days).

    2.  What are typical values for the humidity adjustment factor during the times of year when
    most of the days with poor visibility occur?

           (Typical values appear to be about "4.0").

    3. Does visibility appear to track well among nearby Class I areas?

           (Yes, but not always).

    4. Does poor visibility seem to occur under any specific meteorological conditions?

           (This information is not readily available).

     5. Does poor visibility seem to coincide with high observed concentrations of any particular
    other pollutant?

           (There seems to be some correspondence with high regional ozone concentrations)

    6. What components of particulate matter appear to have relatively high concentrations on
    days with good visibility?

           (Coarse particulate matter and OC)

    7. What are typical values for the humidity adjustment factor during times of year when
    most of the days with good visibility occur?

           (About "2.3")

    8. Does good visibility appear to occur under any specific meteorological conditions?

           (Don't know.)

       Answers to the preceding  questions suggest that strategies to reduce sulfate
concentrations and, perhaps, regional ozone concentrations might be effective in reducing light
extinction on  days when visibility is currently poor.  The responses suggest that a strategy which
                                           131

-------
focuses on this alone should first be tried for the days with good visibility as well.  Even though
sulfate concentrations appear low on such days, the fact that sulfates scatter light efficiently (see
Equation (6.1)) and relative humidity is still high enough to enhance this effect is worth
considering. Responses suggest that further meteorological analyses would be worthwhile prior
to selecting strategies to simulate with a resource intensive regional model.

       It should be clear from the preceding examples that the initial conceptual description of
an area's nonattainment problem draws on readily available information and need not be
detailed.  It is intended to help launch development and implementation of a modeling/analysis
protocol in a productive direction. It will likely be supplemented by subsequent, more extensive
modeling and ambient analyses performed by or for those implementing the modeling/analysis
protocol discussed in Section 12.0.

       Questions like those posed in Section 11.1 can be addressed using a variety of analyses
ranging in complexity from an inspection of air quality data to sophisticated mathematical
analyses. We anticipate the simpler analyses will often be used to develop the initial conceptual
description.  These will be followed by more complex approaches or by approaches requiring
more extensive data bases as the need later becomes apparent. These analyses are intended to
channel resources available to support modeled attainment demonstrations onto the most
productive paths possible. They will also  provide other pieces of information which can be used
to reinforce conclusions reached with an air quality model, or cause a reassessment of
assumptions made previously in applying the model.  As noted in Section 7, corroboratory
analyses  should be used to help assess whether a simulated control strategy is sufficient to meet
theNAAQS.
                                           132

-------
12.0  What Does A Modeling/Analysis Protocol Do, And What Does
       Developing One Entail?

       Developing and implementing a modeling/analysis protocol is a very important part of an
acceptable modeled attainment demonstration. The protocol should detail and formalize the
procedures for conducting all phases of the modeling study, such as describing the background
and objectives for the study, creating  a schedule and organizational structure for the study,
developing the input data, conducting model performance evaluations, interpreting modeling
results, describing procedures for using the model to demonstrate whether proposed strategies
are sufficient to  attain the NAAQS and/or regional haze goals, and producing documentation to
be submitted for EPA Regional Office review and approval. Much of the information in U.S.
EPA (1991a) regarding modeling protocols remains applicable.  States/Tribes should review the
1991 guidance on protocols. In this document, we have revised the name of the protocol to
"Modeling/Analysis Protocol" to emphasize that the protocol needs to address modeling as well
as other supplemental analyses.

       12.1   What Is The Protocol's Function?

       As noted above, the most important function of a protocol is to serve as a means for
planning and communicating up front how a modeled attainment demonstration will be
performed. The protocol is the means by which States/Tribes, U.S. EPA, and other stakeholders
can assess the applicability of default recommendations and develop alternatives. A good
protocol should  lead to extensive  participation by stakeholders in developing the demonstration.
It should also reduce the risk of spending time and resources on efforts which are unproductive
or inconsistent with EPA policy.

       The protocol also serves several important, more specific functions.  First, it should
identify who will help the State/Tribe or local air quality agency (generally the lead agency)
undertake and evaluate the analyses needed to support a defensible demonstration (i.e., the
stakeholders). Second, it should identify how communication will occur among States/Tribes
and stakeholders to develop consensus on various issues.  Third, the protocol should describe the
review process applied to key steps in the demonstration. Finally,  it should also describe how
changes in methods and procedures or in the protocol itself will be agreed upon and
communicated with stakeholders and  the appropriate U.S. EPA Regional Office(s). Major steps
to implement the protocol should be discussed with the appropriate U.S. EPA Regional Office(s)
as they are being decided. States/Tribes may choose to update the  protocol as major decisions
are made concerning forthcoming analyses.
       12.2   What Subjects Should Be Addressed In The Protocol?

       At a minimum, States/Tribes should address the following topics in their
modeling/analysis protocol:
                                          133

-------
1. Overview of Modeling/Analysis Project
       a. Management structure
       b. Technical committees or other communication procedures to be used
       c. Participating organizations
       d. Schedule for completion of attainment demonstration or uniform rate of progress
analyses
       e. Description of the conceptual model for the nonattainment area (or Class I area(s))

2. Model and Modeling Inputs
       a. Rationale for the selection of air quality, meteorological,  and emissions models
       b. Modeling domain
       c. Horizontal and vertical resolution
       d. Specification of initial and boundary conditions
       e. Episode selection
       f. Description of meteorological model setup90
       g. Development of emissions inputs
       h. Geographic area identified for application of the attainment test(s)
       i. Methods used to quality assure emissions, meteorological, and other model inputs

3. Model Performance Evaluation
       a. Describe ambient data base
       b. List evaluation procedures
       c. Identify possible diagnostic testing that could be used to improve model performance

4. Supplemental Analyses
       a. List additional analyses to be completed to corroborate the model attainment test
       b. Outline plans for conducting a weight of evidence determination, should it be
necessary

5. Procedural Requirements
       a. Identify how modeling and other analyses will be archived and documented
       b. Identify specific deliverables to EPA Regional Office
       90States may want to document detailed meteorological modeling decisions in a separate
document.  At least a brief description of the meteorological model and model setup should be
contained in the air quality modeling protocol.

                                          134

-------
13.0  What Should I Consider In Choosing An Air Quality Model?

       Photochemical grid models are, in reality, modeling systems in which an emissions
model, a meteorological model and an air chemistry/deposition model are applied. In this
guidance, we use the term "air quality model" to mean a gridded photochemical modeling
system. Some modeling systems are modular, at least in theory. This means that it is possible to
substitute alternative emissions or meteorological models within the modeling system. Often
however, the choice of an emissions or meteorological model or their features is heavily
influenced by the chosen air quality model (i.e., an effort is needed to develop software to
interface combinations of components differing from the modeling system's default
combination). Thus, choosing an appropriate air quality model is among the earliest decisions
to be made by those implementing the protocol.  In this section, we identify a set of general
requirements which an air quality model should meet in order to qualify for use in an attainment
demonstration or regional haze application. We then identify several factors which will help in
choosing among qualifying air quality models for a specific application.  We conclude this
section by identifying several air quality models which are available for use in attainment
demonstrations or regional haze assessments.  Meteorological and emissions models are
discussed in Sections 16 and 17, respectively.
       13.1   Types of Air Quality Models

       States should use a photochemical grid model to simulate the effects of strategies to
reduce ozone and the secondary components of particulate matter (i.e., mass associated with
SO4, NO3 and secondary OC). Because of the regional nature of "regional haze" and relatively
high efficiency with which secondary particulate matter scatters light, we believe that
photochemical grid models are needed to perform uniform rate of progress assessments.  Based
on its conceptual description of a PM2 5 nonattainment problem, a State could conclude that the
PM25 problem can be addressed solely by reducing primary components of measured PM2591

       Greater flexibility is possible in choosing a modeling approach to address primary
components  of PM25 (i.e., OPP, EC and the primary portion of OC) and coarse particulate matter
(i.e., needed for regional haze-related applications) than is true for secondary components.  That
is, it is not necessary to use a model which considers atmospheric chemistry in addressing
changes in primary components.  Either a numerical grid or a Lagrangian (such as a Gaussian
dispersion) model may be used.  In general, modeling primary PM components with a grid
model is acceptable, but further refined dispersion modeling may be necessary in areas with
large spatial  gradients of primary PM. If a dispersion model is used to estimate RRF's for
       91If this is the case, the State may not need to use a photochemical grid model in their
attainment demonstration if it can present convincing qualitative arguments that an increase in
the secondary components of PM will not cause reductions in the primary components to be
insufficient to meet the NAAQS.

                                          135

-------
primary particulate matter, these estimates should be made at the monitoring location (see
Section 5.3).  Thus, a State may use a regional photochemical grid model to address both
primary and secondary components of particulate matter or they may use a photochemical model
to address secondary PM and an inert model applied over a more limited domain to address
primary components.

       In regional haze-related applications it may suffice when modeling primary particulate
matter to use grid cells which are the same size as those used to model secondary components of
particulate matter. If there is no  reason to believe that there are major individual sources of
primary PM within about 50 km  which affect the monitor site in a Class I area, primary
components can be considered using the same coarse grid used for the photochemical grid
model.  This is generally a good  default assumption.
       13.2   What Prerequisites Should An Air Quality Model Meet To Qualify For Use
             In An Attainment Demonstration?

       A model should meet several general criteria for it to be a candidate for consideration in
an attainment demonstration or uniform rate of progress assessment.  These general criteria are
consistent with requirements in 40 CFR 51.112 and 40 CFR part 51, Appendix W (U.S. EPA,
2005d). Note that, unlike in previous guidance (U.S. EPA, 1991a), we are not recommending a
specific model for use in attainment demonstration or uniform rate of progress assessment. At
present, there is no single model which has been extensively tested and shown to be clearly
superior than its alternatives.  Thus, 40 CFR Part 51 Appendix W does not identify a "preferred
model" for use in attainment demonstrations of the NAAQS for ozone or PM2 5 or uniform rate
of progress assessments for regional haze92.  Based on the language in 40CFR Part 51 Appendix
W, models used for these purposes should meet requirements for "alternative models".

       States/Tribes should use a non-proprietary model, which is a model whose source code is
available for free (or for a "reasonable" cost).  Furthermore, the user must be able to revise the
code93 to perform diagnostic analyses and/or to improve the model's ability to describe
observations in a credible manner.  Several additional prerequisites should be met for a model to
be used to support an attainment demonstration or uniform rate of progress assessment.
       92Appendix W recommends specific dispersion models for primary pollutants, but does
not recommend specific photochemical models.  Recommended dispersion models should be
used for primary PM25 analyses if dispersion models are used for a "local area analysis"(see
Section 5.3).

       93 Air quality models are generally identified by a version number.  The version of the
model that is used in SIP applications should be identified. Code revisions to standard versions
of models should be noted and documented.

                                          136

-------
    (1) It should have received and been revised in response to a scientific peer review.

    (2) It should be appropriate for the specific application on a theoretical basis.

    (3) It should be used with a data base which is adequate to support its application.

    (4) It should be shown to have performed well in past ozone or PM modeling applications.
    (If the application is the first for a particular model, then the State should note why it
    believes the new model is expected to perform sufficiently.)

    (5) It should be applied consistently with a protocol on methods and procedures.

       An air quality model may be considered to have undergone "scientific peer review" if
each of the major components of the modeling system (i.e., air chemistry/deposition,
meteorological and emissions models) has been described and tested, and the results have been
documented and reviewed by one or more disinterested third parties. We believe that it should
be the responsibility of the model developer or group which is applying an air quality model on
behalf of a State/Tribe to document that a "scientific peer review" has occurred.  States/Tribes
should then reference this documentation to gain acceptance of an air quality model  for use in  a
modeled attainment demonstration.


       13.3   What Factors Affect My Choice of A Model For A Specific Application?

       States/Tribes should consider several factors as criteria for choosing a qualifying air
quality model to support an attainment demonstration or uniform rate of progress assessment.
These factors are:  (1) documentation and past track record of candidate models in similar
applications; (2) advanced science and technical features (e.g., probing tools) available in the
model  and/or modeling system; (3) experience of staff and available contractors; (4) required
time and resources versus available time and resources; and (5) in the case of regional
applications, consistency with regional models applied in adjacent regions. Finally,  before the
results of a selected model can be used in an attainment demonstration, the model should be
shown to perform satisfactorily using the data base available for the specific application.

       Documentation and Past Track Record of Candidate Models.  For a model to be used
in an attainment demonstration, evidence should be presented that it has been found  acceptable
for estimating hourly and eight-hourly ozone concentrations and/or hourly and 24-hour average
PM2 5 and PM component concentrations.  Preference should be given to models exhibiting
satisfactory past performance under a variety of conditions. Finally, a user's guide (including  a
benchmark example and outputs) and technical description of the model should be available.

       Advanced Technical Features. Models are often differentiated by their available
advanced science features and tools. For example, some models include advanced probing tools
that allow tracking of downwind impacts from upwind emissions sources.  Availability of

                                          137

-------
probing tools and/or science algorithms is a legitimate reason to choose one equally capable
model over another.

       Experience of Staff and Available Contractors. This is a legitimate criterion for
choosing among several otherwise acceptable alternatives.  The past experience might be with
the air quality model itself, or with a meteorological or emissions model which can be more
readily linked with one candidate air quality model than another.

       Required vs. Available Time and Resources. This is a legitimate criterion provided the
first two criteria are met.

       Consistency of a Proposed Model with Models Used in Adjacent Regions. This
criterion is applicable for regional model applications. If candidate models meet the other
criteria, this criterion should be considered in choosing a model for use in a regional or nested
regional modeling application.

       Demonstration that an "Alternative Model" is Appropriate for the Specific
Application. If an air quality model meets the prerequisites identified in Section 13.2, a
State/Tribe may use the factors described in this section (Section  13.3) to show that it is
appropriate for use in a specific application. The selection of an "alternative model" needs to be
reviewed and approved by the appropriate U.S. EPA Regional Office.

       Satisfactory Model Performance in the Specific Application.  Prior to use of a
selected model's results in an attainment demonstration, the model should be shown to perform
adequately for the specific application. The approach for evaluating model performance are
discussed in Section 18.
    13.4   What Are Some Examples Of Air Quality Models Which May Be Considered?

        Air quality models continue to evolve and have their own strengths and weaknesses
(Russell, 2000).  Table 13.1 lists several current generation air quality models which have been
used to simulate ambient ozone, PM, and regional haze concentrations. Table 13.2 lists several
air quality models which have been used for various ozone and PM applications over the past
decade, but are not widely used at this time.  Table 13.3 lists several current dispersion models
that have used to model primary PM. Table 13.4 lists several receptor models that have been
used to identify sources of PM. The list is not intended to be comprehensive.  Exclusion of a
model from the list does not necessarily imply that it cannot be used to support a modeled
attainment demonstration or uniform rate of progress assessment. In the same way, inclusion on
the list does not necessarily imply that a model may  be used for a particular application.
States/Tribes  should follow the guidance in Sections 13.1 and  13.2 in selecting an air quality
model for a specific application.
                                          138

-------
Table 13.1   Current Air Quality Models Used To Model Ozone and PM
Air Quality Model
CAMx
CMAQ
UAM-V
References
Environ (2006a)
U.S. EPA (1998a)
Systems Applications
International (1996)
Table 13.2  Other Air Quality Models Used to Model Ozone and PM
Air Quality Model
CALGRID
MAQSIP
SAQM
URM
References
Scire. et al. (1989)
'
MCNC (1999)
Odman. et al. (1996)
'
Chans, et al.. (1997)
CARB (1996)
Kumar, et al.. (1996)
Table 13.3 Current Dispersion Models Used to Model Primary PM
Air Quality Model
AERMOD
CALPUFF
ISC3
References
U.S. EPA (2004f)
U.S EPA (2004g)
Scire (2000)
U.S. EPA (1995a)
U.S EPA (1995b)
Table 13.4 Current Receptor Models
Air Quality Model
PMF
UNMIX
CMB
References
Eberly (2005)
Hopke (2001)
Lewis (2003)
U.S. EPA (2004d)
                                   139

-------
14.0  How are the Meteorological Time Periods (Episodes) Selected?

       Historically, attainment demonstrations have been based on a limited number of episodes
consisting of several days each.  In the past, the number of days modeled has been limited by the
speed of computers and the ability to store the model output files.  With the advancement in
computer technology over the past decade, computer speed and storage issues are no longer an
impediment to modeling long time periods. In fact, many groups have recently modeled entire
summers and/or full years for ozone, PM25, and regional haze (Baker, 2004a) (U.S. EPA,
2005b).

       Ozone based research has shown that model performance evaluations and the response to
emissions  controls need to consider modeling results from relatively long time periods, in
particular, full synoptic cycles or even full ozone seasons (Hogrefe, 2000).  In order to examine
the response to ozone control strategies, it may not be necessary to model a full ozone season (or
seasons), but, at a minimum, we recommend modeling "longer" episodes that encompass full
synoptic cycles. Time  periods which include a ramp-up to a high ozone period and a ramp-down
to cleaner  conditions allow  for a more complete evaluation of model performance under a variety
of meteorological conditions.

       More limited research has been conducted related to PM2 5 modeling. Most model
applications for the annual PM2 5 NAAQS have modeled a full year (CAIR, RPOs, etc).  This is a
logical goal since every day of the year is included in the calculation of the annual NAAQS.
Several tests have also  been made using less than a full year of meteorological data. One study
(Environ, 2004) found  that modeling 14 days from each quarter provided results with a
reasonable amount of uncertainty compared to  modeling the full year. The following sections
contain further recommendations for choosing  appropriate time periods to model for attainment
demonstrations and glidepath assessments.

       At a minimum,  four criteria should be used to select time periods which are appropriate
to model:

       1) Simulate a variety of meteorological conditions:
             a) 8-Hour Ozone- Choose time periods which reflect a variety of meteorological
             conditions which frequently correspond with observed 8-hour daily maxima > 84
             ppb at multiple monitoring sites.
             b) 24-Hour PM25- Choose time periods  which reflect a variety of meteorological
             conditions which frequently correspond with observed 24-hour averages > 65
             ug/m3 at violating monitoring sites.
             c) Annual PM2 5- Choose time periods from each quarter which reflect the variety
             of meteorological conditions which represent average concentrations for that
             quarter and year
             d) Regional Haze- .Choose time periods which reflect the variety of
             meteorological conditions which represent visibility impairment on the 20% best
             and 20% worst days in the Class I areas being modeled.

                                          140

-------
       2) Model time periods in which observed concentrations are close to the appropriate
       baseline design value or visibility impairment.

       3) Model periods for which extensive air quality/meteorological data bases exist.

       4) Model a sufficient number of days so that the modeled attainment test applied at each
       monitor violating the NAAQS is based on multiple days (see section 14.1.1).

       These four criteria may sometimes conflict with one another.  For example, there may
only be a limited number of days with intensive data bases, and these may not cover all of the
meteorological conditions which correspond with monitored concentrations close to site-specific
design values during the base period.  Thus, tradeoffs among the four primary criteria may be
necessary in specific applications.

       Those implementing the modeling/analysis protocol may use secondary episode selection
criteria on a case by case basis. For example, prior experience modeling an episode or year, may
result in its being chosen over an alternative.  Another consideration should be to choose time
periods occurring during the 5-year period which serves as the basis for the baseline design value
(DVB). If observed ozone or 24-hour PM25 exceedances occur on weekends, weekend days
should be included within some of the selected time periods. If it has been determined that there
is a need to model several nonattainment  areas simultaneously (e.g., with a nested regional scale
model application), a fourth secondary criterion is to choose time periods containing days of
common interest to different nonattainment areas.

       In this section,  we first discuss each of the four identified primary criteria for choosing
meteorological episodes to model.  We then discuss the secondary criteria, which may be
important in specific applications.
       14.1  What Are The Most Important Criteria For Choosing Episodes?

       In the following section, we identify how the four selection criteria apply to ozone.  We
then discuss issues related to time period selection for PM2 5 and regional haze. We also
recommend minimum time periods to model for 8-hour ozone, 24-hour PM2 5, annual PM2 5, and
regional haze. As a matter of convenience, section 14.1 is organized by pollutant, but it could
also be organized by the nature of the ambient data.  For example, 8-hour ozone and  24-hour
PM2 5 episode selection criteria are similar because both standards are based on short term peak
concentration periods. Regional haze calculations are based on an average of 20 or more days
per year (20% best or worst days).  Therefore regional haze episode selection will likely include
more days throughout the year. The annual PM2 5 NAAQS is unique because each and every
ambient observation is included in the average.  It is likely that a full year (or more) will be
modeled to represent the annual PM2 5 standard.
                                          141

-------
              14.1.1 8-Hour Ozone NAAQS

Simulate a variety of meteorological conditions This criterion is important, because we want
to be assured that a control strategy will be effective under a variety of conditions leading to
elevated ozone concentrations.  The time periods chosen should reflect the variety of conditions
that lead to 8-hour ozone concentrations > 85 ppb.

       Those implementing the modeling/analysis protocol should describe the rationale for
distinguishing among episodes which are modeled.  The selection may reflect a number of area-
specific considerations. Qualitative procedures such as reviewing surface and aloft weather
maps, and observed or modeled wind patterns may suffice for distinguishing episodes with
distinctively different meteorological conditions. More quantitative procedures, such as a
Classification and Regression Tree (CART) analysis or a principal component analysis (PCA), to
identify distinctive groupings of meteorological/air quality parameters corresponding with high
daily maxima or averages, may sometimes be desirable. An example of a CART analysis
applied to select episodes is described by Deuel  (1998). LADCO used CART to rank historical
years for Midwestern cities by their conduciveness to ozone formation (Kenski, 2004).  A PCA
may also be used to characterize predominant meteorological conditions and relate those
conditions to ozone concentrations (Battelle, 2004). This information can be used to quantify the
relative "ozone forming potential" of different days, regimes, and years.

       The interpretation of results of a wind rose analysis or a statistical analysis such as PCA
or CART should focus on episodic time periods, rather than individual days.  The winds may be
blowing from different directions on consecutive days, but that does not necessarily  mean that
those days represent different meteorological regimes.  Preference should be given to modeling
episodic cycles.

       Additionally, statistical analyses such as  PCA normally limit the number of identified
meteorological regimes to a relatively small number of generalized patterns.  The analysis may
indicate that only one or two of these patterns are responsible for most or all of the exceedance
days in an area. But no two days and no two episodes are  exactly the same.  Further analysis
should be performed on potential episode periods to differentiate subtle, but often important,
differences between episodes. For this reason, it may be beneficial to model more than one
episode from the most frequently occurring meteorological regimes which lead to ozone
exceedances. Modeling a continuous time period which encompasses several episodes or a full
ozone season will make it easier to adequately account for all of the potential meteorological
conditions which correspond to high measured concentrations.

Choose episodes having days with monitored  8-hour daily maxima close to observed
average 4th high daily maximum ozone concentrations.

       We want to use episodes whose severity  is comparable to that implied by the form of the
NAAQS (i.e., an episode whose severity is exceeded, on average, about 3 times/year at the time
of the selected episode). The objective is to choose episodes with days which are approximately

                                          142

-------
as severe as the average 4th high 8-hour daily maximum concentration specified in the NAAQS.
As such, even if a full summer is being modeled, it is important to analyze the ambient data to
ensure that an adequate number of ozone conducive days are modeled.

       Air quality measurements recorded during the baseline/current period can be used to
characterize episode severity. This is done by selecting a 5-year period which "straddles" a
modeled episode.  For example, if an episode from 2002 were modeled, we recommend looking
at measured 8-hour daily maxima at each site in the nonattainment area during 2000-2004.
Using this information it should be  possible to assess the relative severity of the days chosen for
modeling at each site. Limiting this characterization to the five years straddling an episode
avoids problems posed by long term trends in emissions in assessing episode severity. However,
it leaves unanswered the question of whether the 5-year period selected to assess severity of a
modeled day is typical or atypical.  If there is an underlying long term trend in ambient ozone
attributable to meteorological cycles or emissions changes, it may not be appropriate to compare
different periods with one another using air quality observations.  Thus, if one uses a 10-year old
episode with an exceptional data base, there is greater uncertainty in ranking its severity relative
to the current period of interest than if the episode were drawn from the current period.

       Note that if the episode is drawn from a recent time period (especially the three years
upon which the nonattainment designation is based), days which are chosen are likely to have
monitored observations very close to the baseline design value. In the absence of such
information, we suggest "+ 10 ppb" as a default recommendation for purposes of prioritizing
choice of episodes94.  If the base and baseline/current periods do not coincide, "close to" is
within +10 ppb of the design value during the base period straddling the episode. If it is not
feasible to meet this default criterion for all monitoring sites, meeting it at sites with
baseline/current design values > 85 ppb should receive greatest priority.

Choose days  with intensive data bases.

       Preference should be given to days with measurements aloft, available measurements of
indicator species (see Section 18) and/or precursor measurements. These preferences result from
a desire to incorporate a rigorous model performance evaluation as a part of the attainment
demonstration.  This  reduces the likelihood of "getting the right answer for the wrong reason".
Thus, the likelihood of mischaracterizing ozone/precursor sensitivity is reduced.
       94The analysis in section 4.1 showed that relative model response may differ on "low"
ozone days compared to days that are at or above the level of the NAAQS.  Therefore, ambient
(and modeled) concentrations that are more than 10 ppb above the design value are preferable to
episodes with ambient concentrations that are more than 10 ppb below the design value.

                                           143

-------
Choose a sufficient number of days to enable the monitored attainment test to be based on
multiple days at each monitoring site violating the NAAQS.

       Figure 4.1 indicates that the relative response factor computed at any given site appears
to be affected by the minimum threshold value.  Based on an analysis of modeled data, the
recommended minimum (baseline) threshold value is 85 ppb. The minimum threshold value
analysis (detailed in section 4.1) was also used to examine how the number of days contained in
the mean RRF calculation influences the mean RRF. It was found that, on average, a minimum
of 10 modeled days (in the mean RRF calculation) produces mean RRFs that are relatively
robust.

       The analysis cited earlier in the guidance (Timin, 2005b) was used to help  determine the
minimum number of days to use in a mean RRF calculation. The dataset consisted of 206
monitoring sites which had at least 10 days with predicted 8-hour daily baseline maximum ozone
concentrations > 85 ppb.  In the analysis we assumed that a mean RRF calculated from a "large"
set of days is more stable than an RRF  calculated from a small set of days. Using information on
the variability of the model response on individual days, we are able to measure the variability of
the mean RRF on any subset of days. The analysis used datasets of 25, 50, and 100 days95.  The
standard deviation of the daily RRFs was used to create the datasets and measure the variability
of the RRFs.

       Figure 14.1 shows an example of the variability of the mean RRF as a function of the
number of days in the mean RRF calculation.  The example plot is for a monitoring site in
Harford County, MD. The  mean RRF for a 50 day sample size is 0.90 (10% ozone reduction).
The standard deviation of the daily RRFs was 0.034 (3.4%)96.  The plot shows the  range of the
mean RRFs calculated using a sample size ranging from 3 to 25 days (a subset of the 50 days).
Each subset (3 days, 4 days, 5 days, etc.) was  sampled  1000 times. As can be seen in the plot,
the range of mean RRFs varies widely for a small sample size (3 days) and is relatively stable for
a large sample size (25 days). As the number of days increases, the variability of the mean RRF
decreases.  A  similar conclusion was reached in a different study (Hogrefe, 2000) which found
that the RRF is more variable when based on a small number of days.
       95The 25, 50, and 100 day datasets were created by calculating the standard deviation of
the daily RRFs for the monitoring sites with at least 10 days > 85 ppb.  The distribution of the
RRF was calculated from the standard deviation. The original dataset had an actual maximum
number of 30 days.

       96The standard deviation is in "RRF units". For example, an RRF of 0.90 is equal to
alO% ozone reduction.  A standard deviation of 3.4% is a measure of the variability such that
ฑ3.4% is equal to a range in mean RRF of 0.866-0.934.

                                         144

-------
       The ability to accurately capture a mean RRF with a small number of days is dependent
on the variability of the daily RRFs (as measure by the standard deviation).  Sites with a small
standard deviation of the daily RRFs will be able to replicate the large dataset mean RRF with
relatively few days.

       Using the available information, we were able to calculate, for each monitoring site, the
number of days needed to provide a mean RRF calculation that is within ฑ1% and ฑ2% of the
"large dataset" mean, with a 95% confidence interval.  The number of days needed to produce a
robust mean RRF is dependent on the variability of the daily RRFs (as measured by the standard
deviation).  Therefore,  more days are needed to produce a stable RRF if the standard deviation of
the daily RRFs is high.

       Table 14.1 summarizes the results for the 25th, 50th (median) and 75th percentile of the
standard deviation for the 206 monitoring sites.  The table presents results for a range of standard
deviations, a range of large datasets (25, 50, and 100 days), and both  ฑ1% and ฑ2% accuracy.
The table shows that for the median standard deviation of the monitoring sites (2.4%), a
minimum number of 10-16 days is needed to replicate the mean RRF  to within ฑ1% (95% of the
time) and a minimum number of 5-6 days is needed to replicate the mean RRF to within ฑ2%
(95% of the time). The table also shows that a smaller standard deviation requires fewer days
and a larger standard deviation requires more days.
Value (206 sites)
25th Percentile
Median
75th Percentile
Standard
Deviation
1.9%
2.4%
3.1%
ฑ1% (25/50/1 00 days)
9/10/12
10/13/16
12/17/23
ฑ 2% (25/50/100 days)
3/4/4
5/5/6
6/8/9
Table 14.1- Number of days needed to replicate the 25/50/100 day dataset mean RRF to
within ฑ1% and ฑ2%, with a 95% confidence interval.
       Based on these results, we recommend a minimum number of 10 days to be included in
the mean RRF calculation for each monitoring site. This will ensure a relatively robust mean
RRF value that is within ฑ1% (on average) of the large dataset mean. If relatively few ozone
days are being modeled or certain monitors have relatively few exceedances (above 85 ppb),
then we recommend using an absolute minimum number of 5 days in the calculation.

       The minimum number of days recommendations can be combined with the minimum
threshold recommendation to create a hierarchy of number of days/threshold combinations that
                                         145

-------
can address any situation.  The recommended minimum concentration threshold identified in
section 4.1 is 85 ppb. But similar to the minimum number of days, there may be situations
where there are relatively few "high" modeled ozone days at certain monitors. Therefore, when
possible, we recommend using the 85 ppb threshold97, but it is acceptable to use a threshold as
low as 70 ppb.

       Therefore, the following criteria should be applied to determine the number of days and
the minimum threshold at each ozone monitor:

•      If there are 10 or more days with daily maximum 8-hour average baseline modeled ozone
       > 85 ppb then use an 85 ppb threshold.
•      If there are less than 10 days with daily maximum 8-hour average baseline modeled
       ozone > 85 ppb then reduce the threshold down to as low as 70 ppb until there are 10
       days in the mean RRF calculation.
•      If there are less than 10 days with daily maximum 8-hour average modeled ozone > 70
       ppb then use all days > 70 ppb.
•      Don't calculate an RRF for  sites with less than 5 days > 70 ppb98.

The following table illustrates several examples of the recommended hierarchy of choosing the
number of days vs. the minimum threshold.
Number of Days
> 70 ppb
50
20
12
11
9
6
3
Number of Days
> 85 ppb
15
12
7
3
6
1
0
Number of Days in
Mean RRF
15
12
10
10
9
6
N/A
Threshold < 85 ppb?
No
No
Yes
Yes
Yes
Yes
N/A
Table 14.2- Examples of the recommended hierarchy in choosing the number of days in the
mean RRF calculation vs. the minimum threshold.  The "number of days" refers to the
number of days (at each monitor) when the daily modeled 8-hour ozone maximum is > 70
or 85 ppb.
       97As noted in section 4.1, for areas with a very high design value (>110 ppb), a higher
threshold might be appropriate.

       98Any situation where there are less than 5 days available for RRF calculations at
monitoring sites with relatively high concentrations, (above the NAAQS and/or close to the area-
wide design value) should be discussed with the appropriate U.S. EPA Regional office(s).
                                          146

-------
       In summary, States should try to model enough episode days so that the mean RRF
calculation at each monitor contains a minimum of 10 days with a modeled concentration > 85
ppb. If there are less than 10 days > 85 ppb, then the threshold should be lowered until 10 days
are included in the calculation.  The threshold should not go below 70 ppb and the number of
days should always be at least 5. In trying to meet these recommendations, the greatest priority
should be given to identifying episode days with appropriate ozone concentrations at the
monitoring sites with the highest design values.  Sites with design values below the NAAQS
should be given a  low priority.

              14.1.2 Recommendations for the 24-Hour PM2 5 NAAQS

       The episode selection criteria for the 24-hour PM2 5 NAAQS are similar to the ozone
criteria. The complicating factor for the 24-hour NAAQS is that exceedances of the standard can
occur under a variety of meteorological conditions at different times of the year and can be due
to a variety of different PM components. Therefore the universe of episodes could be large and
much more varied. The form of the standard is a 3 year average of the 98th percentile
concentration from each year. Therefore, modeling 3  full years of meteorological data would
best capture the full range of conditions that lead to 24-hour design values.  But the standard is
only influenced by the very high end of the distribution of PM2 5 days (top 2% from each year).
This argues for modeling a smaller set of days on an episodic basis.  We recommend one of two
possible approaches:

1) Model every day for a full year (or multiple years).  This is recommended for both dispersion
modeling of primary PM25 components and photochemical modeling of secondary and primary
components.  Many areas that violate the 24-hour PM2 5 NAAQS will also violate the annual
PM2 5 NAAQS. Therefore, full  year modeling may already exist or is being planned for the
annual NAAQS attainment test. States that have nonattainment areas that violate both PM2 5
standards should take advantage of planning opportunities to address both standards at the same
time.  Modeling at least a full year will also help ensure that a sufficient number of days are
included  in the RRF calculations.

2) Model episodes when high PM2 5 concentrations occur. Similar to ozone, episodes should be
selected where PM2 5 concentrations are greater than the NAAQS (> 65 ug/m3) and are close to
the baseline design value. Similar to ozone, data analyses can be completed to help  select a
variety of meteorological episodes which lead to high PM25 concentrations.  In some cases, there
may be very limited conditions  which lead to high 24-hour average PM2 5 concentrations,  and in
other cases there may be a wide variety of cases. The specific situation in each nonattainment
area will  determine the number of episodes and the time periods which need to be modeled. For
example, if exceedance level PM2 5 concentrations in an area only  occur in the winter, then a
limited number of winter days can be modeled.  In other areas, exceedance days may occur in all
seasons.
                                          147

-------
       The 98th percentile PM2 5 values may be highly variable from year to year. Therefore, it
is important to model a year or time periods when high PM2 5 episodes occur in the area.  The
modeling should include a year and/or episodes that are representative of the 24-hour design
values.

