Guidelines for Developing an Air Quality
(Ozone and PM2 5) Forecasting Program
-------
EPA-456/R-03-002
June 2003
Guidelines for Developing an Air Quality (Ozone and PM2 5) Forecasting Program
U.S. Environmental Protection Agency
Office of Air Quality Planning and Standards
Information Transfer and Program Integration Division
AIRNow Program
Research Triangle Park, North Carolina
-------
DISCLAIMER
This report was prepared as a result of work sponsored, and paid for, in whole or part, by
the U.S. Environmental Protection Agency (EPA). The opinions, findings, conclusions, and
recommendations are those of the authors and do not necessarily represent the views of the EPA.
The EPA, its officers, employees, contractors, and subcontractors make no warranty, expressed
or implied, and assume no legal liability for the information in this report. The EPA has not
approved or disapproved this report, nor has the EPA passed upon the accuracy of the
information contained herein.
-------
ACKNOWLEDGMENTS
Information contained in this document is the culmination of literature searches and
interviews with colleagues at local, state, and federal agencies and universities and in the private sector
who either forecast air pollution or use air quality forecasts. Their ideas and suggestions have been
instrumental in producing these guidelines.
The authors especially wish to thank the following individuals for giving us their time
and the benefit of their experience: Mr. Mike Abraczinskas, Mr. Lee Alter, Mr. Craig B. Anderson,
Mr. Rafael Ballagas, Mr. Mark Bishop, Mr. George Bridgers, Mr. Robert Browner, Mr. Chris Carlson,
Mr. Joe Casmassi, Mr. Alan C. Chan, Mr. Joe Chang, Mr. Aaron Childs, Mr. Lyle Chinkin, Dr. Geoffrey
Cobourn, Dr. Andrew Comrie, Ms. Lillie Cox, Ms. Laura DeGuire, Mr. Timothy S. Dye, Mr. Sean
Fitzsimmons, Mr. Mike Gilroy, Ms. Beth Gorman, Hilary R. Hafner, Ms. Sheila Holman, Mr. Michael
Koerber, Mr. Larry Kolczak, Mr. Bryan Lambeth, Mr. Eric Linse, Ms. April Linton, Mr. Fred Lurmann,
Mr. Clinton P. MacDonald, Mr. Michael Majewski, Mr. Cliff Michaelson, Ms. Eve Pidgeon, Ms.
Katherine Pruitt, Mr. Chris Roberie, Dr. Paul Roberts, Mr. Bill Ryan, Mr. Kerry Shearer, Mr. Till
Stoeckenius, Ms. Susan Stone, Mr. Troy Stuckey, Mr. Bob Swinford, Mr. Richard Taylor, Mr. Brian
Timan, Mr. Alan VanArsdale, Mr. Chet Wayland, Mr. John E. White, Mr. Lew Weinstock, Ms. Leah
Weiss, Mr. Neil J.M. Wheeler, and Mr. Robert Wilson.
in
-------
This page is intentionally blank.
-------
TABLE OF CONTENTS
Section Page
ACKNOWLEDGMENTS iii
LIST OF FIGURES vii
LIST OF TABLES xi
LIST OF ACRONYMS xiii
1. INTRODUCTION AND GUIDE TO DOCUMENT 1-1
1.1 Introduction 1-1
1.2 Document Objectives 1-2
1.3 Guide to This Document 1-2
2. PROCESSES AFFECTING AIR QUALITY CONCENTRATIONS 2-1
2.1 Ozone 2-1
2.1.1 Basic Ozone Chemistry 2-1
2.1.2 Ozone Precursor Emissions 2-2
2.2 Particulate Matter 2-5
2.2.1 Basic Particulate Matter Chemistry 2-5
2.2.2 PM2.s Emissions and Sources 2-11
2.2.3 Monitoring Issues 2-14
2.2.4 Unusual PM Events 2-17
2.3 Meteorological Conditions That Influence Air Quality 2-20
2.3.1 Aloft Pressure Patterns 2-23
2.3.2 Temperature Inversions and Vertical Mixing 2-23
2.3.3 Winds and Transport 2-26
2.3.4 Clouds, Fog, and Precipitation 2-26
2.3.5 Weather Pattern Cycles 2-28
3. FORECASTING APPLICATIONS AND NEEDS 3-1
3.1 Public Health Notification 3-1
3.2 Episodic Control Programs 3-1
3.3 Specialized Monitoring Programs 3-2
4. DEVELOPING OZONE AND PM2.5 FORECASTING METHODS 4-1
4.1 Forecasting Methods 4-1
4.1.1 Persistence 4-1
4.1.2 Climatology 4-6
4.1.3 Criteria 4-10
4.1.4 Classification and Regression Tree (CART) 4-13
4.1.5 Regression Equations 4-16
4.1.6 Artificial Neural Networks 4-19
4.1.7 Deterministic Air Quality Modeling 4-22
4.1.8 The Phenomenological/Intuition Method 4-29
4.2 Selecting Predictor Variables 4-31
-------
TABLE OF CONTENTS (Concluded)
Section Page
5. STEPS FOR DEVELOPING AN AIR QUALITY FORECASTING PROGRAM 5-1
5.1 Understanding Forecast Users' Needs 5-1
5.2 Understanding the Processes That Control Air Quality 5-3
5.2.1 Literature Reviews 5-3
5.2.2 Data Analyses 5-3
5.3 Choosing Forecasting Methods 5-17
5.4 Data Types, Sources, and Issues 5-18
5.5 Forecasting Protocol 5-22
5.6 Forecast Verification 5-22
5.6.1 Forecast Verification Schedule 5-23
5.6.2 Verification Statistics for Discrete Forecasts 5-24
5.6.3 Verification Statistics for Category Forecasts 5-27
5.6.4 Methods to Further Evaluate Forecast Performance 5-32
6. REFERENCES 6-1
VI
-------
LIST OF FIGURES
Figure Page
2-1. Average diurnal profile of ozone, NO, and VOC concentrations for August 1995 at
an urban site in Lynn, Massachusetts 2-2
2-2. 1996 VOC emissions from anthropogenic sources by county 2-4
2-3. 1996 NO emissions from anthropogenic sources by county 2-4
2-4. 1996 VOC emissions from biogenic sources by county 2-5
2-5. Volume size distribution measured in traffic showing fine and coarse
particle modes 2-6
2-6. Distribution of particle number, surface area, and volume or mass
with respect to size 2-7
2-7. Relationship between light scattering, absorption, and particle diameter 2-8
2-8. Sources of precursor gases and primary particles, PM formation processes,
and removal mechanisms 2-9
2-9. Seasonal maps of PM2.5 mass for 1994-1996 2-13
2-10. Ambient PM2.5 composition at urban sites in the United States 2-14
2-11. Idealized distribution of ambient PM showing fine-mode particles
and coarse-mode particles and the fractions collected by size-selective samplers 2-15
2-12. Schematic of the typical meteorological conditions and air quality often
associated with an aloft ridge of high pressure 2-21
2-13. Schematic of the typical meteorological conditions and air quality often
associated with an aloft trough of low pressure 2-21
2-14. 500-mb heights on the morning of July 17, 1999 and on the morning of
September 21, 1999 2-24
2-15. Schematic showing diurnal cycle of mixing, vertical temperature profiles, and
boundary layer height on a day with a weak temperature inversion and on a day
with a strong temperature inversion 2-25
2-16. 500-mb heights and surface pressure on the afternoon of January 7, 2002 2-27
2-17. 500-mb heights and surface pressure on the afternoon of January 22, 2002 2-27
vn
-------
LIST OF FIGURES (Continued)
Figure Page
2-18. Life cycle of synoptic weather events at the surface and aloft at 500 mb for
Ridge — high pressure, Ridge — back side of high, and Trough — cold
front patterns [[[ 2-29
4- 1 . Monthly distribution of the average number of days that PIVb.s concentrations
fell into each AQI category from 1999 to 2001 based on the peak 24-hr average
PM2.5 concentrations measured at 1 2 sites in the greater Pittsburgh region .................... 4-8
4-2. Summertime day-of-week distribution of the average number of days
per year that PM2.5 concentrations fell into each AQI category from 1999 to 2001
based on the peak 24-hr average PM2.5 concentrations measured at 12 sites in the
greater Pittsburgh region [[[ 4-9
4-3. Scatter plot of maximum surface temperature and regional maximum 8-hr ozone
concentration in Charlotte, North Carolina in 1996 [[[ 4-12
4-4. Decision tree for daily maximum ozone concentrations in the South Coast
Air Basin in the Los Angeles, California, area [[[ 4-15
4-5. A schematic of an artificial neural network [[[ 4-20
4-6. Schematic showing the component models of an air quality modeling system ............ 4-24
4-7. Schematic illustration of the processes in an Eulerian photochemical model cell ........ 4-26
5-1 . Distribution of the average number of days with 8-hr and 1-hr exceedances
by month for the New Jersey and New York City region from 1993-1997 .................... 5-5
5-2. Distribution by hour of daily maximum 1-hr ozone concentration on days that
exceeded 125 ppb in the New Jersey and New York City region from 1993-1997 ........ 5-6
5-3. Average annual frequency of ozone episode length for the 8-hr and 1-hr standards
in the New Jersey and New York City region from 1993-1997 ...................................... 5-7
5-4. Distribution of the average number of 8-hr and 1-hr exceedances by day of week
for the New Jersey and New York City region [[[ 5-8
-------
LIST OF FIGURES (Concluded)
Figure Page
5-7. A surface synoptic pattern associated with high ozone in
Pittsburgh, Pennsylvania 5-13
5-8. Scatter plot of 0200 EST ozone concentrations at a mountainous site in
Haywood County, North Carolina, versus North Carolina daily regional
maximum ozone concentrations for June to September, 1996 5-14
5-9. Back trajectories during an ozone episode in the northeastern United States
showing possible transport of pollutants from regions to the west 5-15
5-10. A 24-hr back trajectory from Crittenden County, Arkansas, starting at 1400 EST
on August 25, 1995, and ending at 1300 EST on August 26, 1995 5-16
5-11. Example outline of a forecast retrospective 5-24
5-12. Contingency table for a two-category forecast 5-27
5-13. Hypothetical verification statistics for a two-category forecast for Program LM
that has many ozone exceedances and Program SC with fewer exceedances 5-29
5-14. Contingency table for random forecast 5-30
5-15. Hypothetical verification statistics for a two-category forecast for Program LM
and its random forecast 5-31
5-16. Contingency table for a four-category forecast 5-32
5-17. An example of forecast bias for a 24-hr ozone forecast 5-33
IX
-------
This page is intentionally blank.
-------
LIST OF TABLES
Table Page
2-1. Summary of total anthropogenic VOC and NOX emissions in the United States
during 1994 2-3
2-2. Summary of formation pathways, composition, sources, and atmospheric
lifetimes of fine and coarse particulate matter 2-7
2-3. Major PM2.5 components 2-11
2-4. Types of unusual events, how they affect PM concentrations, and a list of
resources for acquiring data and information to forecast these events 2-19
2-5. Meteorological phenomena and their influence on PM2.5 and ozone
concentrations 2-22
4-1. Comparison of forecasting methods 4-2
4-2. Peak 8-hr ozone concentrations for a sample city for 30 consecutive days 4-5
4-3. Annual summaries of 1 -hr ozone exceedance days for New York State 4-7
4-4. Criteria for 1-hr ozone exceedances in Austin, Texas, used by the
Texas Natural Resource Conservation Commission 4-11
4-5. Variables used in regression Equation 4-3 4-17
4-6. Common predictor variables used to forecast ozone 4-32
4-7. Common predictor variables used to forecast PM2.5 4-32
5-1. Data products for developing forecasting methods and for forecasting
weather and ozone 5-19
5-2. Major sources of air quality and meteorological data 5-20
5-3. Example of a forecasting protocol schedule 5-23
5-4. Verification statistics computed on discrete concentration forecasts 5-25
5-5. Hypothetical forecasts for an 11-day period showing a human forecast,
observed values, and forecasts using the Persistence method 5-26
5-6. Verification statistics used to evaluate two-category forecasts 5-28
XI
-------
This page is intentionally blank.