       Calculation of RRFs- Regardless of the time periods modeled, component specific
RRFs should be calculated for all four quarters or for all quarters in which high 24-hour PM2 5
concentrations occur (following the recommendations in section 5.2).  We do not have specific
analyses that have analyzed an appropriate minimum threshold concentration for 24-hour
average PM25. If there are very few days that exceed the 65 ug/m3  standard, then a secondary
recommendation is to include all days with concentrations > 50 ug/m3, or calculate RRFs based
on the highest modeled PM2 5 days from each quarter.  A reasonable estimate of the high end of
the distribution might be between the top 10% of days to the top 25% of days. We recommend
that RRF calculations contain a similar number of days as for ozone. This means that each
quarterly RRF calculation should contain no less than 5 days".

       EPA will be re-examining these recommendations as they relate to the new 35 ug/m3 24-
hour NAAQS as more information becomes available regarding the typical features of areas
violating the revised standard. Additional guidance will be issued at a later date.

              14.1.3 Annual PM25

       Similar to episode selection for the daily standards, the most important consideration for
the annual PM2 5 NAAQS is to model a variety of meteorological conditions throughout the year.
The easiest way to accomplish this is to model a full year. Since all days count toward the
annual average, each season should be represented in the modeling. If less than a full year is
modeled then a minimum number of days should be modeled from each quarter.

       One study  (Environ, 2004) examined how well a subset of modeled days in each quarter
represented modeled quarterly and annual averages. The study calculated the uncertainty
associated with absolute model predictions and RRFs that were based on 7, 14, or 28 days from
each quarter (the periods were contiguous).  They found that the relative uncertainty of the
annual average concentrations was in the range of 40-70% for SO4, 30-50% for anthropogenic
organics, and 60-90% for nitrates (the low percentage is for 28 days and the high percentage is
for 7 days). The uncertainty was lowered by about half when considering the uncertainty in
RRFs for the 7-28 day periods vs.  the annual average. This  shows the advantage of using the
models in a relative sense. But it also shows that modeling less than a full year introduces
       "A careful analysis is needed for the days that exceed the 24-hour NAAQS. It may be
necessary to divide days by species as well as by quarter. For example, exceedence days with
high nitrates may respond differently than exceedence days with high organics (in the same
quarter).  Averaging the response from many dis-similar days may give an unrealistic mean
response.

                                          148

-------
additional uncertainty into the model results (compared to modeling a full year). The conclusion
from the study was that at least 14 days are needed to represent each quarter. This assumes that
the model results are being used in a relative sense. The study also found that nitrate is the most
difficult of the species to adequately represent with a reduced number of modeling days.

       Based on current information, for annual average PM2 5, we recommend modeling a full
year (or more).  Reduced modeling of 15 or more days per quarter100 may be acceptable in some
cases, but areas where nitrate is an important component of PM25 should consider modeling a
full year (or at least the time periods when nitrates are high). Modeling less than a full year (e.g.
one month per quarter)  can be useful for initial strategy development and sensitivity analyses. In
this way, information can be learned from partial year runs and can be supplemented with full
year model runs later in the process. This may save resources and allow more model runs to be
completed (relatively) quickly.
              14.1.4 Regional Haze

       The goals for regional haze focus on the 20% of days with best and worst visibility.
Sampling in Class I areas occurs once every three days. Thus, each year there will be about 24
"worst" days and 24 "best" days to choose from.  Since the base period against which reasonable
progress is to be gauged is 5 years long, there could be as many as 120 "best" and "worst" days
to choose among for modeling.

       It is likely that numerous Class I areas will be considered simultaneously in a modeled
uniform rate of progress assessment. Thus, the preferred approach for regional haze-related
model applications is to simulate an entire, representative year (i.e., one whose mean derived
deciview values for "20% worst" and "20% best" days approximates mean values for deciviews
averaged over  5 years for the best and worst days).  States can then base the RRF values on the
best and worst days in each Class I area for the modeled year (the -24 best and worst days from
the  modeling year). For areas included in modeling for the annual PM2 5 NAAQS, particularly if
that modeling simulates an entire representative year (or more), the modeling for PM2 5 should
also provide adequate results to analyze for regional haze.

       In some cases, the meteorology for a particular year may be representative of the five
year period, but due to large emissions changes during the period, a single year of ambient
visibility is not representative of the five year mean.  One solution is to compile the emissions
inventory using five year average emissions for the base period. Because the base year and
future year visibility represents a five year period, it is appropriate to average both base and
future year emissions in a similar manner.
       100The recommendation is for 15 or more total days per quarter, not 15 or more days with
monitored data.

                                          149

-------
       If it is not feasible to model an entire year, then RRFs are likely to be relatively stable if
at least 10 worst days (and 10 best days) are modeled (similar to the ozone recommendations). If
the worst days occur throughout the year, then more time periods may need to be modeled.
Worst days that occur in the same season are more likely to have similar PM components and are
therefore likely to respond in a similar way.
       14.2          What Additional, Secondary Criteria May Be Useful For Selecting
                    Time Periods to Model?

       In Section 14.0, we noted that there may often be conflicts among the four primary
criteria recommended as the basis for choosing episodes to model. Several additional, secondary
selection criteria may be helpful for resolving these conflicts.

       Choose time periods which have already been modeled. That is, of course, provided
that past model performance evaluation for such a time period was successful in showing that the
model worked well in replicating observations. Given that the four primary criteria are met
approximately as well by such time periods they are by other candidate time periods, a
State/Tribe could likely save a substantial amount of work in evaluating model performance.
However, large changes in emissions and precursor levels or ratios may make the use of older
time periods undesirable.

       Choose time periods which are drawn from the period upon which the baseline
design value is based.  As we note in Section 3,  fewer emission estimates and fewer air quality
model simulations may be needed if the base case period used to evaluate model performance,
and the baseline period used in the recommended  modeled attainment test are the same.
Following this criterion could also make the second primary criterion more straightforward.

       Choose episodes having observed concentrations "close to" the NAAQS on as many
days and at as many sites as possible. This criterion mainly applies to the standards with short
term averaging times (8-hour ozone and 24-hour PM2 5). It is related to the modeled attainment
test and to the fourth primary criterion for episode selection. The more days and sites for which
it is reasonable to apply the test, the greater the confidence in the outcome of the modeled
attainment test.

       It is desirable to include weekend days among those chosen, especially if
concentrations greater than the NAAQS are observed on weekends. Weekend days often
reflect a different mix of emissions than occurs on weekdays101.  This could also lead to different
spatial patterns of 8-hour ozone and/or 24-hour PM2 5 concentrations. Thus, for increased
confidence that a control strategy is effective it needs to be tested on weekends as well as on
weekdays. If emissions and spatial patterns of high ozone or PM25 do differ on weekends versus
       101 http://www.arb.ca.gov/aqd/weekendeffect/weekendeffect.htm

                                          150

-------
weekdays, including weekend days in the choice of episodes will provide a mechanism for
evaluating the accuracy of a model's response to changes in emissions.

       If it has been determined that there is a need to model several nonattainment areas
simultaneously, choose time periods which meet the primary and secondary criteria in as
many of these nonattainment areas as possible. As discussed in Section 15, a State/Tribe or
group of States/Tribes may decide to apply a model on a regional or a nested regional scale to
demonstrate attainment or assess uniform rate of progress in several nonattainment areas or
Class I areas at the same time. Time and resources needed for this effort could be reduced by
choosing time periods which meet the primary and secondary criteria in several nonattainment
areas (or Class I areas) which are modeled.  Several organizations are modeling much of the
Eastern U.S. for an entire year.  This type of application should allow for application of the
attainment tests at many locations using the appropriate set of days for each area.
                                          151

-------
15.0  What Should Be Considered When Selecting The Size And
       Horizontal/Vertical Resolution Of The Modeling Domain?

       A modeling domain identifies the geographical bounds of the area to be modeled. The
appropriate domain size depends on the nature of the strategies believed necessary to meet the
air quality goal. This, in turn, depends on the degree to which air quality observations suggest
that a significant part of an observed exceedance is attributable to regional concentrations which
approach or exceed levels specified in the NAAQS. The choice of domain size is also affected
by data base management considerations.  Generally, these are less demanding for smaller
domains.

       Horizontal resolution is the geographic size of individual grid cells within the modeling
domain. Vertical resolution is the number of grid cells (i.e., layers) considered in the vertical
direction.  The choice of suitable horizontal and vertical resolution depends on spatial variability
in emissions, spatial precision of available emissions data, temporal and spatial variation in
mixing heights, the likelihood that mesoscale or smaller scale meteorological phenomena will
have a pronounced effect on precursor/pollutant relationships, data base management constraints,
and any computer/cost constraints.

       We begin this section by discussing factors States/Tribes should consider in choosing
domain size. Next, we address the selection of horizontal grid cell size and the number of
vertical layers. We conclude by discussing factors affecting the decision on the size and
resolution of coarse scale and fine scale grids within a nested model.

       15.1          How is the Size of the Modeling Domain Chosen?

       Historically (until -1995), ozone attainment demonstrations used urban scale modeling
domains which were typically several hundred kilometers (or less) on a side. With the advent of
nested grid models, most model applications began to use either relatively fine regional grids, or
urban-scale inner grids nested within relatively coarse regional-scale outer grids.  Recent PM
modeling has generally used nationwide coarse grid domains with regional scale nested  domains.
We expect that most urban scale ozone and PM2 5 attainment demonstrations will utilize a
national or regional nested grid modeling approach.

       The principal determinants of model domain size are the nature of the ozone and/or PM25
problem and the scale of the emissions which impact the nonattainment area. Isolated
nonattainment areas that are not impacted by regional transport of ozone and/or PM and its
precursors may be able to use a relatively  small domain102. Some areas of the western U.S. may
fall into this category.  Most nonattainment areas in the eastern U.S. have been shown to be
impacted by transported ozone and/or PM and ozone and/or PM precursors from hundreds of
       102Due to relatively long lifetimes of PM25 aerosols (compared to ozone), PM modeling
domains may generally need to be larger than ozone domains.

                                          152

-------
miles or more upwind of the receptor area (U.S EPA, 1998b) (U.S. EPA, 2005b).  The modeling
domain should be designed so that all major upwind source areas that influence the downwind
nonattainment area are included in the modeling domain. The influence of boundary conditions
should be minimized to the extent possible.

       The inner domain of a nested model application should include the nonattainment area
and surrounding counties and/or States. The size of the inner domain depends on  several factors.
Among them are:

1) The size of the nonattainment area.

2) Proximity to other large source areas and/or nonattainment areas.
       -Relatively isolated areas may be able to use a smaller fine grid domain.
       -Nearby source areas should be included in the fine grid.

3) Proximity of topographical features which appear to affect observed air quality.

4) Whether the model application is intended to cover multiple nonattainment areas.

5) Typical wind speeds and re-circulation patterns.
       -Very light wind speeds and re-circulation patterns may obviate the need for a large fine
       grid domain.

6) Whether the photochemical model utilizes one-way or two-way nested grids.
       -The fine grid domain of a model with one-way nested grids may need to be larger
       (compared to a model with two-way nested grids).  The grid needs to be large enough to
       capture re-circulation due to shifting wind directions. A two-way nested grid model
       allows for continuous feedback from the fine grid to the coarse grid and vice versa.

7) Computer and time resource  issues.
       15.2   How are the Initial and Boundary Conditions Specified?

       Air quality models require specification of initial conditions for model species in each
grid cell in the model domain (in all layers) and boundary conditions for all grid cells along each
of the boundaries (in all layers).  Generation of initial and boundary conditions for individual
model species include gas-phase mechanism species, aerosols, non-reactive species and tracer
species. There is no satisfactory way to specify initial conditions in every grid cell.  Thus, we
recommend using a "ramp-up" period by beginning a simulation prior to a period of interest to
                                          153

-------
diminish the importance of initial conditions103. We recommend a ramp-up period of at least 2-3
days for ozone modeling. Due to longer lifetimes of fine particulate matter, we recommend a
ramp-up period of at least 5-10 days for PM25. In this way, the choice of initial conditions are
not likely to influence the modeling results.  For nested model applications, initial conditions can
be specified using model predictions from the outer grid if the nested grids are started after the
beginning of the simulation for the outer grid.

       Boundary conditions can be specified in several ways. One option is to nest the area of
interest within a much larger domain using nested regional models, as described previously. As
noted above in Section 15.1, use of a large regional domain acts to diminish the importance of
boundary conditions. Alternatively, initial and boundary conditions can be derived from another
regional nested modeling application or from a global model104.  Another option is to use default
initial and boundary concentration profiles representing relatively clean conditions which have
been formulated from available measurements or results obtained from prior modeling studies,
(e.g. prescribing 35 ppb for annual ozone). Another approach to consider is using the model's
simulated pollutant values  (generated for emissions) averaged over one or more of the upper
layers to specify a value for the lateral and top boundary conditions. This approach allows for
better representation of spatial and temporal variation of boundary conditions (i.e.,
diurnal/seasonal/vertical profiles). This method also avoids arbitrarily guessing at the future-
year boundary conditions.

       If there is no larger regional model application available, then it is recommended that
background boundary conditions be used to  specify initial and boundary concentrations for the
attainment demonstration modeling.  However, concentration fields derived from a larger
domain regional or global chemistry model (i.e. nesting approach) is considered more credible
than the assumption of static concentrations, since the pollutant concentration fields reflect
simulated atmospheric chemical and physical processes driven by assimilated meteorological
observations. Therefore, we recommend using boundary conditions derived from a regional or
global  scale model, whenever possible.

       Diagnostic testing which indicates a large impact on the model results from initial or
boundary conditions may indicate that the domain is not large enough or the ramp-up period is
too short.  In either case, it should generally  be assumed that initial and boundary conditions do
       103Sensitivity simulations can be completed to determine the necessary length of the
ramp-up period.  A longer ramp-up period may be needed for very large domains where the
characterization of long range transport is important.

       104One atmosphere model applications for ozone and PM may commonly use global
models to specify boundary conditions. This is especially important for PM species due to their
long lifetimes. A number of recent studies show that long-range, intercontinental transport of
pollutants is important for simulating seasonal/annual ozone and parti culate matter (Jacob, 1999;
Jaffe, 2003; Fiore, 2003).

                                           154

-------
not change in the future. The use of altered initial or boundary conditions in the future year
should be documented and justified.
       15.3         What Horizontal Grid Cell Size Is Necessary?

       As we discuss in Section 16, most applications will use a prognostic meteorological
model to provide meteorological inputs needed to make air quality estimates.  Typically, these
models are set up to produce meteorological fields for nested grids with a 3:1 ratio. In past
modeling applications, the most commonly used grid cell sizes have been 108, 36, 12 and 4 km
cells.  In this section we provide recommendations for choosing the grid size to use as an upper
limit for regional and urban scale models or for fine portions of nested regional grids.

       This section only applies to the use of numerical grid models. If a Lagrangian dispersion
model is used, estimates at (or near) the monitor locations should be used to calculate RRF
values for primary PM.

       In past guidance, we have recommended using horizontal grid cell sizes of 2-5 km in
urban scale ozone modeling analyses (U.S.  EPA, 1991a).  Sensitivity tests performed by Kumar
(1994) in the South Coast Air Basin compared hourly base case predictions obtained with 5 km
versus 10 km versus 20 km grid cells. Results indicate that use of finer grid cells tends to
accentuate higher hourly ozone predictions and increase localized effects of NOx titration during
any given hour. However, statistical comparisons with observed hourly ozone data in this
heavily monitored area appear comparable with the 5 and 20 km grid cells in this study.
Comparisons between hourly ozone predictions obtained with 4 km vs. 12 km grid cells have
also been made in an Atlanta study (Haney, 1996).  As in Los Angeles, the use of smaller (i.e., 4
km) grid cells again leads to higher domain wide maximum hourly ozone concentrations.
However, when reviewing concentrations at specific sites, Haney (1996) found that for some
hours concentrations obtained with the 12 km grid cells were higher than those obtained with the
4 km cells. Other studies have shown that model performance does not necessarily improve with
finer resolution modeling.  In one study, the model performance at 12 km resolution was equal
to or better than performance at 4 km resolution (Gego,  2005).

       Another important aspect in choosing the horizontal grid cell size is the relative response
of the model at various grid cell resolutions. Recent sensitivity tests comparing relative response
factors in predicted 8-hour daily maxima ozone near sites in the eastern United States indicate
relatively small unbiased differences (< .04, in 95% of the comparisons) using a grid with  12 km
vs. 4 km grid cells (LADCO, 1999; Arunachalam, 2006).  The largest differences in the  relative
response of models at varying resolution is  likely to occur in oxidant limited areas. The
horizontal resolution may have a large impact on the spatial distribution and magnitude  of NOx
"disbenefits" (i.e., ozone increases in oxidant  limited areas when NOx emissions are reduced).
                                          155

-------
       A recent study of the Atlanta area (Cohan, 2006) used the Decoupled Direct Method
(DDM) to examine the sensitivity of ozone production efficiency to horizontal resolution.  The
efficiency of ozone production should be closely related to model response to emisions controls.
The analysis found that 12km resolution is sufficient (in Atlanta) to predict ozone production
efficiency.  However, 4km resolution is needed to identify small areas of local NOx inhibition.
They also found that model performance for ozone at 4km and 12km resolution was similar.

       Horizontal resolution will also have an impact on primary particulate matter.  Because
there is no chemistry, we expect spatial gradients in concentrations to be much higher for
primary PM components compared to secondary components.  The model's ability to capture
these gradients will depend in large part upon the horizontal resolution of the model.  Therefore,
emissions density is an important factor in considering grid resolution.  Additionally, terrain
features (mountains, water) should be considered when choosing grid cell size.

       Intuitively, one would expect to get more accurate results in urban applications with
smaller grid cells (e.g., 4 km) provided the spatial details in the emissions and  meteorological
inputs support making such predictions. Thus, using 4 km grid cells for urban or fine portions of
nested regional grids and 12 km cells in coarse portions of regional grids are desirable goals.
However, model performance at 4 km resolution may not be any better than at 12 km resolution.
Additionally,  extensive  use of urban grids with 4 vs. 12 km grid cells and regional grids with  12
vs.  36 km grid cells greatly increases computer costs, run times and data base management
needs.  Further, elsewhere in this guidance we identify needs to model large domains, many
days, and several emission control scenarios. We also identify a number of diagnostic tests
which would be desirable and suggest using more vertical layers than has commonly  been  done
in the past. Also, there may be ways of dealing with potential problems posed by using larger
than desired grid cells. For example, use of plume-in-grid algorithms for large point  sources of
NOx should be considered as  an alternative with coarser than desired grid cells. And dispersion
modeling of primary PM can be used to offset the need  of fine  spatial resolution in a  grid model.

       The relative importance of using a domain with grid cells as small as 4 km should be
weighed on a case by case basis by those implementing the modeling/analysis  protocol.  The
most important factor to consider is model response to emissions controls.  Analysis of ambient
data, sensitivity modeling, and past modeling results can be used to evaluate the expected
response to emissions controls at various horizontal resolutions for both ozone and PM2 5. If
model response is expected to be different (and more accurate) at higher resolution, then higher
resolution modeling should be considered. If model response is expected to be similar at both
high and lowe(er) resolution, then high resolution modeling may not be necessary.

       In this guidance,  we identify upper limits for horizontal grid cell size which may be
larger than desired for some applications.  This is intended to provide flexibility for considering
competing factors (e.g., number of modeled days versus grid cell size) in performing  a modeling
analysis within the limits of time and resources.
                                           156

-------
              15.3.1 Horizontal Resolution Recommendations for Ozone

       For coarse portions of regional grids, we recommend a grid cell size of 12 km if feasible,
but not larger than 36 km.  For urban and fine scale portions of nested regional grids, it may be
desirable to use grid cells about 4 km, but not larger than 12 km.  States/Tribes should examine
past model applications and data analyses for their area when choosing the fine grid resolution.
Past model applications and data analyses may help determine whether a grid cell size as small
as 4 km (or smaller) is necessary for a particular area.  Model performance and the relative
response to emissions controls should be considered in the decision.  States/Tribes should
consider diagnostic tests to assess the difference in model performance and response from
varying model grid resolution, particularly in oxidant-limited areas.

       All ozone monitor locations within a nonattainment area should ordinarily be placed
within the fine scale portion of a nested regional grid if nested models are used. States/Tribes
choosing an urban grid or fine portion of a nested grid with cells 12 km or larger should consider
applying plume-in-grid algorithms to major point sources of NOx. The use of plume-in-grid
should be discussed with the appropriate EPA Regional Office.

              15.3.2 Horizontal Resolution Recommendations for the PM2 5 NAAQS

       The formation of secondary particulate matter shares the same basic photochemical
processes as ozone.  In fact, they are closely interconnected in many respects. Since there
appears to be little difference in RRF estimates for 8-hour ozone with 12 km vs. 4 km grid cells,
we would expect this finding to hold for secondary components of particulate matter. But what
about primary components of particulate matter? Because there is no chemistry, we expect
spatial gradients in concentrations to be much higher for primary particulate  matter than for
secondary particulate matter. This follows, because it takes  some time (less than 24 hours) for
the chemistry to occur.  Although it is clear that spatial resolution of primary PM will impact the
predicted concentrations, it is not clear how it will impact the relative change in concentrations
due to controls (RRFs).  Areas without strong gradients in primary PM will likely have little
benefit from fine scale resolution.  Areas that have large gradients in primary PM may need to
use finer resolution grid models or may need to supplement the grid modeling with dispersion
modeling (see Section 5.3) .  This particularly true if violating monitors  are strongly impacted by
one or more large sources of primary PM.

       The bottom line of the preceding discussion is that we feel comfortable recommending
that States may use grid cell sizes as large as 12 km for urban scale applications addressing
secondary components of particulate matter. We are less sure about an acceptable upper limit
for cell size in applications addressing primary components. We believe it is prudent to assume
that, in some cases, cells as small as 4 km (or possibly smaller) are needed.  Those implementing
the modeling/analysis protocol may wish to perform a diagnostic test using a grid model without
chemistry to see whether estimated RRF's for primary components are affected if one decreases
the grid cell size from  12 km to 4km.  Alternatively, large sources of primary PM can be

                                           157

-------
modeled with a dispersion model or a combination of grid and dispersion models (see Section
5.3).

       We expect that modeling analyses for nonattainment areas will use grid cell sizes of 12
km or less. If a regional scale model is applied, most of the domain will likely cover
rural/remote areas or locations which are not out of compliance with the NAAQS.  For the
regional outer nest of the domain,  grid cells as large as 36 km may be used.

              15.3.3 Horizontal Resolution Recommendations for Regional Haze

       The glidepath analysis for regional haze focuses on Class I areas. Most of these are in
remote or rural locations. Regional haze is not likely to be dominated by local sources.  Further,
light extinction is more likely to be dominated by secondary particulate matter, due to the
efficiency of light scattering by secondary particles.  All these attributes indicate that it will be
necessary to model a regional scale domain for regional haze related applications.  Because of
the remoteness of Class I areas, grid cell  sizes up to 36 km on a side should suffice for regional
haze-related modeling.  States may wish to perform diagnostic tests using plume-in- grid
analyses, as well as finer horizontal resolution to determine if results may differ using more
finely resolved emissions and meteorology.  Guidance on vertical resolution presented in Section
15..4 is also applicable for regional haze-related applications.

       15.4         How Should the Vertical Layers Be Selected?

       As described in Section 16, the preferred approach for generating meteorological data
fields for input to air quality simulation models is to use a prognostic meteorological model with
four dimensional data assimilation (FDDA).  Such models normally use more than 30 vertical
layers. To minimize a number of assumptions needed to interface meteorological and air quality
models, it is better to use identical vertical resolution in the  air quality and meteorological
models. However, application of air quality models with as many as 30 vertical layers may not
be feasible or cost effective. In this section we identify factors to consider in choosing the
number of vertical layers chosen for the air quality model applications.

       In the past,  short ozone episodes of only several days usually encompassed periods of
mostly clear skies with very little precipitation. As such, ozone models often did not explicitly
model clouds or precipitation. However, we are recommending modeling longer episodes (or
even a full ozone season) for ozone and a full year for PM2 5 and regional haze applications.
Therefore the photochemical model needs to account for cloud processes and a full range of
precipitation types. In order to adequately parameterize these processes, the top of the modeling
domain should typically be set at thelOO  millibar level (-16,000 meters). In turn, this means that
many more vertical layers will be needed to capture the meteorological processes both below and
above the boundary layer, up to the top of the model.
                                           158

-------
       The accuracy of predicted base case ozone and PM concentrations will be affected by
how well the model is able to characterize dilution of ozone, PM, and precursors. This, in turn,
depends in part on how precisely the model can estimate mixing heights (i.e., the PEL). The
precision of mixing height estimates is affected by the thickness of the model's vertical layers
aloft which are near the anticipated mixing height (Dolwick, 1999). Ozone concentrations are
most heavily influenced by the rate of rise in the morning mixing height and the maximum
afternoon mixing height.  Since PM is measured as a 24-hour average concentration, PM is also
influenced by the strength of nighttime mixing and the presence of a low level inversion.
Because mixing heights may vary on different days and it is necessary to simulate numerous
days and locations, model predictions can be influenced by the number of vertical layers
considered by the model.

       Placement of vertical layers within the planetary boundary layer is also an important
issue.  For practical reasons, it is best to have an air quality model's vertical layers align with the
interface between layers in the meteorological model.  In view of the importance of carefully
specifying the temporal variation in mixing height, we recommend high precision below and
near the anticipated maximum afternoon mixing height.  In addition, specifying the vertical
extent of mixing overnight during stable conditions is also an important consideration in
determining the vertical layer structure.  In this regard, we recommend that the lowest layer  in
the air quality model be no more than 50 meters thick. In general, layers below the daytime
mixing height should not  be too  thick, or large unrealistic step increases in mixing may occur.
Layers above the boundary layer are important for characterizing clouds and precipitation, but
are less important to the daily mixing processes of pollutants.  Therefore, vertical resolution
above the boundary layer is typically much coarser.

       There is no correct minimum number of vertical layers needed in an attainment
demonstration.  The vertical resolution will vary depending on the application. Recent
applications of one atmosphere models (with model tops at 100mb) have used anywhere from 12
to 21 vertical layers with  8-15 layers approximately within the boundary layer (below 2500m)
and 4-6 layers  above the PEL  (Baker, 2004b), (Hu, 2004).

       There are also ozone model applications which may not need to consider the full set  of
meteorological data through the  tropopause. These applications typically use vertical domains
which extend up to 4 or 5 km. These applications are most appropriate for short ozone episodes
that occur under high pressure conditions (little cloud cover or precipitation). In these cases,
fewer vertical layers are needed  to represent the atmosphere up to the top of the domain (4-5
km). However, where appropriate, EPA encourages the use of full-scale one-atmosphere models
which account for all atmospheric processes up to -100 mb.
                                           159

-------
16.0   How are the Meteorological Inputs Prepared for Air Quality Modeling?

       In order to solve for the change in pollutant concentrations over time and space, air
quality models require certain meteorological inputs that help simulate the formation, transport,
and removal of pollutant material.  The required meteorological inputs can vary by air quality
model, but consistently involve parameters such as wind, vertical mixing, temperature, moisture,
and solar radiation. While model inputs can be derived strictly from ambient measurements, a
more credible technical approach is to use meteorological grid models to provide the necessary
inputs. When these models are applied retrospectively (i.e., for historical time periods) they are
able to blend ambient data with model predictions via four-dimensional data assimilation
(FDDA), thereby yielding temporally and spatially complete data sets that are grounded by
actual observations.

       This section provides recommendations for generating, or otherwise acquiring, the
meteorological data sets sufficient for air quality modeling purposes.  Additional suggestions are
provided to assist in the configuration of standard meteorological modeling analyses. The last
section outlines procedures for evaluating whether the meteorological input is of sufficient
quality for input into the air quality model.  In general, it is recommended that States/Tribes
spend appropriate effort in accurately characterizing the meteorological fields in view of several
sensitivity runs which show that relatively small perturbations in meteorological inputs can  have
large impacts on resultant air quality modeling results (Dolwick, 2002).
       16.1   What Issues are Involved in the Generation and Acquisition of
             Meteorological Modeling Data?

       The recommended approach for generating the meteorological data needed to conduct the
attainment demonstration is to apply dynamic meteorological models with FDDA.  These models
use the fundamental equations of momentum, thermodynamics, and moisture to determine the
evolution of specific meteorological variables from a given initial state. When modeling past
events, the use of data assimilation helps to "nudge" solutions so that they do not diverge greatly
from the actual observed meteorological fields.  A major benefit of using dynamic
meteorological models is that they provide a way of consistently characterizing meteorological
conditions at times and locations where observations do not exist. Examples of frequently used
meteorological models are listed below:

•      The Penn State University / National Center for Atmospheric Research mesoscale model
       known as MM5 (Grell, 1994),
       The Regional Atmospheric Modeling System (RAMS) (Pielke, 1992), and
•      The Weather Research and Forecasting Model (WRF) (Skamarock, 2005).

       Recent advances in relatively low-cost computational power have resulted in widespread
use of MM5 and similar models for air pollution applications over the past decade (Olerud,
2000; Doty,  2001; Johnson, 2003; Baker, 2004b). EPA expects that the large majority of future

                                          160

-------
attainment demonstration analyses will be based on meteorological data from these types of
dynamic meteorological models.  However, there are other acceptable means of generating
meteorological data for an air quality modeling simulation. Over the next several years, EPA
further expects that more meteorological input data sets will be developed from archived
National Weather Service (NWS) model simulations. It is possible that data from archived
model simulations (such as WRF) could be used to feed air quality simulations. Some of these
prognostic meteorological models are already being used to drive real-time air quality forecasts
(McQueen, 2004). Additionally, in some cases the dynamic meteorological model may not
adequately capture key meteorological elements of an airshed's conceptual model (e.g.,
source-receptor transport vectors to key monitoring locations).  In cases such as these,  it may be
appropriate to blend the dynamic model data with wind data from an objective analysis of
observed wind fields.

       A description of the methods used to generate the meteorological fields should  be
included in the modeling protocol. In cases in which standard meteorological modeling (e.g.,
MM5, RAMS, or WRF in a retrospective analysis mode) is not used, it is recommended that a
detailed description of the technique that is used to generate the three-dimensional
meteorological fields be shared with the appropriate EPA regional office(s) prior to conducting
the air quality modeling analysis.
       16.2 How Should the Meteorological Modeling Analysis be Configured?

       As with other parts of the air quality modeling system, choices made regarding how to
configure the meteorological modeling can affect the quality and suitability of the air quality
model predictions. Decisions regarding the configuration of complex dynamic meteorological
models can be particularly challenging because of the amount of flexibility available to the user.
The following are recommendations on how to establish the configuration of a meteorological
model for air quality analyses. The focal point in establishing the proper meteorological
modeling configuration should be to get the best possible meteorological model performance.

       Selecting a Model Domain: As noted in Section 15, it is expected that most attainment
demonstrations will cover large areas and use nested grids. The outermost grid should capture
all upwind areas that can reasonably be expected to influence local concentrations of ozone,
and/or PM2 5. In terms of selecting an appropriate meteorological modeling domain, one should
extend the grid 3 to 6 cells beyond the domains of each air quality modeling grid to avoid
boundary effects.  For example, if 4 km grid cells are to be used in the fine portion of a nested
regional air quality model, then the meteorological fields at this detail would need to extend
12-24 km beyond the bounds of the 4 km grid used for air quality predictions. In terms of grid
resolution, EPA recommends that the dynamic meteorological models use the same grid
resolution as desired for the air quality model applications. In some cases, however, this may
not always be feasible.  One possible reason for modeling with meteorology using a different
grid resolution is in the case of unacceptable model performance from the  meteorological model
at the desired grid resolution.  In other instances, the need for finer resolution may be

                                           161

-------
emissions-driven more than meteorologically-driven and the costs do not warrant the generation
of additional resolution in the meteorological data. In these specific situations it is
recommended that the air quality model application use available results from meteorological
models on the next coarser scale (i.e., 36 km for a desired 12 km estimate,  12 km for a desired 4
km estimate). The coarse grid meteorological fields can be mapped to the  more finely resolved
air quality modeling domain.

       Selecting Physics Options: Most meteorological models have a suite of "physics
options" that allow users to select how a given feature will be simulated. For example, there may
be several options for specifying the planetary boundary layer scheme or the cumulus
parameterization.  In many situations, the "optimal" configuration cannot be determined without
performing an initial series of sensitivity tests which consider various combinations of physics
options over specific time periods and regions. While these tests may not ultimately conclude
that any one configuration is clearly superior at all times and in all areas, it is recommended that
these sensitivity tests be completed, as they should lead to a modeling analysis that is best-suited
for the domain and period being simulated.  Examples of sensitivity analyses can be found in
McNally (2002), Olerud, (2003), and Huang, (2005). Typically, the model configuration which
yields predictions that provide the best statistical match with observed data over the most cases
(episodes, regions, etc.) is the one that should be chosen, although other more qualitative
information can also be considered. Additionally, model configurations should be designed to
account for the pollutants and time periods that are of most interest. As an example, a
wintertime PM simulation in the Midwest (with high measured nitrate concentrations) may need
a meteorological model configuration that employs a land-surface model that properly handles
snow cover fields and their effects on boundary layer humidities and temperatures.

       Use of Data Assimilation:  As noted above, the use of FDD A helps to keep the model
predictions from widely diverging from  what was actually observed to occur at a particular point
in time/space. However, if used improperly, FDD A can significantly degrade overall model
performance and introduce computational artifacts (Tesche and McNally, 2001). Inappropriately
strong nudging coefficients can distort the magnitude of the physical terms in the underlying
atmospheric thermodynamic equations and result in "patchwork" meteorological fields with
strong gradients between near-site grid cells and  the remainder  of the grid. Additionally, if
specific meteorological features are expected to be important for predicting the location and
amount of pollution formed, based on an area's conceptual model, then the meteorological
modeling should be set up to ensure that FDDA does not prevent the model from forming these
features (e.g. nocturnal low-level wind jets). In general, analysis nudging  strengths should be no
greater than 1.0 x 10"4 for winds and temperatures and 1.0 x 10"5 for humidity. In the case of
observation nudging (i.e., FDDA based  on individual observations as opposed to analysis fields),
it is recommended that the resultant meteorological fields be examined to ensure that the results
over the entire domain are still consistent. Further, based on past experience, we recommend
against using FDDA below the boundary layer for thermodynamic variables like temperature and
humidity because of the potential for spurious convection. If the dynamic  model is applied
without FDDA, it is suggested that the simulation durations be  shorter than 24 hours.
                                           162

-------
       Conversion of Meteorological Outputs to Air Quality Model Inputs:  Even before
determining how the meteorological model is configured, careful thought should be given to the
compatibility between candidate meteorological models and the air quality model(s) chosen for
use.  A variety of post-processors exist to convert the outputs from the meteorological models
into the input formats of the air quality models.  Some examples include: the Meteorology
Chemistry Interface Processor (MCIP) (Otte, 2004), MMSCAMx (Environ, 2006a), and
RAMSCAMx (Environ, 2006a).  These meteorological preprocessors provide a complete set of
meteorological data needed for the air quality simulation by accounting for issues related to:  1)
data format translation, 2) conversion of parameter units, 3) extraction of data for appropriate
window domains, 4) reconstruction of the meteorological data on different grid and layer
structures, and 5) calculation of additional variables.
       16.3 How Should the Performance of the Meteorological Modeling Be Evaluated?

       While the air quality models used in attainment demonstrations have consistently been
subjected to a rigorous performance assessment, in many cases the meteorological inputs to these
models are accepted as is, even though this component of the modeling is arguably more
complex and contains a higher quantity of potential errors that could affect the results of the
analysis (Tesche, 2002).  EPA recommends that States/Tribes devote appropriate effort to the
process of evaluating the meteorological inputs to the air quality model as we believe good
meteorological model performance will yield more confidence in predictions from the air quality
model. One of the objectives of this evaluation should be to determine if the meteorological
model output fields represent a reasonable approximation of the actual meteorology that
occurred during the modeling period. Further, because it will never be possible to exactly
simulate  the actual meteorological fields at all points in space/time, a second objective of the
evaluation should be to identify and quantify the existing biases and errors in the meteorological
predictions in order to allow for an downstream assessment of how the air quality modeling
results are affected by issues associated with the meteorological data.  To address both
objectives,  it will be necessary to complete both an operational evaluation (i.e., quantitative,
statistical, and graphical comparisons) as well as a more phenomenological assessment (i.e.,
generally qualitative comparisons of observed features vs. their depiction in the model data).

       Operational Evaluation:  The operational evaluation results should focus on the values
and distributions of specific meteorological parameters as paired with and compared to observed
data.  It is recommended that the observation - model matching be paired as closely as possible
in space and time. Typical statistical comparisons of the key meteorological parameters105 will
include: comparisons of the means, mean bias, mean normalized bias, mean absolute error, mean
absolute  normalized error, root mean square error  (systematic and unsystematic), and an index of
       105 It is difficult to say which meteorological parameters will most affect any particular modeling exercise
and it will vary by parameter and location. However, in general, it is expected that following key variables should
be most closely evaluated: temperature, water vapor mixing ratio, wind speed, wind direction, clouds/radiation,
precipitation, and the depth and evolution of vertical mixing over the modeling periods.

                                           163

-------
agreement. For modeling exercises over large domains and entire ozone seasons or years, it is
recommended that the operational evaluation be broken down into individual segments such as
geographic subregions and/or months/seasons to allow for a more comprehensive assessment of
the meteorological strengths and weaknesses.  Other useful ways of examining model
performance include: aloft, surface, individual episodes (e.g., high ozone /PM25 days), diurnal
cycle, as a function of synoptic regimes, or combinations of the above. It is recommended that
the ambient data used in these statistical comparisons be quality checked by doing standard
range check and buddy analyses. To the extent that modelers can set aside a portion of the
ambient data strictly for evaluation purposes (i.e., data not used in the FDD A), that is also
encouraged. Figure 16.1 shows a sample operational evaluation summary. This graphic was
developed using the Atmospheric Model Evaluation Tool (AMET) (Gilliam, 2005), though other
evaluation tools are available.
                                          164

-------
                             Model  Performance Report
    Model :
    Model Rvin : 30o2_i3km— EU
    Variable  :   T
                              Station :
                               uer^ ID asosass
Date Range :  300EQ401 to aooeoeai
Time of Day :
Forecast Hoii
               Observation—Model Scatter
     ?70
                 280
                         Observed
                              290
                                           300
     Statistical Metrics versus Observation Range
                     Obi served Range
                                                StDev
                                                MAE
                                                BIAS
                       285
                              29O
                                                            Model Performance Statistics
                                                          Data count
                                                          Correlation
                                                          Sta.nda.rd Deviation
                                                          Mean Absolute Error
                                                          Mean Bias
                                                          *Mean Fra.ctiona.1 Bias (
                                                          'Mean Normalised Bias
                                                          'Mean Normalized Error
                                                          ^Normalized Mean Bias
                                                          ^Normalized Mean Error
                                                          Root—Wean—Sqr—Error
                                                          Index of Agreement
                           2152
                           O.B6G11
                           2.1DD1
                           1.7745
                           —11 731
                           -0.40641
                           —3.0059
                           4.E752
                           —3.039
                           4.E9B7
                           2.4 052
                           0.95568
     Stats are normalized by observation range
                                                                Histogram of Model Error
                                                                   Obs
                                                                                 Mod
Figure 16.1 Sample operational evaluation summary for a 2002 MM5 case. This particular
analysis looked at April through June temperature performance at a single site.  The upper
left quadrant shows a scatter plot of observed versus predicted temperatures, the lower left
traces show how model performance varies as a function of the observation range, the upper
right table lists some of the statistical outputs, and the lower right compares the spectrum of
predictions against the observational spectrum.
                                             165

-------
       It may be helpful when calculating domainwide and/or regional summary statistics to
compare the results against previously generated meteorological model performance
"benchmarks" (Emery et al., 2001).  However, because of concerns about potentially misleading
comparisons of model performance findings across different analyses with differing model
configurations and FDDA strengths, EPA does not recommend using these benchmarks in a
"pass/fail" mode, but only as a means of assessing general confidence in the meteorological
model data.  Statistical results that are outside the range of the compiled benchmarks may
indicate an issue that should be given further examination. Figure 16.2 shows a sample analysis
in this vein.
2002 MM5 performance for temperature at three sites, by season
^ nn
o.uu
• /
r^ f\f\
A
i_
O
i_
LU

— A.LHJ
A
A
•
1 nn
I .UU
r\ r\r\

•
•



l \\j.\j\j i l
-2.50 -1.50 -0.50 0.50 1.50 2.50
Bias

• DET 1st Q
DET2ndQ
• DETSrdQ
• DET4thQ
ABhMlstQ
BHVl2ndQ
ABhMSrdQ
ABHVI4thQ
• SEA 1st Q
SB\2ndQ
• SEA 3rd Q
• SEMthQ

 Figure 16.2: Sample "soccer plot" summary of temperatures from a 2002 MM5 case.  This
 particular analysis looked at seasonal temperature performance at three sites.  The benchmarks
 in this case were set at +/- 1.0 K for bias and 2 K for error. This analysis called attention to a
 relatively large cold bias in the winter.
                                          166

-------
       In most cases the performance evaluation will be completed on the raw meteorological
fields.  However it is also important to compare the results before and after the meteorological
post-processing to ensure that the meteorological fields going into the air quality model have not
been adversely affected.

Phenomenological Evaluation:  As discussed in Chapter 11, it is recommended that a
conceptual description of the area's air quality problem be developed prior to the initiation of any
air modeling study. Within the conceptual description of a particular modeling exercise, it is
recommended that the specific meteorological parameters106 that influence air quality be
identified and  qualitatively ranked in importance. When evaluating meteorological models or
any other source of meteorological  data, the focus of the phenomenological evaluation should be
on those specific meteorological phenomena that are thought to strongly affect air pollution
formation and transport within the scope of a specific analysis. It is expected that this
event-oriented evaluation will need to summarize model performance in terms of statistical
metrics such as probability of detection and false alarm rate. As an example of a potential
phenomenological analysis, many regional air quality modeling exercises attempt to assess the
effects of transport of emissions from one area to a downwind area with an intent to establish
source-receptor relationships. For these types of modeling analyses, accurate transport wind
trajectories are needed to properly establish these source-receptor linkages. In this type of model
application, a useful event-based meteorological evaluation would be to compare model-derived
trajectories versus those based on ambient data to determine what error distance can be
associated with the model fields.
       106 Possible examples include: lake/sea breezes, low-level jets, or amount of convection
on a given day.

                                           167

-------
17.0   How Are the Emission Inputs Developed?

       Air quality modeling for 8-hour ozone, PM2 5, and regional haze requires emission inputs
for base case, baseline, and future modeling years. As explained in the EPA Emission Inventory
Guidance (U.S. EPA, 2005e), 2002 is designated as a new base year for 8-hour ozone and PM25
SIPs and regional haze plans; therefore, wherever possible, 2002 should be used for baseline
modeling36.  The use of an alternative baseline year should be discussed with the appropriate
EPA regional office. The future-year depends on the nonattainment classification of the
individual State or Tribe, as described in Section 3.6. Note that emissions data should be
consistent with the data used in the modeled attainment test, described in  Sections 3-6.

       Preparation of emissions data for air quality models for the base and future years requires
several steps. First, States and Tribes need to compile base-year inventories for their modeling
region (e.g., the States and Tribes in the modeling grid). For ozone model applications, emission
inventories should include a complete accounting of anthropogenic and biogenic VOC
(speciated), NOx, and CO. For PM and regional haze model applications, emission inventories
should also include emissions of SO2, NH3, PM2 5 (speciated), and PM coarse (PMC). Second,
modelers must collect "ancillary data" associated with the inventories, which prescribes the
spatial, temporal, and chemical speciation information about the emission inventory. Third,
modelers use the ancillary data for "emissions modeling". Emissions models spatially,
temporally, chemically, and vertically allocate emission inventories to the resolution needed by
AQMs.  Fourth, modelers must collect data on growth rates and existing control programs for
use in projecting the base year emission inventories to the future year, and then use  an emissions
model to prepare that future year inventory data for input to the air quality model. Fifth,
emissions inventories that reflect the emissions reductions needed for attainment will have to be
prepared for air quality modeling.

       Sections 17.1 and 17.2 summarize the issues for preparing emission inventories. Section
17.3 describes the needs for ancillary data. Section 17.4 summarizes the emissions  modeling
steps. Section 17.5 and Section 17.6 summarizes the issues associated with modeling of future
year emissions data. Finally, Section 17.7 provides guidance on emissions to be used for local
area primary PM2 5 modeling.
       17.1  What Emission Inventory Data are Needed to Support Air Quality Models?

       Air quality models require hourly emissions for each grid cell and model layer in the
modeling domain. Depending on the application, models may be run for episodic periods or for
       36 2002 is the recommended inventory year for the baseline modeling (the starting point
for future year projections). Other years may be modeled for the base case modeling (for
performance evaluation) if episodes are chosen from years other than 2002. If all of the episodes
are from 2002, then the base case and baseline inventories will be for the same year. See Section
14.1  for more details on selecting modeling time periods.

                                          168

-------
up to a full year or multiple years.  The emissions need to be allocated to the chemical species
required for the chemical mechanism and version of the model being used. The speciated
emissions of VOC and PM25 depend on the air quality model and chemical mechanism being
used. For primary PM2 5 emissions, the models typically divide the PM into (at a minimum)
organic carbon, elemental carbon, primary sulfate, primary nitrate,  and "all other" PM2 5
emissions. Calculating the speciated emissions is part of the emission modeling step, described
in Section 17.4 based on data described in Section 17.3.

       Emission inventory data from five general categories are needed to support air quality
modeling: stationary point-source emissions, stationary area-source emissions (also called non-
point), mobile emissions for on-road sources, mobile emissions for nonroad sources (including
aircraft, railroad, and marine vessels), and biogenic/geogenic emissions.  The emission inventory
development and emissions modeling steps can be different for each of these categories.

Point Sources- Point source inventories for modeling should be compiled at a minimum by
country, State/Tribe, county, facility, stack, and source category code (SCC) but often are further
subdivided by "point" and "segment"(see references below to point-source inventory
development). The point source data must include information on the location of sources (e.g.,
latitude/longitude coordinates); stack parameters  (stack diameter and height, exit gas temperature
and velocity); and operating schedules (e.g., monthly, day-of-week, and diurnal).

Stationary Area Sources- Stationary-area source emissions data should be compiled by
country, State/Tribe, county, and SCC. The most current stationary source emissions methods
can be found at http://www.epa.gov/ttn/chief/.

On-Road Mobile- On-road mobile source emissions should be estimated using the most current
version of the U.S. EPA MOBILE model ( http://www.epa.gov/omswww/m6.htm), and for
California, the most current version of EMFAC (http://www.arb.ca.gov/msei/msei.htm) in
concert with vehicle miles traveled (VMT) data representative of the time periods modeled.  The
MOBILE model allows modelers to override default settings to obtain a local-specific on-road
inventory, and modelers  should consider using these options to improve their inventories.  On-
road emissions and VMT should be compiled at least at the country, State/Tribe, county, and
SCC level, though modelers may optionally compile and use data for individual road segments
(called "links"). The link approaches  requires starting and ending coordinates for each road  link.
As noted in Section 17.3, link-based emissions can be based on Travel Demand Models (TDMs).

Nonroad Mobile- For nonroad mobile sources, the emissions should be compiled as country,
state/tribe, county and SCC totals.  For activities  other than aircraft, locomotives, and marine,
inventory developers should use the most current version of EPA's NONROAD model
(http://www.epa.gov/oms/nonrdmdl.htm) or alternative model. The California OFFROAD
model  should be preferentially used for emissions in California, available at
http://www.arb.ca.gov/msei/msei.htm. Since the  NONROAD model does not include emissions
of aircraft, locomotives, and commercial marine vessels, these emissions should be estimated
using other approaches described as follows:


                                          169

-------
          Aircraft emissions should ideally be calculated based on airport-specific landing and
          takeoff activity data and the Federal Aviation Administrations (FAA) Emissions Data
          and Dispersion Model (EDMS), available at
          http://www.faa.gov/about/office_org/headquarters_offices/aep/models/edms_model/.
          Although aircraft emissions may be at the county and SCC level, point
          (latitude/longitude) or shape file spatial resolution may be desirable (but possibly
          more challenging to use in emissions models). County emissions of aircraft at airports
          can be allocated to point locations using emissions modeling features, such as the
          SMOKE area-to-point feature.

   •      Railroad emissions should ideally be calculated at the county and SCC level using
          fuel usage activity data distributed among train types (e.g., long haul, yard,
          passenger) and rail ton/miles data.

   •      Commercial marine vessel emissions should ideally be calculated at the county and
          SCC level by fuel/cargo activity, or by growth from previously determined inventory
          efforts. The emissions are distributed to ports and cargo lanes by activity data.
          County emissions of port-based emissions can be allocated to point locations using
          emissions modeling features, such as the  SMOKE area-to-point feature.

The most current mobile source emissions methods can be found at
http ://www. epa.gov/ttn/chief/.

Biogenic Emissions - Biogenic emissions from plants and soil contribute VOC (including
terpenes that are imported for PM and regional haze modeling), NOx, and CO emissions, which
are best calculated as part of the emissions modeling step described in Section  17.4.  These
emissions require land use data, which is described as an ancillary dataset in Section 17.3.
Geogenic emissions (such as volcanic emissions) are often not relevant for many areas, but if
geogenic sources in the modeling domain are expected to contribute in a significant way to air
quality problems, they should be included.

Other Considerations - For all sectors, emissions data must be compiled at a minimum as an
annual total for the base modeling year.  Emissions can also be compiled as monthly total
emissions or (for ozone) an average summer day inventory. In any case, the temporal allocation
step during emissions modeling (see Sections 17.3 and 17.4) must adjust the inventory resolution
for the modeling time period. Additionally, we encourage the use of more temporally specific
data where it is available and can be expected to improve model performance.  For example,
hour-specific Continuous Emissions Monitoring (CEM) data may be used as a  source for hour-
specific NOx and SO2 emissions and exit gas flow rates37, and the hourly heat-input CEM data
       37Day specific emissions data, such as CEM data or wildfire data may be useful in the
base case modeling to improve model performance.  But in some cases, it may not be appropriate
to project day-specific data to the future.  For example, if a large power plant was shutdown for
maintenance during the base case period, it would not be logical to project zero emissions for
that source in the future (if it is assumed that the plant will be operating in the future year).

                                          170

-------
can be used to allocate emissions of other pollutants. CEM data are provided by EPA's Clean
Air Markets Division in a format suitable for use in emissions processing at
http://cfpub.epa.gov/gdm/index.cfm?fuseaction=emissions.prepacksmoke.

       Additionally, inventory developers are encouraged to focus their emissions data
collection, quality assurance, and emissions modeling efforts on the emissions sources that will
have the greatest impact on their modeling results. In other words, inventory developers should
devote the greatest attention to the sources that are most important: those in the immediate
nonattainment areas, or large sources that may be responsible for transport into the
nonattainment areas. Such a hierarchical approach to inventory development is described in
(Smith, 1998).

       Inventories should be built using the most current, accurate, and practical methods
available.  Several references are available for guidance on building emission inventories.  The
first is the "Emissions Inventory Guidance for Implementation of Ozone and Particulate Matter
NAAQS and Regional Haze Regulations" (U.S. EPA, 2005e).  Additionally, modelers may want
to consider EPA's approaches used for developing the 2002 NEI.  Available NEI documentation
may be used to help guide the development of the modeling inventory
(http://www.epa.gov/ttn/chief/net/2002inventory.html).  Lastly, 10 volumes (many with several
chapters comprised of separate documents) have been issued by the Emission Inventory
Improvement Program (EIIP) for use in inventory development. Nine of these volumes are
relevant to creating emission inventories for ozone, PM, and regional haze modeling (U.S. EPA,
2004h):

       Volume I:    Introduction
       Volume II:    Point Sources
•      Volume III:   Area Sources and Area Source Method Abstracts
•      Volume IV:   Mobile Sources
•      Volume V:    Biogenics Sources
•      Volume VI:   Quality Assurance/Quality Control
       Volume VII:  Data Management Procedures
       Volume IX:   Particulate Emissions
•      Volume X:    Emission Projections

       The EIIP documents are available electronically through the EPA website at
http://www.epa.gov/ttn/chief/eiip/techreport/. The quality assurance procedures contain
essential steps for inventory preparation, which help assure that the emission inventory is
appropriate for SIP air quality modeling.
Therefore, before projecting the baseline to the future, it may be necessary to remove certain
day-specific inventory information and replace it with average data. This issue is not a concern
for day-specific mobile source or biogenic emissions data which are dependent on day specific
(or even hourly) meteorological data for the time periods modeled.

                                          171

-------
       17.2   Can The National Emission Inventory Be Used As a Starting Point?

       We recommend that States and Tribes start with available inventories suitable for air
quality modeling of the selected time period(s). If no such inventory is available, States/Tribes
may derive an inventory suitable for use with models starting from the National Emission
inventory (NEI), available from http://www.epa. gov/ttn/chief/net/. The 2002 NEI can be used as
a starting for inventory development. However, the detail and accuracy of the emissions in the
NEI may not be sufficient for use in attainment demonstration modeling.  NEI data may be
sufficient for estimating emissions in States that are hundred of kilometers downwind of the
nonattainment area, but may not be sufficient for estimating local emissions that are critically
important in estimating ozone or PM in the nonattainment area.  Thus, States/Tribes should
review the contents of the NEI for accuracy and completeness for regional and local scale
modeling and amend the data where it is insufficient to meet the needs of the local air quality
model application. While some benefits can be realized for states using the same inventory
(such as the NEI) as a starting point for their work, the more important goal for SIP inventory
development is to ensure that the data are representative and accurate enough for making good
control strategy decisions.
       17.3   What Other Data are Needed to Support Emissions Modeling?

       The emission inventories must be converted (through emissions modeling) from their
original resolution (e.g., database records) to input files for air quality models. These input files
generally require emissions to be specified by model grid cell, hour, and model chemical species.
This section describes the ancillary data that modelers should collect that allow emissions
models to convert the emission inventory data.

       Ancillary data for temporal allocation are necessary for stationary point, stationary area,
and all mobile sources. To facilitate temporal allocation of the emissions, factors (called
profiles) must be created to convert annual emissions to specific months (monthly profiles),
average-day emissions to a specific day of the week (weekly profiles), and daily emissions to
hours of the day (hourly profiles). Additionally, a cross-reference file is needed to assign the
temporal profiles to the inventory records by SCC, facility, or some other inventory
characteristics.  Where available, the operating information that may be available from the point-
source inventory should be used to create inventory-specific temporal factors.  EPA provides a
starting point for the temporal profiles and cross-reference files, available  at:
http://www.epa.gov/ttn/chief/emch/temporal/.

       For point sources, hourly Continuous Emissions Monitoring data are recommended for
use in model-evaluation runs. For future-year runs, we recommend creating an "average-year" or
"typical year" temporal allocation approach that creates representative emissions for the
"baseline inventory" but that also includes similar daily temporal variability as could be
expected for any given year. Care should be taken to not reduce or increase day-to-day
                                           172

-------
variability in the averaging approach, with the exception of eliminating year-specific outages or
other year-specific anomalies38 within the years used for the model-attainment test.

       The emissions models also need information about the chemical species of the VOC and
PM2 5 emissions for stationary point, stationary area, and all mobile sources.  These data are used
to disaggregate the total VOC and PM2 5 emissions to the chemical species expected by the air
quality model and are called speciation "factors" or "profiles". EPA provides a starting point for
the VOC and PM2 5 speciation data, which are available at:
http://www.epa.gov/ttn/chief/emch/speciation/. For large or critical VOC and PM2 5 sources in
the modeling domain, States/Tribes should consider determining the individual chemical
compounds contributing to the total VOC and PM25. If collected, this information should then
be used to compile improved speciation profiles for the critical facilities or source categories.
These speciation profiles should be assigned to the inventory  by a speciation cross-reference file,
which also needs to be created or updated from the available defaults.  The cross-reference files
typically assign speciation profiles based on SCC code, though facility-specific assignments for
point source code is also possible if plant-specific data are available.

       For all  source sectors that are compiled at a county resolution, the emissions models also
need information about allocating the countywide emissions to individual grid cells that intersect
the county.  Such sectors include stationary area, nonroad mobile, and non-link on-road mobile
sources.  The spatial allocation process assigns fractions of county-total emissions to the model's
grid  cells intersecting the county based on a "surrogate" data type (e.g., population or housing
data). The appropriate types of surrogate data to use for each SCC in the inventories should be
identified for this processing step. Spatial surrogates can be created using Geographic
Information Systems (GISs) to calculate the fraction of countywide emissions to allocate to each
grid  cell based on the surrogate type. These calculations can also be made using EPA's Surrogate
Tool (http ://www.cep.unc.edu/empd/projects/mims/spatial/srgtool/), which is based on the
Multimedia Integrated Modeling System (MEVIS) Spatial Allocator
(http://www.cep.unc.edu/empd/projects/mims/spatial/). Additionally, all SCCs needing spatial
surrogates should be assigned a surrogate in a cross-reference file. Point sources do not need
spatial surrogates, since the emissions models assign the grid  location based on the latitude and
longitude of the point sources.  Additionally, EPA provides spatial surrogates and cross-
references for a limited set of modeling grids and surrogate types at:
http://www.epa.gov/ttn/chief/emch/spatial/.  Finally, for future-year modeling, emissions
developers can choose to change their spatial surrogate data based on predicted changes in land
       38The baseline inventory should contain emissions estimates that are consistent with
emissions patterns which might be expected in the future year(s).  If unplanned outages or
shutdowns at point sources occurred in the base year, then it may not be appropriate to project
those emissions to the future. Using a typical or average year baseline inventory provides an
appropriate platform for comparisons between the base year and future years. However, certain
emissions strategies may depend on capturing temporal emissions patterns on days with peak
ambient concentrations. This should be considered when creating typical or average year
inventories.

                                           173

-------
use patterns, population growth, and demographics, however, the impact and utility of such
approaches is not well characterized39.

       For biogenic emissions modeling, the Biogenic Emission Inventory System, version 3
(BEIS3) (http://www.epa.gov/asmdnerl/biogen.html) model comes with all needed ancillary
data, except the gridded land-use data and meteorology data for a specific air quality modeling
domain and grid. Other biogenic models are acceptable as well, provided that the users have
demonstrated their suitability for the modeling application. Land use and meteorology data that
are compatible with BEIS3 are needed for the specific grid and grid-cell resolution that is being
used.  For BEIS3, land use data can be created with the MIMS Spatial Allocator
(http://www.cep.unc.edu/empd/projects/mims/spatial/), using raw Biogenic Emissions
Landcover Data (BELD), version 3 (http://www.epa.gov/ttn/chief/emch/biogenic/). For future-
year modeling, emissions developers can choose to change their land cover data based on
predicted changes in land use patterns, however, the impact and utility of such approaches is not
well characterized. The impact and model characterization of biogenic terpene and sequiterpene
on secondary aerosol format for PM modeling is currently evolving. Emissions modelers should
consider the latest available information on this issue where biogenic sources are key sources in
their modeling region.

       On-road emissions for fine-scale model grids (e.g., 4-km grid cells or smaller) may be
based on a link-based mentioned in Section 17.2. The VMT and speed data needed for a link-
based approach can be provided using a Travel Demand Model (TDM). These models require
their own sets of inputs, which depend on the specific TDM used. Details on using TDMs for
link-based on-road mobile emissions are available from the EIIP document "Use of Locality-
Specific Transportation Data for the Development of Source Emission Inventories"
(http://www.epa.gov/ttn/chief/eiip/techreport/volume04/iv02.pdf).

       Emissions models have other input files that must be created. For example, criteria may
be needed for selecting elevated from non-elevated point sources. Each model has a large
number of files and settings which work together in fairly complex ways; therefore, care must be
taken  to determine the files needed for the emissions model, and to prepare all needed input files
in a way that will support using the emissions model for the specific air quality modeling
episode.
       17.4   How Are Inventory Data Converted Into Air Quality Model Input?

       Emissions models are used to convert inventory data to inputs for air quality modeling.
As described in Section 17.3, additional ancillary data is needed to complete the process. The
emissions data from each of the five emissions sectors are temporally allocated, chemically
speciated, and spatially allocated.  The resulting hourly, gridded, and speciated emissions from
       39At the time this document was written, tools to readily predict future-year land use
patterns are not readily available in a form for use in emissions modeling.

                                          174

-------
all sectors are then combined before being used by an air quality model. In this section, we will
provide information on several emissions models and summarize some additional issues that are
key for emissions modeling.

Emissions models Several emissions models are available for use in SIPs. While no single
model has been specifically created for all situations of SIP development, each model is
generally capable of performing the temporal, chemical, and spatial allocation steps as well as
various other steps. Users of these models are responsible for ensuring that the emissions
processing steps are transforming the emission inventories as intended and are not changing the
emissions in any unexpected way.  The models each have different capabilities, limitations, and
nuances. Therefore, when choosing an emissions model, it is worthwhile to discuss the choice
with the developers of these systems and/or with EPA to  establish which model is best for the
specific application.

       Currently there are three main  emissions models being used to process emissions for
input into photochemical grid models and a fourth model that is under construction at the time
that this document was written.  They  are:  Sparse Matrix Operator Kernel Emissions (SMOKE);
Emissions Modeling System (EMS-2001); and Emissions Preprocessor System - Version 2.5
(EPS 2.5) and version 3.0 (EPS-3). The Consolidated Community Emissions Processing Tool
(CONCEPT) is under development by the Regional Planning Organizations.

       The Sparse Matrix Operator Kernel Emissions (SMOKE),  software and User's Guide are
available through the University of North Carolina,  Carolina Environmental Program
(http://www.cep.unc.edu/empd/products/smoke). SMOKE supports processing of criteria,
mercury, and toxics inventories for stationary point, stationary  area, mobile, and biogenic
emissions. It can create input files for the CMAQ, CAMX, UAM-V, and REMSAD air quality
models. SMOKE was the basis for development of the BEIS3  system, so BEIS3 in its original
form can be used easily with SMOKE. Applications of SMOKE have been presented at several
of the International Emissions Inventory Workshops (Houyoux, 2000; Strum, 2003).  SMOKE is
available for UNIX and Linux operating systems and is not recommended for use on a PC.  It
does not require third party software.  It does not include utilities for creating speciation profiles,
biogenic land use, or spatial surrogates, though the latter two datasets can be built using the
Multimedia Integrated Modeling System (MEVIS) Spatial Allocator Tool
(http://www.cep.unc.edu/empd/projects/mims/spatial/).  Support for the SMOKE system is
available through the Community Modeling and Analysis System  (CMAS) help desk
(http ://www. cmascenter.org/html/help .html).

       The Emissions Modeling System, (EMS  2001,
http://www.ladco.org/tech/emis/ems_2001/) is a later version of EMS-95, which was used in the
modeling underlying the U.S. EPA NOx SIP call rule to reduce regional NOx emissions (U.S.
EPA 1998b), as well as in other applications of nested regional air quality models.  It can create
inputs for the CAMX and UAM-V models. It includes the BIOME3 model, which provides
access to similar calculations of biogenic emissions as are available in the BEIS3 system.  EMS-
2001 can be run on either Linux or Windows NT, and users must purchase a license for the SASฎ
                                          175

-------
software to use it. It includes utilities for creating speciation profiles, biogenic land use, and
spatial surrogates. An updated version has new spatial processors, which limit the need for GIS
software.

       The Emissions Preprocessor System - Version 2.5 (EPS-2.5), software and User's Guide
are available through Systems Applications International/ICF Consulting (www.uamv.com).
EPS-2.5 is a comprehensive emissions processing system that supports processing of stationary
point, stationary area, and mobile emissions for the development of base and future-year
modeling inventories for input to the CAMX, UAM-V, and REMSAD models.  EPS-2.5 consists
of a set of stand-alone FORTRAN programs that do not require third-party software. The system
is capable of preparing local, regional, and continental-scale emission inventories for ozone,
particulate matter, mercury, and air toxics modeling applications. EPS 2.5 is available for UNIX
and Linux operating systems. It  includes utilities for creating source-specific temporal and
chemical speciation profiles based on locally provided detailed information for episode specific
emission inventories. It also includes utilities for preparing spatial surrogates.  In addition, EPS-
2.5 has the capability of creating modeling inventories required for the application of source
apportionment techniques such as UAM-V's Ozone and Precursor Tagging Methodology
(OPTM). A newer version of EPS is also available, called EPS-3.

       The Consolidated Community Emissions Processing Tool (CONCEPT) is an open source
model written primarily in PostgreSQL.  Users are encouraged to make additions and
enhancements to the model. The database structure of the model makes the system easy to
understand, and the modeling codes themselves are documented to encourage user participation
in  customizing the system for specific modeling requirements.  The CONCEPT model  structure
and implementation allows for multiple levels of QA analysis during every step of the emissions
calculation process.  Using the database structures, an emissions modeler can easily trace a
process or facility and review the calculation procedures and assumptions for any emissions
value. CONCEPT includes modules for the major emissions source categories: area source,
point source, on-road motor vehicles, non-road motor vehicles and biogenic emissions, as well as
a number of supporting modules, including spatial allocation factor development, speciation
profile development, growth and control for point and area sources, and CEM point source
emissions handling.  Additional work by the emissions modeling community has begun
development of CEM preprocessing software, graphical QA tools, and an interface to the traffic
demand models for on-road motor vehicle emissions estimation.

Biogenic emissions  Estimates for biogenic emissions can be made using the BEIS emissions
model (Geron, et al., 1994). The BEIS3  model estimates  CO, NOx, and VOC emissions from
vegetation and soils in the gridded, hourly, and model-species forms  needed for air quality
modeling. Guenther, et al., (2000) contains the technical  development and review of BEIS3.
Vukovich, et al.,  (2002) summarizes new input data and algorithms as implemented within
SMOKE. Arunachalam, et al., (2002) presents the impact of BEIS3 emissions on ozone. For
more detailed local estimates, a State or Tribe should review the biogenic emissions on a county
basis and update as needed the spatial patterns of land use data. BEIS3.13 was the latest version
of BEIS3 in use by EPA at the time this document was written.
                                          176

-------
        Other models for biogenic emissions include the BIOME model that is a part of EMS-
 2001, the Global Biosphere Emissions and Interactions System (GloBEIS), and the Model of
 Emissions of Gases and Aerosols from Nature (MEGAN). GloBEIS estimates emissions from
 natural (biogenic) sources and is designed for use in combination with photochemical modeling
 systems, such as CAMx (http://www.globeis.com/). The MEGAN model is available from
 http ://bai. acd.ucar. edu/Megan/.

        All of the biogenic models require a mix of land uses to be specified for each county or
 grid cell, as well as hourly temperature and in some cases other meteorological data. If a State or
 Tribe believes the average land use mix characterized for a county is inappropriate for certain
 gridded locations within a county, this may be overridden for the grid cells in question on a case
 by case basis.

 Agricultural Ammonia Emissions- Ammonia emissions from agricultural operations are
 typically estimated using standalone models.  A commonly used model to estimate ammonia is
 the Carnegie Mellon University Ammonia Model (http://www.emu.edu/ammonia/). A promising
 model that is under development is the CONCEPT Farm Model (http://www.conceptmodel.org/
 nh3/nh3_index.html).
        17.5   Are there Other Emissions Modeling Issues?

       In addition to the general emissions modeling steps and the biogenic emissions modeling,
there are several other issues of which the air quality modeler should be aware.  These are:

       Elevated point sources
       Advanced approaches for mobile source modeling
       Quality assurance

In the remainder of this section, we briefly address each of these issues.

Elevated Point Sources Point sources need to be assigned to an appropriate model layer40 (the
vertical modeling dimension). Depending on the air quality model that is being used, different
emissions modeling steps can be taken. Models such as UAM-V and CAMX, expect input
emissions files separately for layer-1 emissions and elevated point-source emissions.
Additionally,  elevated point sources may be flagged for treatment with a plume-in-grid (PinG)
approach. For these models, emissions modelers must supply a criteria for specifying which point
sources will be treated as elevated and as PinG sources.  In this case, the air quality model
calculates the plume rise of the point source emissions when the model is run.
        40Point sources generally comprise most of the elevated emissions (above layer 1),
 although other area sources, such as fires and aircraft, may also have emissions assigned to
 elevated layers.

                                           177

-------
       Models such as CMAQ expect (1) a 3-D emissions input file, for which the elevated
plume rise has already been calculated and (2) a separate optional PinG emissions file. Emissions
modelers may optionally specify which sources to treat as elevated and PinG. Since the
emissions model must calculate plume rise in advance, it must use meteorological data such as
temperature, pressure and wind speeds. For the sake of consistency, the meteorological data that
is used in calculating plume rise is usually the same as what is used by the air quality model.

Mobile Source Modeling  Mobile source emissions modeling takes two primary approaches.
The first approach is to compute emissions from VMT and emission factors from the MOBILE
model prior to use by an emissions model. The second approach is to allow an emissions model,
such as SMOKE, to drive the MOBILE model using gridded meteorology data.  Many more
assumptions must be made about using average temperatures in the  first approach, since
emissions are calculated on an annual total, monthly, or average-day basis and therefore do not
include the day-to-day temperature variability that the second approach includes. It is widely
assumed that the second approach is more robust for local-scale modeling, though we do not
recommend one approach over the other.  States/Tribes are encouraged to choose an approach
that gives sufficient model  performance for their attainment demonstration modeling.