-------
LIST OF ACRONYMS
Term
AQI
Aref
babs
BAM
CALGRID
CAMx
CART
CBL
CMAQ
CSI
DGV
EC
EMS-95
EPA
EPS 2.0
FAR
FRM
GOES
H2O
HY-SPLIT CheM
MAQSIP
MM5
MODIS
MOPITT
MOS
NAAQS
NBL
NCEP
NH3
NO
NO2
NOAA
NOX
NWS
O2
03
oc
OC/EC
PAMS
PM
PM10
Meaning
Air Quality Index
Accuracy of a reference forecast
Particle absorption
Beta Attenuation Monitor
California Grid Model
Comprehensive Air Quality Model with Extensions
Classification and Regression Tree
Convective Boundary Layer
Community Multiscale Air Quality Model
Critical Success Index
Geometric mean diameter by volume
Elemental Carbon
Emissions Modeling System
U.S. Environmental Protection Agency
Emission Processing System
False Alarm Rate
Federal reference method
Geostationary Operational Environmental Satellites
Water vapor
Hybrid Single-Particle Lagrangian Integrated Trajectories with a
generalized non-linear Chemistry Module
Multiscale Air Quality Simulation Platform
Penn State/NCAR Mesoscale Model Version 5
Moderate Resolution Imaging Spectroradiometer
Measurements of Pollution in the Troposphere
Model output statistics
National Ambient Air Quality Standards
Nocturnal Boundary Layer
National Center for Environmental Prediction
Ammonia
Nitric oxide
Nitrogen dioxide
National Oceanic and Atmospheric Administration
Nitrogen oxides
National Weather Service
Oxygen
Ozone
Particulate organic carbon
Organic and elemental carbon
Photochemical Assessment Monitoring Stations
Particulate matter
Particulate matter with an aerodynamic diameter less than 10 micrometers
Xlll
-------
Term
PM2.5
POD
POES
ppb
RAMS
RH
RL
SAQM
SMOKE
S02
ss
SVOCs
TEOM
UAM-AERO
UAM-IV
UAM-V
VOCs
WRAC
WSPs
LIST OF ACRONYMS (Concluded)
Meaning
Particulate matter less than 2.5 jim in diameter
Probability of Detection
Polar Orbiting Satellites
Parts per billion
Regional Atmospheric Modeling System
Relative humidity
Residual Layer
SARMAP Air Quality Model
Sparse Matrix Operator Kernel Emissions
Sulfur dioxide
Skill Score
Semi-volatile organic compounds
Tapered Element Oscillating Microbalance
Urban Airshed Model with Aerosols
Urban Airshed Model with Carbon Bond IV Chemistry
Variable Grid Urban Airshed Model
Volatile organic compounds
Wide range aerosol classifier
Weather Service Providers
xiv
-------
1. INTRODUCTION AND GUIDE TO DOCUMENT
1.1 INTRODUCTION
Air pollution is a contamination of the atmosphere by gaseous, liquid, or solid wastes or
by-products that have a serious affect on human health and the biosphere, reduce visibility, and
damage materials. The major pollutants affecting the United States and other countries
throughout the world are ozone, particulate matter, lead, carbon monoxide, nitrogen dioxide,
sulfur dioxide, and toxic compounds.
Air quality forecasts provide the public with air quality information with which they can
make daily lifestyle decisions to protect their health. This information allows people to take
precautionary measures to avoid or limit their exposure to unhealthy levels of air quality. In
addition, many communities use forecasts for initiating air quality "action" or "awareness" days,
which seek voluntary participation from the public to reduce pollution and improve local air
quality. Current air quality forecasting efforts focus on predicting ozone and PM2.5.
Ozone is a reactive oxidant that forms in trace amounts in two parts of the atmosphere:
the stratosphere (the layer between 20-30 km above the earth's surface) and the troposphere
(ground level to 15 km). Stratospheric ozone, also known as "the ozone layer," is formed
naturally and shields life on earth from the harmful effects of the sun's ultraviolet radiation. Near
the earth's surface, ground-level ozone can be harmful to human health and vegetation and is
created in part by pollution from man-made (anthropogenic) and natural (biogenic) sources.
Because ground-level ozone accumulates in or near large metropolitan areas during certain
weather conditions, it typically exposes tens of millions of people every week during the summer
to unhealthy ozone concentrations (Paul et al., 1987).
Particulate matter (PM) is a complex mixture of solid and liquid particles that vary in size
and composition, and remain suspended in the air. Over the past decade, many health effect
studies have shown an association between exposure to PM and increases in daily mortality and
symptoms of certain illnesses (Dockery and Pope, 1994; Health Effects Institute, 2002;
Schwartz, 1994). Sources of PM are numerous; naturally occurring processes and human
activities all contribute to total PM concentrations. Some sources are natural, such as dust from
the earth's surface (crustal material), sea salt in coastal areas, and biologic material (pollen,
spores, and plant and animal debris). Periodic events like forest fires and dust storms can
produce large amounts of PM. In cities, PM is mainly a product of combustion from mobile
sources such as cars, buses, ships, trucks, and construction equipment, and from stationary
sources such as heating furnaces, power plants, and factories. Some PM is emitted directly into
the atmosphere as particles (primary particles), while some particles are produced by chemical
reactions in the atmosphere (secondary particles).
The size of ambient air particles ranges over a wide scale, from approximately 0.005 to
100 |im in aerodynamic diameter (from the size of just a few atoms to about the thickness of a
human hair). Particles fall into three basic size modes: ultrafine particles (smaller than about
0.1 |im in diameter), fine particles (between 0.1 and 2.5 jim), and coarse particles (larger than
2.5 |im). PMio is defined as particulate matter with an aerodynamic diameter less than
1-1
-------
10 micrometers. PM2.5 is a subset of PMio and includes those particles with an aerodynamic
diameter less than 2.5 jim. Cut points (2.5 jim and 10 jim) are not perfectly sharp for these PM
indicators; instruments that collect PM2.5 and PMio samples collect some particles larger than the
cut point while some particles smaller than the cut point are not retained. PM can vary greatly in
size, composition, and concentration depending on the sources generating the particles and such
factors as geographic location, season, day, time of day, and weather conditions.
In light of the health effects of ground-level ozone, many air quality agencies have been
forecasting ozone concentrations to warn the public of unhealthy air and to encourage people to
avoid exposure to unhealthy air and voluntarily reduce emissions. Fewer agencies have
forecasted PMio or PM2.5. From 1978 to 1997, ozone forecasts were based on the 1-hr National
Ambient Air Quality Standard (NAAQS) for ozone, which was 125 parts per billion (ppb). In
1997, the U.S. Environmental Protection Agency (EPA) revised the NAAQS to reflect more
recent health-effects studies that suggest that respiratory damage can occur at lower ozone
concentrations. Under the more-stringent revised standard, regions exceed the NAAQS when the
three-year average of the annual fourth highest 8-hr average ozone concentrations is at or above
85 ppb. Likewise in 1999, the EPA implemented a new NAAQS for PM2.5. The NAAQS for
PM2.5 is a 24-hour average concentration of 65 |ig/m3 and an annual standard of 15 |ig/m3.
1.2 DOCUMENT OBJECTIVES
This document provides guidance to help air quality agencies develop, operate, and
evaluate ozone and PM2.5 forecasting programs. This guidance document provides:
• Background information about ozone and PM2.5 and the weather's effect on these
pollutants.
• A list of how air quality forecasts are currently used.
• A summary and evaluation of methods currently used to forecast ozone and PM2.5.
• Steps to develop and operate an air quality forecasting program.
• Information on the level of effort needed to set up and operate a forecasting program.
The intended audience of this document is project managers, meteorologists, air quality
analysts, and data analysts. The information presented in this document is based on literature
reviews and on interviews with air quality forecasters throughout the country.
1.3 GUIDE TO THIS DOCUMENT
This document is divided into six sections with the following contents:
Section 2: Processes Affecting Air Quality Concentrations describes the principal chemical
processes that produce ozone, PM2.5, and their precursor emissions. It also describes
how atmospheric phenomena affect ozone and PM2.5 concentrations.
1-2
-------
Section 3: Forecasting Applications and Needs discusses how agencies throughout the United
States use air quality forecasts.
Section 4: Developing Forecasting Methods explains the different approaches used to forecast
air quality. It also describes each method and compares its strengths and limitations,
thus allowing forecasters to select the methods that meet their agency's needs and
resources.
Section 5: Steps for Developing an Air Quality Forecasting Program identifies the steps to
develop, operate, and evaluate an ozone or PM2.5 forecasting program.
Section 6: References provides a list of references cited in this report.
1-3
-------
2. PROCESSES AFFECTING AIR QUALITY CONCENTRATIONS
Air quality concentrations are strongly affected by weather. Developing a basic
understanding of how ozone and PM forms and where emissions originate will help air quality
agencies forecast the effects of weather on ozone, PM, and their precursor emissions.
This section provides a background on ozone (Section 2.1) and PM (Section 2.2)
including a summary of chemical reactions and sources of precursor emissions. Section 2.3
explains generally how weather affects pollutant formation, transport, and dispersion. A
discussion of how to develop a more detailed understanding of the chemical and meteorological
processes that control air pollution is presented in Section 5.2.
2.1 OZONE
Ozone (O3) is not emitted directly into the air; instead it forms in the atmosphere as a
result of a series of complex chemical reactions between oxides of nitrogen (NOX) and
hydrocarbons, which together are precursors of ozone. Ozone precursors have both
anthropogenic (man-made) and biogenic (natural) origins. Motor vehicle exhaust, industrial
emissions, gasoline vapors, and chemical solvents are some of the major sources of NOX and
hydrocarbons. Many species of vegetation including trees and plants emit hydrocarbons, and
fertilized soils release NOX.
2.1.1 Basic Ozone Chemistry
In the presence of ultraviolet radiation (hv\ oxygen (O2) and nitrogen dioxide (NO2)
react in the atmosphere to form ozone and nitric oxide (NO) through the reactions given in
Equations 2-1 and 2-2.
NO2 + hv -> NO + O (2-1)
O + O2 -» O3 (2-2)
Resultant ozone, however, is quickly reacted away to form nitrogen dioxide by the
process given in Equation 2-3. This conversion of ozone by NO is referred to as titration. In the
absence of other species, a steady state is achieved through the reactions shown by Equations 2-1
through 2-3. Even without anthropogenic emissions, these reactions normally result in a natural
background ozone concentration of 25 to 45 ppb (Altshuller and Lefohn, 1996).
O3 + NO -» NO2 + O2 (2-3)
Ozone cannot accumulate further unless volatile organic compounds (VOCs), which
include hydrocarbons, are present to consume or convert NO back to NO2 as shown by
Equation 2-4.
VOC + NO -» NO2+other products (2-4)
2-1
-------
This equation is a simplied version of many complex chemical reactions (see National Research
Council, 1991, for details). As NO is consumed by this process, it is no longer available to titrate
ozone. When additional VOC is added to the atmosphere, a greater proportion of the NO is
oxidized to NO2, resulting in greater ozone formation. Additionally, anthropogenic sources of
NO result in greater levels of NO2 in the atmosphere. This NO2 is then available for photolysis
to NO and O (Equation 2-1) and, ultimately, for conversion to NO2 (Equation 2-4) and ozone
(Equation 2-2).