Quality Assurance The third additional emissions modeling topic we have summarized here is
emissions modeling quality assurance (QA).  A brief synopsis of appropriate quality assurance
(QA) approaches for emissions modeling is available in Section 2.19 of the SMOKE manual
(http://cf.unc.edU/cep/empd/products/smoke/version2.l/html/ch02sl9.html).  The purpose of QA
for emissions modeling is to ensure that the inventories are correctly processed using the
information the modeler intended. (It is assumed here that the inventory itself has already been
QA'd through inventory QA procedures, as referenced in Section 17.3.)  Emissions modeling QA
includes such activities as reviewing log files for errors and warnings and addressing problems;
comparing emissions between each of the processing stages (e.g., data import, speciation,
temporal allocation) to ensure mass is conserved; checking that the correct speciation, temporal
allocation, and spatial allocation factors were applied; ensuring that emissions going into the air
quality model are consistent with expected results; and checking modeling-specific parameters
such as stack parameters. In addition, the process of running emission inventories through
emissions models and air quality models often provides insights into the emission inventories.
These insights can lead to inventory changes that improve the quality of inventories for additional
modeling iterations.

       In general, this guidance also encourages the use of graphical analysis and Geographic
Information Systems (GIS) for improved QA of emissions data and processing. A commonly used
analysis tool for model-ready emissions data is the Package for Analysis and Visualization of
Environmental Data (PAVE), available at
http ://www. cep.unc. edu/empd/EDS S/pave_doc/index. shtml.

       Lastly, at the time this document was written, the EPA is developing the Emissions
Modeling Framework (EMF), which includes extensive data management and quality assurance
                                          178

-------
features intended to assist emissions modelers with the enhanced quality of their data.  This tool
will be available for use in 2007 for support of ozone, PM2 5, and regional haze SIPs .

       17.6          How Are Emissions Estimated for Future Years?

       Emissions estimates for future years are called "emissions projections".  These projections
include both emissions growth (due to increased or decreased activities) and emissions controls
(due to regulations that reduce emissions in specific ways in the future). The goal in making
projections is to obtain a reasonable estimate of future-year emissions that accounts for the key
variables that will affect future emissions.  Each State/Tribe is encouraged to incorporate in its
analysis the variables that have historically been shown to drive its economy and emissions, as
well as the changes in growth patterns and regulations that are expected to take place between the
time of their base year and future years. For details on which future year(s) should be  modeled
for attainment demonstrations, refer to Section 3.6.

       A report entitled "Procedures For Preparing Emissions Projections" describes emissions
projections issues and approaches (U.S. EPA, 1991b; also available at
www.epa.gov/ttn/chief/eiip/techreport/volumelO/x01.pdf). It is currently the most
comprehensive resource available about emissions projections. In this section, we will address an
overall approach for tackling emissions projection issues that incorporates use of this document
and points out where other approaches or new data are available to amend the information
included in that report.

Developers of projection-year emissions are advised to take the steps in the bulleted list below.
Each of these steps corresponds to a subsequent subsection.

•      Identify sectors of the inventory that require projections and sectors for which projections
       are not advisable, and prioritize these sectors based on their expected impact on the
       modeling region (Section 17.6.1).

•      Collect the available data and models that can be used to project emissions for each of the
       sectors (Section 17.6.2).

•      For key sectors, determine what information will impact the projection results the most,
       and ensure  that the data values reflect conditions or expectations of the modeling region
       (Section 17.6.3)

•      Create inputs needed for emissions models, create future year inventories,  quality assure
       them, and create air quality model inputs from them (Section 17.6.4)

The remainder of this section provides additional details about these steps.
                                            179

-------
              17.6.1 Identify and prioritize key sectors and regions for projections

       Emissions modelers should evaluate their inventories for those sectors and regions that are
most critical to their modeling results.  The purpose of identifying these key sectors is to direct
more resources and efforts to estimating future-year emissions for these sectors in the subsequent
steps.

       Sectors can be subsets of the larger groups of stationary area, on-road mobile, nonroad
mobile, and point sources. For ozone modeling, in regions that are NOx-limited, sectors with
higher NOX emissions will be more important for ensuring correct projections, and for VOC-
limited regions, sectors with higher VOC emissions will be more important. Mobile sources are
typically large contributors to both VOC and NOx and are therefore usually a priority in
estimation of future-year emissions, particularly for urban areas.  For PM and regional haze
modeling, the relative importance of primary PM25 emissions and precursors will depend on the
area of the country. Primary PM2 5 emissions may be important in urban areas. In the East, SO2
emissions may be of primary importance.  And NOx, VOC, and NH3 will vary in importance
depending on the particular situation.

       In some cases, there are sectors that are very difficult or highly uncertain  to project into
the future.  Wildfire emissions is a common example of such a sector, since models are not
readily available that estimate the magnitude and location of future-year emissions. In these
cases, it may be advisable to create an "average year" inventory and temporal distribution of the
emissions.

       Emissions modelers can consider the proximity  of the emissions and the expectation of
long-range  transport when reviewing projection assumptions. Those emissions that have an
impact on the nonattainment status (whether transported from outside areas or locally generated)
should be given highest priority in evaluating the projection assumptions. For example, less
priority can be given to emissions from distant states for pollutants that do not transport from
those states to  the nonattainment area.

              17.6.2 Collect available data and models for each sector needing projections

       The EIIP projections document (referenced previously) provides a starting point for this
step. The EIIP document provides the majority  of information about sources of data on growth
and controls.  This section supplements that document.

       For each  sector needing projections (and in particular for the priority sectors), emissions
modelers should collect and review growth and  control  information.  As noted in the EIIP
document, the  sources of such information depend upon the sector of interest. New data and
approaches beyond those in the EIIP document are available and provided here; likewise, new
data and approaches will become available subsequent to  the publication of this document.  The
information provided here is a snapshot of information that will grow and adapt over time, and
emissions modelers should seek out such new information as part of this step.
                                           180

-------
       For sectors and regions with lower priority for consideration (as described in Section
17.6.1), emissions modelers can rely on available default data. For sectors and regions that are
important to modeling the nonattainment area, emissions modelers should spend additional effort
obtaining improved information from other modeling applications (such as modeling done by
RPOs or other States) or developing the improved information themselves.

       In addition to the sources of control information about non-EGU point and stationary area
sources listed in the EIIP projection document, EPA has provides control assumptions with its
modeling data used for promulgating air quality rules.  For example, emissions growth and
control assumptions were provided with the Clean Air  Interstate Rule (See
http://www.epa.gov/air/interstateairquality/pdfs/fmaltech01.pdffor additional information on
obtaining actual data files) More current information is usually available than what is available
from EPA, which is best obtained by working with state and local regulatory officials for the SIP
modeling region.  The remainder of this section  focuses on sources of growth information.

Point sources.
       Section 2 of the EIIP projections document gives details about projection of point-source
emissions.  There are two major subsets of point sources: electric generating utilities (EGUs) and
non-EGUs.  The  Clean Air Markets Division (CAMD) of the U.S. EPA uses the Integrated
Planning Model (IPM) to model emissions trading programs that now dominate the  prediction of
future-year emissions from EGUs. More information on IPM is available at
(http://www.epa.gov/airmarkt/epa-ipm/). Additionally, IPM-based emissions are posted by
CAMD on EPA's website (http://www.epa.gov/airmarkets/epa-ipm/iaqr.htmn. Other trading
models may exist and could be used for estimation of future-year emissions.

       To prevent double-counting of emissions sources, emissions modelers who use  IPM
should be careful to match the sources of emissions from IPM with the base year emissions
sources.  This helps to ensure that the EGU part of their point source inventory is separated from
the non-EGU part based  on the facilities included in IPM. The facilities included in IPM are
defined using the National Electric Energy Database System (NEEDS). The NEEDS dataset
should be compared to the point inventory using the Office of Regulatory Information Systems
(ORIS) Plant ID  field from both.  In some cases (e.g., co-generation facilities), only some of the
units at these facilities are included in IPM; therefore, the separation of EGUs from  non-EGUs
should be done by unit and not by facility.

       Since base-year emissions of EGU point sources can be based on hour-specific  emissions
from the CEM program,  emissions modelers must choose a temporal allocation approach for
estimated future-year emissions. Emissions modelers should choose an approach that is
representative of future expected behavior and not limited to any single year's closures and
maintenance schedule. Ideally, types of EGUs that are run during high demand conditions would
have temporal allocation approaches that reflected those peaks, and units that are run
continuously would be temporally allocated in a more uniform way. Analysis of several years of
CEM data could  be helpful to determine these unit-by-unit trends and develop profiles
accordingly.
                                           181

-------
       Additional considerations for projecting point source emissions without the IPM model
are provided in the EIIP projection document.  However, several of the references in that
document are now out of date, as follows.

•      The latest SCC codes are available at http://www.epa.gov/ttn/chief/codes/ and not the
       website provided in the document.
•      The AP-42 emission factor website is now http://www.epa.gov/ttn/chief/ap42/index.html.
•      The website with further information about point source inventories is
       http://www.epa.gov/ttn/chief/eiip/techreport/
•      The references in Table 2.1-1 web link is no longer applicable. Instead, the following
       information can be used to help provide similar information:
              - Energy consumption by industry report, 2002:
              http://www.eia.doe.gov/oiaf/analysispaper/industry/consumption.html
              - Annual Energy outlook report, 2005:
              http://www.eia. doe. gov/oiaf//aeo/pdf/03 83 (2005ypdf

The references in Table 2.1-2 are also out of date and should be updated as follows:
          - US EPA Emissions trends reports. These reports can be used to provide the historical
              trends of emissions sectors based on the National Emission Inventory:
              http ://www.epa. gov/air/airtrends/
          - Integrated Planning Model:
              http://www.epa.gov/airmarkt/epa-ipm/
          - Multiple Projections System - is no longer in use
          - California Emission Forecasting System:
              http ://www. arb. ca. gov/html/databases.htm

The tools and data sources described in the "Additional resources" section below should also be
evaluated for use for projection of point source emissions.

Stationary area sources.
       Section 3 of the EIIP projections document provides details about projecting area-source
(a.k.a nonpoint emissions). However, several of the references or statements in that document are
now out of date,  as follows.

•      The web  sites for SCCs, AP-42 emission factors, and more information about area-source
       inventories are out of date. These should be updated to the same sites as were listed in the
       point sources section above.

•      The ASEM model mentioned in Section 3.2 of the EIIP projection document was not
       completed and is not available.

•      The references in Table 3.1-2 are out of date and should be updated as follows:
          - US EPA Emissions trends reports. These reports can be used to provide the historical
              trends of emissions sectors based on the National Emission Inventory:
                                           182

-------
             http://www.epa.gov/air/airtrends/
         - Multiple Projections System - is no longer in use
         - California Emission Forecasting System:
             http ://www. arb. ca.gov/html/databases.htm

The tools and data sources described in the "Additional resources" section below should also be
evaluated for use for projection of point source emissions.

On-road mobile sources.
       Section 3 of the EIIP projections document provides details about projecting on-road
mobile-sources. It is  still a very relevant snapshot of on-road mobile projection approaches;
however, several of the references or statements in that document are now out of date, as follows.
Additionally, several  new capabilities are available, which also are included in the list below.

•      The newer versions of the MOBILE model support 28 vehicle types instead of the 8 types
       listed in this document, however, the NEI and other inventories often group these 28 types
       to the same 8 described in the document.
•      The web sites for SCCs, AP-42 emission factors, and more information about area-source
       inventories are out of date.  These should be updated to the same sites as were listed in the
       point sources section above.
•      The EPA Office of Mobile Sources is now called the Office of Transportation and Air
       Quality (OTAQ). Its website is http://www.epa.gov/otaq/index.htm.
•      The new website for the MOBILE model is http://www.epa.gov/otaq/mobile.htm.
•      The PARTS model is no longer used, because it has been superceded  by the MOBILE
       model, starting with version 6 (MOBILE6).
•      The increased popularity  of sport utility vehicles has continued, but may be affecting more
       than only the percentage of light duty gasoline vehicles. Additionally, since MOBILE6
       includes 28 vehicle types, the increases in larger and heavier vehicles can be more
       carefully accounted for using MOBILE6 than the EIIP projection document mentions.
•      Updates to the SMOKE model  and other emissions models permit the temperature
       information used in MOBILE6 to come from the prognostic meteorological models (e.g.,
       MM5 and MCIP) that are used  to prepare inputs to air quality models.
•      OTAQ has created the National Mobile Inventory Model, which drives both the
       MOBILE6 model and the NONROAD model to create base-year and future-year
       inventories. Additional information on this tool is available at:
       http ://www. epa. gov/otaq/nmim.htm.
•      The references in Table 4.1-1 are out of date and should be updated as follows:
         - MOBILE6 - is an operational  model and can be used
             See http://www.epa.gov/otaq/mobile.htm
         - The PARTS model is no longer needed because MOBILE6 replaces it
         - The California mobile sources emissions page is now:
             http ://www. arb. ca. gov/msei/msei .htm
         - The EIIP  mobile source document is now available at:
             http://www.epa.gov/ttn/chief/eiip/techreport/volume04/index.html
                                          183

-------
         - Emission Inventory Guidance documentation is now available at:
             http://www.epa.gov/ttn/chief/eidocs/eiguid/index.html

Additionally, the Economic Growth Analysis System (EGAS) contains VMT growth information
and can be used to grow VMT by SCC. This system is further described in the "Additional
resources" section below.

Nonroad mobile sources.
       Section 4 of the EIIP projections document provides details about projecting nonroad
mobile sources. Several of the references or statements in that document are now out of date, as
follows. Additionally, several new capabilities are available, which also are included in the list
below.

•      As noted above for  on-road sources, the Office of Mobile Sources has changed to the
       Office of Transportation and Air Quality and has a different web site.
•      The EPA's NONROAD model has been completed and has been used in a variety of ways
       by EPA, states, and local agencies. The website for the NONROAD model is:
             http ://www. epa. gov/otaq/nonrdmdl. htm
•      California's OFFROAD model website is now:
             http ://www. arb .ca. gov/msei/off-road/updates.htm
•      OTAQ has created the National Mobile Inventory Model, which drives both the
       MOBILE6 model and the NONROAD model to create base-year and future-year
       inventories. Additional information on this tool is available at:
       http ://www. epa. gov/otaq/nmim.htm.
•      For aircraft, locomotive, and marine emissions, the best source of data on the impact of
       national rules from EPA's Office of Transportation and Air Quality (OTAQ) is the
       Federal Register notices including these rules.  Since each rule is different, there is no
       single guidance that can be made about how these should be incorporated. As EPA
       OAQPS incorporates these data into their work, it is possible in the future that it will be
       available on EPA's  Emissions Modeling Clearinghouse at
       http://www.epa.gov/ttn/chief/emch/index.html.
•      For aircraft emissions, the Federal Aviation Administration data can be used to assist with
       growth estimates. More information is provided in the "additional resources" section,
       below.

Additional resources.
       In addition to sector-specific resources, there are other sources of information that can be
used for more than a single sector as information to help determine the most applicable growth
rates for the modeling region.  These include the Economic Growth Analysis System (EGAS),
Bureau of Economic Analysis data, Bureau of Labor Statistics "Employment Outlook", Census
Bureau data, Trade Organizations and individual  facilities, and the Chemical Economics
Handbook. The following paragraphs provide a brief description and additional references for
each of the sources of data just listed.  Section 17.6.3 provides information about how to choose
the best information for a given sector.
                                          184

-------
       Multiple sources of information can be used to better inform emissions modelers. For
example, if a state and/or industry has shown steady decreases in gross product over the past five
years, that information could temper (or be used in place of) a projected 30% growth predicted
from EGAS over the next 10 years.  Such a discrepancy would simply highlight the fact that all
models, including REMI or those used by DOE can produce results that are not consistent with
other expectations. In addition, because of changes in technologies and fuel efficiency, economic
growth is not necessarily a sufficient indicator of emissions growth. While this guidance does not
provide a prescriptive approach for combining such information, the authors recognize that
predicting the future of emissions is not an exact science and therefore multiple sources of
information should be used to develop an informed  approach.

EGAS.  The latest EGAS model at this time is version 5 (http://www.epa.gov/ttn/ecas/egas5.htm)
The default version of EGAS relies primarily on three sources of data: (1) state-specific economic
data from the Regional Economic Model, Inc. (REMI) Policy Insight ฎ model (version 5.5) that
includes population growth estimates, (2) Region-specific fuel-use projections from the U.S.
Department of Energy (DOE) Annual Energy Outlook 2004, and (3) VMT projections included in
the MOBILE6.0 model. EGAS outputs growth factors that do not include control information for
a user-define base year and multiple future years through 2035.  EGAS uses default cross-walks
from these data sources to SIC, SCC and MACT codes.  The DOE data are used to compute
growth factors assigned by default to stationary source fuel combustion sectors, the VMT
projections are used for on-road mobile, and the REMI data are used for the remaining sectors.
Additionally, EGAS5 supports emissions modelers adding new data sources, changing
assignments of data sources to SICs, SCCs, and MACT codes, and creating custom configurations
of data that best represent the growth expected in their modeling region. The additional sources
of information listed below can be input into EGAS to develop such custom  scenarios. It should
be noted that these sources have very little, if any, projection information. Instead, they provide
historical data that could be extrapolated to produce projections or to derive projections using
other methods.

       EGAS results should be assigned to the inventory to give the best mapping of the raw data
to the inventory.  Generally, SIC is the preferred cross-walk for point sources and SCC is the
preferred cross-walk for non-point sources.  However, the DOE-based fuel consumption forecasts
(assigned by SCC) can also be useful for some sectors that use fuel combustion.

Bureau of Economic Analysis. The Bureau of Economic Analysis (BEA) provides historical
data about macroeconomic factors at the state level  such as gross state product, personal income
by industry, employment by industry, wages and earnings by industry, and population by state.
The website to obtain state and/or state-industry data is
http://www.bea.doc.gov/bea/regional/gsp/. Such information may be used to verify emissions
growth forecasts in EGAS by comparing available historical BEA  data (1997-2004) with the
REMI economic activity data driving emissions in EGAS for the same set of years. A good use
of the gross  state product data is to verify growth in sectors/codes that use REMI output and
value-added data.  Additionally, employment and population statistics can be compared with
sectors/codes using REMI employment and population data. The BEA also provides industry-
                                          185

-------
specific national information at http://www.bea.doc.gov/bea/dn2/home/annual_industry.htm.
However, since these data are at the national level, the use of the data should be limited to
comparisons with national simulations in EGAS.

Bureau of Labor Statistics. The Bureau of Labor Statistics (BLS) provides historical state-
specific and national employment data.  The BLS database is fairly large and provides more detail
than needed for verifying the growth factors generated in EGAS using REMI employment data.
However, BLS employment data is likely to be as accurate if not more accurate than employee
data provided by BEA and US Census.  The data are available at
http://www.bls.gov/sae/home.htm.

Census Bureau. The US Census Bureau provides data on total employees, total payroll, number
of establishments, and value of shipments by NAICS code (2 digit - 6 digit) for 1997 and 2002.
This database is ideal for examining changes in the number of establishments from 1997 - 2002.
However, it does not provide data from  1998-2001 so its trend information is limited to two
"snapshots".  Therefore, this data is useful primarily to evaluate the state/SCC and state/SIC
combinations from EGAS that show zero emissions growth, which indicates an assumption by
EGAS or its input data that those processes or industries do not exist in the state or region of
interest. The census data are available at
http://www.census.gOv/http://www.census.gov/econ/census02/guide/geosumm.htm.

Federal Aviation Administration.  The Federal Aviation Administration (FAA) provides aviation
data users with summary historical and forecast statistics on passenger demand and aviation
activity at U.S. airports.  The summary level forecasts are based on individual airport projections.
The Terminal Area Forecast (TAP) includes forecasts for  active airports in the National Plan of
Integrated Airport Systems (NPIAS).  The TAP can be accessed at http://www.apo.data.faa.gov/.
The TAP model allows users to create their own forecast scenarios. The TAP database, which
contains a query data application, allows the public to access and print historical and forecast (up
to 2025) aviation activity data by individual  airport, state or FAA region.

Trade Organizations and individual facilities. Trade organizations can also be a helpful
resource because they often have projections for future-year growth of their industry and may be
aware of pending facility closures or new facility construction.  This resources is most relevant
for  large industrial sources that usually are included in point source inventories. In addition,
emissions modelers should consider contacting large sources of emissions in their modeling
region for their expectations of emissions growth or reduction.  In many cases, large industries
may be willing to provide such information when presented with what will be used if no
additional information can be provided.  For these types of data sources, emissions modelers
should be aware of possible conflicts of interest, ie. industry groups or facilities may feel an
incentive to under-represent future-year emissions in hopes of avoiding additional requirements
for  reducing emissions; however, many  industrial representations have been valuable stakeholders
who are interested in providing accurate information.
                                           186

-------
Chemical Economics Handbook.  The Chemical Economics Handbook, produced by Access
International, Inc., is a series of reports on prices, production and consumption of hundreds of
chemical industry products and commodities. Past and current information on chemical products
and commodities is available, and projections of future prices, production and consumption are
often available. Reports on specific industries are also available. Reports at an industry level can
often be used to verify the efficacy of future industry modeling results. Each report is updated
every 3 years. Projections, some up to 5 years from  the current day,  are often prepared using
proprietary methods. Reports are available by subscription, and can be obtained as hard copy,
CD,  or through the Internet at http://www.sriconsulting.com/CEH/Public/index.html.

EPA base-year control assumptions. EPA OAQPS creates and uses data that quantify assumed
base year controls on non-EGU point and stationary  area sources. The most  recent data published
(at the time of this writing) are described with the Clean Air Interstate Rule (CAIR) Technical
Support Document (TSD). The Emission Inventory  chapter of this TSD contains a description of
the base year controls assumed and how to get the data files from an FTP site; it is available
online at http://www.epa.gov/air/interstateairquality/pdfs/finaltech01.pdf. Additionally, two
possible future references for these types of data include (1) the Emissions Modeling Framework,
which will include a Controls Programs Database and (2) the EPA OAQPS Emissions Modeling
Clearinghouse, available at http://www.epa.gov/ttn/chief/emch/index.html.

              17.6.3 Evaluate and refine data for projections in modeling region

       For key sectors, emissions modelers should determine what information will impact the
projection results the most, and ensure that the data values reflect conditions or expectations of
the modeling region. The key information is identified based on a combination of base-year
emissions data and growth and control rates.  An iterative process is helpful to complete this step,
as follows:

•      Estimate future-year emissions by combining the base-year emissions with the projection
       information.

•      Review future-year emissions data to identify large sectors of key pollutants (e.g. NOx,
       SO2, and VOC). Emissions summaries by state/SCC or state/SIC can be helpful for this
       step to consolidate the data to review.

•      For the largest future-year emissions sectors, review the growth and control assumptions,
       compare among all available data sources, and/or collect additional information to refine
       the assumptions. A representative  growth rate should be identified from the available data
       sources and all information known about the  sources and sectors. Stakeholder review of
       the data can be helpful during this step; for example, an industrial facility with large
       projected emissions may be able to review the data and provide additional information for
       a more informed future-year emissions estimate.
                                           187

-------
•      Additionally, emission modelers should also compare the future-year emissions to base-
       year emissions to identify overall growth rates for industrial sectors and reconsider
       excessively high growth rates, especially when associated with significant emissions.

•      Using the new information, repeat step 1 and continue these steps until a satisfactory
       future-year projection across the entire inventory has been completed.

       Emissions models (e.g., SMOKE, MOBILE6, NONROAD,  etc.) provide the capability to
create future-year inventories using base year inventories and projection information. Emissions
modelers will need to convert the projection data into specific formats for input to these models.
Prior to starting this step,  emissions modelers should determine which emissions model will be
used to perform the calculations and make sure that the type of information needed by the model
is being collected.

              17.6.4 Create future-year inventories and air quality model inputs

       Using the final projection data determined in the previous step, create the final inputs to
emissions models being used. Then, use the emissions model to create the future-year inventory.
Other inputs to emissions models, such as spatial surrogates, speciation profiles, or temporal
allocation factors may also need to be adjusted to reflect conditions expected in the future.

       Once a future-year inventory and other data have been created, it must undergo the same
steps as for the base-year  modeling, such as temporal allocation, speciation, spatial allocation,
elevated source selection, special on-road mobile processing, and quality assurance. Every
attempt should be made to use consistent approaches between the future year and the base year
for all of these modeling steps.  Inconsistencies in approaches between the future-year modeling
and the base-year modeling can lead to artificial differences in air quality modeling results that
can affect conclusions.  Therefore, it is critical to avoid such differences whenever possible.

       17.7   Emissions Preparation for a Local Area Analysis

       In  section 5.3 we discuss modeling of local primary PM25 sources with a dispersion
model. The emissions for these sources should closely match those used in photochemical
models. For example, every effort should be made to temporalize and speciate the primary
emissions using available profiles. Information on input and formatting of the emissions for
dispersion models can be  found in the respective users' guides. Pre-processing of the primary
emissions with an emissions model may be needed to derive the detailed information needed to
temporalize and speciate the emissions.

       In  section 5.3 we recommend using "actual emissions" in the local area analysis, as
compared to using "allowable emissions". By "allowable emissions", we mean the maximum
amount of emissions from a source which is consistent with emission limitations contained in an
                                           188

-------
applicable operating or construction permit for that source. We define "actual emissions"as the
best estimate of the actual emissions that occurred during the baseline period being modeled41.

       Modeling with allowable emissions is sometimes warranted. For example, for permit
modeling, we generally compare the absolute predicted modeled concentrations against the
NAAQS or the PSD increments. In these cases we want the modeling to be conservative.  That
is, we want to be sure that a maximum permitted level of emissions cannot cause of violation of
the NAAQS or PSD increment. Therefore, in the case of permit modeling, it is sometimes
appropriate to model with allowable emissions.  But for a local area analysis, we are trying to
determine the actual (or typical) contribution from sources to a monitor and we are using the
model results in a relative sense. Therefore, it is only appropriate (and in fact conservative) to use
actual emissions. Using actual emissions should lead to a  reasonable estimate of of air quality
improvements from reducing emissions at the flagged sources. Using allowable emissions might
lead to an overestimate of benefits from emissions controls at flagged sources.
        41In the emissions baseline modeling, the actual emissions may sometimes be replaced
 with "typical" emissions (e.g. EGU and fire emissions).

                                           189

-------
18.0  What are the Procedures for Evaluating Model Performance and
       What is the Role of Diagnostic Analyses?

       The results of a model performance evaluation should be considered prior to using
modeling to support an attainment demonstration or regional haze assessment.  The performance
of an air quality model can be evaluated in two ways: (1) how well is the model able to replicate
observed concentrations of PM25 components, ozone and/or precursors (surface and aloft), and
(2) how accurate is the model in characterizing the sensitivity of ozone and/or PM25 to changes in
emissions? The first type of evaluation can be broadly classified as an "operational evaluation"
while the second type of evaluation can be classified as a "diagnostic evaluation".  The modeled
attainment tests recommended in Sections 3-6 use models to predict the response of ozone and
PM2 5 to changes in emissions and then applies the resulting relative response factors to observed
(rather than modeled) ambient data.  Thus, while historically, most of the effort has focused on
the operational evaluation, the relative attainment test makes the diagnostic evaluation even more
important.

       In addition to the model performance evaluation, diagnostic analyses are potentially useful
to better understand whether or not the predictions are plausible.  Diagnostic analyses may also be
able to provide: (1) information which helps prioritize efforts to improve and refine model inputs,
(2) insight into which control strategies may be the most effective for meeting the NAAQS, and
(3) an indication  of the "robustness" of a control strategy.  That is, diagnostic tests may help
determine whether the same conclusions would be reached regarding the adequacy of a strategy if
alternative, plausible, assumptions were made in applying the model for the attainment test.

       In this section, we first discuss the general concept of operational and diagnostic
evaluations. We then identify  and discuss issues related to the ambient measurements of ozone
and PM2 5. We then discuss methods which may be useful for evaluating model performance. It
is recommended that performance be assessed by considering a variety of methods. The section
continues by identifying potentially useful diagnostic tests which States/Tribes should consider at
various stages of the modeling analysis to increase the confidence in the model predictions of
future ozone and/or PM2 5 levels. The  section concludes with a discussion on interpreting the
results of the operational and diagnostic evaluations.
       18.1   How Do I Assess Model Performance And Make Use of Operational and
             Diagnostic Evaluations?

       As noted above, model performance can be assessed in one of two broad ways: how
accurately does the model predict observed concentrations for specific cases, and how accurately
does the model predict responses of predicted air quality to changes in inputs (e.g. relative
response factors)?  Given existing data bases, nearly all analyses have addressed the first type of
performance evaluation. The underlying rationale is that if we are able to correctly characterize
changes in concentrations accompanying a variety of meteorological conditions, this gives us
some confidence that we can correctly characterize future concentrations under similar

                                           190

-------
conditions. Typically, this type of operational evaluation is comprised principally of statistical
assessments of model versus observed pairs. Operational evaluations are generally accompanied
by graphical and other qualitative descriptions of the model's ability to replicate historical air
quality patterns.  The robustness of an operational evaluation is directly proportional to the
amount and quality of the ambient data available for comparison.

       The second type of model performance assessment, a diagnostic evaluation, can be made
in several ways.  One way to evaluate the response of the model is to examine predicted and
observed ratios of "indicator species". Indicator species techniques have been developed for both
ozone and secondary PM species (in particular nitrate) (Sillman, 1995 and 1998; Ansari and
Pandis, 1998; Blanchard et al., 2000). If ratios of observed indicator species are very high or very
low, they provide a sense of whether further ozone or secondary PM25 production at the
monitored location is likely to be limited by availability of NOx or VOC (or NH3). Agreement
between paired observed and predicted ratios suggests a model may correctly predict the
sensitivity of ozone or secondary PM25 at the monitored locations to emission control strategies.
Thus, the use of indicator species has the potential to evaluate models in a way which is most
closely related to how they will be used in attainment demonstrations.  A second way for
assessing a model's performance in predicting the sensitivity of ozone or PM2 5 species to changes
in emissions is to perform a retrospective analysis. This involves  comparing model predicted
historical trends with observed trends. Retrospective analyses provide potentially useful means
for diagnosing why a strategy did or did not work as expected.  They also provide an important
opportunity to evaluate model performance in a  way which is closely related to how models are
used to support an attainment demonstration. More types of diagnostic analyses are provided in
Section 18.5.  We recommend that diagnostic analyses be performed during the initial phase of
the model application and during any mid-course review.
       18.2  What Pollutants Are of Interest for Model Performance?

       There are a number of gas phase species that can be used to evaluate both ozone and
secondary PM25. Their presence may affect the response of ozone or secondary PM components
to emissions strategies. Also, the model's performance in predicting certain gaseous species may
provide diagnostic clues which help explain poor performance in predicting ozone or  secondary
PM25.  Gaseous pollutants of interest include:

       - Ozone (O3)
       - Nitric acid (HNO3)
       - Nitric oxide (NO)
       - Nitrogen dioxide (NO2)
       - Peroxyacetyl  nitrate (PAN)
       - Volatile Organic Compounds (VOCs)
       - Ammonia (NH3)
       - NOy (sum of NOx and other oxidized compounds)
       - Sulfur dioxide (SO2)
                                           191

-------
       - Carbon monoxide (CO)
       - Hydrogen peroxide (H2O2)

       Some of the species listed above are currently available from either existing monitoring
network systems and/or special studies. However, it is important to note that many of the species
listed above are not routinely measured, (e.g. H2O2, HNO3, and NH3). Model performance can
best be assessed using extensive data bases, such as those obtained in major field studies.
However, we recognize that such data may not be available for every model application which
needs to be performed.  At a minimum, a State should supplement its performance tests with
available data with a review of results from performance tests for the model(s) which were
completed elsewhere where extensive data bases were used.

       Because PM2 5 is a mixture, a meaningful performance evaluation should include an
assessment of how well the model is able to predict individual  chemical components that
constitute PM25.  Components of PM25 of interest include:

       - Sulfate ion (SO4) and/or Sulfur (S)
       - Nitrate ion (NO3)
       - Ammonium ion (NH4)
       - Elemental Carbon (EC)
       - Organic Carbon (OC) and/or Organic Carbon Mass42
       - Crustal (weighted average of the  most abundant trace elements in ambient air)
       - Mass of "other" primary particulate matter43
       - Mass of individual constituents of inorganic PM
       18.3   What Are the PM2 5 Measurement versus Model Concentration Issues?
       Ambient measurements are needed to perform a model evaluation. Historically, ozone
model evaluations have used routinely measured ambient data (e.g. ozone) collected on a
continuous basis.  The measurements tend to be of high frequency and quality.  Conversely, PM25
and PM2 5 component measurements are (mostly) filter based measurements which have relatively
long sampling times (usually 24 hours) and aren't collected every day (usually once every three or
       42 For predicted/observed comparisons, organic carbon is preferred over organic carbon
 mass to allow for transparency between model predictions and observed measurements. Organic
 carbon mass may include hydrogen, oxygen, and other components.

       43 Other inorganic parti culate matter can contribute up to 10% of PM25 mass in certain
 locations. Unspeciated inorganic PM (sometimes referred to as "other PM") is often used
 interchangeably with "crustal" or "soil" PM.  The term "soil" usually refers to the specific
 definition of soil from the IMPROVE program. "Other primary (inorganic) PM" includes heavy
 metals and other inorganic species in addition to those commonly found in soil.

                                          192

-------
six days).  Speciated data play an especially important role, as the deterministic models predict
exact chemical components which can be compared to some of the corresponding measured
analytes. In addition, there are known positive and negative sampling artifacts as well as analysis
artifacts associated with FRM and speciation monitors.  For example, FRM monitors do not
measure all of the PM2 5 in the air, and the speciation samplers don't measure PM2 5 in the same
way as the FRM monitors (Frank, 2006). Due to the complex nature of the measurements, a basic
understanding of the various PM25 and speciation monitoring technology and measurements are
needed before undertaking a model evaluation.  This section attempts to address the range of
measurements and measurement issues associated with PM2 5 and its components.

       Table  18.1 summarizes many of the PM25 species measurements and networks and lists
sampling time and measurement issues.  This data will commonly be used as part of an
operational evaluation.
Table 18.1 PM2 5 Species Measurements by Monitoring Network and Related Sampling Time
Issues.44
Network
FRM
STN
Measurement
PM2.5
PM2.5
SO4
NO3
NH4
Notes
- Negative artifacts- nitrate, organic
carbon
- Positive artifacts- organic carbon,
water
-STN total measured mass may not be
reliable (cutpoint issues)

- Differences among sampler
measurements across network.
Different sampler types use different
measurement techniques.
- Negative artifacts may occur on the
Nylon filters due to volatilization of
ammonium from the filter because the
alkaline filter may not retain NH3
released in the decomposition of
ammonium nitrate.
Sampling Time
24-hour average 1/3
days or 1/6 days
(some daily sites)
24-hour average 1/3
days or 1/6 days
       44 Most of the issues identified in Table 18.1 are discussed in (Frank, 2006).