The formation and increase in ozone concentrations occur over a period of a few hours as
shown in Figure 2-1. Shortly after sunrise, NO and VOCs react in sunlight to form ozone.
Throughout the morning, ozone concentrations increase while NO and VOCs are depleted.
Eventually, either the lack of sunlight, NO, or VOCs limit the production of ozone. This diurnal
cycle varies greatly depending on site location, emission sources, and weather conditions.
Q.
Q.
C
O
0)
O
C
O
O
C
re
o>
C
O
N
O
200
180
160
140
120
100
80
60
40
20
0
o"
si
Q.
Q.
C
O
e
•4-i
C
0)
o
C
o
o
O
O
0 1 2 34 5 6 78 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time (LST)
Figure 2-1. Average diurnal profile of ozone, NO, and VOC concentrations for August 1995 at
an urban site in Lynn, Massachusetts.
2.1.2 Ozone Precursor Emissions
Precursor emissions of NO and VOC are necessary for ozone to form in the troposphere.
Understanding the nature of when and where ozone precursors originate may help forecasters
factor day-to-day emissions changes into their forecasts. For example, if a region's emissions are
dominated by mobile sources, emissions, and hence ozone that forms, may depend on the
day-of-week commute patterns. This section provides a brief overview of the sources and spatial
distribution of VOC and NOX (NO and NO2) emissions.
2-2
-------
Table 2-1 summarizes the total anthropogenic VOC and NOX emissions in the
United States for 1994. The dominant NOX producers are combustion processes, including
industrial and electrical generation processes, and mobile sources such as automobiles. Mobile
sources also account for a large portion of VOC emissions. Industries such as the chemical
industry or others that use solvents also account for a large portion of VOC emissions.
Table 2-1. Summary of total anthropogenic VOC and NOX emissions in the United States
during 1994 (U.S. Environmental Protection Agency, 1996). Note that 1 short ton
equals 2000 pounds.
Source Type
Fuel Combustion, Electric
Utility
Fuel Combustion,
Industrial
Fuel Combustion, Other
Chemical & Allied
Product Manufacturing
Metals Processing
Petroleum & Related
Industries
Other Industrial Processes
Solvent Utilization
Storage & Transport
Waste Disposal &
Recycling
On-Road Vehicles
Non-Road Sources
Miscellaneous
Total Emissions
NOX Emissions
Emissions
(thousand short tons)
7795
3206
727
291
84
95
328
3
3
85
7580
3095
374
23,666
Percentage of Total
33.0
13.6
3.1
1.2
0.4
0.4
1.4
0.01
0.01
0.4
31.9
13.1
1.6
—
VOC Emissions
Emissions
(thousand short tons)
36
135
715
1577
77
630
411
6313
1773
2273
6295
2255
685
23,175
Percentage of Total
0.2
0.6
3.1
6.8
0.3
2.7
1.8
27.2
7.7
9.8
27.2
9.7
3.0
—
Anthropogenic VOC and NOX emissions are highest near urban areas. Figures 2-2
and 2-3 show the anthropogenic VOC and NO emissions by county across the United States.
Notice that emissions levels correlate well with population levels, which are larger in the eastern
third of the United States and near metropolitan areas.
Along with anthropogenic emissions, the EPA also estimates annual biogenic emissions.
Figure 2-4 shows that biogenic VOC emissions occur mostly in the forested regions of the
United States (Southeast, Northeast, and West Coast regions). Biogenic VOC emissions include
the highly reactive compound isoprene. Biogenic VOC emissions from forested and vegetative
areas may impact urban ozone formation in some parts of the country. Biogenic NOX emissions
levels are typically much lower than anthropogenic NOX emissions levels.
2-3
-------
Emission Density
(short tons/sq, ml.)
£3 > 11
•I 5toM
3 to 5
2 to 3
Otc 2
Figure 2-2. 1996 VOC emissions from anthropogenic sources by county
(U.S. Environmental Protection Agency, 1997a).
Emission Density
(short tons/sq. mi.)
if! > 12
• 5 to 12
3 to 5
2 to 3
Oto2
Figure 2-3. 1996 NO emissions from anthropogenic sources by county
(U.S. Environmental Protection Agency, 1997a).
2-4
-------
Figure 2-4. 1996 VOC emissions from biogenic sources by county (U.S. Environmental
Protection Agency, 1997a).
2.2 PARTICIPATE MATTER
Particulate matter is the general term used for a mixture of solid particles and liquid
droplets found in air. Numerous studies show association between morbidity/mortality and high
PM concentrations. Studies also indicate that short-term exposure to acute PM concentrations
can lead to long-term health effects. The negative health effects associated with high PM
concentrations and the public's desire for accurate air quality information has produced a need
for PM forecasting programs that warn the public one or two days in advance of high PM
concentrations. Since the EPA promulgated a new NAAQS for PM2.5 (PM less than 2.5 jim in
diameter) in 1997 this guidance document has been updated to include information about
forecasting PM2.5 concentrations.
Much of the material summarized in the following sections was drawn from the PM2.5
Data Analysis Workbook (Main and Roberts, 2001), the PM criteria document (U.S.
Environmental Protection Agency, 2001), Seinfeld and Pandis (1998), and documents referenced
therein.
2.2.1 Basic Particulate Matter Chemistry
Particulate matter, unlike ozone, is not a specific chemical entity but is a mixture of
particles of different sizes, shapes, compositions, and chemical, physical, and thermodynamic
properties. Atmospheric PM2.5 results from primary fine particle emissions (emitted directly
from sources) and emissions of gaseous compounds that form secondary aerosols. Secondary
particles are formed from gases through chemical reactions in the atmosphere involving
atmospheric oxygen (62) and water vapor (F^O); reactive species such as ozone (63); radicals
such as the hydroxyl and nitrate radicals; and pollutants such as sulfur dioxide (802), nitrogen
2-5
-------
oxides (NOX), ammonia (NHs), and volatile organic compounds (VOCs) from natural and
anthropogenic sources. 862 forms sulfates, NOX forms nitrates, NHs forms ammonium
compounds, and VOCs form organic carbon compounds. Some particles are liquid; some are
solid. Others may contain a solid core surrounded by liquid. Atmospheric particles contain
inorganic ions (e.g., nitrate, sulfate, sodium), metallic compounds, elemental carbon (EC),
organic compounds, and crustal compounds (e.g., iron, calcium). Some atmospheric particles are
hygroscopic and contain particle-bound water. The organic portion of PM is especially complex,
containing hundreds of organic compounds.
The particle formation process includes nucleation1 of particles from gases emitted from
sources or formed in the atmosphere by chemical reactions, condensation of gases on existing
particles, and coagulation of particles (Figure 2-5). Formation, transport, and removal rates are
all a function of the particle size, chemical constituents of the particles, and meteorological
processes (see Table 2-2).
IX
Ci"
Vapor
Mechanically
Generated
Condensation
i Nucleation
D--G
DC¥ =
0,002
C 1 •
Particle Diameter. D I'L
10
wo
Figure 2-5. Volume size distribution measured in traffic showing fine (including nuclei
and accumulation modes) and coarse particle modes (Wilson and Suh, 1997).
The geometric mean diameter by volume (DGV), equivalent to volume median
diameter, and geometric standard deviation (02) are shown for each mode.
Also shown are transformation and growth mechanisms (e.g., nucleation,
condensation, and coagulation).
1 The formation of new particles.
2-6
-------
Table 2-2. Summary of formation pathways, composition, sources, and atmospheric
lifetimes of fine and coarse particulate matter (from Seinfeld and Pandis,
1998).
Formation pathway
Composition
Sources
Atmospheric lifetime
Travel distance
Fine
Chemical reaction,
nucleation, condensation,
coagulation, cloud/fog
processes
Sulfate, nitrate, ammonium,
hydrogen ion, elemental
carbon, organics, water,
metals (lead, cadmium,
vanadium, nickel, copper,
zinc, manganese, iron...)
Combustion (coal, oil,
gasoline, diesel, wood);
gas-to-particle conversion of
NOX, SOX, and VOCs;
smelters, mills, etc.
Days to weeks
100 to 1000+ km
Coarse
Mechanical disruption, suspension
of dust
Resuspended road dust; coal and oil
flyash; crustal elements (silicon,
aluminum, titanium, iron, usually as
oxides); calcium carbonate, salt;
pollen, mold, spores; plant and
animal debris
Resuspension of industrial soil and
dust; suspension of soil (farming,
mining, unpaved roads), biological
sources, construction/demolition,
ocean spray
Minutes to days
Generally < 100 km
The size of PM ranges from about tens of nanometers (nm) (which corresponds to
molecular aggregates) to tens of microns (70 jim = the size of a human hair) (see Figure 2-6).
The smallest particles are generally more numerous, and the number distribution of particles
generally peaks below 0.1 jim. The size range below 0.1 jim is also referred to as the ultrafme
range. The largest particles (0.1-10 jim) are small in number but contain most of the PM volume
(mass). The peak of the PM surface area distribution is always between the number and the
volume peaks.
0.001
0.01
0.1 1
Particle diamefpr,
Figure 2-6. Distribution of particle number, surface area, and volume or mass with respect to
size (adapted from Husar, 1998).
2-7
-------
PM in the 0.1 to 1 micron size range has the longest residence time (days to weeks)
because it neither settles nor coagulates quickly. Particles in this size range are the most efficient
at penetrating deep into the lung. In addition, the light scattering efficiency per PM mass is
highest at about 0.5 |j,m (see Figure 2-7). This is why, for example, 10 |j,g of fine particles scatter
over 10 times more than 10 |j,g of coarse particles. Thus, PM2.s is important to investigations of
both human health and visibility impairment. Figure 2-7 also shows the absorption efficiency;
there is little variation of the absorption efficiency as a function of particle size.
•ID
IS)
CO
i_
CD
Q.
O
t 5
.a
•a
CO
O)
CD
3=1
w 0
*t—'
O)
t I
SCATTERING/
ABSORPTION^
5 iCT1 2
DIAMETER,
to1
2 4,00
Figure 2-7. Relationship between light scattering, absorption, and particle diameter
(Husar, 1998).
Most secondary fine PM is formed from condensable vapors generated by chemical
reactions of gas-phase precursors (i.e., vapors generated by chemical reactions condense to form
particles). Secondary formation processes can result in either the formation of new particles or
the addition of particulate material to pre-existing particles. Most of the sulfate and nitrate and a
portion of the organic compounds in atmospheric particles are formed by chemical reactions in
the atmosphere. Secondary aerosol formation depends on numerous factors including:
• The concentrations of precursors (which are a function of the proximity of emissions,
wind speed, and mixing height).
• The concentrations of other gaseous reactive species such as ozone, hydroxyl radical,
peroxy radicals, or hydrogen peroxide (which are a function of solar radiation and
temperature).
• Atmospheric conditions, including solar radiation and relative humidity (RH).
• The interactions of precursors and pre-existing particles within cloud or fog droplets or in
the liquid film on solid particles.
2-8
-------
As a result, it is considerably more difficult to relate ambient concentrations of secondary
species to sources of precursor emissions than it is to identify the sources of primary particles. A
diagram of these interactions is provided in Figure 2-8. Details about the chemical reactions can
be found in Seinfeld and Pandis (1998), for example.
Sources Emissions
Chemical Processes
PM
Formation
Sample
PM Loss Collection
Mechanical + Particles \
• Sea salt • NaCI \
• Dust • Crustal \
gases condense
cloud/fog processes
• Motor vehicles
•
•
Industrial
Fires
Other gaseous
•
•
Biogenic
Anthropogenic
• Soot /
• Metals /
• OC /
— > Gases \
• NOX \
• so2 \
• VOCs
• NH3 i
/
—*• Gases /
• VOCs /
• NH3 /
• NOX /
\
/
/
condensation /
coagulation /
' photochemical production
cloud/fog processes
Measurement
\ sedimentation ^Issues:
) (dry deposition) . inlet cut points
' • Vaporization of
wet deposition nitrate> H2°, VOCs
• Adsorption of VOCs
• Absorption of H2O
Meteorological Processes
Winds
Temperature
Solar radiation
Vertical mixing
Clouds, fog
Temperature
Relative humidity
Solar radiation
Winds Temperature
Precipitation Relative humidity
Winds
Figure 2-8. Sources of precursor gases and primary particles, PM formation processes, and
removal mechanisms. Important meteorological measures are provided for each
process.
Sulfates
Sulfates constitute about half of the PM2.5 in the eastern United States. Virtually all the
ambient sulfate is secondary, formed within the atmosphere from 862. About half of the
conversion of 862 to sulfate occurs in the gas phase through photochemical oxidation in the
daytime. NOX and VOC emissions tend to enhance the photochemical oxidation rate. At least
half of the SO2 oxidation takes place in cloud droplets as air molecules pass through convective
clouds. Within clouds, the soluble pollutant gases, such as 862, combine with water droplets and
2-9
-------
rapidly oxidize to form sulfate. SO2-to-sulfate transformation rates peak in the summer due to
enhanced summertime photochemical oxidation and 862 oxidation in clouds.
Conversion of 862 to sulfate occurs at about 1% per hour in cloud-free air, but can
convert to sulfate at 50% per hour in clouds and fog. Removal rates for SC>2 (mostly by dry
deposition) and sulfate (mostly by wet deposition) are about 2 to 3% per hour each. This gives
sulfur (as 862 and sulfate) an atmospheric residence time of from 1 to 5 days, depending on
season, geography, and weather conditions.
Nitrates
Nitrates are a principal component of PM2.5 in the western United States and, like
sulfates, are nearly all formed within the atmosphere from nitrogen oxide emissions. About one-
third of anthropogenic NOX emissions in the United States are estimated to be removed by wet
deposition. NC>2 is converted to nitric acid by reaction with hydroxyl radicals during the day
(oxidation). The reaction of hydroxyl radical with NC>2 is 10 times faster than the oxidation of
SC>2. The peak daytime conversion rate of NC>2 to nitric acid in the gas phase is about 10 to
50% per hour. During the nighttime, NC>2 is converted into nitric acid by a series of reactions
involving ozone and the nitrate radical. Nitric acid reacts with ammonia to form particulate
ammonium nitrate. Thus, PM nitrate can be formed at night and during the day.
Thermodynamically, nitrate formation is favored in cold, moist conditions. Thus, nitrate
formation is enhanced in the winter compared to the summer.
Organic and elemental carbon compounds
Elemental carbon (EC), also called black carbon, is emitted directly into the atmosphere
through combustion processes. Particulate organic carbon (OC) is both directly emitted and
formed in secondary reactions. OC comprises a significant portion of the PIVb.s throughout the
United States. Atmospheric reactions involving VOCs yield organic compounds with low vapor
pressures at ambient temperature (i.e., the vapor or gas condenses to form a liquid). These
reactions can occur in the gas phase, in fog or cloud droplets or in aqueous aerosols. Reaction
products from the oxidation of VOCs also may nucleate to form new particles or condense on
existing particles to form secondary organic PM. Both biogenic and anthropogenic sources
contribute to primary and secondary organic particulate matter. Although the mechanisms and
pathways for forming inorganic (i.e., sulfate, and nitrate) secondary particulate matter are fairly
well known, those for forming secondary organic PM are not as well understood.
Ozone and the hydroxyl radical are thought to be major contributing reactants. However,
other radicals, including nitrate and organic radicals, are thought to contribute to secondary
organic PM formation. Experimental studies of the production of secondary organic PM in
ambient air have focused on the Los Angeles Basin. Evidence shows that secondary PM
formation occurs during periods of photochemical ozone formation in Los Angeles and that as
much as 70% of the organic carbon in ambient PM was secondary in origin during a smog
episode in 1987 (see Turpin et al., 1991). Other experiments showed that 20 to 30% of the total
organic carbon in fine PM in the Los Angeles airshed is secondary in origin on an annually
averaged basis (Schauer et al., 1996). Thus, photochemical reactions are important to secondary
organic carbon PM formation.
2-10
-------
Another formation pathway is the adsorption of semi-volatile organic compounds
(SVOCs, e.g., including polycyclic aromatic hydrocarbons)2 onto existing solid particles. This
pathway can be driven by diurnal and seasonal temperature and humidity variations at any time
of the year. Higher temperatures generally favor the gaseous phase of the SVOCs.
2.2.2 PMi.s Emissions and Sources
The major constituents of atmospheric PM are sulfate, nitrate, ammonium, and hydrogen
ions; particle-bound water; elemental carbon; a variety of organic compounds; and crustal
material (see Table 2-3). These constituents can be primary or secondary. PM is called
"primary" if it is in the same chemical form in which it was emitted into the atmosphere. PM is
called "secondary" if it is formed by chemical reactions in the atmosphere. Primary fine
particles are emitted from sources either directly as particles or as vapors that rapidly condense
to form ultrafme particles (diameters < 0.1 micron) including: soot from diesel engines, a variety
of organic compounds condensed from incomplete combustion or cooking, and metal
compounds that condense from vapor formed during combustion or smelting.
Table 2-3. Major PM2.5 components.
Geological Material - suspended dust consists
mainly of oxides of aluminum, silicon,
calcium, titanium, iron, and other metal oxides.
NaCl - salt is found in PM near sea coasts,
open playas, and after de-icing materials are
applied. The chloride ion can be replaced by
nitrate as a result of reaction during long-range
transport.
Sulfate - results from conversion of 862 gas to
sulfate-containing particles.
Organic Carbon (OC) - consists of hundreds
of separate compounds containing mainly
carbon, hydrogen, and oxygen.
Nitrate - results from a reversible gas/particle
equilibrium between ammonia, nitric acid, and
particulate ammonium nitrate.
Elemental Carbon (EC) - composed of
carbon without much hydrocarbon or oxygen.
EC is black, often called soot.
Ammonium - ammonium bisulfate, sulfate,
and nitrate are most common.
Liquid Water - soluble nitrates, sulfates,
ammonium, sodium, other inorganic ions, and
some organic material absorb water vapor from
the atmosphere.
There are both anthropogenic and natural sources of PM2.5. Anthropogenic emissions
that contribute to ambient PM2.5 concentrations include the following:
• Mobile sources. Gasoline- and diesel-fueled vehicles; resuspended road dust from
vehicle activity on paved and unpaved roads; vehicle tire and brake wear; and off-road
mobile sources such as trains, marine vessels, and farm machinery.
2 The phase in which an organic compound exists in the atmosphere is largely dependent on its vapor pressure. Nonvolatile compounds, such as
polychlorinated biphenyls, exist almost exclusively on particulate matter (i.e., in the "particle phase"), whereas highly volatile compounds, such
as small alkanes, remain in the gas phase. However, due to their intermediate volatility, the SVOCs partition between the gas and particle phases.
2-11
-------
• Stationary sources. Fuel combustion for electric utilities and industrial processes; fuel
combustion for home heating; construction and demolition; metals, minerals,
petrochemical, and wood products processing; mills and elevators used in agriculture;
erosion from tilled lands; food cooking; and waste disposal and recycling.
Numerous natural sources also contribute primary and secondary particles to the
atmosphere:
• Primary sources. Windblown dust from undisturbed land; sea spray; and plant and insect
debris.
• Secondary sources. Oxidation of naturally emitted biogenic hydrocarbons, such as
terpenes, leads to formation of secondary organic PM2.5 and can accelerate the formation
of inorganic secondary PM25, such as ammonium sulfate and ammonium nitrate.
• Natural and man-made sources. Ammonia gas (precursor) and wood-burning particles
(fuel wood burning and forest fires) are potentially important sources of PM25.
Since the precursor gases and PM2.s are capable of long-range transport, it is often
difficult to identify individual sources of PM2.5. It is especially difficult to distinguish
contributions from natural and anthropogenic sources.
The chemical composition of PM2.5 varies both geographically and seasonally
(Figures 2-9 and 2-10). The composition may also vary with the magnitude of the PM2.5 mass
concentrations. There are some relatively common characteristics across much of the United
States, however, including:
• The crustal component is relatively small (<10%) in PM2.5.
• The carbonaceous aerosol component, which consists of both organic material and
elemental carbon, is either the most abundant or second most abundant class of
compounds (35 to 60%).
• Higher sulfate content is observed in eastern United States PM2.5 than in the western
United States PM2.5.
• The nitrate content of the PM2.s in the eastern United States is consistently lower than in
the western United States.
• PM2 5 ammonium concentrations are generally present in sufficient quantities to buffer
the sulfate and nitrate in the aerosol (i.e., to keep the sulfate and nitrate in the particles,
and thus to keep it from volatilizing to form gaseous sulfuric or nitric acids).
• The relative amounts of secondary constituents tend to increase under air pollution
episode conditions.
• Seasonal (and geographical) variations in primary emissions and secondary formation
rates lead to seasonal and geographical differences in composition and concentrations.
For example, wood burning for home heating is an important source of PM2 5 OC in the
winter compared to the summer and in northern states compared to southern states.
2-12
-------
PM2.5Q1 1994-1996
to
PM2.5Q2 1994-1996
V
Figure 2-9. Seasonal maps of PM2.5 mass for 1994-1996 (Falke, 1999).
-------
Carbonaceous
Nitrite
C'rust a I
Sulfate
I Not Chemically
Characterized
Urban
Son Jcoqum Y
4Site- A\-g • ,?•? ug.'rn Ji
Figure 2-10. Ambient PM2.5 composition at urban sites in the United States (U.S.
Environmental Protection Agency, 1998).
2.2.3 Monitoring Issues
Monitor types
Historically, PM mass and composition have been monitored using filter-based methods
with 24-hr (or longer) averages. Inlets to the sampling systems have traditionally been set with
particle diameter cut points of 2.5 or 10 microns (see Figure 2-11). The federal reference
method (FRM) collects PM onto a 47-mm Teflon filter which is later weighed for mass.
Deployment of the continuous PM2.5 monitoring network began in 1999 and is producing
a more detailed understanding of PM2.5 mass characteristics. Two continuous mass
measurements are listed below; both can be operated with either a PMio or a PM2.5 inlet.
• BAM - beta attenuation monitor; beta gauge. Beta ray transmission is measured across a
clean section of filter tape and advanced to the sampling inlet. Air is then drawn into the
sample inlet and PM is deposited on the filter tape. After a set interval (typically 1 hour),
the filter tape is returned to its original location and the beta ray transmission is
re-measured to obtain a difference which is proportional to the PM mass concentration.
2-14
-------
5
b 40
,8f
"u
to
to
Fine-Mode Particles
/ \
0.2
Coarse-Mode Particles
TSP
,.- HiVoi
PWU
WRAC
p i " K / S * t!
Aerodynamic Panicle Diameter i
;--J
- T«ai (TSP:
PW-..
PNU»
Figure 2-11. Idealized distribution of ambient PM showing fine-mode particles and
coarse-mode particles and the fractions collected by size-selective samplers
(Wilson and Suh, 1997). WRAC is the wide range aerosol classifier which
collects the entire coarse mode.
• TEOM - tapered element oscillating microbalance. An inertial mass measurement
technique for making a direct measurement of the particle mass collected on a filter in
real time. The instrument works with an inertial balance that directly measures the mass
collected on a filter by monitoring the frequency changes of a hollow tapered element.
As more mass accumulates on the filter on top of the tapered element, the oscillating
frequency changes, and that change is related to PM mass concentration.