                                          193

-------
Network

CASTNet
IMPROVE
Measurement
Organic Carbon
Elemental
Carbon
SO4
NO3
NH4
SO2
HNO3
PM25
SO4 and S
NO3
NH4
Organic Carbon
Elemental
Carbon
Notes
- Blank correction is needed to account
for positive artifact.45
- OC/EC split is operationally defined
using TOT method (NIOSH).
- OC/EC split is operationally defined
using Thermal Optical Transmittance
(TOT) method (NIOSH)

Negative artifact


Split between NO3 and HNO3 may not
be reliable in warm weather due to
conversion from NO3 to HNO3

- Sulfur measurements are usually used
to estimate sulfate

- Only measured at 3 Eastern sites.
- OC/EC split is operationally defined
using Thermal Optical Reflectance
(TOR) analysis.
- Blank corrections are applied to all
data before they are reported.
- OC/EC split is operationally defined
using Thermal Optical Reflectance
(TOR) analysis.
Sampling Time

weekly average
24-hour average 1/3
days
       In general, the speciation data can be directly compared to the model outputs. Of note,
care should be taken in examining issues with blank corrections for organic carbon (especially
when measured OC is low) as well as the EC/OC split.  The STN and IMPROVE networks use
different analysis techniques when estimating EC and OC. This will generally have a larger
impact on EC concentrations.  It may not be appropriate to compare EC concentrations between
networks.  The Clean Air Scientific Advisory Committee (CASAC) has recommended to
transition the STN network to the IMPROVE carbon protocol (TOR) in order to provide data
       45
 ug/m3).
         Network average blank corrections for many of the STN samplers are large (up to 1.53
                                         194

-------
consistency among the two networks (U.S. EPA, 2005f). Over the next two to three years (2007-
2009), EPA plans to convert all of the STN sites (approximately 200 trends and supplemental
sites) to IMPROVE carbon sampling and analysis protocols.

       Due to the previously stated positive and negative artifacts, the FRM data may not be
directly comparable with model outputs without some adjustments.  The models predict PM
species as they might be measured in the ambient air. But the FRM measurements may not
always measure what is in the air (due to known positive and negative artifacts). As part of the
PM2 5 attainment test in Section 5 we recommended default methodologies to adjust speciation
(STN and IMPROVE) data to better approximate the FRM data. A similar adjustment could be
made to the model output data to allow the direct comparison to FRM measurements.  But due to
the relatively limited information to be gained by evaluating daily average total measured PM2 5
mass (as compared to PM2 5 components),  it may not be worth the effort. More meaningful
information may be gleaned from continuous PM2 5 ambient data.  Comparisons between model
output and continuous  data can be made on an hourly basis. This may provide important
information with regards to the daily temporal patterns predicted by the model.

Ambient Measurement Networks

       Provided below is an overview of some of the various ambient air monitoring networks
currently available. Network methods and procedures are subject to change annually due
systematic review and/or updates to the current monitoring network/program.  Please note, there
are other available monitoring networks which are not mentioned here and more details on the
networks and measurements should be obtained from other sources.

AQS  The Air Quality System (AQS) is not an air quality monitoring network.  However it is a
repository of ambient air pollution data and related meteorological data collected by EPA, state,
local and tribal air pollution control agencies from tens of thousands of monitors. The monitoring
data in AQS are the result of the various Clean Air Act requirements to provide a national
database of ambient air pollution data. This information management system contains over 1000
pollutants from 1957 through the present day. AQS contains all the routine hourly gaseous
pollutant data collected from State and Local Air Monitoring Stations (SLAMS) and National Air
Monitoring Stations (NAMS) sites.  SLAMS is a dynamic network of monitors for state and local
directed monitoring objectives (e.g., control strategy development) which consists of
approximately 4,000 monitoring stations. A subset of the SLAMS network, the NAMS has an
emphasis on urban and multi-source areas (i.e, areas of maximum concentrations and high
population density).

       The AQS database includes criteria pollutant data (SO2, NO2, O3, CO, PM10, PM25  and
Pb) and speciation data of paniculate matter (SO4, NO3, NH4, EC,  OC,  and crustal Material), air
toxic data, photochemical assessment data, and meteorological data. The data are measured and
reported on an hourly or daily average basis.  The AQS system continues to expand as regulations
consider to include more ambient air pollutants (U.S. EPA, 2006b).  An overview of the AQS can
be found at http://www.epa.gov/ttn/airs/airsaqs/index.htm.  For querying the database or viewing

                                          195

-------
the data one can use EPA's Air Explorer tool (http://www.epa.gov/mxplorer/aboutthedata.htm)
which is a collection of user-friendly visualization tools for air quality analysts.  The tools
generate maps, graphs, and data tables dynamically.

IMPROVE The Interagency Monitoring of PRQtected Visual Environments (IMPROVE)
network began in 1985 as a cooperative visibility monitoring effort between EPA, federal land
management agencies, and state air agencies (IMPROVE, 2000). Data are collected at Class I
areas across the United States mostly at National Parks, National Wilderness Areas, and other
protected pristine areas.  Currently, there are approximately 160 IMPROVE rural/remote sites
that have complete annual PM2 5 mass and/or PM2 5 species data. Of the total sites, 110 sites are
for regional haze and visibility (and are located at or near Class I areas) and 50 are protocol sites
(determined by the needs of state  and local air pollution control agencies to meet their respective
State Implementation Plan (SIP) requirements). The website to obtain IMPROVE documentation
and/or data is http://vista.cira.colostate.edu/improve/.

CASTNet  Established in 1987, the Clean Air Status and Trends Network (CASTNet) is a dry
deposition monitoring network where data are collected and reported as weekly average data
(U.S. EPA, 2002b). CASTNet provides atmospheric data on the dry deposition component of
total acid deposition, ground-level ozone and other forms of atmospheric pollution. The data are
collected in filter packs that sample the ambient air continuously during the week. CASTNet now
comprises over 70 monitoring stations across the United States.  The longest data records are
primarily at eastern U.S. sites. More information can be obtained through the CASTNet website
at http://www.epa.gov/castnet/.

STN  The Speciation Trends Network (STN) began operation in 1999 to provide nationally
consistent speciated PM25 data for the assessment of trends at representative sites in urban areas
in the U.S.  The STN was established by regulation and is a companion network to the
mass-based Federal Reference Method (FRM) network implemented in support of the PM2 5
NAAQS. As part of a routine monitoring program, the STN quantifies mass concentrations and
PM2 5 constituents, including numerous trace elements, ions (sulfate, nitrate, sodium, potassium,
ammonium), elemental carbon, and organic carbon. There are currently 54 Trends sites
nationally in operation. Trends sites are largely static urban monitoring stations with protocols
for sampling methods; dedicated to characterizing aerosol mass components in urban areas of the
U.S. to discern long-term trends and provide an accountability mechanism to access effectiveness
of control programs.  In addition,  there are approximately 181 supplemental speciation sites
which are part of the STN network and are SLAMS sites. The STN data at trends sites are
collected 1 in every 3 days, whereas supplemental sites collect data either 1 in every 3 days or 1
in every 6 days.  Comprehensive information on the STN and related speciation monitoring can
be found at http://www.epa.gov/ttn/amtic/specgen.html and
http://www.epa.gov/aqspubll/select.html.

NADP Initiated in the late 1970s, the National Acid Deposition Program (NADP) monitoring
network began as a cooperative program between federal and state agencies, universities, electric
utilities, and other industries to determine geographical patterns  and trends in precipitation

                                           196

-------
chemistry in the U.S. NADP collects and reports wet deposition measurements as weekly
average data (NADP, 2002).  The network is now known as NADP/NTN (National Trends
Network) with nearly 200 sites in operation. The NADP analyzes the constituents important in
precipitation chemistry, including those affecting rainfall acidity and those that may have
ecological effects.  The NTN measures sulfate, nitrate, hydrogen ion (measure of acidity),
ammonia, chloride, and base cations (calcium, magnesium, potassium). Detailed information
regarding the NADP/NTN monitoring network can be found at http://nadp.sws.uiuc.edu/.

SEARCH   The South Eastern Aerosol Research and CHaracterization (SEARCH) monitoring
network was established in 1998 and is a coordinated effort between the public and private sector
to characterize the chemical and physical composition as well as the geographical distribution and
long-term trends of PM2 5 in the Southeastern U.S (Edgerton, 2005; Hansen, 2003; Zheng, 2002).
SEARCH data are collected and reported on an hourly/daily basis.  There are currently six
SEARCH measurement sites:  Birmingham, Alabama (urban), Centreville, Alabama (rural),
Gulfport, Mississippi (urban), Jefferson Street, Atlanta, Georgia (urban), Oak Grove, Mississippi
(rural), Yorkville, Georgia (rural), suburban Pensacola, Florida (suburban), and downtown
Pensacola, Florida (urban). Background information regarding standard measurement
techniques/protocols and data retrieval  can be found at
http://www.atmospheric-research.com/studies/SEARCH/index.html.

Supersites   The Supersite (SS) monitoring program began in 2001 as an EPA OAQPS and
ORD co-funded project which measures speciated PM25 at eight air shed sites.  Specifically, these
eight monitoring supersites are located in Baltimore, Maryland; Pittsburgh, Pennsylvania; New
York City, New York; Atlanta, Georgia; Houston, Texas; St. Louis, Missouri; Fresno, California
and Los Angeles, California.  These supersites were developed as special study platforms for
research and measurement methods to address the scientific uncertainties associated with PM2 5.
The goals of the  SS program include (1) fine particulate characterization to understand processes
and source-receptor relationships and to support PM2 5 and ozone SIPs, (2) testing of advanced
sampling methods to enable a smooth transition from established routine techniques, and (3)
development of monitoring data and samples shown in (1) to support health effects and exposure
studies. PM Supersites information can be obtained at
http: //www. epa. gov/ttn/amti c/sup ersites. html.
       18.4   How Should the Operational Evaluation of Performance Be Completed?

       As noted above, an operational evaluation is used to assess how accurately the model
predicts observed concentrations. Therefore, an operational  evaluation can provide a benchmark
for model performance and identify model limitations and uncertainties that require diagnostic
evaluation for further model development/improvement.  Computer graphics, ozone and PM
metrics/statistics, and observational models are all potentially useful for evaluating a model's
ability to predict observed air quality. This section describes the recommended statistical
measures, graphical displays, and other analytical techniques which should be considered as part
                                           197

-------
of an operational evaluation of ozone and PM2 5 model performance. Note that model predictions
from the ramp-up days should be excluded from the analysis of model performance.

             18.4.1  Ozone Operational Evaluation

       Statistics:  It is recommended that, at a minimum, the following three statistical measures
be calculated for hourly ozone and 8-hourly maxima ozone over the episode days in an attainment
demonstration.

•      Mean Normalized Bias (MNB): This performance statistic averages the model/observation
       residual, paired in time, normalized by observation, over all monitor times/locations. A
       value of zero would indicate that the model over predictions and model under predictions
       exactly cancel each other out. The calculation of this measure is shown in Equation 18.1.
   MNB  =  -        - — -   -100%                          (181)
       Mean Normalized Gross Error (MNGE): This performance statistic averages the absolute
       value of the model/observation residual, paired in time, normalized by observation, over
       all monitor times/locations.  A value of zero would indicate that the model exactly
       matches the observed values at all points in space/time. The calculation of this measure is
       shown in Equation 18.2.
    MNGE   =   iff ^ฐ^].100%
                      1
                                Obs
       Average Peak Prediction Bias and Error: These are measures of model performance that
       assesses only the ability of the model to predict daily peak 1-hour and 8-hour ozone.  They
       are calculated essentially the same as the mean normalized bias and error (Equation 18.1
       and 18.2), except that they only consider daily maxima data (predicted versus observed) at
       each monitoring location. In the attainment test, models are used to calculate relative
       response factors near monitoring sites by taking the ratio of the average 8-hour daily
       maximum concentrations calculated for the future and current cases.  Thus, the model's
       ability to predict observed mean 8-hour daily maxima is  an important indicator of model
       performance.

       Appendix A provides other statistics such as mean bias, mean error, mean fractional bias,
mean fractional error, root mean square error, correlation coefficients, etc. which should also be
                                          198

-------
calculated to the extent that they provide meaningful information. Wherever possible, these types
of performance measures should also be calculated for ozone precursors and related gas-phase
oxidants (NOx, NOy, CO, HNO3, H2O2, VOCs and VOC species, etc.) and ozone (and
precursors) aloft.

       Thresholds: EPA recommends that the three metrics above be calculated two ways: 1)
for pairs in which the 1-hour or 8-hour observed concentrations are greater than 60 ppb46, and 2)
for all pairs (no threshold)47.  This will help to focus the evaluation on the models ability to
predict NAAQS-relevant ozone and minimize the effects of the normalization. In terms of pairing
model predictions with monitored observations, EPA recommends that the grid cell value in
which the monitor resides be used for the calculations. It would also be acceptable to consider
bi-linear interpolation of model predictions to specific monitoring locations48. States/Tribes
should recognize that, even in the case of perfect model performance, model-observed residuals
are unlikely to result in exact matches due to differences between the model predictions which are
volume averages and the observations which are point values.

       Averaging:  The statistics should initially be calculated for individual days (averaged
over all sites) and individual sites (averaged over all days). As appropriate, States/Tribes should
then aggregate the raw statistical results into meaningful groups of subregions or subperiods. For
example, examination of model performance within and near the non-attainment area(s) of
interest may be more important than examination of performance in other parts of the modeling
domain. Similarly, priority may be placed on examination of the days that are potentially used in
the attainment test (base period days with 8-hour ozone > 70 or 85 ppb). That is not to say that
model performance evaluations should ignore performance on lower ozone  days or in areas
outside of the nonattainment areas. Simply, additional factors should be considered when
prioritizing more detailed analyses.
        46 Past ozone modeling applications have used a minimum cutoff of either 40 ppb or 60
 ppb. Due to the interest in predicted ozone concentrations at or above the 8-hour standard (85
 ppb), the higher cut off (60 ppb) is recommended.

        47 The use of a 0 ppb threshold can add valuable information about the ability of the
 model to simulate a wide range of conditions. Because of the tendency of the MNB and MNGE
 metrics to inflate the importance of biases at the lowest observed values (which are in the
 denominator), it is recommended that the alternate metrics of normalized mean bias (NMB) and
 normalized mean gross error (NMGE) be used as substitutes for evaluations with no minimum
 threshold.

        48 In certain instances, States/Tribes may also want to conduct performance evaluations
 using the "near the monitor" grid cell arrays. A "near the monitor" analysis may be useful when
 strong ozone gradients are observed, such as in the presence of a sea breeze or in strongly
 oxidant limited conditions. Furthermore, a "near the monitor" performance evaluation is
 consistent with the RRF methodology.

                                           199

-------
       Plots/Graphics: Along with the statistical measures, EPA recommends that the following
four sets of graphical displays be prepared and included as part of the performance analysis.

       Time series plots of model and predicted hourly ozone for each monitoring location in the
       nonattainment area, as well as key sites outside of the nonattainment area. These plots can
       indicate if there are particular times of day or days of the week when the model performs
       especially poorly. Figure 18.1  shows a sample time series plot. This graphic was
       developed using the air quality analysis module of the Atmospheric Model Evaluation
       Tool (AMET) (Appel, 2005), though other evaluation tools exist and can be used.

                  2001   12km Modeled vs. Observed O3 for AQS Ihrmax Site: All Sites
                                                    Sep
                                                                      Oct
                                          Date
                     2001  12km O3 Bias and Error for AQS Ihrmax Site: All Sites
                Jul
                                  Aug
                                                                      Oct
                                          Date
Figure 18.1.  Sample time series analysis plot for a 12-km model vs. observed 2001 case.  The
upper plot shows a particular analysis for 1-hour maxima ozone, July through September
performance averaged over five sites in Cleveland, Ohio. The bottom plot shows the bias and
error (percentage) calculated based on the predicted 1-hour maxima ozone from the upper plot.

•      Scatter plots of predicted and observed ozone at each site within the nonattainment area
       (and/or an appropriate subregion). These plots should be completed using: a) all hours
       within the modeling period for hourly ozone, and b) all 8-hour daily maxima within the
       modeling period.  It may also be useful to develop separate plots for individual time
       periods or key subregions. These plots are useful for indicating if there is a particular part
                                           200

-------
       of the distribution of observations that is poorly represented by the model49. See Figure
       18.2 below for a scatter plot example, which was generated using the AMET.
                           2001 - 12km O3 for 20010701 to 20010731
                           20
40
   60       80
Observation
100
120
Figure 18.2.  Scatter plot analysis for a 12-km model vs. observed 2001 case. The scatter plot
shows a performance analysis for daily 8-hour maxima ozone pairs for the month of July.
       Daily tile plots of predicted ozone across the modeling domain with the actual
       observations as an overlay. Plots should be completed for both daily 1-hour maxima and
       daily 8-hour maxima. These plots can reveal locations where the model performs poorly.
       Superimposing observed hourly or daily maximum concentrations on the predicted
       isopleths reveals useful information on the spatial alignment of predicted and observed
       plumes.
        49 Quantile-quantile (Q-Q) plots may also provide additional information with regards to
 the distribution of the observations vs. predictions.  But due to the fact that Q-Q plots are not
 paired in time, they may not always provide useful information.  Care should be taken in
 interpreting the results.

                                           201

-------
       Animations of predicted hourly ozone concentrations for all episode days or for certain
       periods of interest.  Animations are useful for examining the timing and location of ozone
       formation. Animations may also reveal transport patterns (especially when looking at
       ozone aloft). Animations can also be used to qualitatively compare model outputs with
       the conceptual model of particular ozone episodes.
              18.4.2 PM/RH Operational Evaluation

       An operational evaluation for PM2 5 and regional haze is similar to that for ozone.  Some
important differences are that PM25 consists of many components and is typically measured with
a 24-hour averaging time.  The individual components of PM25 should be evaluated individually.
In fact, it is more important to evaluate the components of PM2 5 than to evaluate total PM25 itself.
Apparent "good performance" for total PM2 5 does not indicate whether modeled PM2 5 is
predicted for "the right reasons" (the proper mix of components).  If performance of the major
components is good, then performance for total PM2 5 should also be good.

       This section contains some additional recommended statistics that have typically not been
calculated for ozone performance, but have been found to be particularly useful for PM analyses
(such as fractional bias and error). We also show examples of some new types of display plots
such as "soccer plots" and "bugle plots". Soccer plots provide a convenient way to display a
summary of model performance (including bias and error at the same time). Bugle plots have
variable bias and error goals, based on ambient concentrations. This allows for a higher
percentage error and bias at very low concentrations. This recognizes the fact that models often
have difficulty in accurately predicting near background concentrations50.

       Similar to ozone evaluations, it may be useful to prioritize examination of model
performance within and near the non-attainment area(s) and/or Class I areas of interest.
Additionally, priority may be placed on examination of the days that are potentially used in the
attainment test (20% best and worst days for visibility and the days  > 65 ug/m3  for 24-hour
PM2,).

       Statistics: Statistics for PM and regional haze performance are similar to those calculated
for ozone. We recommend calculating statistics for components of PM25, and PM precursors.
Useful metrics include mean fractional bias and mean fractional error, normalized mean bias, and
normalized mean error. Formulas for estimating these metrics at an individual monitoring site, j,
are shown in Equations (18.3) -(18.6).  Other statistics such as mean bias, mean error, root mean
square error, correlation coefficients, etc. should also be calculated to the extent that they provide
meaningful information (see  Appendix A for definitions).
        50Predicting very low concentrations is not critical for modeling the PM25 NAAQS.
 However, predicting low concentrations of PM species can be important for regional haze
 modeling, where in the cleanest Class I areas, PM concentrations may be less than a few ug/m3
 (especially on the 20% best days).

                                           202

-------
   Mean Fractional Bias (percent): Normalized bias can become very large when a
   minimum threshold is not used. Fractional bias is used as a substitute. The fractional bias
   for cases with factors of 2 under- and over-prediction are -67 and + 67 percent,
   respectively (as opposed to -50 and +100 percent, when using normalized bias).
   Fractional bias is a useful  indicator because it has the advantage of equally weighting
   positive and negative bias estimates. The single largest disadvantage is that the predicted
   concentration is found in both the numerator and denominator.

   Equation 18.3:
            2  ฃ((Model-Obs)
FBIAS = — Y   7	-4  • 100%
            N^ ((Mode  + Obs)
   Mean Fractional Error (percent):  Normalized error can become very large when a
   minimum threshold is not used. Therefore fractional error is used as a substitute. It is
   similar to the fractional bias except the absolute value of the difference is used so that the
   error is always positive.

   Equation 18.4:
                2  ฃ( \Model-Obs\
 FERROR = —Y   7	—	H   ' 100%
                        ( Model + Obs)
   Normalized Mean Bias (percent): This performance statistic is used as a normalization
   to facilitate a range of concentration magnitudes.  This statistic averages the difference
   (model - observed) over the sum of observed values. Normalized mean bias is a useful
   model performance indicator because it avoids over inflating the observed range of values.
   Equation 18.5:

              N
                ( Model -Obs)

                   N
                  I (Obs)
NMB =	r,	•  100%
                                     203

-------
       Normalized Mean Error (percent): This performance statistic is used as a
       normalization of the mean error to facilitate a range of concentration magnitudes. This
       statistic averages the difference (model - observed) over the sum of observed values.
       Normalized mean error is a useful model performance indicator because it avoid over
       inflating the observed range of values.

       Equation 18.6:
                       | Model - Obs\
          NME = -—^	• 100%
                        I (Obs)
       Averaging Times:  Note that units of time associated with model and observed
concentrations can be days (i.e., usually for paniculate matter and its species), hours (i.e., usually
for species with continuous measurements, like ozone) or sometimes weeks (CASTNet filter pack
measurements). Also note  that the preceding metrics may not be meaningful if the number of
modeled days with monitored data is limited at a site.

       Since modeling for  the PM2 5 NAAQS and regional haze will likely require modeling
different times of year, season-specific statistics and graphic displays are helpful for evaluating
and diagnosing model performance.  Statistics and graphics can be averaged for various time
scales. For example, statistical metrics and scatterplots can show daily averaged ambient-
modeled pairs, monthly averaged pairs, quarterly (or seasonal averaged) pairs, or annual average
pairs. Each of these averaging times can provide useful information. We recommend a range of
different averaging times for annual or seasonal modeling. At a minimum, States should examine
daily averaged pairs and seasonal (or quarterly) averaged pairs.  It should be noted that statistics
and plots tend to look "better" as the averaging time increases from daily to monthly to quarterly
to annual. As such, daily pairs should always be examined to ensure a detailed look at model
performance on the time scale of the FRM and STN measurements (24-hour average).
                                          204

-------
       Plots/Graphics: As mentioned above in the ozone operational evaluation, graphics are a
useful means for understanding how predictions and observations differ. For PM evaluations we
recommend creating time series plots, tile plots, and scatter plots. Time series plots tell whether
there is any particular time of day, day(s) of the week, or months (seasons) of the year when the
model performs poorly.  Tile plots reveal geographic locations where the model performs poorly.
Information from tile plots and time series may provide clues about where to focus quality
assurance efforts for model inputs.  Scatter plots show whether there is any part of the distribution
of observations for which the model performs poorly.  These plots are also useful for helping to
interpret calculations of bias between observations and predictions. For example, they could
show large differences between component PM25 observations and predictions which just happen
to balance, producing low estimated aggregated bias. As mentioned above, since the NAAQS for
PM25 and the regional haze goals will likely require modeling different times  of year, season-
specific graphic displays are helpful for evaluating and diagnosing model performance.

       These above mentioned graphical plots have been provided as  examples in the ozone
operational evaluation section.  Other graphical analysis displays can also be developed to better
inform a model performance  evaluation.  Additional types of graphical plots are shown below.
These plots were generated using the AMET tool.
                                           205

-------
       Figure 18.3 shows an example of spatial plots of normalized mean bias and mean bias by
monitoring site. These plots allow an analysis of regional differences in model performance (e.g.
inland vs. coastal areas).
  _03 HMBJEa> for ryn J3a b313_12km from June !o August for Slate Aฃland_Srte AH    Modeled - Observed SO4 for ryn J3aj3313J^ฃri from Jurte to August for RPO None; Stste: All; ฃ

                                   < -35
                                   J610-35
                                   '610-25
v\

m
1 •
1510 15
261035
>35
(a)
               (b)
                                                     ^—*^^       \  \
                                                     s^ienc Mom^Evaiu.ii^n i^MEIi ProskKt L  \
                                                       square=IMPROVE; circle=STN; t
Figure 18.3.  Example of (a) Spatial plot of NMB; (b) model-observed difference plot (mean
bias).
        Figure 18.4 shows an example of a "soccer plot" and "bugle plots" (Boylan, 2006)
(Tesche, 2006). The soccer plot is so named because the dotted lines resemble a soccer goal.  The
plot is a convenient way to visualize model performance, as measures of both bias and error are
shown on a single plot. As bias and error approach zero, the points are plotted closer to or within
the "goal", represented here by the dashed boxes.  The "bugle plot", named for the shape formed
by the criteria and goal lines, is another plot available for model performance evaluation.  The
bugle plots are shaped as such because the goal and criteria lines are adjusted based on the
average concentration of the observed species. As the average concentration becomes smaller,
the criteria and goal lines become larger to adjust for the model's ability to predict at low
concentrations.  We recommend allowing for larger bias and error when the ambient
concentration falls below 2 ug/m3.
                                             206

-------
(a)
(b)
                                Jls W13 1ปm S04 I,am 20000101 lo 20011231; SUK.MI: SllesAI




g




K



f;



8 -























•
9

c
X
A

fl








'




o
A

	 #



O3
SOI
NQ3 B
PMZ.S
oc
EC
HMO3
TNO3
S02
NO
CO
PTK*
X
IMPROVE
STN
CASTNet








I




E



1



1

                          -75     -SO
                                        >   -15   0    15   30

                                         t*xmalizซa Mean Bias (%)
(c)
              BuqkploL tor JSa bj 13_13un S  Goal
                                       IMPROVE
                                     •i  STN
                                     x  CASTNel
          246

               Anrtgg Ccnonlratrai (ugAn*3)
Figure 18.4.  (a) The "soccer goal" plot measures both bias and error. The "goal" is represented
by the dashed boxes, (b) A  "bugle plot"  of bias and (c) a "bugle plot" of error. As the average
concentration becomes smaller, the criteria and goal lines become larger to adjust for the model's
ability to predict at low concentrations.
                                                   207

-------
       Figure 18.5 shows an example of two types of box plots can be developed for model
performance evaluation.  One displays monthly (left) and one that displays hourly (right)
information.  The box plots show several quantities: the shading represents the 25% to 75%
quartiles, while the lines represent the median values, for both model and observation
concentrations. The monthly box plot can be used to quickly visualize model performance across
the entire year, highlighting the seasonal change in model performance. The hourly or "diurnal"
box plot is used with hourly  data, and shows how the model predictions compare against
observations throughout an entire day.
(a)
           (b)
      J3a 5313 12km SO4 tor Improve from 20000101 to 20011231 : All Sites
                                                           J3a_b313_12km O3 from June to August
       All Sites, 43a b313 12km , Improve
       0~rt*v UM - WX '
IMPROVE
CMAQ
                                               I s
                                               8  ฐ
                                                 8 -I
                                                 o
RPO: None
Slate: All
Site: All
                       678
                       Months
                               9  10 11  12
                        6  8  10  12 14 16 18  20  22
                               Hours
Figure 18.5. The box plots show several quantities: the shading represents the 25% to 75%
quartiles, while the lines represent the median values, for both model and observation
concentrations, (a) displays monthly data (left) and (b) displays hourly data (right).
       18.5   How Do I Assess Model Performance and What Types of Analyses Can be
              Done to Evaluate the Accuracy of the Model Response: Diagnostic
              Evaluations?

       This section lists possible analyses that could be performed to investigate the ability of the
model to accurately forecast changes in predicted concentrations of ozone and PM resulting from
changes in ozone and PM2 5 precursor emissions. Diagnostic evaluations/tests are used to explain
model performance and to provide clues about how to improve the reliability of predictions.
                                            208

-------
       The diagnostic analyses discussed in this section include observational models, probing
tools, receptor models, retrospective analyses, and sensitivity analyses. This represents a large
variety of tools that can aid in diagnostic analyses.  Some of these techniques are based on
analysis of ambient data, some require running the photochemical many times, and some of the
techniques involve coding of the photochemical model to allow additional outputs. States/Tribes
are encouraged to complete as many of these types of diagnostic analyses as possible throughout
the modeling process in order to help understand model performance and to help develop
information which may help improve performance, as well as increase confidence in the modeled
attainment projections.

              18.5.1 Observational models

       A performance evaluation which includes comparisons between modeled and observed
ratios of indicator species carries with it a large potential advantage.  For ozone, measurements of
certain "indicator species ratios" are a potentially useful way to assess whether local ozone
formation is VOC- or NOx-limited at any particular point in space and time as well as, help reveal
whether the model is correctly predicting the sensitivity of ozone to VOC and/or NOx controls
(e.g., comparisons between modeled and observed ratios:  O3/NOy, O3/HNO3) (Sillman, 1995,
1997, 1998, and 2002; Lu and Chang, 1998; Pun and  Seigneur, 1999). For PM, such a
comparison may reveal whether the model is predicting sensitivity of secondary components of
PM25 to changes in SO2, NH3, VOC and/or NOx controls  correctly (Ansari and Pandis,  1998;
Blanchard et al.,  2000; Pun and Seigneur, 2001). If a model accurately predicts observed ratios of
indicator species, then one can conclude with additional confidence that the predicted change in
future year ozone or PM may be accurate. One precaution with respect to the use of indicator
species is that there may be a range of observed ratios for which the preferred direction of control
is not clear.  When this occurs, agreement between predictions and observations does not
necessarily imply that the response to controls, as predicted by the model is correct (especially for
secondary particulate matter to changes in precursors).  If a model  predicts observed ratios of
indicator species such that observed and predicted ratios fall within the same range of ratios, this
provides some confidence that the predicted change in particulate matter may be accurate.  A
second precaution is that application of this method often requires more measurements than are
commonly made. In some cases, it may be difficult to achieve the  required precision with current
routine monitoring. Finally, much of the work done to date with indicator species has focused on
peak hourly concentrations of ozone. Despite these precautions, comparing predicted and
observed ratios of indicator species provides a means of assessing a model's ability to accurately
characterize the sensitivity of predicted ozone and predicted secondary components of PM25 to
changes in precursors.

       Many studies and applications of indicator species approaches have addressed ozone-
related problems. Sillman (1995), Sillman  (1998) and Lu and Chang (1998) provide good
descriptions of the method, identify key ratios and illustrate application of the approach. Even
though the preceding ratios are oriented toward ozone, they can provide insight into why modeled
concentrations of secondary particulate matter are sensitive to changes in VOC, or NOx emissions
(Pun and Seigneur, 1999). Ansari and Pandis (1998)  have developed an indicator ratio of species
                                           209

-------
and applied it to several combinations of secondary paniculate matter present under different
environmental conditions.  They use this ratio to predict how mass of particulate matter will
respond to reductions in sulfate, nitrate and ammonia.  Blanchard et al., (2000) have also
examined how indicator species might be used to assess whether particulate nitrate concentrations
are limited by NOx or by ammonia emissions using mechanisms which incorporate reactions
dealing with secondary particulate matter. These authors identify two ratios of indicator species
which appear potentially useful for identifying limiting precursors for secondary nitrate
particulate matter: (1) the ratio of particulate ammonium plus gas-phase ammonia over the sum of
nitric acid plus particulate nitrate plus particulate sulfate, and (2) the ratio of particulate to total
nitrate.  It is likely that additional indicator species approaches will be identified as the user
community continues to gain more experience with chemical mechanisms incorporating
secondary particulate formation and more speciated particulate and gas phase ambient
measurements become available.

       Other observational methodologies exist and can be used in a similar manner. The Smog
Production (SP) algorithm is another means by which ambient data can be used to assess areas
that are NOx or VOC-limited (Blanchard, 1999).  Additionally, it has been  postulated that
differences in weekend-weekday ozone and PM2 5 patterns may also provide real-world
information on which precursors are most responsible for ozone and PM2 5 formation in any given
area (Pun, 2001 and 2003; Heuss, 2003; Fujita, 2003; Blanchard and Tanenbaum, 2003 and 2005;
Chinkin, 2003; Yarwood, 2003; Qin, 2004).  For example, activity levels and patterns, leading to
PM2 5 and precursor emissions from mobile, area and some point sources, may differ on weekends
vs. week days. In areas where there are large differences between average weekend and weekday
ambient ozone and PM2 5 concentrations over the span of several seasons, it would  be useful to
compare statistical model performance for weekends versus weekdays (e.g., does the model
accurately reflect observed differences in component concentrations in summer vs. winter?).  This
would allow one to assess whether the model is capturing the effect of the emissions differences
which are presumably driving the real-world concentration differences.

              18.5.2 Probing Tools

       Recently, techniques have been developed to embed procedures within the code of an air
quality model which enable users to assess the contributions of specific source categories or of
specific geographic regions to predicted model concentrations at specified sites (e.g. for ozone
predictions, Zhang, 2003).  Various techniques have been implemented into various air quality
models, but three of the most commonly used probing tools axe photochemical source
apportionment (Dunker, 2002a; ENVIRON, 2006a; Tonnesen and Wang, 2004; SAI, 2005;
Yarwood, 2005; Douglas, 2006), the decoupleddirect method (DDM) (Dunker, 1980 1981, and
1984; Yang, 1997a and 1997b; Dunker, 2002b; ENVIRON, 2006a;  Cohan,  2002, 2004, and 2005;
Hakami, 2003 and 2004), process analysis (Jeffries, 1994 and 1997; Jeffries, 1996; Jang,  1995;
Lo, 1997; Byun and Ching, 1999; Morris, 2001 and 2003; ENVIRON, 2006a; Henderson, 2006).
In the context of model performance evaluation, these attribution procedures are useful in that
they allow one to "track" the importance of various emissions categories or phenomena
contributing to predicted ozone and particulate matter at a given location.  This can provide


                                          210

-------
valuable insight into whether the model is adequately representing the conceptual description of
ozone and PM patterns in the nonattainment area. In cases where model performance is subpar,
these analyses can be useful for indicating where model input or model algorithm improvements
are most needed.