While the above instruments provide the mass of PM in the air during a given period,
chemical speciation is useful to determine the constituents that comprise the PM. Chemical
speciation is still largely filter-based. In the FRM, PM is collected onto Teflon and Quartz filters
which are later analyzed for mass, metals, soluble ions (e.g., sulfate, nitrate), and organic and
elemental carbon. Continuous monitors are now available for black carbon (Aethalometer),
organic and elemental carbon (OC/EC), sulfate, nitrate, and particle scatter (nephelometer -
measures can be related to PM2.5 mass).
Continuous real-time data serve as the cornerstone of a forecasting program and are
needed to develop forecasting methods, monitor current conditions, amend forecasts, and
evaluate forecasting performance. As the new instruments are being deployed and compared to
historical filter samplers, uncertainties and inconsistencies with the data are emerging (see U.S.
2-15
-------
Environmental Protection Agency, 2001, for details). Issues to consider when comparing data
include the following:
• FRM - As PM is deposited on the filters, temperature, humidity, and PM compositional
changes occur throughout the 24-hr sample period and thus some constituents may
volatize (e.g., nitrates, organic carbon compounds, water) or adsorb (e.g., organic carbon
compounds, water) with the changing weather conditions.
o Organic gas adsorption (positive bias) comprised up to 50% of the organic carbon
measured on quartz-fiber filters in southern California (Turpin et al., 1994). This
study also indicated that adsorption was much more important than organic particle
volatilization (negative bias).
o Sampling losses on the order of 30% of the annual federal standard for PM2.5 may be
expected due to volatilization of ammonium nitrate in those areas of the country
where nitrate is a significant contributor to the fine particle mass and where ambient
temperatures tend to be warm (Hering and Cass, 1999).
• TEOM - the TEOM is typically operated at either 30°C or 50°C in an attempt to
minimize the changes in PM mass with respect to humidity. But by heating the PM,
other chemical components may be volatilized (e.g., nitrates, organic carbon compounds)
(Allen et al., 1997) leading to the TEOM under-reporting the PM mass concentration.
This bias for reporting low concentrations typically is worse during colder months when
the ambient temperature is low.
Data quality
According to the EPA, "The purpose of data validation is to detect and then verify any
data values that may not represent actual air quality conditions at the sampling station" (U.S.
Environmental Protection Agency, 1984). Data validation is critical because serious errors in
data analysis and modeling results can be caused by erroneous individual data values. The
EPA's PM2.5 speciation guidance document provides quality requirements for sampling and
analysis. The guidance document also discusses data validation, including the suggested four-
level data validation system. It is the monitoring agency's responsibility to prevent, identify,
correct, and define the consequences of difficulties that might affect the precision and accuracy,
and/or the validity, of the measurements. Once the quality-assured data are provided to data
analysts, additional data validation steps need to be taken. Given the newness and complexity of
the PM2.s mass and speciation monitoring and sample analysis methods, errors are likely to pass
through the system despite rigorous application of quality assurance and validation measures by
the monitoring agencies. Therefore, data analysts should also check the validity of the data
before conducting their analyses.
For a forecasting program, historical data are required to investigate relationships
between meteorology and PM2.5 (or ozone) concentrations. These data need general and specific
checks of consistency and validity prior to their use in analysis (e.g., Main and Roberts, 2001).
To complete these checks, all relevant data need to be gathered including collocated PM2.5
measurements, collocated gaseous (e.g., ozone, NOX) measurements, collocated other PM (or
PM-related) measurements (e.g., bscat, babs, black carbon), speciated PM2.5 and PMio data, and
PM2.5 and PMio mass data (by all monitoring methods).
2-16
-------
General checks include:
• Assess data completeness. Generally, 75% of the data samples are required to make a
valid average (24-hour, seasonal, or annual) to be used in a statistical analysis.
• Graphically review time series of pollutant concentrations. Inspect data spikes, dips, and
outliers. Plot complementary data together (e.g., PM2.5 and PMio, PM2.5 and light
scattering). The first assumption upon finding a measurement that is inconsistent with
physical expectations is that the unusual value is due to a measurement error. If, upon
tracing the path of the measurement, nothing unusual is found, the value can be assumed
to be a valid result of an environmental cause.
• Apply screening criteria.
o PM2.5 mass concentrations should be above 0 or below 200 |ig/m3 in most cases
(concentrations outside this range are considered suspect and require further
inspection including investigation of possible unusual events).
o Three consecutive mass concentrations in continuous data should not be equal. If so,
the data require additional inspection.
o PM2.5 mass concentrations should be less than or equal to PMio mass concentrations.
• Compare data from collocated samplers - between the same sampler type and different
sampler types. Sometimes, seasonal differences (i.e., biases) between samplers are
observed because of differences in sampling methods.
Specific checks include:
• The sum of individual chemical species concentrations should be less than the total
PM2.5 mass concentrations.
• Particle absorption (babs) should correlate well with elemental carbon.
• Cations (sodium, potassium, and ammonium) should compare well to (i.e., balance) the
anions (chloride, nitrate, and sulfate) when measured by equivalents.
• The total of front and back-up filter nitrate should compare well to the front filter nitrate.
A scatter plot of these two measurements will help analysts to better understand possible
nitrate volatilization losses (probably higher in the summer than in winter).
• The front and back-up filter organic carbon should compare well to the front filter
organic carbon. A scatter plot of these two measurements will help analysts to better
understand possible organic carbon adsorption (positive filter artifact).
Once the data have been validated, the analyst can then proceed with more confidence
when generating the relationships between ambient concentrations and meteorology.
2.2.4 Unusual PM Events
Unusual emission events can produce high PM concentrations and reduce visibility.
Sometimes called "exceptional events" for regulatory purposes (U.S. Environmental Protection
2-17
-------
Agency, 1986), these events are caused by uncontrollable and often infrequent activities from
either natural or man-made activities, at different locations, and with different source strengths
and thus are often difficult to predict. Typical unusual events include agricultural burning,
wildland fires, and windblown dust. This section discusses the causes of unusual events and
provides resources and suggestions for trying to understand and forecast these events.
Air quality conditions may be affected in the following ways by these unusual emission
events:
• Local PM2.5 concentrations at 1 or 2 monitors can increase due to localized fires,
agricultural burning, or other activities that may not be representative of the entire
forecast area.
• Transported PM2.5 (smoke) from large wildland fires can be transported hundreds or
thousands of kilometers to a forecast region and can increase the background levels of
PM2.5, thus combining transported PIVb.s with locally-generated PIVb.s to produce a more
severe episode. Depending on concentrations, this transported PM2.5 could produce an
exceedance of the NAAQS and/or degrade visibility.
• Visibility can be significantly affected by fine particles from dust and smoke, which are
efficient scatterers of light.
Table 2-4 further describes the common causes of unusual PIVb.s events, how they affect
PM2.5 mass, and resources to help understand and forecast these events. Due to the infrequent
nature of these events, they are not predicted well by air quality models or other air quality
forecasting tools (e.g., statistical methods). Instead, forecasters should monitor observations
(surface data and satellite images) and wildfire information for evidence of unusual emission
events. Then, once an event has been detected and located, forward trajectories can be used to
estimate the transport direction and potential time the smoke or dust might enter a particular
forecast region (see www.arl.noaa.gov/ready/hysplit4.html for calculating forward and backward
trajectories on the Internet).
To better prepare and forecast for these infrequent and unusual events, consider the
following:
• Determine if these unusual events are likely to occur in a region by examining historical
data and literature.
• Review the data and information sources in Table 2-4 to determine if the real-time data
sources can offer early detection of an event and provide some warning.
• Develop methods and ways to anticipate when these events might occur (for example,
wildland fires are more likely to occur in late summer and fall, African dust has been
shown to impact the southeastern United States during June, July, and August; Prospero,
1999).
Examine satellite data to detect PM from dust storms and wildland fires
(meted.ucar.edu/npoess/nrlsat). Real-time satellite data can be viewed at www.goes.noaa.gov
and historical satellite data can be found at National Oceanic and Atmospheric Administration
(NOAA's) satellite archive (www.saa.noaa.gov).
2-18
-------
Table 2-4. Types of unusual events, how they affect PM concentrations, and a list of resources for acquiring data and
information to forecast these events.
Event
Definition
Burning of farm lands and
farming byproducts (rice
straw, orchard primings, etc.).
How it affects PM
Data and Information Resources on the Internet
Agricultural
burning
Often burned at the end of growing seasons,
smoke from these fires can increase local or
regional PM2 5 concentrations.
EPA resources on agricultural burning
(www. epa.gov/agriculture/tburn. html)
California Air Resources Board's smoke management plan
(www.arb.ca.gov/smp/smp.htm)
Wildland fires
Large fires of 100-1000+ acres
that burn all or most biomass
(trees, shrubs, grasses, etc.).
Biomass burning produces substantial amounts of
PM2 5. Transport of this PM2 5 from ten to
thousands of miles can occur. Smoldering
combustion releases several times more particles
than flaming combustion (Ward, 1999).
NOAA's Operational Significant Event Imagery
(www.osei.noaa.gov/)
National Interagency Fire Center with real-time and historical fire
data and statistics (www.nifc.gov/)
National Fire Weather Center (www.boi.noaa.gov/firewx.htm)
NOAA's Air Resources Laboratory - Wildfire/Forest Fire Smoke
Forecasting (www.arl.noaa.gov/ss/transport/fires.html)
to
Windblown
dust
Local
Locally generated airborne
dust from winds blowing
across agricultural and barren
land.
Strong winds (~6 m/s or more; Saxton et al.,
2000) can cause dust to become airborne. Many
factors influence the amount of PM2 5 and PM10
produced by windblown dust: vegetation cover,
soil moisture, soil particle size distribution,
surface roughness, and changes in wind direction.
Most of the PM generated during dust storms is
larger than 2.5 microns; but Claiborn et al. (2000)
found that PM2 5 was about 30% of the PM10 mass
during windblown dust events in Spokane,
Washington.
NOAA's Operational Significant Event Imagery
(www.osei.noaa.gov/)
NASA site with satellite observations of dust and smoke
(earthobservatory.nasa.gov/NaturalHazards/)
NASA's MODIS Land Rapid Response system for real-time satellite
images (rapidfire.sci.gsfc.nasa.gov/production)
Windblown
dust
Global
High winds cause soil dust to
become airborne. While the
larger dust particles (>10 um)
have high settling velocities
and fall back to earth in 100's
of kilometers, fine particles
have a long lifetime and low
gravitational settling (and
without precipitation) can
remain airborne for weeks and
over 1000's of kilometers
African dust can be transported by the easterly
trade winds to the eastern and southeastern
United States (Prospero, 1999) and can increase
PM2 5 concentrations at the surface and degrade
visibility. The western United States can be
affected by dust transported from Asia (Falke et
al., 2001).
NOAA's Operational Significant Event Imagery
(www.osei.noaa.gov/)
SeaWIFS satellite imagery for tracking large dust storms
(seawifs. gsfc. nasa. gov/SE A WIF S. html)
Naval Research Laboratory's global aerosol modeling system
(www.nrlmry.navy.mil/aerosol/#currentaerosolmodeling)
-------
The types of satellite imagery useful for detecting smoke and dust include:
o Geostationary Operational Environmental Satellites (GOES) visible images with a 1-km
resolution (www.oso.noaa.gov/goes).
o Polar Orbiting Satellites (POES) visible images from satellites that orbit close to the earth
provide higher resolution images than GOES, but less frequently (twice per day).
o Measurements of Pollution in the Troposphere (MOPITT) sensor on NASA's Terra
satellite provide global, yet coarse (22-km horizontal resolution) measurements in the
lower part of the atmosphere (www.eos.ucar.edu/mopitt). Also, NASA's Moderate
Resolution Imaging Spectroradiometer (MODIS) provides an historical archive of global
aerosol measurements from satellite-based instruments (modis-atmos.gsfc.nasa.gov).