       Caution should be applied to the tools and techniques mentioned here as they are limited
by the accuracy of the underlying air quality photochemical model, when the tools utilize the
same algorithms/modules and assumptions. An overview of these aforementioned probing tools
are provided below.

       Photochemical source apportionment:  A photochemical source apportionment tool tags
and tracks the release, transport, chemical transformation, and deposition of precursor species
from primary emission sources, source categories, source regions, initial conditions and/or
boundary conditions from a photochemical grid model. Therefore, the contribution of tagged
sources to simulated concentrations (including secondary pollutants) and deposition can be
measured in a single model run. This source apportionment tool can be used to estimate how
emissions from individual source areas and regions affect modeled ozone and PM concentrations
over space and time.  For example, this is achieved by using multiple tracer species to track the
fate of ozone precursor emissions (VOC and NOx) and the ozone formation caused by these
emissions within an air quality model simulation.  There are challenges to consider during
implementation of this methodology, i.e., tracking the  spatial and temporal relationships between
separate groups of emission sources and ozone and paniculate matter formation (Douglass, 2006),
(Environ, 2006a), (Environ, 2006b).  Among them are insuring compatibility with the underlying
air quality model formulation  so that derived source-receptor relationships will be consistent with
model response to emission changes.

       The information gleaned from source apportionment can be useful in considering potential
control strategies. However, a key consideration in weighing alternative strategies is which types
of emissions are cost effective and practical to control, a consideration not addressed by source
apportionment modeling.

       Decoupled Direct Method (DDM):  The DDM calculates the local, first-order sensitivity
coefficients of gas-phase pollutant concentrations to perturbations of model inputs,  e.g. initial
conditions, boundary conditions, emissions.  Theoretically, the DDM is applicable to the
calculation of higher order sensitivity coefficients as well, but these are much more numerous
than the first order coefficients and generally will be less useful (Environ, 2006a). DDM can also
be used to explore model sensitivities to perturbations  in meteorological parameters and chemical
rate constants, although resource intensive and at present somewhat less useful. The decoupled
direct method is advantageous both in terms of computational efficiency and stability of the
numerical solution. DDM has been used as a sensitivity analysis tool in several studies. A single
model run using DDM can potentially provide the information equivalent to tens, if not hundreds
of brute force emissions  reductions model runs. DDM has been  incorporated into several models
including CAMx (Environ, 2006a) and CMAQ (Cohan, 2005).
                                           211

-------
       Process analysis: Process analysis techniques allows one to quantitatively track (mass
balancing) the contributions of individual physical and chemical atmospheric processes to the
overall change in pollutant concentrations at a particular time and location. Thus, process
analysis will assess the relative importance of each physical and chemical process for a particular
model simulation as well as explain how model predictions are obtained, which is valuable when
modeling nonlinear systems such as atmospheric photochemistry.  Since models used to simulate
ozone and secondary particulate matter are similar, process analysis should also be useful for
addressing both ozone and PM2 5 (and visibility related) applications.

       Three of the most common process analysis tools implemented in grid models are the
integrated process rate (IPR) analysis, the integrated reaction rate (IRR) analysis, and the
chemical process analysis (CPA).  The IPR analysis is used to determine the relative contributions
of individual physical (e.g., advection, diffusion, emissions) and chemical (e.g., chemical
reaction, aqueous chemistry, aerosol production) processes, potentially revealing unexpected low
or high process contributions. The IRR analysis is used to identify and explain the details of the
chemical transformations (chemical pathways and key chemical characteristics) in a particular
model simulation. Typically, IRR analyses have been used to determine and understand model
prediction differences between chemical mechanisms (e.g., CB-IV, SAPRC). Likewise, the CPA
is an improvement on the IRR analysis technique where  part of the processing of IRR information
is internalized within the air quality grid model to yield key chemical parameters directly (e.g.,
budget terms for ozone and NOx).

       Process analysis requires a substantial amount of expertise to be interpreted to full
advantage.  However, useful insights are also possible with less detailed analyses. An analysis
can focus on selected grid cells in a small area. Process analysis then takes advantage of the fact
that a numerical grid model addresses physical and chemical factors affecting secondary
pollutants in a sequential manner.  For example, a typical sequence followed in a model for each
time step (e.g., 1 hour) might be (1) advection of PM25 components and precursors present at the
beginning of the time step, (2) PM25 and precursor emissions added during the time step, (3)
vertical diffusion of the advected material and fresh emissions, (4) estimated cloud cover and its
effects on photolysis  rates, (5) atmospheric chemistry involving advected and diffused material
with fresh emissions, and (6) deposition of certain compounds. Process analysis examines
incremental effects on changes in component and/or PM2 5 predictions from hour to hour
attributable to each of the processes described above. In this way, one gets a sense of how
important each process is as a contributor to predicted air quality at a specified time and location.

       If a focused diagnostic analysis, such as one obtained with process analysis, suggests a
particular model prediction may be an artifact of a model assumption rather than a result of real
chemical/physical atmospheric processes, States may wish to go back to the meteorological or
emissions model to verify that the inputs and assumptions that have been used are correct. If a
prediction is the result of an apparent artifact which cannot be resolved, States may wish to
discount that prediction in the attainment demonstration  or uniform rate of progress assessment.
                                           212

-------
              18.5.3 Receptor Models (Multi-variate Source Apportionment)

       Receptor models (also referred to as multi-variate source apportionment models) are
mathematical or statistical procedures for identifying and quantifying the sources of air pollutants
at a receptor location (Henry, 1991; Hopke, 1991, 1999, 2001; Lewis, 2003; Coulter, 2000;
Willis, 2000; Watson, 2002; Maykut, 2003; Mukerjee, 2004). Receptor models are generally
used to examine primary pollutants such as the primary components of PM25.  They are not
typically used to examine ozone or secondary PM2 5 species. Unlike photochemical and
dispersion air quality models, receptor models do not use pollutant emissions, meteorological data
and chemical transformation mechanisms to estimate the contribution of sources to receptor
concentrations. Instead, receptor models use the chemical and physical characteristics of gases
and particles measured at source and receptor to both identify the presence of and to quantify
source contributions to receptor concentrations. Receptor models can identify possible air quality
management strategy solutions especially when used in conjunction with examination of the local
emission inventory (Brook, 2004). These models are therefore a natural complement to other air
quality models and can be used as part of State Implementation Plans (SIPs) for identifying
sources contributing to air quality problems.

       Receptor models provide scientific support for current ambient air quality standards and
for implementation of those standards by identifying and quantifying contributions of various
source types. As mentioned in sections 18.2 and 18.3, the richness of the ambient air quality data
sets has been increasing, due to more species being measured, species being stratified by particle
size, shorter durations of sampling, and measurements not only at the surface but also aloft. To
take advantage of these richer data sets the receptor models have become more complex. To
ensure that receptor modeling tools, both simple and complex, are available for use in the
development and implementation of air quality standards, the United States Environmental
Protection Agency's Office of Research and Development (ORD) has and continues to develop a
suite of receptor models (multivariate statistical techniques) that are freely distributed to the air
quality management community.  The EPA has developed the Chemical Mass Balance (CMB)
and UNMIX models as well as the Positive Matrix Factorization (PMF) method for use in air
quality management. CMB fully apportions receptor concentrations to chemically  distinct
source-types depending upon the source profile database, while  UNMIX and PMF internally
generate source profiles from the ambient data. Details of CMB, UNMIX and PMF are described
below.

       The Chemical Mass Balance (CMB) Model (most recent version, EPA-CMBv8.2) is one
of several receptor models that has been applied to air quality problems over the last two decades
(U.S. EPA, 2004d and 2004e; Seigneur, 1997, Coulter, 2000). Based on an effective-variance
least squares method (EVLS), EPA has supported CMB as a regulatory planning tool through its
approval of numerous State Implementation Plans (SIPs) which have a source  apportionment
component.  The chemical mass balance model is probably the most directly applicable
observational approach  for this purpose, since it can focus on the same day(s) considered with the
air quality model. Cautions raised previously about representativeness of the monitored data
continue to apply. CMB requires speciated profiles of potentially contributing sources and the
                                          213

-------
corresponding ambient data from analyzed samples collected at a single receptor site. CMB is
ideal for localized nonattainment problems and has proven to be a useful tool in applications
where steady-state Gaussian plume models are inappropriate, as well as for confirming or
adjusting emissions inventories.

       UNMIX is named for its function, which is to "unmix" the concentrations of chemical
species measured in the ambient air to identify the contributing sources. The particular
mathematical approach used by UNMIX is based on a form of Factor Analysis, but its novelty is
that physically-meaningful constraints are imposed which are intended to remove the undesirable
ambiguity of the multiple solutions that are characteristic of ordinary Factor Analysis. For a
given selection of species, UNMIX estimates the number of sources, the source compositions, and
source contributions to each sample.  Chemical profiles of the sources are not required, but
instead are generated from the ambient data.

       The PMF technique  is a form of factor analysis where the underlying co-variability of
many variables (e.g.,  sample to sample variation in PM species) is described by a smaller set of
factors (e.g., PM sources) to which the original variables are related. The structure of PMF
permits maximum use of available data and better treatment of missing and below-detection-limit
values.  Also available is a document which discusses the PMF methodology:  A Guide to
Positive Matrix Factorization (PDF).

              18.5.4  Retrospective Analyses

       A retrospective analysis is intended to examine the ability of the model to respond to
emissions changes by comparing recent trends in observed ozone or PM2 5 concentrations to the
model-predicted trend over the same time period.  The approach is a direct assessment of what we
are most interested in—does the model accurately predict  changes in air quality as a result of
changes in emissions? As part of this analysis the model is run for current episodes or time
periods and episodes  in one or more historical time periods using the emissions and
meteorological inputs appropriate for each time  period modeled. While retrospective analyses
may be useful, it may be difficult to obtain meteorological and emissions inputs for the historical
time period(s) that are calculated using techniques and assumptions which are consistent with the
calculation of these same inputs for the current time period. Using inconsistent inputs will
confound the interpretation  of the predicted trend.

       Because differences in meteorology between years can confound the apparent change in
pollutants, EPA recommends that, in most cases, retrospective analyses use constant meteorology
and backcasted emissions. The model should respond in a predictable way if the emissions
changes are large enough. Hence, backcasting an emissions change of only a few percent would
not be viable.  However, if NOx or SO2 emissions have been reduced by a large percentage (e.g.
30% or more), then that may be a good test of the response of the model.  In Section 8, we noted
that a retrospective analysis can be a useful tool for diagnosing why an area has not attained the
NAAQS.  Therefore,  retrospective analyses are one of the few tools that can be used to determine
                                           214

-------
if the model is responding "adequately" to control measures. Retrospective analyses may be an
important tool to show that the models are responding "correctly" to emissions changes.

              18.5.5  Alternative Base Cases (Sensitivity tests)

       In some cases it may be useful to evaluate how the response of the model to emissions
reductions varies as a function of alternative model inputs or model algorithms. Outcomes of
these types of sensitivity tests are useful for several purposes.  First, the tests can be used to see
whether model performance is especially sensitive to a particular input or combination of inputs
(Arnold,  2003; Gilliland, 2006; Morris, 2006).  Second, sensitivity tests may help prioritize
additional data gathering efforts so that a subsequent review/diagnosis can be performed at the
time of a mid-course review or required attainment date.  Third, the tests can be used to assess
the robustness of a control strategy. As an example, States/Tribes could consider the effects of
assumed  boundary conditions on predicted effectiveness of a control strategy.  If the model
response does not differ greatly over a variety of alternative plausible  configurations, this
increases confidence in the model results.

       The parameters for sensitivity tests can include, but are not limited to: different chemical
mechanisms,  finer or coarser grid resolution, meteorological inputs from alternative, credible
meteorological model(s), different initial/boundary conditions, and multiple sets of reasonable
emission projections.  Sensitivity tests can and should be applied throughout the modeling
process, not just when model performance is being evaluated.  In cases where the operational
model performance is  considered to be poor, these tests may help indicate where base case
input/algorithm changes are warranted.

       A new technique has been developed to more efficiently examine a large range of
emissions control scenarios.  Response Surface Modeling (RSM), has been developed by utilizing
advanced statistical techniques to characterize the relationship between model outputs and input
parameters in a highly economical manner (U.S. EPA, 2006c).  The RSM is a metamodel  of a
model (i.e., air quality model); it is a reduced-form prediction model using statistical correlation
structures to approximate model functions through the design of complex multi-dimension
experiments.

       The RSM is based on a new approach known as air quality metamodeling that aggregates
numerous pre-specified air quality modeling simulations into a multi-dimensional air quality
"response surface". Simply, this metamodeling technique is a "model of the model" and can be
shown to reproduce the results from an individual modeling simulation with little bias or error
(U.S. EPA, 2006c, Hubbell, 2005).  The RSM is based on statistical relationships between model
inputs and outputs to provide real-time estimate of these air quality changes. The RSM provides
a wide breadth of model outputs, which can be used to develop emissions control scenarios. The
RSM approach can inform the selection and evaluation of various control scenarios. This
approach allows for the rapid  assessment of air quality impacts and changes of different
combinations of emissions reductions. While the RSM may not provide a complete picture of all
changes necessary to reach various alternative standards nationwide, it is highly useful in  the
                                           215

-------
context of providing illustrative control scenarios for selected areas, and understanding the
contribution of different source categories, source regions and pollutant emissions to air quality
across the U.S. The RSM can be used in a variety of ways: (1) strategy design and assessment
(e.g. comparison of urban vs. regional controls; comparison across sectors; comparison across
pollutants); (2) optimization (develop optimal combinations of controls to attain standards at
minimum cost); (3) model sensitivity (systematically evaluate the relative sensitivity of modeled
ozone and PM levels to changes in emissions inputs.
       18.6   How Should the Results of the Model Evaluation be Assessed (Interpreted)?

       In EPA guidance for the 1-hour ozone attainment demonstrations (U.S. EPA, 199la),
several statistical goals were identified for operational model performance. These goals were
identified by assessing past modeling applications of ozone models and determining common
ranges of bias, error, and accuracy (Tesche et al., 1990). The 1-hour guidance noted that because
of differences in the quality of the applications considered, it was inappropriate to establish "rigid
criterion for model acceptance or rejection" (i.e., no pass/fail test). It was recommended that
these ranges should be used in conjunction with the additional qualitative procedures to assess
overall model performance.51

       With the additional experience of more than a decade of photochemical modeling, it is
clear that there is no single definitive test for evaluating model performance.  All  of the tests
identified in Sections 18.2 and 18.3 have strengths and weaknesses.  Further, even within a single
performance test, it is  not appropriate to assign "bright line" criteria that distinguish between
adequate and inadequate model performance. In this regard, EPA recommends that a "weight of
evidence" approach (like that described in Section 7) be used to determine whether  a particular
modeling application is valid for assessing the future attainment status of an area. EPA
recommends that States/Tribes undertake a variety of performance tests and weigh them
qualitatively to assess model performance.  Provided suitable data bases are available, greater
weight should be given to those tests which assess the model capabilities most closely related to
how the model is used in the modeled attainment test. Generally, additional confidence should be
attributed to model base case applications in which a variety of the tests described above are
applied and the results indicate that the model is performing well.  From an operational
standpoint, EPA recommends that States/Tribes compare their evaluation results against similar
modeling exercises to ensure that the model performance approximates the quality of other
applications.  To aid in this comparison, we have summarized performance metrics  (including
information on models and notes on inputs) of recent ozone, PM25, and regional haze applications
in Appendix B.
        51 In practice, however, most 1-hour ozone modeling applications using the 1991
 guidance tended to focus almost entirely on meeting the three statistical "goals" for bias, error,
 and accuracy at the expense of more diagnostic evaluation.

                                           216

-------
       In summary, we continue to believe that for ozone, PM25, and regional haze; 1) there
should be no bright line performance criteria; 2) the evaluation of statistical performance
measures should be compared to current and past modeling applications; and 3) statistical
measures should be used in conjunction with diagnostic tests and other qualitative analyses to
assess overall model performance.
                                           217

-------
REFERENCES

Abt Associates, (2003), "Environmental Benefits Mapping and Analysis Program (BenMAP),
       User's Manual: http://www.epa.gov/ttn/ecas/benmodels.html.

ABT Associates, (2007), "Model Attainment Test Software (MATS) User's Manual".

Air Quality Management Work Group, (2005), "Recommendations to the Clean Air Advisory
       Committee: Air Quality Management Work Group, Phase 1 and Next Steps", available at
       http ://www. epa.gov/air/caaac/pdfs/report 1-17-05.pdf

Ansari, A.S., and S.N. Pandis, (1998), "Response of Inorganic PM to Precursor Concentrations",
       Environ. Sci.  Technol, 32, 2706-2714.

Appel, W., S. Howard, R. Gilliam, and S. Phillips, (2005), "The Atmospheric Model Evaluation
       Tool: Air Quality Module",  4th Annual CMAS Models-3 Users' Conference, September
       26-28, 2005, Chapel Hill, NC.

Arnold, J.R., R.L. Dennis, and G.S. Tonnesen,  (2003), "Diagnostic Evaluation of Numerical Air
       Quality Models with Specialized Ambient Observations; Testing the Community
       Multiscale Air Quality Modeling System (CMAQ) at Selected SOS 95 Ground Sites",
       Atmospheric Environment, 37, 1185-1198.

Arunachalam, S., Z. Adelman, R. Mathur, and J. M. Vukovich, (2002), "The Impacts of Biogenic
       Emissions Estimates from BEIS-3 on Ozone Modeling in the Southeastern U.S.", 11th
       Annual Emissions Inventory Conference of the U.S. EPA",
       http://www.epa.gov/ttn/chief/conference/eill/index.html

Arunachalam, S., A. Holland, B. Do, and M. Abraczinskas,  (2006), A Quantitative Assessment of
       the Influence of Grid Resolution on Predictions of Future-Year Air Quality In North
       Carolina, USA, Atmospheric Environment, 40 (26), 5010-5026.

Atmospheric and Environmental Research, Inc., (2002), "Evaluation of Probing Tools
       Implemented in CAMX", CRC Project  A-37-1, Document Number CP099-1-2, August,
       http://www.crcao.com/reports/recentstudiesOO-02/A-37-l%20Final%20Report.pdf

Baker, K., (2004a), "Midwest Regional Planning Organization Modeling Protocol", Lake
       Michigan Air Directors Consortium / Midwest Regional Planning Organization, Des
       Plaines IL. http://www.ladco.org/tech/photo/docs/RPOModelingProtocolV2_dec2004.pdf

Baker K., (2004b), "Meteorological Modeling Protocol For Application to
       PM2s/Haze/OzoneModeling Projects", Lake Michigan Air Directors Consortium /
       Midwest Regional Planning Organization, Des Plaines IL.
                                         218

-------
Battelle, 2004, "Characterizing Ozone Episodes to Inform Regulatory Modeling", Technical
       Report, U.S. EPA Contract No. 68-D-02-061, Work Assignment 2-07

Biswas, J., and S.T. Rao, (2001), "Uncertainties in Episodic Ozone Modeling Stemming from
       Uncertainties in the Meteorological Fields", J. Applied Met.., 40, 117-136

Blanchard, C.L., and P.M. Roth, (1997), "User's Guide Ozone M.A.P.P.E.R., Measurement-based
       Analysis of Preferences in Planned Emission Reductions, Version 1.1", Final Report
       prepared for the U.S. EPA pursuant to Contract 68-D6-0064 Work Assignment,
       W.M.Cox, EPA Work Assignment Manager.

Blanchard, C.L., F.W. Lurmann, P.M. Roth, H.E. Jeffries and M. Korc, (1999), "The Use of
       Ambient Data to Corroborate Analyses of Ozone Control Strategies", Atmospheric
       Environment, 33, 369-381.

Blanchard, C. L., (2000), "Ozone process insights from field experiments- Part III: Extent of
       reaction and ozone formation", Atmospheric Environment,  34, 2035-2043.

Blanchard, C. L., and T. Stoeckenius, (2001), "Ozone response to precursor controls: comparison
       of data analysis methods with the predictions of photochemical air quality simulation
       models", Atmospheric Environment, 35, 1203-1216.

Blanchard, C.L., and S.t. Tanenbaum, (2003), "Differences  between Weekday and Weekend Air
       Pollutant Levels  in South California", J. Air & Waste Management Association, 53, 816-
       828.

Blanchard, C.L., and S.T. Tanenbaum,  (2005), "Weekday/Weekend Differences in Ambient
       Ozone and Particulate Matter Concentrations in Atlanta and the Southeastern United
       States", Envair, Final Report, CRC Project No. A-47, pp. 78.

Boylan, J.W. and Russell, A.G, (2006), "PM and Light Extinction  Model Performance Metrics,
       Goals, and Criteria for Three-Dimensional Air Quality Models", Atmos. Environ., 40,
       4946-4959.

Brook, J., E. Vega, and J. Watson,  (2004), "Receptor Methods" Chapter 7, In: Particulate Matter
       Science for Policy Makers,  A NARSTO Assessment,  P.H. McMurry, M.F. Shepherd, J.S.
       Vickery, ed., Cambridge University Press, Cambridge, UK, 235-281.

Byun, D.W., and J.K.S.  Ching, (1999), "Science Algorithms of the EPA Models-3 Community
       Multiscale Air Quality (CMAQ) Modeling System," Office of Research and
       Development, USEPA, EPA/600/R-99/030.
                                         219

-------
CARB, (1996), "Performance Evaluation of SAQM in Central California and Attainment
      Demonstration for the August 3-6, 1990 Ozone Episode", California Air Resources Board,
      Sacramento, CA.

Cabada, J.C.,  S. N. Pandis, R. Subramanian., A. L. Robinson, A. Polidori, and B. Turpin, (2004),
      "Estimating the Secondary Organic Aerosol Contribution to PM2 5 Using the EC Tracer
      Method", Aerosol Science and Technology, 38, 140-155.

Cardelino, C.A., and W.L. Chameides, (1995), "An Observation Based Model for Analyzing
      Ozone Precursor Relationships in the Urban Atmosphere", J. Air & Waste Mgmt. Assoc.,
      45,    161-181.

Chang, J.S., S. Jin, Y. Li, M. Beauharnois, C.L. Lu, and H. Huang, (1997), "SARMAP Air
      Quality Model", prepared by Atmospheric  Sciences Research Center, State University of
      New York (Albany).

Chinkin, L.R., D.L. Coe, T.H. Funk, H.R. Hafner, P.T. Roberts, P. A. Ryan, and D.R. Lawson,
      (2003), "Weekday versus Weekend Activity Patterns for Ozone Precursor Emissions in
      California's South Coast Air Basin", J. Air & Waste Management Association, 53, 829-
      843.

Chow, J. C., J. G. Watson, L.-W. A. Chen, W.P. Arnott, and H. Moosmuller, (2004),
      "Equivalence of Elemental Carbon by Thermal/Optical Reflectance and Transmittance
      with Different Temperature Protocols", Environmental Science and Technology, 38(16),
      4414-4422.

Chu, S.-H., (2005), "Stable estimate of primary OC/EC ratios in the EC tracer method",
      Atmospheric Environment, 39, 1383-1392.

Clegg, S.L., P. Brimblecombe, and A. S. Wexler, (1998), "Aerosol Inorganics Model; A
      Thermodynamic Model of the System H+ - NH4+ - SO42- - NO3- - H2O at Tropospheric
      Temperatures.",/. Phys. Chem., 102A, 2137-2154.
      http://www.aim.env.uea.ac.uk/aim/project.htm

Clinton, W.J., (July 16, 1997), Subject: "Implementation of Revised Air Quality Standards for
      Ozone and Particulate Matter",  Memorandum to the Administrator of the Environmental
      Protection Agency.

Croes, B.E., L.J. Dolislager, L.C. Larsen, and J.N.  Pitts, (2003), "The O3 'Weekend Effect' and
      NOx Control Strategies—Scientific and Public Health Findings and Their Regulatory
      Implications"; EM, July 2003, 27-35.
                                         220

-------
Cohan, D.S., (2002), "Implementation of a Direct Sensitivity Method into CMAQ", Paper 5.3,
       Presented at 1st CMAS Models-3 User's Conference - Models-3 Performance and
       Applications, RTF, NC, (October 2002).

Cohan, D.S., (2004), "Applicability of CMAQ-DDM to Source Apportionment and Control
       Strategy Development", Paper 5.3, Presented at 3rd Annual CMAS Models-3 Users'
       Conference, RTF, NC, (October 2004).

Cohan, D. S., A. Hakami, Y.T. Hu, and A.G. Russell, (2005), "Nonlinear Response of Ozone to
       Emissions: Source Apportionment and Sensitivity Analysis." Environ. Sci. Technol.  39,
       6379-6748.

Cohan, D.S., Y. Hu, A.G. Russell, (2006), "Dependence of ozone sensitivity analysis on grid
       resolution", Atmospheric Environment, 40, 126-135.

Collett, Jeffrey, (2004), Technical note on IMPROVE study.

Coulter, T., R.A. Wagoner, and C.W. Lewis, (2000), "Chemical Mass Balance Software: EPA-
       CMB8.2", Submitted for inclusion in Proceedings of the AWMA/EPA Symposium,
       Measurement of Toxic and Related Air Pollutants, September 12-14, 2000, Research
       Triangle Park, NC.

Cox, W.M., and S. Chu, (1993), "Meteorologically Adjusted Ozone Trends in Urban Areas: A
       Probabilistic Approach", Atmospheric Environment, 27B, (4), 425-434.

Cox, W.M., and S. Chu, (1996), "Assessment of Interannual Ozone Variation in Urban Areas
       from a Climatological Perspective", Atmospheric Environment, 30, 2615-2625.

Dennis, R.L., (1994), "Using the Regional Acid Deposition Model to Determine the Nitrogen
       Deposition Airshed of the Chesapeake Bay Watershed", Ch.21, Atmospheric Deposition
       of Contaminants to the Great Lakes and Coastal Waters, I.E. Baker (ed.), SET AC Press,
       Pensacola, FL.

Deuel, H.P., and S.G. Douglas, (1998), "Episode Selection for the Integrated Analysis of Ozone,
       Visibility and Acid Deposition for the Southern Appalachian Mountains", Draft Technical
       Report Systems Applications International, Inc. (SYSAPP-98/07), prepared for the
       Southern Appalachians Mountains Initiative.

Dolwick, P.O., C. Jang, and B. Timin, (1999), "Effects of Vertical Grid Resolution on
       Photochemical Grid Modeling", Paper 99-309, Presented at 99th AWMA Meeting,
       St.Louis, MO, (June 1999).

Dolwick, P.O., (2002), "USEPA/OAQPS Meteorological Modeling Analyses", presented at the
       2002 Ad Hoc Meteorological Modeling Meeting, Des Plaines IL.
                                         221

-------
Doty. K.G., T.W. Tesche, D.E. McNally, B. Timin, and S. F. Mueller, (2001), "Meteorological
       Modeling for the Southern Appalachian Mountains Initiative (SAMI)", Final Report,
       Southern Appalachian Mountain Initiative. Asheville, NC.

Douglas, S.G., R.C. Kessler, and E.L. Carr, (1990), "User's Guide for the Urban Airshed Model,
       Volume  III: User's Manual for the Diagnostic Wind Model", EPA-450/4-90-007C, U.S.
       EPA, Research Triangle Park, NC 27711, (NTIS No.: PB91-131243).

Douglas, S., T. Myers, and Y. Wei, (2006). "Implementation of Sulfur and Nitrogen Tagging in
       the Community Multi-scale Air Quality (CMAQ) Model." Prepared for the U.S.
       Environmental Protection Agency, Office of Air Quality Planning and Standards,
       Research     Triangle Park, North Carolina. Prepared by ICF International, San Rafael,
       California (06-076).

Dunker, A.M., (1980), "The response of an atmospheric reaction-transport model to changes in
       input functions", Atmospheric Environment, 14, 671-679.

Dunker, A.M., (1981), "Efficient calculations of sensitivity coefficients for complex atmospheric
       models", Atmospheric Environment, 15, 1155-1161.

Dunker, A.M., (1984), "The Decoupled Direct Method For Calculating  Sensitivity Coefficients In
       Chemical-Kinetics",,/. Chem. Phys. 81, 2385-2393.

Dunker, A.M., G. Yarwood, J.P. Ortmann, and G.M. Wilson, (2002a), "Comparison of source
       apportionment and source sensitivity of ozone in a three-dimensional air quality model",
       Environ. Sci.  Techno!., 36, 2953-2964.

Dunker, A.M., G. Yarwood, J.P. Ortmann, and G.M. Wilson, (2002b), "The decoupled direct
       method for sensitivity analysis in a three-dimensional air quality model-implementation,
       accuracy, and efficiency", Environ. Sci. TechnoL, 36, 2965-2976.

Eberly, S., (2005), "EPAPMF 1.1 User's Guide", June 30, 2005. USEPA, National Exposure
       Research Laboratory, http://www.epa.gov/heasd/products/pmf/pmf.htm .

Eder, B.K., J.M. Davis, and P. Bloomfield, (1993a), "A Characterization of the Spatiotemporal
       Variability of Non-urban ozone concentrations over the Eastern  United States",
       Atmospheric Environment, 27A, No. 16, 2645-2668.

Eder, B.K., J.M. Davis, and P. Bloomfield, (1993b), "An automated Classification Scheme
       Designed to Better Elucidate the Dependence of Ozone on Meteorology", J. Applied Met.,
       33, 1182-1198.

Edgerton E.S., B.E. Hartsell, R.D. Saylor, J.J. Jansen, D.A. Hansen, and G.M. Hidy, (2005), "The
       Southeastern Aerosol Research and Characterization Study: Part II. Filter-based


                                          222

-------
       measurements of fine and coarse paniculate matter mass and composition", J. Air Waste
       Mange. Assoc., 55, 1527-1542.

Emery, C., and E. Tai, (2001), "Enhanced Meteorological Modeling and Performance Evaluation
       for Two Texas Ozone Episodes ", prepared for the Texas Near Non-Attainment Areas
       through the Alamo Area Council of Governments", by ENVIRON International Corp,
       Novato, CA., http://www.tnrcc.state.tx.us/air/aqp/airquality_contracts.html#met01

ENRIRON International Corporation, (2004), "Annual PM Modeling and Analysis of Episode
       Aggregation Uncertainties for Annual Average PM2 5 and Visibility Modeling", CRC
       Project A-44a/b, http://www.crcao.com/.

ENVIRON/Alpine/UCR, (2005), "White Paper- Visibility Projection Approaches- Preliminary
       Approaches for VISTAS 2018 Modeling".

ENVIRON International Corporation, (2006a), "User's Guide to the Comprehensive Air Quality
       Model with Extensions (CAMx) version 4.30", Novato, CA. Http://www.camx.com .

ENVIRON International Corporation, (2006b), "Guidance for the Application of the CAMx
       Hybrid Photochemical Grid Model to Assess Visibility Impacts of Texas BART Sources
       at Class I Areas", Novato, CA. Prepared for Texas Commission on Environmental
       Quality (TCEQ)
       (http://www.tceq. state.tx.us/assets/public/implementation/air/sip/bart/BART_TX_CAMx_
       PSAT_Guid.pdf).

Frank, N., 2006:  "Retained Nitrate, Hydrated Sulfates, and Carbonaceous Mass in Federal
       Reference Method Fine Particulate Matter for Six Eastern U.S. Cities" J. Air Waste
       Mange. Assoc., 56, 500-511.

Fiore, A.M., D.J. Jacob, H. Liu, R.M. Yantosca, T.D. Fairlie,  and Q. Li,(2003),  "Variability in
       surface ozone background over the United States: Implications for air quality policy", J.
       Geophys.  Res.,  108, 4787, 2003.

Flaum, J. B., S.T. Rao, and I.G. Zurbenko, (1996), "Moderating the influence of
       meteorological conditions on ambient ozone concentrations", Journal of the Air and
       Waste Management Association 46, 35-46.

Gao, D., W.R. Stockwell and J.B. Milford, (1996), "Global Uncertainty Analysis of a Regional
       Scale Gas-Phase Chemical Mechanism", J. Geophys. Res., 101, 9107-9119.

Gego, E., H. Hogreffe, G. Kallos, A. Voudouri, J.  Irwin, and S. T. Rao, (2005),  "Examination of
       Model Predictions at Different Horizontal  Grid Resolutions", Environmental Fluid
       Mechanics, 5, 63-85.
                                          223

-------
Geron, C., A. Guenther and T. Pierce, (1994), "An Improved Model for Estimating Emissions of
       Volatile Organic Compounds from Forests in the Eastern United States", J. Geophys. Res.,
             99, 12,773-12,791.

Gilliam R., (2005), The Atmospheric Model Evaluation Tool: Meteorology Module, 4th Annual
       CMAS Models-3 Users' Conference, September 26-28, 2005, Chapel Hill, NC.

Gilliland, A., K.W. Appel, R.W. Finder, and R. Dennis, (2006),  "Seasonal NH3 Emissions for the
       Continental United States: Inverse Model Estimation and Evaluation", Atmospheric
       Environment, 40 (26), 4986-4998.

Grell, G.A., J. Dudhia and D.R. Stauffer, (1994), "A Description of the Fifth-Generation Penn
       State/NCAR Mesoscale Model (MM5)", NCAR/TN-398+STR, 138 pp.

Guenther, A., C. Geron, T. Pierce, B. Lamb, P. Harley, and R. Fall, (2000), "Natural emissions of
       non-methane volatile organic compounds, carbon monoxide, and oxides of nitrogen from
       North America", Atmospheric Environment., 34, 2205-2230.

 Hakami, A., Odman M. T., and Russell, A. G., (2003), "High-order, direct sensitivity analysis of
       multidimensional air quality models" Environ. Sci. Technol, 37, 2442-2452.

Hakami, A.,  Odman, M. T., and Russell, A. G., (2004), "Nonlinearity in atmospheric response: A
       direct sensitivity analysis  approach",. J. Geophys. Res., 109, D15303.

Hand, J.L. and W.C. Malm, (2006), "Review of the IMPROVE Equation for Estimating Ambient
       Light Extinction Coefficients- Final Report", March 2006,
       http://vista.cira. colostate.edu/improve/Publicati ons/GrayLit/016_IMPROVEeqReview/IM
       PROVEeqReview.htm

Haney, J.L.,  and S.G. Douglas, (1996), "Analysis of the Effects  of Grid Resolution on UAM-V
       Simulation Results for the Atlanta Nonattainment Area", Systems Applications
       International Technical Memorandum SYSAPP-96/68, prepared for Southeast Modeling
       Center participants, October 21, 1996.

Hansen, D.A., E.S. Edgerton, B.E. Hartsell, JJJansen, N. Kandasamy, G.M. Hidy, and C.L.
       Blanchard, (2003), "The Southeastern Aerosol Research and Characterization Study: Part
       1 - Overview", J. Air Waste Mange. Assoc., 53, 1460-1471.

Henderson, B., W. Vizuete, and H. Jeffries, (2006), "Development of a Lagrangian pyPA
       (Process Analysis in python) tool", Presented at the 5th Annual CMAS Conference,
       October 16-18, 2006, http://www.cmascenter.org/conference/2006/agenda.cfm

Henry, R.C., (1991), "Multivariate Receptor Models", In: Receptor Modeling for Air Quality
       Management, P.K. Hopke, ed., Elsevier Science Publishers, Amsterdam, 117-147.
                                         224

-------
Hering, S. and G. Cass, (1999),  "The Magnitude of Bias in Measurement of PM25 Arising from
       Volatilization of Paniculate Nitrate from Teflon Filters." JAWMA 49:725-733.

Hogrefe, C., S.T. Rao, I.G. Zurbenko, and P.S. Porter, (2000), "Interpreting the information in
       time series of ozone observations and model predictions relevant to regulatory policies in
       the eastern United States", Bull. Amer. Met. Soc., 81, 2083-2106.

Holland, D.M., W.M. Cox, R. Scheffe, AJ. Cimorelli, D. Nychka, and P.K. Hopke, (2003),
       "Spatial Prediction of Air Quality Data", EM,  August 2003, 31-35.

Hopke, P.K., ed. (1991), "Receptor Modeling for Air Quality Management", Elsevier Science,
       Amsterdam.

Hopke, P.K., ed. (1999), "An Introduction to Source Receptor Modeling", In: Elemental Analysis
       of Airborne Particles, Landsberger, S., Creatchman, M., Eds. Gordon and Breach Science,
       Amsterdam, 273-315.

Hopke, P.K., ed. (2001), "Advances in Receptor Modeling", Journal of Aerosol Science, 32, 363-
       368.

Houyoux, M.R., J.M. Vukovich, CJ. Coats Jr., N.W.  Wheeler, and P.S. Kasibhatla, (2000),
       "Emission Inventory Development and Processing for the Seasonal Model for Regional
       Air Quality (SMRAQ) project." Journal of Geophysical Research, Atmospheres
       105:07:9079-9090.

Hu, Y., D. Cohen, M.T. Odman, and T. Russell, (2004), "Fall Line Air Quality Study- Draft Final
       Report- Air Quality Modeling of the August 11-20, 2000 Air Pollution Episode."

Huang J-P., J. Fung, Y. Zhang, A. Lau, R. Kwok, and J. Lo, 2005: Improvement of Air Quality
       Modeling in Hong Kong by Using MM5 Coupled with LSM,  4th Annual CMAS
       Models-3 Users' Conference, September 26-28, 2005, Chapel Hill, NC.

Hubbell, B.J., P.O. Dolwick, D. Mooney, and M. Morara, (2005), "Evaluating the relative
       effectiveness of ozone precursor controls: Design of computer experiments applied to the
       comprehensive air quality model with extensions (CAMx)", Air and Waste Management
       Association Conference Proceedings

IMPROVE, (2000), "Spatial and Seasonal Patterns and Temporal Variability of Haze and its
       Constituents in the United States: Report III." Cooperative Institute for Research in the
       Atmosphere, ISSN: 0737-5352-47. IMPROVE Web Site:
       http://vista.cira.colostate.edu/improve.
                                          225

-------
IMPROVE, (2006), "Revised IMPROVE Algorithm for Estimating Light Extinction from Particle
       Speciation Data", January 2006,
       http://vista.cira.colostate.edu/improve/Publications/GrayLit/gray _literature.htm

Isakov, V., J.S. Irwin, and J. Ching, (2007), Using CMAQ for exposure modeling and
       characterizing the sub-grid variability for exposure estimates", Journal of Applied
       Meteorology and Climatology, (in press).

Jacob, D.J., J. A. Logan, and P.P. Murti, (1999), "Effect of Rising Asian Emissions on Surface
       Ozone in the United States", Geophy. Res. Lett., 26, 2175-2178.

Jaffe D., McKendry I, Anderson T., and Price H.,(2003), "Six 'new' episodes of trans-Pacific
       transport of air pollutants", Atmos. Envir. 37, 391-404.

Jang,  J.C., H.E. Jeffries, and S. Tonnesen, (1995), "Sensitivity of Ozone Model to Grid
       Resolution Part II: Detailed Process Analysis for Ozone Chemistry", Atmospheric
       Environment,  29 (21), 3101-3114.

Jeffries, H.E., (1994), "Process Analysis for UAM Episode 287", memo to Brock Nicholson, NC
       DEHNR, April 8, 1994, ftp://airsite.unc.edu/pdfs/ese_unc/ieffries/projprpt/uammod.
       panel287.pdf.

Jeffries, H.E., T. Keating, and Z. Wang, (1996), "Integrated Process Rate Analysis of the Impact
       of Nox Emission Height on UAM-modeled Peak Ozone Levels", Final  Topical Report,
       Prepared for Gas Research Institute, Chicago, IL,
       ftp://airsite.unc.edu/pdfs/ese_unc/jeffries/reports/gri. uncgri.pdf.

Jeffries, H.E., (1997), "Use of Integrated Process Rate Analyses to Perform Source Attribution
       for Primary and Secondary Pollutants in Eulerian Air Quality Models", Presented at U.S.
       EPA Source Attribution Workshop, Research Triangle Park, NC, July 16-18, 1997,
       ftp://airsite.unc.edu/pdfs/ese_unc/jeffries/ipradocs. sourcewksp.pdf, and
       http://www.epa.gov/ttn/faca/stissu.htmL Source Att WS-Proc.analysis/mass track--
       Jeffries.

Johnson, M., (2003),  "Iowa DNR 2002 Annual MM5 Modeling Project", presented at the 2003
       Ad Hoc Meteorological Modeling Meeting, Des Plaines IL.

Jones, J. M., C. Hogrefe, R. Henry, J. Ku, and G. Sistla, (2005), "An Assessment of the
       Sensitivity and Reliability of the Relative response factor Approach in the Development of
       8-hr Ozone Attainment Plans", JAWMA, 55 (1), 13-19.

Kenski, D., (2004), "Analysis of Historic Ozone Episodes using CART",
       http://www.ladco.org/tech/monitoring/docs_gifs/CART%20Analysis%20of%20Historic%
       20Ozone%20Episodes.pdf

                                           226

-------
Koerber, M. and D. Kenski, (2005), "The Case for Using Weight of Evidence Demonstrations in
       State Implementation Planning"; EM, April 2005, 24-28.

Ku, I, H. Mao, K. Zhang, K. Civerolo, S.T. Rao, C.R. Philbrick, B. Doddridge and R. Clark,
       (2001), "Numerical Investigation of the Effects of Boundary-Layer Evolution on the
       Predictions of Ozone and the Efficacy of Emission Control Options in the Northeastern
       United States/Tribes", Environmental Fluid Mechanics, 1, 209-233

 Kumar, N., M.T.Odman, and A.G. Russell, (1994), "Multiscale Air Quality Modeling:
       Application to Southern California",./. Geophysical Research, 99, 5385-5397.

Kumar, N., and A.G. Russell, (1996), "Multiscale Air Quality Modeling of the Northeastern
       United States", Atmospheric Environment, 30, 1099-1116.

LADCO, (1999), Personal communication.

Lefohn, A.S.,  D.S. Shadwick, and S.D. Ziman, (1998), "The Difficult Challenge of Attaining
       EPA's New Ozone Standard", Environmental Science and Technology, Policy Analysis,
       32, No. 11, (June 1, 1998), 276A-282A.

Lehman, J., K. Swinton, S. Bortnick, C. Hamilton, E. Baldridge, B. Eder and B. Cox, (2004),
       "Spatial-Temporal characterization of Tropospheric Ozone Across the Eastern United
       States'", Atmospheric Environment, 38:26, 4357-4369, www.sciencedirect.com

Lewis, C.W.,  G.A. Norris, R.C. Henry, and T.L. Conner, (2003), "Source Apportionment of
       Phoenix PM-2.5 Aerosol with the Receptor Model", Journal of the Air & Waste
       Management Association, 53, 325-338.

Lo, C.S., and  H.E. Jeffries, (1997), "A Quantitative Technique for Assessing Pollutant Source
       Location and Process Composition in Photochemical Grid Models", presented at annual
       AWMA Meeting, Toronto, Ontario (1997),
       ftp://airsite.unc.edu/pdfs/ese_unc/j effries/ipradocs. PCA_awma.pdf.

Lu, C.H., and J.S. Chang, (1998), "On the  Indicator-Based Approach to Assess Ozone
       Sensitivities and Emissions Features, J. Geophysical Research, 103, 3453-3462.

Lyons, W.A.,  C.J. Tremback, and R.A. Pielke, (1995), "Applications of the Regional
       Atmospheric Modeling System (RAMS) to Provide Input to Photochemical Grid Models
       for the Lake Michigan Ozone Study (LMOS)", J. Applied Meteorology, 34, 1762-1786.

Maykut, Lewtas,  Kim, and Larson, (2003), "Source Apportionment of PM25 at an Urban
       IMPROVE Site in Seattle, Washington", Environmental Science and Technology, 37,
       5135-5142. (This paper uses UNMIX and PMF  )


                                          227

-------
Meyer, E.L., K.W. Baldridge, S. Chu, and W.M. Cox, (1997), "Choice of Episodes to Model:
       Considering Effects of Control Strategies on Ranked Severity of Prospective Episode
       Days", Paper 97-MP112.01, Presented at 97th Annual AWMA Meeting, Toronto,
       Ontario.

MCNC, (1999), MAQSIP air quality model, http://www.iceis.mcnc.org/products/maqsip/.

McNally, D.E., (2002),  "A Comparison of MM5 Model Estimates for February and July 2001
       Using Alternative Model Physics", report prepared for the United States Environmental
       Protection Agency, prepared by Alpine Geophysics, LLC, Arvada CO.

McQueen, I, P. Lee, M. Tsidulko, G. DiMego, R. Mathur, T. Otte, J. Pleim, G. Pouliot, D.Kang,
       K. Schere, J. Gorline, M. Schenk, and P.  Davidson, (2004), "Update of The Eta-CMAQ
       Forecast Model run at NCEP operations and its performance for the Summer 2004",
       presented at the 2004 CMAS Workshop,  Chapel Hill, NC.

Milanchus, M. L., S.T. Rao, and I.G. Zurbenko,  (1998), "Evaluating the effectiveness of
       ozone management efforts in the presence of meteorological variability", Journal of
       the Air and Waste Management Association 48, 201-215.

Morris R.E., G. Yarwood, C.A. Emery, and G.M. Wilson, (2001), "Recent Advances in CAMx
       Air Quality Modeling", Presented at the 2001 Air and Waste Management Association
       Annual Conference, http://www.camx.com/publ/pdfs/camx934_AWMA_2001.pdf

Morris R.E., G. Yarwood, C.A. Emery, and G.M. Wilson, (2003), "Recent Advances in
       Photochemical Air Quality Modeling Using the CAMx Model: Current Update and Ozone
       Modeling of Point Source Impacts", Presented at the 2003 Air and Waste Management
       Association Annual  Conference,
       www. camx.com/publ/pdfs/camx_advances_AWMA_2002_rev3.pdf

Morris, R.E., B. Koo, A.Guenther, G. Yarwood,  D. McNally, T.W. Teshe, G. Tonnesen, J.
       Boylan, and P. Brewer, (2006), "Model Sensitivity Evaluation for Organic Carbon using
       Two Multi-pollutant Air Quality Models that Simulate Regional Haze in the Southeastern
       United States", Atmospheric Environment, 40 (26), 4960-4972.

Mukerjee, S., G.A. Norris, L.A. Smith, C.A Noble, L.M. Neas, H.A. Ozkaynak, andM. Gonzales,
       (2004), "Receptor Model Comparisons and Wind Direction Analyses of Volatile Organic
       Compounds and Submicrometer Particles in an Arid, Binational, Urban Air Shed",
       Enviromental Science and Technology, 38(8), 2317-2327.

Na, K. S., A.A. Sawant, C. Song, and D.R. Cocker, (2004), "Primary and secondary carbonaceous
       species in the atmosphere of Western Riverside County, California", Atmospheric
       Environment, 38, 1345-1355.
                                         228

-------
NADP, (2002), National Acid Deposition Program 2002 Annual Summary.

Nielsen-Gammon, John W., (2002), "Evaluation and Comparison of Preliminary Meteorological
      Modeling for the August 2000 Houston-Galveston Ozone Episode (Feb. 5, 2002 interim
      report)", http://www.met.tamu.edu/results/.

Odman, M.T., and C.L. Ingram, (1996), "Multiscale Air Quality Simulation Platform (MAQSIP):
       Source Code Documentation and Validation", MCNC Technical Report ENV-96TR002-
      vl.O, 83pp.

Olerud, D., K. Alapaty, and N. Wheeler, (2000), "Meteorological Modeling of 1996 for the
      United States with MM5", MCNC-Environmental Programs, Research Triangle Park, NC.

Olerud, D., and A. Sims, (2003), "MM5 Sensitivity modeling in support of VISTAS", prepared
      for Mike Abraczinskas of the VISTAS Technical Analysis Workgroup by Baron
      Advanced Meteorological Systems, LLC, Raleigh NC.

Otte, T., (2004), "What's New in MCIP2", presented at the 2004 CMAS Workshop, Chapel Hill,
      NC.

Pacific Environmental  Services, Inc., (1997), "Draft Technical Memorandum Analysis of 8-hour
      Ozone Values:  1980-1995", prepared under EPA Contract No.68D30032, Work
      Assignment 111-88, Edwin L. Meyer, Jr., Work assignment manager.

Pielke, R.A., W.R. Cotton, R.L. Walko, C.J. Tremback, W.A. Lyons, L.D.  Grasso, M.E.
      Nicholls, M.D.  Moran, D.A. Wesley, T.J. Lee, and J.H. Copeland, (1992), "A
      Comprehensive Meteorological Modeling System - RAMS", Meteor. Atmos. Phys., 49,
      69-91.

Pierce, T., C. Geron, G. Pouliot, J. Vukovich, and E. Kinnee, (2004), "Integration of the Biogenic
      Emissions Inventory System (BEIS3) into the Community Multiscale Air Quality
      (CMAQ) Modeling System", 12th Joint Conference on the Application  of Air Pollutant
      Meteorology with the Air and Waste Management Association,
      http://ams.confex.com/ams/AFMAPUE/12AirPoll/abstracts/37962.htm.

Pitchford, M.L., and W.C. Malm, (1994), "Development and Applications  of a Standard Visual
      Index", Atmos. Environ., 28(5), 1049-1054.

Poirot, Wishinski, Hopke, and Polissar, (2001), "Comparative Application  of Multiple Receptor
      Methods to Identify Aerosol Sources in Northern Vermont", Environmental Science and
       Technology, 35, 4622-4636 (This paper uses Unmix, PMF, and Ensemble back
      trajectories (CAPITAs Monte  Carlo, PSCF, and RTA)).
                                         229

-------
Pun, B.K., and C. Seigneur, (1999), "Understanding Particulate Matter Formation in the
       California San Joaquin Valley: Conceptual Model and Data Needs", Atmospheric
       Environment, 33, 4865-4875.

Pun, B.K., and C. Seigneur, (2001), "Sensitivity of Particulate Matter Nitrate Formation to
       Precursor Emissions in the California San Joaquin Valley", Environ. Sci. Technol., 35
       (14), 2979-2987.

Pun, B.K, C. Seigneur, and W. White, (2001), "Data Analysis for a Better Understanding of the
       Weekday /Weekend Ozone and PM Differences", Final report prepared by Atmospheric
       and Environmental Research, Inc., San Ramon, CA, under CRC contract Number A-36B
       for the Coordinating Research Council, Alpharetta, GA, June 2001.

Pun, B.K., C. Seigneur, and W. White, (2003), "Day-of-Week Behavior of Atmospheric Ozone in
       Three U.S. Cities", J. Air & Waste Management Association, 53, 789-801.

Qin, Y., G.S. Tonnesen, and Z. Wang, (2004), "Weekend/Weekday Differences of Ozone, Nox,
       CO, VOCs, PMio and the Light Scatter during Ozone Season in Southern California",
       Atmospheric Environment, 38 (19), 3069-3087.

Rao, S. T., E. Zalewsky, E., and I.G. Zurbenko, (1995), "Determining temporal and spatial
       variations in ozone air quality", Journal of the Air and Waste Management
       Association 45, 57-61.

Reynolds,  S.D., H.M. Michaels, P.M. Roth, T.W. Tesche, D. McNally, L. Gardner and G.
       Yarwood, (1996), "Alternative Base Cases in Photochemical Modeling: Their
       Construction, Role and Value", Atmospheric Environment, 30, 1977-1988 .

Roth, P.M., S.D. Reynolds, and T.W. Tesche, (2005), "Air Quality Modeling and Decisions for
       Ozone Reduction Strategies", Journal of the Air and Waste Management Association, 55,
       1558-1573.

Russell, A., and R. Dennis, (2000), "NARSTO critical review of photochemical models and
       modeling", Atmospheric Environment, 34, 2283-2324.

Saylor, R. D., E. S. Edgerton, and B. E. Hartsell, (2006), "Linear regression techniques for use in
       the EC tracer method of secondary organic aerosol estimation", Atmospheric
       Environment, 40, 7546-7556.

Saxena, P. and Hildemann, L, (1996), "Water Soluble Organics in Atmospheric Particles: A
       Critical Review of the Literature and Application of thermodynamics to Identify
       Candidate
       Compounds", J. Atmos. Chem., 24:57-109.
                                         230

-------
Scire, J.S., RJ. Yamartino, G.R. Carmichael, and Y.S. Chang, (1989), "CALGRID: AMesoscale
      Photochemical Grid Model", Volume II: User's Guide, Sigma Research Corp., Concord,
      MA.

Scire, J.S., F.R. Francoise, M.E. Fernau, and RJ. Yamartino, (1998), "A User's Guide for the
      CALMET Meteorological Model (Version 5.0)", EarthTech., Inc., Concord, MA.

Scire, J.S., D.G. Strimaitis, RJ. Yamartino, (2000), "A User's Guide for the CALPUFF
      Dispersion Model", Jan 2000, http://www.src.com/calpuff/calpuffl.htm.

Seaman, N.L., (1997), "Use of Model-Generated Wind fields to Estimate Potential Relative Mass
      Contributions from Different Locations", Prepared for U.S. EPA Source Attribution
      Workshop, Research Triangle Park, NC, July 16-18, 1997,
      http://www.epa.gov/ttn/faca/stissu.htmL Source Attribution Workshop Materials.

Seaman, N.L., (2000), "Meteorological modeling for air quality assessments", Atmospheric
      Environment, 34, 2231-2259.

Seigneur, C., P. Pai, J. Louis, P. Hopke, and D. Grosjean, (1997), "Review of Air Quality Models
      for Particulate Matter", Document Number CPO15-97-la , prepared for American
      Petroleum Institute.

Seigneur, C., P. Pai, P.K. Hopke and D. Grosjean, (1999), "Modeling Atmospheric Particulate
      Matter", Environmental Science and Technology/News, Feb.l, 1999, pp.80-86.

Sillman, S., (1995), "The Use of NOy, H2O2, and HNO3 as Indicators for O3-NOx-ROG
      Sensitivity in Urban Locations", J.  Geophys. Res. 100, 14,175-14,188.

Sillman, S., D. He, C. Cardelino, and R.E.  Imhoff, (1997), "The Use of Photochemical Indicators
      to Evaluate Ozone-NOx-Hydrocarbon Sensitivity:  Case Studies from Atlanta, New York
      and Los Angeles",  J. Air and Waste Mgmt. Assoc. , 47 (10), 1030-1040. (Oct. 1997)

Sillman, S., (1998), "Evaluating the Relation Between Ozone, NOX and Hydrocarbons: The
      Method of Photochemical Indicators", EPA/600R-98/022,
      http://www-personal.engin.umich.edu/~sillman/publications.htm.

Sillman, S., and D. He, (2002), "Some theoretical results concerning O3-NOX-VOC chemistry and
      NOx-VOC indicators", J. Geophys. Res., 107, D22, 4659, doi: 10.1029/2001JDOO1123,
      2002, http://www-personal.engin.umich.edu/~sillman/publications.htm.

Sisler, J.F. (Editor), (1996), "Spatial and Seasonal Patterns and Long Term Variability of the
      Composition  of the Haze in the United States: An Analysis of Data from the IMPROVE
      Network",  Cooperative Institute for Research in the Atmosphere Report ISSN: 0737-
      5352-  32,  Colorado State University, Fort Collins, CO.


                                          231

-------
Sistla, G., C. Hogrefe, W. Hao, J. Ku, R. Henry, E. Zalewsky, and K. Civerolo, (2004), "An
       Operational Assessment of the Application of the Relative response factors in the
       Demonstration of Attainment of the 8-Hour Ozone National Ambient Air Quality
       Standard", JAWMA, 54 (8), 950-959.

Skamarock, W. C., J. B. Klemp, J. Dudhia, D. O. Gill, D. M. Barker, W. Wang and J. G. Powers,
       (2005), A Description of the Advanced Research WRF Version 2,
       http://www.wrf-model.org/wrfadmin/docs/arw_v2.pdf.

Smith,  James H.(1998),  "Hierarchical Emissions Inventory Development for Nested
       Photochemical Modeling, Part I: Conceptual Overview", Proceedings of the Air & Waste
       Management Association  Specialty Conference, Emissions Inventory: Living in a Global
       Environment, New Orleans, La., December 1998.

Stehr, J., (2004), "How's the Air Up There? A Guide to Mid-Atlantic Air Quality." Presentaion
       at MANE-VU/MARAMA Science Meeting, January 27, 2004.

Strader, R., F. Lurmann, and S.N. Pandis, (1999), "Evaluation of secondary organic aerosol
       formation in winter", Atmospheric Environment, 33, 4849-4863.

Strum,  M., G. Gipson, W. Benjey, R., M. Houyoux, C.  Seppanen, and G. Stella, (2003), "The Use
       of SMOKE to Process Multipollutant Inventories - Integration of Hazardous Air Pollutant
       Emissions with Volatile Organic Compound Emissions," 12th International Emissions
       Inventory Conference, Session 10,
       http ://www. epa. gov/ttn/chief/conference/ei 12/index.html.

Systems Applications International, (1996), "Users Guide to the Variable Grid Urban Airshed
       Model (UAMV)", SYSAPP 96-95/27r. http://uamv.saintl.com/.

Systems Applications Internation, Inc. (SAI), (2005), "Overview of the REMSAD Modeling
       System", November 15, 2005, see website: http://remsad.saintl.com/overview.htm

Tang, I, (1996), "Chemical and Size Effects of Hygroscopic Aerosols on Light Scattering
       Coefficients", J. Geophysical Research 101 No.D14, pp. 19,245-19,250.

Tesche, T.W., P. Georgopoulos, F.L. Lurmann and P.M. Roth, 1990, "Improvement of Procedures
       for Evaluating Photochemical Models", Draft Final Report, California Air Resources
       Board,
       Sacramento, CA.

Tesche, T. W., and D. E. McNally, (2001), "Evaluation of the MM5, RAMS, and SAIMM
       Meteorological Model for the 6-11 September 1993  COAST and 25 August-1 September
       2000 TexAQS2000 Ozone SIP Modeling Episodes", report prepared for the Business
                                         232

-------
       Coalition for Clean Air-Appeals Group, prepared by Alpine Geophysics, LLC, Ft. Wright,
       KY.

Tesche T. W., D.E. McNally, and C. Tremback, (2002), "Operational evaluation of the MM5
       meteorological model over the continental United States: Protocol for annual and episodic
       evaluation." Submitted to USEPA as part of Task Order 4TCG-68027015.  (July 2002)

Tesche T.W., R. Morris, G. Tonnesen, D. McNally, J. Boylan, and P. Brewer, (2006),
       CMAQ/CAMx annual 2002 performance evaluation over the eastern US", Atmospheric
       Environment, 40, 4906-4919.

Timin, B., (2005a), "Draft Final Ozone Guidance Comments and Proposed Changes",  presented
       at the 3rd PM/RH/O3 Modeling Workshop, New Orleans, LA, pages 25-27.
       http://www.cleanairinfo.com/modelingworkshop/presentati ons/O3_Guidance_Timin.pdf.

Timin, B., (2005b), "Draft Final Ozone Guidance Comments and Proposed Changes", presented
       at the 3rd PM/RH/O3 Modeling Workshop, New Orleans, LA, pages 8-22.
       http://www.cleanairinfo.com/modelingworkshop/presentati ons/O3_Guidance_Timin.pdf.

Tonnesen, G., and B. Wang, (2004), "CMAQ Tagged Species Source Apportionment", July 22,
       2004; http://www.wrapair.org/forums/aoh/meetings/040722/UCR_tssa_tracer_v2.ppt

Turpin, B.J., and J.J Huntzicker, (1995),." Identification of secondary organic aerosol episodes
       and quantitation of primary and secondary organic aerosol concentrations during
       SCAQS", Atmospheric Environment, 29, 3527-3544.

Turpin, B. J. and H-J Lim, (2001), "Species Contributions to PM25 Mass Concentrations:
       Revisiting Common Assumptions for Estimating Organic Mass." Aerosol Science and
       Technology 35: 602-610.

U.S. EPA, (199 la), "Guideline for Regulatory Application of the Urban Airshed Model", EPA-
       450/4-91-013, http://www.epa.gov/ttn/scram/guidance_sip.htm.

U.S. EPA, (1991b), "Procedures for Preparing Emissions Projections", EPA-450/4-91-014.

U.S. EPA, (1993), "Volatile Organic Compound (VOC)/Particulate Matter (PM) Speciation Data
       System (SPECIATE), Version 1.5", EPA/C-93-013,
       www.epa.gov/ttn/chief/software.htmltfspeciate

U.S. EPA, (1994a), Office of Transportation and Air Quality (OTAQ),
       http://www.epa.gov/otaq/models.htm. MOBILE Model: MOBILES, MOBILE6

U.S. EPA, (1994b), "Guidance on Urban Airshed Model (UAM) Reporting Requirements for
       Attainment Demonstration", EPA-454/R-93-056.
                                         233

-------
U.S. EPA, (1995a), "User's Guide for the Industrial Source Complex (ISC3) Dispersion Models:
       Volume I - User Instructions", Office of Air Quality Planning and Standards, Emissions,
       Monitoring, and Analysis Division; September 1995; EPA-454/B-95-003a, RTF, NC.

U.S. EPA, (1995b), "User's Guide for the Industrial Source Complex (ISC3) Dispersion Models:
       Volume II - Description of Model Algorithms", Office of Air Quality Planning and
       Standards, Emissions, Monitoring, and Analysis Division; September 1995, EPA-454/B-
       95-003b, RTF, NC.

U.S. EPA, (1996a), "Guidance on Use of Modeled Results to Demonstrate Attainment of the
       Ozone NAAQS",   http://www.epa.gov/ttn/scram/guidance_sip.htm, EPA-454/B-95-007.

U.S. EPA, (1997a): 40 CFR Part 50 National Ambient Air Quality Standards for Particulate
       Matter; Final Rule, Federal Register / Vol. 62, No. 138 /Friday, July 18, 1997.
       http ://www. epa.gov/ttn/amtic/files/cfr/recent/pmnaaqs.pdf

U.S. EPA, (1997b), "Guidance For Network Design and Optimum Site Exposure For PM2 5 and
       PM10", OAQPS, Research Triangle Park, NC, EPA report no. EPA-454/R-99-022.

U.S. EPA, (1997c), "EIIP Volume I, Introduction and Use of EIIP Guidance for Emission
       Inventory Development", July 1997, EPA-454/R-97-004a.

U.S. EPA, (1998a), "EPA Third-Generation Air Quality Modeling System, Models-3 Volume 9b
       Users Manual", EPA-600/R-98/069(b), http://www.cmascenter.ore./, and
       http://www.epa.gov/asmdnerl/models3/cmaq.html.

U.S EPA, (1998b)"Finding of Significant Contribution and Rulemaking for Certain States in the
       Ozone Transport Assessment Group Region for Purposes of Reducing Regional Transport
       of Ozone; Final Rule," 63 FR 57,356 (October 27, 1998).

U.S. EPA, (1999a), "Draft Guidance on the Use of Models and Other Analyses in Attainment
       Demonstrations for the 8-Hour Ozone NAAQS", EPA-454/R-99-004,
       http://www.epa.gov/ttn/scram. Modeling Guidance, DRAFT8HR. (May 1999)

U.S. EPA, (1999b), "Implementation Guidance for the Revised Ozone and Particulate Matter
       (PM) National Ambient Air Quality Standards  (NAAQS) and Regional Haze Program".

U.S. EPA, (1999c), "Guidance on the Reasonably Available Control Measures (RACM)
       Requirement and Attainment Demonstration Submissions for Ozone Nonattainment
       Areas.", http://www.epa.gov/ttn/oarpg/tlpgm.html .

U.S. EPA, (2001),  "(Draft) Guidance for Demonstrating Attainment of Air Quality Goals for
       PM2 5 and Regional Haze",  http://www.epa.gov/ttn/scram/guidance_sip.htm.
                                         234

-------
U.S. EPA, (2002a), "3 Year Quality Assurance Report- Calendar Years 1999, 2000, and
       2001-PM2 5 Performance Evaluation Program."

U.S. EPA, (2002b): Clean Air Status and Trends Network (CASTNet), 2001 annual report.
       http ://www. epa. gov/castnet/

U.S. EPA, (2003a), "Federal Reference Method Quality Assurance Report".

U.S. EPA, (2003b), "Guidance for Tracking Progress Under the Regional Haze Rule", September
       2003, EPA-454/B-03-004,  http://www.epa.gov/ttn/oarpg/tl/memoranda/rh_tpurhr_gd.pdf

U.S. EPA, (2004a), "The Ozone Report: Measuring Progress Through 2003." EPA 454/k-04-001.
       http://www.epa.gov/air/airtrends/aqtrnd04/ozone.html

U.S. EPA, (2004b), "The Particle Pollution Report: Current Understanding of Air Quality and
       Emissions Through 2003 "  EPA 454/R-04-002.
       http://www.epa.gov/air/airtrends/aqtrnd04/pm.html

U.S. EPA, (2004c), "Developing Spatially Interpolated Surfaces and Estimating Uncertainty",
       EPA-454-R-04-0004, http://nsdi.epa.gov/oar/oaqps/pm25/docs/dsisurfaces.pdf.

U.S. EPA, (2004d), "EPA-CMB8.2 Users Manual", December 2004, EPA-452/R-04-011, Office
       of Air Quality Planning and Standards RTF, NC,
       .http://www.epa.gov/scram001/models/receptor/EPA-CMB82Manual.pdf

U.S. EPA, (2004e), "Protocol for Applying and Validating the CMB Model for PM2.5 and
       VOC", December 2004, EPA-451/R-04-001, Office of Air Quality Planning and
       Standards, RTF, NC, http://www.epa.gov/scramOO 1 /models/receptor/CMB_Protocol.pdf

U.S. EPA, (2004f),  "User's Guide  for the AMS/EPA Regulatory Model - AERMOD",
       EPA-454/B-03-001, September 2004, Office of Air Quality Planning and Standards, RTF,
       NC.

U.S. EPA, (2004g), "AERMOD: Description of Model Formulation", EPA-454/R-03-004,
September 2004, Office of Air Quality Planning and Standards, RTF, NC.

U.S. EPA, (2004h), Emission Inventory Improvement Program (EIIP), series of technical
documents, http://www.epa.gov/ttn/chief/eiip.

U.S. EPA, (2005a), "Regulatory Impact Analysis for the Final Clean Air Interstate Rule",
       http ://www. epa. gov/interstateairqualitv/pdfs/finaltech08 .pdf
                                         235

-------
U.S. EPA, (2005b), "Technical Support Document for the Final Clean Air Interstate Rule- Air
       Quality Modeling", March 2005. Docket number OAR-2003-0053-0162.
       http ://www. epa. gov/C AIR/pdfs/fmaltech02. pdf

U.S. EPA, (2005c), "Evaluating Ozone Control Programs in the Eastern United States: Focus on
       the NOx Budget Trading Program, 2004",
       http ://www. epa. gov/air/airtrends/2005/ozonenbp/

U.S. EPA, (2005d), 40CFR, Part 51, Appendix W, Revision to the Guideline on Air Quality
       Models, 68 FR 68235-68236, November 9, 2005.
       http://www.epa.gov/scram001/guidance/guide/appw 05.pdf
U.S. EPA, (2005e), "Emission Inventory Guidance For Implementation Of Ozone And Particulate
       Matter National Ambient Air Quality Standards (NAAQS) and Regional Haze
       Regulations, http://www.epa.gov/ttn/chief/eidocs/eiguid/eiguidfmal_nov2005.pdf.

U.S. EPA, (2005f), "EPA's Final Draft National Ambient Air Monitoring Strategy", EPA-SAB-
       CASAC-05-006, May 2005; posted at http://www.epa.gov/sab/pdf/casac-05-006.pdf

U.S. EPA, (2006a), "SPECIATE 4.0 Speciation Database Development Documentation- Final
       Report", EPA/600/r-06-161, November 2006, Office of Research and Development, RTF,
       NC,
       http ://www. epa.gov/ttn/chief/software/speciate/speciate4/documentation/speciatedoc_l 20
       6.pdf

U.S. EPA, (2006b), "AQS Data Dictionary", Version 2.12,
       http://www.epa.gov/ttn/airs/airsaqs/manuals/manuals.htm and
       http://www.epa.gov/ttn/airs/airsaqs/manuals/AQS%20Data%20Dictionary%20v2.12.pdf

U.S. EPA, (2006c), "Technical Support Document for the Proposed PM NAAQS Rule- Response
       Surface Modeling", February 2006, Office of Air Quality Planning and Standards, RTF,
       NC, http://www.epa.gov/scramOOl/reports/pmnaaqs_tsd_rsm_all_021606.pdf

U.S. EPA. (2007) "Guidance for Setting Reasonable Progress Goals Under the Regional Haze
       Program".

Vukovich, J.M., and T. Pierce, (2002), "The Implementation of BEIS3 within the SMOKE
       modeling framework", 11th Annual Emissions Inventory Conference of the U.S.  EPA,
       http://www.epa.gov/ttn/chief/conference/eillA Session 10, or
       http://www.epa.gov/ttn/chief/conference/eill/modeling/vukovich.pdf

Watson, J.G., (1997), "Observational Models: The Use of Chemical Mass Balance Methods",
       Prepared for the U.S. EPA  Source Attribution Workshop, Research Triangle Park, NC,
                                         236

-------
       July 16-18, 1997. http://www.epa.gov/ttn/faca/stissu.httnl Source Attribution Workshop
       Materials.

Watson, J.G., J.C. Chow, and E.M. Fujita, (2001), "Review of volatile organic compound source
       apportionment by chemical mass balance", Atmospheric Environment, 35, 1567-1584.