2.3 METEOROLOGICAL CONDITIONS THAT INFLUENCE AIR QUALITY
This section presents the types of weather conditions that have a strong influence on
PM2.5 or ozone concentrations. Since daily weather variations best explain the day-to-day
changes in air quality concentrations, understanding how weather influences air quality in a
region is critical for producing accurate air quality forecasts.
Different scales of weather phenomena are important to air quality. The weather
phenomena range from large storm systems that can encompass thousands of kilometers to small
turbulent eddies that are a few meters in size. In general, large-scale weather phenomena are
easier to characterize compared to small ones. In addition, weather forecast models typically do
a better job of predicting large weather phenomena as opposed to small-scale, short lived
phenomena. Therefore, to understand and predict air quality, it is usually best to use a large-
scale to small-scale approach by first understanding the relationship between large-scale weather
features and local air quality, and then understanding the relationship between local weather and
air quality.
Meteorological conditions that strongly influence air quality include: transport by winds,
recirculation of air by local wind patterns, and horizontal dispersion of pollution by wind;
variations in sunlight due to clouds and season; vertical mixing and dilution of pollution within
the atmospheric boundary layer; temperature; and moisture. The variability of these processes,
which affects the variability in pollution, is primarily governed by the movement of large-scale
high- and low-pressure systems, the diurnal heating and cooling cycle, and local and regional
topography.
Figures 2-12 and 2-13 show the general relationships among meteorological phenomena
and air quality. Table 2-5 describes how specific meteorological conditions directly influence
PM2.5 and ozone concentrations. The remainder of this section discusses the key meteorological
phenomena in these figures and tables. Educational resources on basic meteorology are available
on the Internet (Cooperative Program for Operational Meteorology, 2002; University of Illinois
Urbana-Champaign, 2002).
2-20
-------
Ridge of High Pressure
Sinking Motion
Warms, Dries, and Stabilizes
Creates
Temperature
Inversion
Surface High
Clear Skies
Photochemistry
Local Flows and/or Light
Winds, Possible
Transport
Reduces Vertical
Mixing
Stagnation/Recirculation
Poor Air Quality
Figure 2-12. Schematic of the typical meteorological conditions and air quality often associated
with an aloft ridge of high pressure.
Trough of Low Pressure
Rising Motion
Cools, Moistens, and Destabilizes
Surface Low
Cloudy Skies
No
Temperature
Inversion
Reduces Photochemistry
(but may enhance PM25 chemistry)
Enhances Vertical
Mixing
Good Air Quality
Moderate to
Strong Winds
Horizontal Dispersion
Figure 2-13. Schematic of the typical meteorological conditions and air quality often associated
with an aloft trough of low pressure.
2-21
-------
Table 2-5. Meteorological phenomena and their influence on PM2.5 and ozone concentrations.
Phenomena
Aloft Pressure
Pattern
Winds and
Transport
Temperature
Inversions
Rain
Moisture
Temperature
Clouds/Fog
Season
Emissions
No direct impact.
No direct impact.
No direct impact.
No direct impact.
No direct impact.
Warm temperatures are associated
with increased evaporative, biogenic,
and power plant emissions, which act
to increase both PM2 5 and ozone.
Cold temperatures can also indirectly
influence PM2 5 concentrations (i.e.,
home heating on winter nights).
No direct impact.
Forest fires, wood burning,
agriculture burning, field tilling,
windblown dust, road dust, and
construction vary by season.
Chemistry
No direct impact.
In general, stronger winds disperse
pollutants, resulting in a less ideal
mixture of pollutants for chemical
reactions that produce ozone and PM2 5.
Inversions reduce vertical mixing and
therefore increase chemical
concentrations of precursors. Higher
concentrations of precursors can
produce faster, more efficient chemical
reactions that produce ozone and PM2 5.
Rain can remove precursors of ozone
andPM25.
Moisture acts to increase the
production of secondary PM2 5
including sulfates and nitrates.
Photochemical reaction rates for ozone
increase with temperature.
Water droplets can enhance the
formation of secondary PM2 5. Clouds
can limit photochemistry, which limits
ozone production.
The sun angle changes with season,
which changes the amount of solar
radiation available for photochemistry.
Accumulation/Dispersion/Removal
Ridges tend to produce conditions conducive for accumulation of PM2 5 and ozone.
Troughs tend to produce conditions conducive for dispersion and removal of PM and
ozone.
In mountain-valley regions, strong wintertime inversions and high PM2 5 levels may not be
altered by weak troughs. In addition, high PM2 5 and ozone concentrations often occur
during the approach of a trough from the west.
Strong surface winds tend to disperse PM2 5 and ozone regardless of season. However,
strong winds can create dust which can increase PM2 5 concentrations. In the East and
Midwest, winds from a southerly direction are often associated with high PM2 5 and ozone,
due to transport from one region to another.
A strong inversion acts to limit vertical mixing allowing for the accumulation of PM2 5 or
ozone.
Rain can remove PM2 5, but has little influence on existing ozone.
No direct impact.
Although warm surface temperatures are generally associated with poor air quality
conditions, very warm temperatures can increase vertical mixing and dispersion of
pollutants.
Convective clouds are an indication of strong vertical mixing, which disperses pollutants.
No direct impact.
to
to
to
-------
2.3.1 Aloft Pressure Patterns
Aloft large-scale (1000 km or more) atmospheric circulations have a strong influence on
regional and local weather conditions. Meteorologists generally focus on the so-called "500-mb
level" to evaluate the aloft large-scale pressure systems. In particular, they focus on the location,
size, intensity, and movement of 500-mb high-pressure ridges and low-pressure troughs
(mountains of warm air and cold air, respectively). In general, poor air quality conditions are
associated with high-pressure ridges and good air quality conditions are associated with low-
pressure troughs. However, high PM2.5 levels can occur without the existence of aloft ridges,
from a very strong PM2.5 emission source, such as a forest fire. Figure 2-14 shows an example
of a 500-mb ridge over the eastern United States on July 17, 1999, a day with high
PM2.5 concentrations throughout the region, and on September 21, 1999, a day with low
PM2.5 concentrations throughout the region. The existence of ridges and troughs can be
diagnosed by reviewing weather charts, which are widely available as observations and forecasts
on the Internet.
2.3.2 Temperature Inversions and Vertical Mixing
A temperature inversion is a layer of warm air above a layer of relatively cooler air. An
inversion acts to limit the vertical mixing of pollutants, which allows concentrations to build.
Several temperature inversions can exist at different altitudes in the lower part of the atmosphere.
Typically, a temperature inversion can form from 25 to 300 m agl when the ground (and air near
the ground) cools at night, while air above remains warmer. This type of inversion is called a
nocturnal inversion. Nocturnal inversions are strongest when skies are clear at night and in the
winter when nights are long. In the presence of clouds or strong winds, nocturnal inversion are
often weak or do not form at all. Nocturnal inversions trap emissions, released during the
overnight hours, close to the ground. As the ground warms during the day, the air near the
surface warms, which erodes the nocturnal inversion. Typically, a nocturnal inversion
disappears by mid-morning, allowing the trapped pollutants to mix vertically. If a nocturnal
inversion is strong or if solar heating is weak, the inversion may not break until late in the day or
at all. Under these circumstances, pollutants do not mix vertical and high pollutant
concentrations are typical.
When there is an aloft ridge of high pressure over an area, there is often another inversion
above the nocturnal inversion, called a subsidence inversion. A subsidence inversion is caused
by sinking air in the mid- to low-levels of the atmosphere associated with the aloft ridge. As the
air sinks, it warms due to compression. The warmest temperatures associated with the sinking
air are typically found from 500 to 2000 m agl. When there is a strong subsidence inversion as
indicated by aloft temperatures, the daytime heating at the surface may not be strong enough to
break this inversion. Under such circumstances, vertical mixing of pollutants is weak and
pollutants remain trapped near the surface for the entire day. An aloft inversion can also form
when winds transport warm air at a greater rate aloft compared to the surface. This differential
warming typically occurs on the west side of an upper-level ridge, ahead of an upper-level
trough.
2-23
-------
Figure 2-14. 500-mb heights (a) on the morning of July 17, 1999 (1200 UTC) and (b) on the
morning of September 21, 1999 (1200 UTC).
Subsidence inversions do not form when there is an aloft trough over the region. This is
because aloft troughs cause rising motion in the mid- to low-levels of the atmosphere. As the air
rises, it cools due to expansion resulting in cooler air above warmer air. When there is cooler air
above warmer air, the atmosphere is unstable. This instability causes vertical mixing, which
dilutes pollutants whose source is near the surface.
Figure 2-15 shows the diurnal cycle of mixing, vertical temperature profiles, and
boundary layer height on a day with a weak temperature inversion and on a day with a strong
temperature inversion. On the day with the weak inversion, the convective boundary layer grows
rapidly as the sun warms the ground during the day. The rapid growth of the convective
boundary layer is associated with strong vertical mixing and the vertical dispersion of pollutants.
2-24
-------
(a)
D)
Temperature soundings
\
Weak and high inversion
on\3reaks
NBL
_
Midnight
Sunrise
Sunset
to
to
(b)
D)
Strong and low inversion
NBL
Inversion Holds
CBL
RL
NBL
Midnight
Sunrise
RL = Residual Layer
CBL = Convective Boundary Layer
NBL = Nocturnal Boundary Layer
Sunset
= Surface-based mixing depth
= Surface-based vertical mixing
Figure 2-15. Schematic showing diurnal cycle of mixing, vertical temperature profiles, and boundary layer height (a) on a day with a
weak temperature inversion and (b) on a day with a strong temperature inversion. In (a) the pollutants mix into a large
volume resulting in low pollution levels and in (b) pollutants mix into a smaller volume resulting in high pollution levels.
-------
On the day with the strong inversion, the convective boundary layer growth is inhibited. The
limited growth of the convective boundary layer is associated with weak vertical mixing and
limited vertical dispersion of pollutants.
2.3.3 Winds and Transport
Winds can be described as large-, regional-, and local-scale. The large-scale winds are
driven by the pressure gradients between surface high- and low-pressure systems. Light,
regional, surface winds often occur near the center of the surface-high, below the ridge of high
pressure, where pressure gradients are weak. Light winds are not effective at dispersing
pollutants and, therefore, often occur during high pollutant concentrations. Moderate to strong
winds occur between surface high and low pressure systems or near the center of low pressure
systems, provided that moderate to strong pressure gradients exist. Moderate to strong surface
winds act to disperse pollution and thus are typically associated with low pollutant
concentrations. However, high pollutant concentrations can occur during moderate to strong
wind conditions, if the winds transport pollution from one region to another.
In general, surface lows occur under the leading half of aloft troughs (typically on the
eastern side), whereas, surface highs occur under the leading half of aloft ridges. Figures 2-16
and 2-17, respectively, show a 500-mb ridge and an associated surface high and a 500-mb
trough and an associated surface low. The ridge and surface high on January 7, 2002, created
conditions conducive to high PM2.5 concentrations in Salt Lake City, Utah, including light
surface winds and reduced vertical mixing. The trough and surface low on January 22, 2002,
created conditions conducive to low PM2.5 concentrations including strong surface winds, clouds,
and vertical mixing.
Local winds are driven by the interaction between the large-scale pressure patterns and
local forcing mechanisms. The local forcing mechanisms are driven by the diurnal temperature
cycle and topography. Local winds tend to dominate over the large- and regional-scale winds
when the large-scale pressure patterns are weak (i.e., at the center of a surface high pressure).
The local winds may include land breezes, sea breezes, morning downslope flows, afternoon
upslope flows, and terrain channeled flows, which can combine in various ways to recirculate air
and cause stagnation.
2.3.4 Clouds, Fog, and Precipitation
Clouds, rain, and fog all influence pollutant concentrations through a variety of
mechanisms as detailed in Table 2-5. Clouds form when air is cooled and water vapor
condenses. This cooling can be caused by rising motion or contact with a cool surface such as a
body of water or cool land during the night. Rising motion is generated by aloft low-pressure
systems, frontal boundaries, air flowing over mountains, and convective instability (warm air
below cooler air). Clouds are important because they typically reduce the amount of sunlight
available for photochemical reactions that participate in the production of ozone and PM2.5. Fog
is a type of cloud that is in contact with or near the ground. Fog and clouds can dramatically
increase the conversion of sulfur dioxide to sulfate (a secondary type of PM2.s). Precipitation is a
removal mechanism for fine particles.
2-26
-------
Ridge = Sinking = Surface High
500 mb heights on the afternoon
of January 7, 2002 (OOZ Jan 8)
Surface pressure on the afternoon
of January 7, 2002 (OOZ Jan 8)
Figure 2-16. 500-mb heights (left) and surface pressure (right) on the afternoon of
January 7, 2002 (0000 UTC on January 8).
Trough = Rising = Surface low
500 mb heights on the afternoon
of January 22, 2002 (OOZ Jan 23)
Surface pressure on the afternoon
of January 22, 2002 (OOZ Jan 23)
Figure 2-17. 500-mb heights (left) and surface pressure (right) on the afternoon of
January 22, 2002 (0000 UTC on January 23).
2-27
-------
2.3.5 Weather Pattern Cycles
Typically, a region will cycle between a ridge and trough pattern every 2 to 7 days, but
more stationary patterns can develop. Studying and understanding these cycles and their impact
on local weather and air quality will help improve forecasting capabilities. Figure 2-18 shows
the typical life cycle of large-scale weather patterns. The following meteorological descriptions
are generic and may vary from one region to another and between pollutants:
Ridge—high (Figure 2-18a and b) is typically associated with poor air quality.
pressure This pattern occurs about one to two days after a cold front and
pattern trough have passed through an area. As surface high pressure
develops in an area, winds become weak allowing for the
accumulation of pollutants. Warming temperatures increase the
biogenic and evaporative VOCs and lower humidity results in
clearer skies, which are favorable for photochemistry. Sinking
air (subsidence) warms and stabilizes the lower atmosphere,
which suppresses cloud development and mixing. In addition,
an aloft temperature inversion may form that inhibits vertical
mixing and reduces dilution of pollutants. The aloft high
pressure ridge typically occurs west of the surface high and can
be diagnosed using 500-mb height fields.
Ridge—back (Figure 2-18c and d) occurs as the surface high pressure moves
side of high east of the region and the accumulated pollutants are transported
pattern to downwind locations. In some regions, warm air is advected
into the region and winds may increase from a southerly to a
westerly direction depending on the orientation of the high. This
pattern typically produces warm temperatures and relatively
clear skies, even with a low-pressure system approaching from
the west. Pollutant levels can remain high on these types of
days, and the potential for longer-range transport is greater.
Trough— (Figure 2-18e and f) is characterized by a low-pressure system at
cold front the surface and associated cold and warm fronts. Aloft at
pattern 500 mb, a trough of low pressure exists just upstream (west) of
the surface low. This weather pattern produces clouds and
precipitation that reduce photochemistry. Stronger winds and
mixing also reduce pollutant concentrations.
Although aloft ridges and their associated regional and local weather conditions are
generally associated with poor air quality, slight variations in the meteorological processes
described above can have a dramatic affect on the spatial and temporal characteristics of air
quality. It is these variations in meteorological processes that need to be analyzed and
understood for different pollutants, seasons, and regions of interest to better understand the
processes that produce air quality episodes.
2-28
-------
500
I- Isobar I
Study
area
|-Ridge1 -Trough
I Axis ' Axis
I
1500km
Figure 2-18. Life cycle of synoptic weather events at the surface and aloft at 500 mb for (a)
and (b) Ridge—high pressure, (c) and (d) Ridge—back side of high, and (e) and
(f) Trough—cold front patterns. Surface maps show isobars and frontal positions.
The 500-mb maps show contours of equal height.
2-29
-------
3. FORECASTING APPLICATIONS AND NEEDS
The success of an air quality forecasting program depends not only on accurate
predictions, but also on meeting the needs and objectives of forecast recipients. For more than
two decades the public has been warned of unhealthy air in several regions of the United States.
Today, ozone and particulate matter forecasts are used throughout the United States for three
major purposes: (1) public health notification, (2) episodic control programs (such as Ozone
Action Days), and (3) scheduling specialized air monitoring programs. This section describes
how air quality forecasts are used.
3.1 PUBLIC HEALTH NOTIFICATION
Pollution forecasts are typically issued by air quality agencies and communicated via
television, radio, newspapers, the Internet, and fax to the public to give them adequate time to
reduce or avoid exposure to unhealthy air. Forecasts are generally issued each day for the
maximum ozone concentration expected for the current day and next day. For particulate matter,
forecasts typically correspond to a 24-hour averaged concentration, which is standard of the Air
Quality Index (AQI) for reporting PM. For example, the South Coast Air Quality Management
District forecasts daily maximum ozone and average particulate matter concentrations for
40 sub-regions throughout the Los Angeles metropolitan area. For smaller cities, some agencies
forecast the maximum ozone concentrations for the entire city (such as Charlotte, North
Carolina). Air quality forecasts are usually formulated during the morning or early afternoon
and then communicated to the public later the same day.
The needs of public health notification programs vary by region (see Section 5.1), but
generally include:
• Pollution forecasting that errs on the side of public health (i.e., forecasts that tend to
overpredict ozone or PlV^.s rather than underpredict it).
Forecasts that are as localized and specific as possible, particularly for large metropolitan
regions with different geographic, emission, and source areas.
Forecasts that provide the time of day when and location where high levels of ozone
and/or particulate matter are expected.
Forecasts that are completed as early in the day as possible, allowing sufficient time for
public outreach personnel to communicate the forecast and other information to the
public.
3.2 EPISODIC CONTROL PROGRAMS
Reducing air-quality violations and avoiding redesignation to nonattainment or a more
severe classification is the major goal of episodic control programs (Jorquera, 1998; U.S.
Environmental Protection Agency, 1997b). To accomplish this goal, episodic control programs
educate the public about emission-producing activities and seek voluntary action from the public
to reduce emissions on poor air quality days. More than 80 episodic control programs exist
3-1
-------
throughout the United States and have various names, such as Ozone Action Day, Ozone Alert,
Spare The Air, Don't Lite Tonight, and No Burn Days; but the underlying objectives are similar
(see www.italladdsup.gov, www.epa.gov/airnow, and
www.epa.gov/otaq/transp/publicat/pub_volu.htm for links to local programs).
Health officials rely on air quality forecasts (typically ozone and PM during the summer
and particulate matter during the winter) to determine whether or not to call an Action Day and
seek voluntary action from the public to reduce emission-producing activities (e.g., driving,
mowing lawns, residential wood burning, etc.) on forecasted poor air quality days. In addition,
business and industry often participate by offering services that help reduce pollution (e.g., free
bus rides on high ozone days). Since these programs ask the public to reduce pollution
voluntarily, the credibility of the program depends on forecast accuracy.
Typically, forecasters issue ozone forecasts midday or in the afternoon for the next-day's
peak ozone or 24-hr average PM2.5 concentration. Public outreach personnel then communicate
the forecasts and emission-reduction tips for Action Days to the public so they can plan their
activities for the next day (i.e., carpooling). Therefore, forecasters must issue predictions as
early in the day as possible to ensure timely information dissemination.
Episodic control programs typically have the following forecasting needs:
• Minimizing the number of forecasts that falsely alert the public (i.e., minimize
overpredicting). These "false alarms" may cause the public to ignore the warnings and
over time would diminish the effectiveness of the program.
• Receiving forecasts as early as possible to allow sufficient time for public outreach
personnel to communicate the forecast and other information to the public.
• Including a discussion of current and forecasted weather and air quality conditions in the
forecast. Public outreach personnel can use this information to better communicate the
forecast to the media and public.
• Providing an indication of forecast uncertainty so that public outreach personnel can plan
the amount and degree of media and outreach spending necessary given the uncertainties
in the forecast.
3.3 SPECIALIZED MONITORING PROGRAMS
Specialized monitoring programs are field studies run by federal, state, and private
agencies to collect surface and/or aloft air quality and meteorological measurements on poor air
quality days. Personnel for these programs have used ozone forecasts for decades to help
schedule and plan intensive sampling efforts and more recently initiated studies using
PM2.5 forecasts. Since the 1970s, field study personnel have used ozone forecasts to plan when
and where to conduct ozone sampling using expensive measurement equipment (e.g., aircraft,
rawinsondes, etc.). They also use forecasts to help conserve resources by sampling only on high
ozone days and to provide advanced warning to "gear up" for sampling on these days.
Historically, program personnel only needed ozone forecasts for short-term projects lasting
several months during selected study years. Recently, with new continuous monitoring projects
3-2
-------
like Photochemical Assessment Monitoring Stations (PAMS), the need for accurate ozone
forecasts has increased. With the PAMS program, the EPA requires some state agencies to
perform more extensive ozone and ozone precursor monitoring in areas with persistently high
ozone levels. Specialized carbonyl and hydrocarbon monitoring as well as aloft sampling by
aircraft, are performed in many regions only on predicted high ozone days.
Specialized monitoring programs typically have the following forecasting needs:
• Forecasts that are as localized and specific as possible, particularly for large metropolitan
regions so region-specific sampling can be conducted.
• Multi-day forecasts in order to allow sufficient time to prepare monitoring equipment and
personnel.
• Forecast information about when an episode will begin and when it will end, including
the day prior to the episode ("ramp-up" day) when sampling is often conducted to
understand the air quality and meteorological conditions prior to an episode.
In summary, to make air quality forecasts as effective as possible, it is critical that
forecasters understand how air quality forecasts are used in their region. The material provided
in Section 5.1 will help to identify and determine these needs.
3-3
-------
4. DEVELOPING OZONE AND PM2.5 FORECASTING METHODS
Many methods exist for predicting air quality concentrations. Some methods are simple
to develop and easy to operate, yet are not very accurate. Other methods are more difficult to
develop but produce more accurate forecasts. Most forecasters use several methods—some
objective, others subjective—to forecast ozone and PM2.5. Using several methods can balance
one method's strengths with another method's limitations to produce a more accurate forecast.
Since PM2.5 forecasting is new for most agencies, fewer PM2.5 forecasting techniques have been
tested and used.
Section 4.1 describes the most commonly used forecasting methods. Each subsection
defines a method, explains how it works and how to develop it for a particular program, and lists
its strengths and limitations.
All of the methods described here use multiple predictor variables to forecast either
ozone or PM2.5 concentrations. The process of selecting these predictor variables is described in
Section 4.2.
4.1 FORECASTING METHODS
This section presents several of the most common methods used to forecast ozone
concentrations; these can also be used to predict PM2.5 concentrations, provided that long-term
(ideally 3 or more years) of historical PM2.s data are available. For easy comparison, Table 4-1
lists and summarizes the methods.
4.1.1 Persistence
Persistence means to continue steadily in some state. Persistence forecasting is simply
saying that today's or yesterday's pollutant concentration will be the same as tomorrow's
pollutant concentration. Persistence air quality forecasting is best used as a starting point and to
help guide other forecasting methods. In addition, a persistence forecast can be used as a
reference (or baseline) against which to compare forecasts generated from other methods. It
should not be used as the only forecasting method.