Watson, J.G., T. Zhu, J.C. Chow, J.P. Engelbrecht, E.M. Fujita, and W.E. Wilson, (2002),
       "Receptor modeling application framework for particle source apportionment",
Chemosphere,       49(9), 1093-1136.

Willis, R.D., (2000), Workshop on UNMIX and PMF as Applied to PM25, EPA/600/A-00/048.

Wilson, W.E. and H.H.Suh, (1997), "Fine Particles and Coarse Particles: Concentration
       Relationships Relative to Epidemiologic Studies", J.Air & Waste Management
       Association 47, pp. 1238-1249.

Yang, Y.J., W.R. Stockwell, and J.B. Milford, (1995), "Uncertainties in Incremental Reactivities
       of Volatile Organic Compounds", Environmental Science and Technology, 29, 1336-1345.

Yang, Y.J., J.G. Wilkinson and A.G. Russell, (1997a), "Fast, Direct Sensitivity Analysis of
       Multidimensional Air Quality Models for Source Impact Quantification and Area-of-
       Influence Identification", Prepared for the U.S. EPA Source Attribution Workshop,
       Research Triangle Park, NC, July 16-18, 1997. http://www.epa.gov/ttn/faca/stissu.htmL
       SAW-Direct  Sensi Analys.-PowerP.-T.Russell.

Yang, Y.J., J.G. Wilkinson, and A.G. Russell, (1997b), "Fast, Direct Sensitivity Analysis of
       Multidimensional Photochemical Models", Environmental Science and Technology, 31,
       2859-2868.

Yarwood, G., and R. Morris, (1997a), "Description of the CAMx  Source Attribution Algorithm",
       prepared for U.S. EPA Source Attribution Workshop, Research Triangle Park, NC, July
       16-18, 1997.  http://www.epa.gov/ttn/faca/stissu.html.

Yarwood, G., G. Wilson, R.E. Morris, and M.A. Yocke, (1997b),  "User's Guide to the Ozone
       Tool: Ozone  Source Apportionment Technology for UAM-IV", prepared for Thomas
       Chico, South Coast Air Quality Management District, Diamond Bar, CA, March 28, 1997.

Yarwood, G., T.E. Stoeckenius,, J.G. Heiken, and A.M. Dunker, (2003),  "Modeling
       Weekday/Weekend Ozone Differences in the Los Angeles Region for 1997", J.Air &
       Waste Management Association, 53,  864-875.

Yarwood, G., Morris, R. E., and Wilson, G.M., (2005), "Particulate Matter Source Apportionment
       Technology (PSAT) in the CAMx Photochemical Grid Model",
       www.camx.com/publ/pdfs/   Yarwood ITM paper.pdf
                                          237

-------
Yua, J. Z., H. Yanga, H. Zhanga, and A.K.H. Lau, (2004), " Size distributions of water-soluble
       organic carbon in ambient aerosols and its size-resolved thermal characteristics",
       Atmospheric Environment 3 8, 1 06 1-1 07 1 .

Yuan., Z.B., J. Z. Yu, A. K. H. Lau, P. K. K. Louie, and J. C. H. Fung, (2006), " Application of
       positive matrix factorization in estimating aerosol secondary organic carbon in Hong
       Kong and its relationship with secondary sulfate", Atmospheric Chemistry and Physics., 6,
       25-34.

 Zhang, X. and P. H. McMurry, (1992), "Evaporative losses of fine parti culate nitrates during
              g", Atmospheric Environment, 26, 3305-3312.
 Zheng, M., G.R. Cass, JJ. Schauer, and E.S. Edgerton, (2002), "Source apportionment of PM2.5
        in the southeastern United States using solvent-extractable organic compounds as
        tracers", Environ. Sci. Technol, 36,2361-2371.
                                           238

-------
Glossary

Modeled attainment demonstration - A modeled attainment demonstration consists of two parts: an
analysis estimating emission levels consistent with attainment of the NAAQS, and a list of measures that
will lead to the desired emission levels once growth is accounted for. The first (analysis) part consists of
a modeled attainment test.  It may also include an "unmonitored area analysis" and a review of a diverse
set of model outputs and emissions, air quality and meteorological data for consideration in a weight of
evidence determination to assess whether attainment of the NAAQS is likely with the proposed control
strategy.

Modeled attainment test  - This test takes the ratio of mean predicted future and current ozone or PM25
species concentrations averaged over multiple days and multiplies this ratio times the site-specific
monitored design value at each monitoring location. If the product is less than the NAAQS near all
monitoring sites, the test is passed.

Modeling system - This is a group of models used to predict ambient ozone and/or PM25
concentrations.  The group includes an emissions model which converts countywide emission information
into gridded speciated emissions which vary diurnally and reflect environmental conditions. It also
includes a meteorological model which provides gridded meteorological outputs and an air
chemistry/deposition model which takes information provided by the emissions and meteorological
models and uses it to develop gridded predictions of hourly pollutant concentrations.

Relative response factor (RRF)  -  The ratio of predicted ozone and/or PM25 concentrations averaged
over multiple days near a monitoring site with future emissions to corresponding predictions obtained
with current emissions.

Uniform Rate Of Progress (a.k.a Glidepath)- The amount of visibility improvement needed (in each
review period) to reach natural background visibilty conditions by 2064 at each Class I area by achieving
a linear  rate of improvement in visibility between the 2000-2004 base period and 2064.

Unmonitored Area Analysis - An analysis used to ensure  that a proposed control strategy will be
effective in reducing ozone and/or PM2 5 at locations without air quality monitors so that attainment is
shown throughout a nonattainment area. The purpose of the analysis is to use a combination of model
output and ambient data to identify areas that might exceed  the NAAQS if monitors were located there.

Weight of evidence determination (WOE)  -  This is a set of diverse analyses used to judge whether
attainment of the NAAQS is likely.  The credibility of each analysis is assessed and an outcome
consistent with an hypothesis that the NAAQS will be met is identified beforehand. If the set of
outcomes, on balance, is  consistent with attainment, then the WOE can be used to show attainment. A
weight of evidence determination includes results from the modeled attainment test, the unmonitored area
analysis, other model outputs and several recommended analyses of air quality, emissions and
meteorological data.
                                              239

-------
APPENDIX A

Below are the definitions of model performance statistics suggested as part of this ozone
modeling guidance.

Mean Observation: The time-average mean observed value (in ppb)

        1  N
OBS = — Y
       N,
Mean Prediction: The time-average mean predicted value (in ppb) paired in time and space
with the observations.
          1 N
 Model = —/  Model
Ratio of the Means: Ratio of the predicted over the observed values. A ratio of greater than 1
indicates on overprediction and a ratio of less than 1 indicates an underprediction.
         1  ^ Model
Ratw = ฅ4 ~^bT
Mean Bias (ppb):  This performance statistic averages the difference (model - observed) over
all pairs in which the observed values were greater than zero. A mean bias of zero indicates that
the model over predictions and model under predictions exactly cancel each other out.  Note that
the model bias is defined such that positive values indicate that the model prediction exceeds the
observation, whereas negative values indicate an underestimate of observations by the model.
This model performance estimate is used to make statements about the absolute or unnormalized
bias in the model simulation.
         1
BIAS = — V (Model- Obs)
                                         240

-------
Normalized Mean Bias (percent): This statistic averages the difference (model - observed)
over the sum of observed values. Normalized mean bias is a useful model performance indicator
because it avoids over inflating the observed range of values.

          N
         X (Model -  Obs)
  NMB =	• 100%
              I (Obs)
Mean Fractional Bias (percent): Normalized bias can become very large when a minimum
threshold is not used. Fractional bias is used as a substitute.  The fractional bias for cases with
factors of 2 under- and over-prediction are -67 and + 67 percent, respectively (as opposed to -50
and +100 percent, when using normalized bias). Fractional bias is a useful indicator because it
has the advantage of equally weighting positive and negative bias estimates. The single largest
disadvantage is that the predicted concentration is found in both the numerator and denominator.
              i  V (Model
Mean Error (ppb): This performance statistic averages the absolute value of the difference
(model - observed) over all pairs in which the observed values are greater than zero. It is similar
to mean bias except that the absolute value of the difference is used so that the error is always
positive.
 ERR = T7 Z \Model - Obs\


Normalized Mean Error (percent): This performance statistic is used to normalize the mean
error relative to the observations.  This statistic averages the difference (model - observed) over
the sum of observed values. Normalized mean error is a useful model performance indicator
because it avoids over inflating the observed range of values.
         N
        X \Model - Obs\
NME = -—	• 100%
            I (Obs)
                                          241

-------
Mean Fractional Error (percent): Normalized error can become very large when a minimum
threshold is not used. Therefore fractional error is used as a substitute. It is similar to the
fractional bias except the absolute value of the difference is used so that the error is always
positive.
                i V (Model
Correlation Coefficient (R2): This performance statistic measures the degree to which two
variables are linearly related. A correlation coefficient of 1 indicates a perfect linear
relationship; whereas a correlation coefficient of 0 means that there is no linear relationship
between the variables.
                       (Model - Model)(Obs - Obs)
CORRCOEFF=
                      Model- Model)  ฃ (Obs- Obs)
                                          242

-------
Appendix B:  Summary of recent model performance evaluations conducted
by other modeling groups
 1. Boylan, I, VISTAS, "PM Model Performance Goal and Criteria", National RPO Modeling
 Meeting, Denver, CO, 2005a.

 •      Based on benchmarking with a combination of data from a number of PM modeling
       studies (SAMI, VISTAS, WRAP, EPA, MANE-NU, EPRI, and Midwest RPO).
       Proposed performance goals (close to best achievable) and performance criteria
       (acceptable) for these metrics and showed where the modeling studies fall for the
       various components of PM.
 •      Proposed to use mean fractional bias (MFB) and mean fractional error (MFE) as the
       standard performance metrics.
    Goal: MFE <= 50%, MFB <= +-30%
    Criteria: MFE <= 75%, MFB <= +-60%
     Less abundant species should have
     less stringent goal & criteria
 •      Proposed to use asymptotically approaching goals & criteria when data are greater than
       2.5 urn, approaching +200% MFE and +-200% MFB for extreme small model &
       observed data (formula of logarithmic MFB & MFE are proposed)
 •      Based on combined modeling studies described above, for more abundant conditions,
       MFE and MFB are typical in the range of
    Sulfate:  MFE = 30% ~ 77%,  MFB = -45% ~ +51% (> 2 ug/m3)
    Nitrate:  MFE = 55% ~ 125%, MFB = +3% ~ +82% (> 1 ug/m3)
    Organic: MFE = 35% ~ 95%, MFB = -70% ~ +35% (> 1.5 ug/m3)
    EC:     MFE = 50% ~ 95%, MFB = -45% ~ +50% (> 0.5 ug/m3)
    PM 25:  MFE =  50% ~ 85%,   MFB = -55% ~ +60% (> 5 ug/m3)
 •      Suggested to conduct performance evaluation on episode-by-episode basis or month-
       by-month for annual modeling
       Different performance goals &  criteria may be needed for gaseous precursors and wet
       depositions.
 •      Benchmarking should be done for the entire modeling system (meteorology,  emissions
       inventory, and model).
 2. Morris, R., et al., "Application of Multiple Models to Simulation Fine Particulate in the
 Southeastern US", National RPO Modeling Meeting, Denver, CO, 2005a.

 •      Based on model multiple model applications over VISTAS modeling:
       Uuly 1999 & July 2001: CMAQ and CMAQ 36-km & 12-km
       2January 2002 & July 2001: CMAQ and CMAQ MADRID 36-km only
                                       243

-------
       3Used same horizontal & vertical grids, CMAQ-to-CAMx emissions, ICs/BCs, but
different MM5 interface (MCIP vs. MMSCAMx).
•      Both models performed reasonably well, with CMAQ performing better for SO4 and
       CAMx performing better for OM.  Both models did not perform well for NO3, soil,
       and coarse PM, but CAMx seemed to have much higher positive bias in winter nitrate
       than CMAQ
•      Used fractional bias and fractional error (instead of normalized ones) to illustrate
       model performance because these statistics do not exhibit such extreme fluctuation.
•      Demonstrated the usefulness of "soccer goal" plots for MFB & MFE.  Suggested
       15%/35% (MFB/MFE) for O3, and illustrated 50%/75% for PM25 species and bounded
       at 100%/200%
       Based on modeling studies described above (July 1999 & 2001), MFE and MFB are
       typical in the range of
   Sulfate:  MFE  = 25% ~ 70%, MFB = -25% ~ +51%
   Organic: MFE  = 30% ~ 75%, MFB = -50% ~ +45%
   EC:     MFE  = 35% ~ 65%, MFB = -30% ~ +40%
•      CMAQ performance improved for OM when they adjusted the Kz min value to
       1.0 m2/sec (from O.lm2/sec).
       Data (July 2001) of SO4 and total carbon mass (TCM). Both models seemed to
       captured the diurnal and daily trends well for both SO4 & TCM, but the magnitude can
       be significantly off for some of the days.
3. Boylan, J. and Baker, K., "Photochemical Model Performance and Consistency", National
RPO Modeling Meeting, Denver, CO, 2005.

       Illustrated model performance comparisons of 36-km daily modeling results by MRPO
       (CAMx4), VISTAS (CMAQ-CB4), and MANE-VU (CMAQ-SARPC99) for two
       episodes (July 1999 and January 2002) at three IMPROVE sites; and by MRPO
       (CAMx4) and VISTAS (CMAQ-CB4 for one episode (July 1999) for the Pittsburgh
       supersite.
       Indicated that in some cases there was good agreement among the models and selected
       IMPROVE sites, but in others cases there were noticeable variations (even between the
       two RPOs running CMAQ).
•      Sulfate were more consistent among the three models; nitrate and OM have higher
       discrepancies; All three models overpredicted nitrate and underpredicted OM (based on
       model  vs. observed scatter plots comparisons)
•      Suggested that where models  diverge, they may show a different response to control
       strategies. Reasons for the model variations may include differences between CB4 and
       SAPRC99, and potential  differences in emissions inventories, differences in Kz_min
       values, differences in met. & land use/soil methodologies, etc.
•      Also described a comparison of hourly modeling results by MRPO and VISTAS for a
       July 1999 episode at the Pittsburgh super site. In some cases these hourly results were
                                        244

-------
       better than daily results and generally captured the diurnal patterns for PM2 5 species
       and gaseous species (O3, NOx, HNO3, SO2, etc.), although the magnitude was
       considerably off in some cases.
4. Seigneur, C., "Review of CMAQ and REMSAD Performance for Regional PM Modeling",
AER Report for UARG, Doc# CP163-03-01b, March 2003.

•      Describe a review of summary performance of CMAQ, CMAQ-MADRID 1, CMAQ-
       MADRID 2, and REMSAD.  Based on the comparison studies by others given below:
       a. BRAVO 1999 study: CMAQ-MADRID 1 & REMSAD
       b. WRAP 1996 studies: CMAQ & REMSAD
       c. WRAP Aug. 1999 & Jan. 2000 studies: CMAQ & REMSAD
       d. EPA/ORD July 1999 studies: CMAQ & REMSAD
       e. Southeast PM - July 1999 study (Nashville/Atlanta):CMAQ,
        CMAQ-MADRID 1,2 & REMSAD
       AER Suggested normalized errors of 50% as the benchmark for sulfate and PM2 5
       (note: this suggestion was not agreed by other modeling groups, see references #2, #5,
       #7: the MNE is most biased metrics and has high fluctuations). The review showed
       only SE PM study meets this criteria for sulfate and PM2 5, and all the rest of species
       failed. No model showed consistently better performance than the others. Suggested
       that the model inputs has more effect on performance than model formulations.
•      Indicated that "the current performance of air quality models for PM is poor". .."There
       is a dire need for improving model inputs and model formulations in order to obtain
       acceptable model performance"... "3-D air quality models are the best tools available to
       address the PM source-receptors relationships because they take into account the
       non-linearities that affect the formation of secondary PM".
5. Boylan, J., VISTAS, "Calculating Statistics: Concentration Related Performance Goals",
PM Model Performance Workshop, RTF, NC, 2004b.

•      Illustrated a set of standard bias and errors calculations commonly used for model
       performance statistics and proposed model performance goal.
•      Indicated Mean Normalized Bias and Errors (MNBE) are most biased and least useful
       among "MNBE", "NMBE" and "MFBE".  The Mean Fractional Bias and Errors
       (MFBE) is least biased and most useful among the three metrics.
•      Recommended MFB & MBE and proposed performance goal: MFE<=50% and MFB
       <=+-30% for more abundant species (eg., sulfate & PM25) and less stringent for less
       abundant species (eg., nitrate, OC, EC, soil, etc.). Performance goal is not criteria and
       should be prohibit the modeling from being used if it fails to meet the goal.
       Proposed to use "2.5ug/m3" as the "grayline" for 50% MFE & +-30% MFB and
       asymmetrically approaching 200% MFE & +-200% MFB to extremely small
       concentrations.

                                        245

-------
      Recommended to use monthly avg. for annual modeling.
6. Morris, R., et al., "Model and Chemistry Inter-comparison: CMAQ with CB4, CB4-2002,
SAPRC99", National RPO Modeling Meeting, Denver, CO, 2005b.

      Based on US (36-km) and VISTAS (12-km) January 2002 modeling, conducted
      chemistry mechanisms inter-comparisons for CMAQ with CB4, CB4-2002,and
      SAPRC99.
      The performance of CB4 and CB4-2002 was similar for PM, and superior to SAPRC99
      overall (for the Jan02 case).
•      The model performance for CMAQ/CB4, US 36-km domain is in the range of:
   Sulfate:  MFE=42%~73%,   MFB =-21% ~+14%
   Nitrate:  MFE = 62% ~ 105%,   MFB = -21% ~ +46%
   Organic: MFE = 50% ~ 77%,    MFB = +3% ~ +59%
   EC:     MFE =59%-88%,    MFB = +2% ~ +70%
   Soil:    MFE = 165% ~ 180%, MFB =+164% ~ +180%
   PM 2 5: MFE  = 48% ~ 88%,   MFB = +25% ~ +81%
      Given that the computational cost of SAPRC99 is twice that of CB4, suggested to use
      36 and 12 km grids with CB4 chemistry for PM modeling for the time being.
      Noted that both CB4 and SAPRC underpredicted winter O3 significantly.
7. Tonnesen, G., et al., "Regional Haze Modeling: Recent Modeling Results for VISTAS and
WRAP", CMAS Annual workshop, RTF, NC, 2003.

•     Illustrated the WRAP 1996 CMAQ 36-km modeling and performance evaluation in the
      Western U.S. and VISTAS  CMAQ 12-km modeling for 3 episodes: January 2002, July
      1999, July 2001. Recommended to use the performance metrics of Mean Fractional
      Bias (MFB)and Mean Fractional errors (MFE) over mean normalized bias & errors
      (used in earlier WRAP model evaluation).
   Sulfate:  MFB = -47% ~ +48% (1996  WRAP domain)
   Nitrate:  MFB = -95% ~ +30% (1996  WRAP domain)
   Organic: MFB = -20% ~ +70% (1996 WRAP domain)
   OC:     MFB = -45% ~ +3% (1996 WRAP domain)
•     VISTAS modeling key findings (1) sulfate performance reasonably well (2) nitrate
      overpredictions in the winter, underpredictions in summer, may need better NH3
      emissions (3) Kv min =1 improved performance, mixing height is important (4) minor
      differences in 19 vs. 34 layers
8. Zhang, Y., et al., "Performance Evaluation of CMAQ and PM-CAMx for July 1999 SOS
Episode", AER Report for CRC, Doc# CP131-03-01, April 2003.
                                      246

-------
       Conducted CMAQ (2002 version) and PM-CAMx performance evaluation based on
       July 1999 SOS episode (6/29-7/11) modeling study (32-km nested w/ 8-km in the SE
       U.S., including Atlanta & Nashville)
       Ozone performance: use MNB & MNE w/ 60 ppb threshold for O3 (CMAQ performed
       better):
           CMAQ: MNB < 1%,  MNE = 18%
           CAMx: MNB = 27%, MNE = 33%
       PM performance: CMAQ & PM-CAMx are generally consistent in the rural areas (vs.
       IMPROVE); differs significantly over urban/suburban; in general, CMAQ performs
       much better than PM-CAMx
   PM25: CMAQ: MNB = 38%, MNE =-7%
           CAMx: MNB = 55%, MNE = 35%
   Sulfate:  CMAQ: MNB = 9%,  MNE = 45%
           CAMx: MNB = 44%, MNE = 63%
   Nitrate: CMAQ: MNB =-50%, MNE = 98%
           CAMx: MNB =137%, MNE = 158%
       The performance of CMAQ for PM and O3 is consistent with the performance
       expected for air quality models (however, nitrate issue existed); the performance of
       CAMx does not generally meet current expectations for AQM.
9. Morris. R., et al.,"Evaluation of CAMx: Issues Related to Section Models ", PM Model
Performance Workshop,RTP, NC, 2004c.

      Illustrated a WRAP comparison of CMAQ (v4.3), REMSAD (v7), CAMx (bimodal
      PM), and CAMx (4-section PM) based on January and July 1996 and annual 1996 over
      the 36-km WRAP domain.
•     Indicated that all models exhibited variations in performance, but no clear winner
      across all species and periods.  Sulfate predictions were reasonable,  but nitrate was
      significantly overpredicted. For all three models,
   Sulfate  : MFE = 40% ~ 60%,   MFB = -40% ~ +14%  (July 96)
   Nitrate  : MFE = 105% ~ 200%, MFB = +45% ~> +100% (Jan. 96)
   Organic: MFE = 50% ~ 75%,   MFB = -65% ~ +5%   (July 96)
   EC     : MFE = 47% ~ 105%,  MFB = -48% ~ +25%  (Jan. 96)
      The 1996 model performance is less than stellar, indicating potential issues in 1996
      MM5 and emissions.
•     Showed the effects of sectional PM distribution on PM Modeling in the Western US.
      Based on a study in the South Coast Air Basin using CAMx4+, which allows side-by-
      side comparisons of aqueous-phase chemistry modules (bulk vs. variable size
      resolution) and of PM size distribution treatments (bimodal vs. sectional).
10. Tonnesen, G., et al..,"Model Performance Metrics, Ambient Data Sets and Evaluation
Tools", PM Model Performance Workshop, RTF, NC, 2004a.
                                        247

-------
       Illustrated model performance metrics, available PM2 5 and O3 data, and evaluation
       software tool. Suggested that air quality modeling should include model evaluation as
       part of the system.
       Comparing performance metrics may not be enough since performance metrics often
       show mixed response and it is possible for a better model to have poorer metrics (e.g.,
       bias in met & emissions inputs). Diagnostic evaluation is needed to judge finer grid
       performance since coarser grid may have compensating errors. But should we assume
       that finer grid modeling always gives better simulations/physics?
       An example of VISTAS modeling (July  1999) showed differences of hourly sulfate and
       its wet deposition between CMAQ results using 12-km and 36-km grids, possible due
       to regional transport (wind speed & direction), precipitations/clouds, and numerical
       diffusion.
       Recommended bias factor as the best metric for evaluating haze.
11. Wang, Z., et al.," Comparison and Diagnostic Evaluation of Air Quality Models for
Paniculate Matter: CAMx, CMAQ, CMAQ-MADRID", National RPO Modeling Meeting,
2004.

       Conducted model evaluation based on 1999 SOS episode (6/29-7/10) EPRI modeling
       study,
•       CAMx has higher positive bias than CMAQ and CMAQ-Madrid in predicting sulfate;
       CMAQ underestimated nitrate and CAMx and CMAQ-MADRID overestimated
       nitrate.
•       All three models underestimated OM. However, there was no clear winner in model
       performance.
       The three models responded differently to a 50% increase in ammonia emissions,
       indicating a need to further look at the models' responsiveness to changes in emissions.
12. Ku, C., CENRAP, "CMAQ and CAMx Simulations for January and July 2002", National
RPO Modeling Meeting, Denver, CO, 2005.

      Compared CMAQ and CAMx 36-km simulations Based on January and July 2002
      (basB) over a continental US domain.
•      Indicated that the results were mixed for the models over CENRAP generally. Both
      models performed  acceptably for PM2 5 and sulfate in the summer (when sulfate is
      abundant) but they were overpredicted in the winter compared against IMPROVE
      network.
•      CAMx significantly overpredicted nitrate in the winter (higher prediction than
      CMAQ), but has better performance in OM (lower prediction than CMAQ), based on
      performance measures of normalized bias and errors.
•      The study also showed mixed results for the models in three climatically different
      regions in CENRAP that contain Class I areas, with varying performance depending on
      region and season.  This finding points up the difficulty in improving model
      performance over the whole CENRAP domain.

                                        248

-------
13. Eder B. and Yu S., "An Evaluation of the 2003 Release of Models-3/CMAQ", CMAS
Annual workshop, RTF, NC, 2003.

      Illustrated CMAQ 2003 evaluation for two episodes (winter 2002 and Summer 1999)
      for O3 and PM 2 5 species against AIRS, CASTNet, IMPROVE, and STN.  Suggested
      the use of the performance metrics of Normalized Mean Bias (NMB)and Normalized
      Mean errors (NMB) (and correlation coefficient, R).
      Ozone is fairly unbiased and accurate (NMB<10%) and NME=~20%); sulfate
      performance is quite well, even for STN. NO3 was the worst in the winter (NME
      -67% for STN and -96% for IMPROVE)
14. Frank, N.,"Use of National PM25 and Speciation Network Measurements for Model
Evaluation ", PM Model Performance Workshop, RTF, NC, 2004.

•      Compared correlated speciated monitoring sites (e.g., IMPROVE, CASTNet, STN) for
      ambient PM species measurements.
•      The sulfate agrees very well, but nitrate has both positive & negative bias site-by-site.
      OM is more uncertain than other species, but is still somewhat robust for model
      evaluation (ie., uncertainty is relatively small compared to current range of uncertainty
      in modeling). OM is about 50%-80% of urban excess of PM25.
15. Hu, Y., et.al., "Evaluation of CMAQ with FAQS Episode of August 1 lth-20th", 2000,
CMAS Annual workshop, RTF, NC, 2003.

       Conducted CMAQ (36/12/4 km nesting) performance evaluation for O3 based on Fall
       Line Air  Quality Study (FAQS) 8/11-20.
•       Indicated that CMAQ had a good O3 performance, but has a nighttime problem, which
       could be  due to min Kz used.  Analysis suggested that an optimal of min Kz may lie
       between 0.1-1 m2/s.
       The Isoprene emissions may be overestimated in the rural area and CO emissions may
       be underestimated.
16. Tonnesen, G., et al., "Prelim Preliminary Results CMAQ and CMAQ-AEVI with
SAPRC99", National RPO Modeling Meeting, 2004b.

      Compared CMAQ and CMAQ-AIM with SAPRC99 based on 2001 VISTAS modeling
      study. CMAQ-AIM with SAPRC99 has larger negative bias and lower predictions of
      Sulfate PM than standard released CMAQ.
      Conducted model & chemistry inter-comparisons: CMAQ with CB4, CB4-2002,
                                       249

-------
      SAPRC99. Some chemistry differences were observed in the models based on the 36-
      km U.S. and the 12-km VISTAS modeling; the performance of CB4 and CB4-2002
      was similar, and superior to SAPRC99 overall.
      Given that the computational cost of SAPRC99 is twice that of CB4, suggested to use
      36 and 12 km grids with CB4 chemistry for current VISTAS modeling study.
17. Yu, S., et al., "Simulation of Primary and Secondary Organic Aerosols over the US by
CMAQ: Evaluation and Analysis", CMAS Annual workshop, RTF, NC, 2003.

•      Evaluated CMAQ w/ SAPRC99 performance of Primary Organic Aerosols (POA), EC,
      and secondary Organic Aerosols (SOA) based on IMPROVE, SEARCH, and SOS data.

•      Suggested that model captured general trends & patterns of most EC & POA within a
      factor of two (based on IMPROVE & SEARCH).
      Slight underprediction in the Eastern U.S., likely due to underprediction in biogenic
      SOA.
18. Chien, C. Y., et al., "CMAQ Model Performance Evaluation with the Updated CB4-2002",
CMAS Annual workshop, RTF, NC, 2003.

       Evaluated CMAQ w/ CB4 vs. CB4-2002 (and CB4-2002 with removed N2O5 gaseous
       reaction) against IMPROVE, CASTNet and AQS based on January & July 1996 cases.
       Described the key updates of CB4-2002 (HNO3, N2O5 &  PAN rxns).
       CB4-2002 has lower O3, lower N2O5, lower PAN, higher HNO3, higher nitrate,
       slightly higher sulfate, and slightly lower SOA prediction than CB4.
•       The performance of CMAQ CB4 and CB4-2002 was similar for PM species. However,
       CB4-2002 has higher positive bias for winter nitrate (since it predicted higher nitrate).
19. Dennis, R.,"Time-Resolved and In-Depth Evaluation of PM and PM Precursors Using
CMAQ ", PM Model Performance Workshop, RTF, NC, 2004.

      Evaluated the CMAQ (2003 version) performance of inorganic PM, EC, and gaseous
      precursors against Atlanta 1999 supersite summer data and 2002 Pittsburgh supersite
      winter data.
•      Suggested that collapse of evening PEL is too fast and rise of morning PEL too slow
      (based on higher modeled EC, NOy & CO against Atlanta data at sunrise & sunset)
•      Performance for sulfate & ammonium was fairly good;
•      Overprediction for nitrate, but updated chemistry in CMAQ (included in 2004 version)
      has improved the nitrate performance (updated reaction probability for HNO3
      heterogeneous rxn and remove gaseous N2O5->HNO3 rxn)
•      Use ratio to define nitrate PM formation being HNO3- or NEC-limited: (NHx-
                                       250

-------
      2*SO4)/(HNO3+NO3)
      Nitrate PM predictions are very sensitive to NHx, and thus the NH3 emissions need
      serious attention.
20. Kumar, N.,"PM25 Model Performance: Lessons Learned and Recommendations ", PM
Model Performance Workshop, RTF, NC, 2004.

      Illustrated the use of model performance statistics based on two CMAQ-MADRID
      applications, SOS 1999 episode in the SE U.S. and 12-km BRAVO summer 1999
      episode (nested within REMSAD BRAVO 36-km modeling results)
•     Indicated model evaluation issues regarding local vs. regional, daily/weekly vs.
      month/seasonal
   Sulfate :  MNB= +20% / +51%, MNE = 51% / 89% (SEARCH/IMPROVE)
   Nitrate :  MNB=+72% /-25%,  MNE = 72% / 46% (SEARCH/IMPROVE)
   EC    :  MNB=+14% /-8%,   MNE = 52% / 54% (SEARCH/IMPROVE)
   PM 25 : MNB= -19% / -8%,   MNE = 32% / 49% (SEARCH/IMPROVE)
      Described the available performance metrics and illustrated the use of logarithmic and
      fractional bias and errors as potential benchmarks
21. Baker K., Midwest RPO, "Fine Grid Model Performance", National RPO Modeling
Meeting, Denver, CO, 2005.

•     Based on the Midwest RPO 36-km and 12-km Modeling study using CAMx4
•     Indicated that the model performance for PM2 5 species generally was similar for 36
      and 12 km grids; there were some differences in running control scenarios for
      sulfate/nitrate, but we could not determine which one is better
22. Morris, R., et al., "VISTAS Grid Resolution Sensitivity", National RPO Modeling
Meeting, Denver, CO, 2005d.

•     Based on VISTAS Southeast U.S. modeling study for Phase I episode (Jan. 2002, July
      1999, and July 2001), compared results for CMAQ w/ CB4 and SAPRC99 using the 36
      km national RPO domain and the  12 km grid VISTAS domain
•     Effects of grid resolution on model performance was mixed (performance was not
      necessarily improving using 12-km)
•     CMAQ w/ SAPRC99 (mixed for sulfate, worse for nitrate) was not performing better
      than w/ CB4
•     Compared July 1999 episode for ozone using 36-km & 12-km CMAQ and CAMx.
      CMAQ O3 performance degraded to underprediction with finer grid; CAMx is similar
      at 36-km and 12-km.
•     Examined the performance of the  MM5 model configurations using various cloud
      schemes (including Kain-Fritsch and Reisner schemes) for the 12 km (WRAP) and 36
      km domains. The results showed  that cool and moist bias found in the West,
                                       251

-------
      overprediction of convective precipitation using KF.
23. Zhang, Y., et al., "Development and Application of MADRID: A New Aerosol Module in
CMAQ", CMAS Annual workshop, RTF, NC, 2003.

      Evaluated CMAQ-MADRID against SCAQS 1987 episode.  The performance of
      CMAQ-MADRID seemed to perform well in O3, PM25, and sulfates, and comparable
      in other PM 25 species to other models (UAM-IV/CALGRID/UAM-AERO, GATOR,
      SMOG, CIT, SAQM-AERO)
•     Described the development and sciences of MADRID aerosol module included in
      CMAQ. Key features:  sectional (size bins), hybrid equilibrium, more sophisticated
      SOA treatment, CMU aqueous chemistry
24. Morris, R., et al., "WRAP Multi-Model Evaluation Using the 1996 36 km Section 309
Database", National RPO Modeling Meeting, Denver, CO, 2005e.

      Conducted WRAP 1996 multi-model evaluation. Compared CMAQ (v4.3), REMSAD
      (v7), CAMx (bimodal PM), and CAMx (4-section PM) using the 1996 36-km section
      309 database.
•     Indicated that sulfate was the best performing species on an annual basis, with a winter
      overprediction compensating for a summer underprediction; NO3 was predicted poorly
      by all models; OC, EC, and CM were underpredicted by all models, concluded that
      modeling is more challenging in the West than in the Midwest and Southeast.
25. Seigneur, C., "Current Status of Air Quality Models for Particulate Matter", NARSTO
2000—Tropospheric Aerosols: Science and Decisions in an International Community, October
23-26, 2000, Queretaro, Mexico, 2000b.

      Based on a limited set of comparisons with limited temporal and spatial model studies.
      Statistics calculated for aggregated gross error (AGE) and aggregated normalized bias
      (ANB). Note that the statistics for AGE and AGB were not obtained with the same
      models nor the same data base. Thus, performance for the two metrics in the table are
      not associated with each other.
   PM25 :  AGE ~ +30% - +50%, AGB ~ +10% (based on one study)
   Sulfate  : AGE ~ +30% - +50%, AGB ~ -20% - -30%
   Nitrate  : AGE ~ +20% - +70%, AGB--15-+50%
   EC     : AGE-+15%-+60%, AGB = none available
   OC     : AGE --40% -+50%, AGB - 38% (based on one study)
                                       252

-------
United States                        Office of Air Quality Planning and Standards             Publication No. EPA-454/B-07-002
Environmental Protection             Air Quality Analysis Division                           April 2007
Agency                             Research Triangle Park, NC
                                                          253

-------