How persistence forecasting works
Persistence forecasting works because atmospheric variables, including ozone and PM2.5,
exhibit a positive statistical association with their own past or future values (Wilks, 1995). That
is, large values of a variable tend to be succeeded by large values; likewise, small values of a
variable tend to be succeeded by small values. For example, if today's 24-hour
PM2.5 concentration was 10 ug/m3, it is likely that tomorrow's 24-hour PM2.5 concentration will
also be relatively low. Similarly, if today's 24-hour PM2.5 concentration is 75 ug/m3, it is more
likely that tomorrow's 24-hour PM2.5 concentration will be high than low. Some PM events may
4-1
-------
Table 4-1. Comparison of forecasting methods.
Page 1 of2
Development Effort
Operational Effort
Accuracy
Method Description
Development
Expertise1
Software
/Hardware
Operations
Expertise
Persistence
Low
Low
Low
Today's (or
yesterday's) observed
pollutant
concentration is
tomorrow's
forecasted pollutant
concentration.
-
Spreadsheet/PC
Ability to acquire
today's and
yesterday's pollutant
data.
Climatology
Low/Moderate
Low
Low
Historical frequency
of pollutant events
help guide and
bound pollutant
forecast.
-
Spreadsheet/PC
Ability to acquire
pollutant data and
interpret graphs and
tables.
Criteria
Low/Moderate
Low
Moderate
When parameters
that influence
pollution are
forecasted to reach a
pre-determined level
(criteria), high
pollutant
concentrations are
forecasted.
Ability to identify
key predictor
variables.
Statistical
Software/PC
Ability to acquire
observed and
forecasted
meteorological and
air quality data.
CART
Moderate
Low
Moderate/High
A decision tree
predicts pollutant
concentration based
on values of various
meteorological and
air quality
parameters.
Understanding of
statistics and CART.
CART
Software/PC
Ability to acquire
observed and
forecasted
meteorological and
air quality data and
use a decision tree.
Regression
Moderate
Moderate
Moderate/High
A regression equation
predicts pollutant
concentrations using
observed and
forecasted
meteorological and air
quality variables.
Understanding of
statistics and
regression.
Statistical
Software/PC
Ability to acquire
observed and
forecasted
meteorological and air
quality data and use a
computational program
or spreadsheet.
Neural Networks
Moderate/High
Moderate
Moderate/High
A non-linear set of
equations and
weighting factors
predicts pollutant
concentrations using
observed and
forecasted
meteorological and
air quality variables.
Understanding of
statistics and neural
networks.
Statistical and Neural
Network
Software/PC
Ability to acquire
observed and
forecasted
meteorological and
air quality data and
use a computational
program.
3-D Air Quality
Models
Very High
Moderate/High
Moderate/High
A prognostic modeling
system simulates the
physical and chemical
processes that lead to
the formation and
accumulation of air
pollutants.
High level
understanding of
meteorological and air
quality relationships,
and meteorological,
emissions, and air
quality models.
Prognostic
meteorological model,
emissions model, and
air quality models;
compilers/High speed
computer system with
large memory and disk
storage.
Basic understanding of
meteorological and air
quality relationships to
determine
reasonableness of
model results.
Phenomenological
/Intuition
High
Moderate
High
A person synthesizes
meteorological and air
quality information
including pollutant
concentration
predictions from other
methods to produce a
final air quality
forecast.
Experience in air
quality forecasting and
a conceptual
understanding of
meteorological and air
quality processes.
None
Ability to synthesize
meteorological and air
quality information
including pollutant
predictions from other
methods to produce a
pollutant forecast.
to
All methods require a basic understanding of meteorological and air quality relationships and basic data processing skills.
-------
Table 4-1. Comparison of forecasting methods.
Page 2 of2
Forecast production
time
Data needs
Software
/Hardware
Strengths
Potential
Limitations
Persistence
-------
extend over large regions, and PM2.5 concentrations from one day to the next may be similar and
have persistence.
Air quality forecasting using the Persistence method works because air quality
concentrations are highly dependent on synoptic-scale weather, which typically exhibits similar
characteristics for several days, and, therefore, air quality concentrations are also typically
similar for several days. For example, a high-pressure system will usually persist over an area
for several days during which time weather and air quality concentrations exhibit modest
day-to-day variation. Likewise, if an area is under the influence of a low-pressure system, the
area will likely exhibit low air quality concentrations for several days until the synoptic pattern
changes.
An analysis of the data presented in Table 4-2 illustrates how persistence forecasting
works. Table 4-2 shows peak 8-hr ozone concentrations for a sample city for 30 consecutive
days. Seven days during this period had peak ozone concentrations greater than the federal
8-hr standard and five of these days occurred after an exceedance; thus, the odds of an ozone
exceedance occurring on the day after an exceedance are 5 out of 7 days (71.4%). The odds of a
non-exceedance occurring after a non-exceedance are 20 out of 22 days (90.9%). Therefore, in
this example, if the Persistence method was used to forecast a non-exceedance or an exceedance,
the forecast would be accurate 25 out of 29 days, or 86% of the time. Note that the first day of
the forecast period does not count in the forecast statistics because Day 1 is not a forecast day.
As shown in Table 4-2, the Persistence method does not correctly predict the beginning
or end of an episode. However, the Persistence method can be used to help guide forecasts and
predictions from other methods.
Modifying a persistence forecast with forecasting experience can help improve forecast
accuracy. For example, if today's weather conditions (which included clear skies) were ideal for
high ozone concentrations, today's observed peak ozone concentration reached 130 ppb, and
tomorrow's weather conditions are expected to be the same as today's except for partly cloudy
skies, using the Persistence method the first cut at the forecast is 130 ppb, but the forecast is
lowered to 100 ppb to account for the influence of cloud cover. The Persistence method
provides a good starting point for the next-day ozone forecast.
Persistence forecasting development
Although the Persistence method requires no real development, forecasters should be sure
that the method will work in their area. The following steps describe how to test the
effectiveness of persistence forecasting in a particular area.
1. Create a data set containing at least four years of recent ozone or PM2.5 data.
2. From this data set, use each day's maximum air quality concentration to simulate a
forecast for the next day (i.e., use the Persistence method). Compare the forecast and
observed pollutant concentrations for the historical data set and compute the forecast
verification statistics provided in Section 5.6.
4-4
-------
Table 4-2. Peak 8-hr ozone concentrations for a sample city for 30 consecutive
days. Exceedance days are shown in bold.
Day
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Ozone (ppb)
80
50
50
70
80
100
110
90
80
80
80
70
80
90
110
Day
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Ozone (ppb)
120
110
80
80
70
60
50
50
70
80
80
70
80
60
70
3. Keep in mind the following development issues:
• Consider when the forecast will be issued to determine what ozone or PM2.5 data are
available. For example, if a forecast must be issued at 11:00 a.m. for the next day and
the current day's peak ozone or PM2.5 concentration has not yet been observed, the
previous day's peak ozone concentration would be used for the next-day forecast.
• The Persistence method only works well for regions that experience several
continuous days of similar air quality. This approach fails if pollutant episodes
typically last only one day.
Persistence forecasting operations
Using the Persistence method to forecast pollutant concentrations requires very little
expertise and is perhaps the easiest and quickest of all air quality forecasting techniques, yet its
accuracy is the poorest. Effectively using the Persistence method requires forecasters to
recognize when weather patterns are static and when they are changing. Persistence forecasting
can be effective under static conditions, but generally ineffective under changing conditions.
4-5
-------
Persistence forecasting strengths
• Persistence forecasting can be very accurate during several days with similar weather
conditions.
• It provides a starting point for an air quality forecast that can be refined by using other
forecasting methods.
• It is easy to produce and operate and requires little expertise.
Persistence forecasting limitations
• The first and last days of a pollution episode cannot be predicted using persistence
forecasting.
• This method does not work well under changing weather conditions when accurate air
quality predictions can be most critical.
4.1.2 Climatology
Climatology is the study of average and extreme weather conditions at a given location.
Climatological techniques can be applied to air quality forecasting. Although not very accurate
as a predictive tool, climatology can help forecasters bound and guide their air quality
predictions.
How climatology works
Climatology works because history tends to repeat itself, especially when it comes to
seasonal weather. Since pollutant concentrations are highly weather dependent, air quality
climatologies can be used in the same manner as weather climatologies. For example, an initial
forecast is for a maximum temperature of 105°F in downtown Boston for August 13. According
to a climate table, a maximum temperature of 105°F has never occurred in Boston and the
forecast is probably too high. Thus, the forecast is adjusted down to 100°F. The climate data
acted as a bound and a guide to the temperature forecast. Analogously, let's say that ozone is
being forecasted for April 10 for upstate New York, and the forecast techniques indicate that an
exceedance may occur. A climate table (Table 4-3) shows that upstate New York had no
exceedances in April for the 15-year period of records. Based upon the additional information
provided by the climate table, a non-exceedance is forecasted for April 10. The table served as a
complimentary tool to other forecast methods and helped improve forecast accuracy.
Developing climate tables
Complete the following steps to develop ozone climate tables for a particular region:
1. Create a data set containing at least four years of recent ozone or PM2.5 data.
2. Examine the data for quality and be sure to note when emissions changed significantly
due to the use of reformulated fuel in the area, for example. Changes in emissions can
result in the same weather conditions producing lower pollutant concentrations. Also
4-6
-------
Table 4-3. Annual summaries of 1-hr ozone exceedance days for New York State (1983-1997), (Taylor, 1998).
Year
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
Total
Avg/Year
April
Total Downstate Upstate
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
May
Total Downstate Upstate
0
1
2
2
4
3
0
0
2
1
0
0
0
0
0
15
1
0
0
2
1
4
3
0
0
2
1
0
0
0
0
0
13
1
0
1
0
1
2
2
0
0
0
0
0
0
0
0
0
6
0
June
Total Downstate Upstate
11
5
2
o
5
o
5
9
2
2
o
J
0
0
3
1
0
o
J
47
o
6
10
5
1
o
J
o
J
7
2
1
o
J
0
0
2
1
0
3
41
3
1
0
1
0
1
6
0
1
0
0
0
1
1
0
0
12
1
July
Total Downstate Upstate
8
3
8
4
10
14
4
o
J
5
0
6
5
4
3
4
81
6
8
3
7
3
10
12
4
3
5
0
4
4
2
1
4
70
5
1
0
1
1
1
7
1
0
2
0
2
1
2
2
0
21
2
August
Total Downstate Upstate
10
7
1
0
3
7
1
2
5
1
3
1
3
1
1
46
3
10
7
1
0
0
6
1
2
4
1
3
1
3
1
1
41
o
J
1
0
0
0
3
2
0
0
1
0
0
0
0
0
0
7
1
-------
note that changes in the monitoring network can dramatically change the maximum observed
pollutant concentrations and/or the number of exceedances.
3. Create tables or charts for forecast areas containing the following types of information:
• All-time maximum ozone or PM2.5 concentrations (by month, by site).
• Duration of high ozone or PM2.5 episodes (number of consecutive days, hours of high
pollutant each day).
• Average number of days with high ozone or PM2.5 by month and by week.
• Day-of-week distribution of high ozone or PM2.5 concentrations.
• Average and peak PIVb.s concentrations by holidays and non-holidays, weekends and
weekdays.
Examples of such charts are shown in Figures 4-1 and 4-2 for Pittsburgh, Pennsylvania.
4. If significant changes in emissions occurred, it may be useful to divide the climate tables
or charts into "before" and "after" periods.
5. Examine the tables or charts for usefulness. For example, if there are differences
between weekend and weekday ozone or PM2.5 concentrations or exceedance frequency,
then a climate table showing the frequency of high ozone or PM2.s concentrations by day
of week may be quite useful.
n n -
—
•
~\
— i
~i
L
-^
DGood
• Mod.
DUSG
• UH
May
Jun
Jul
Sep
Figure 4-1. Monthly distribution of the average number of days that PM2.s concentrations fell
into each AQI category from 1999 to 2001 based on the peak 24-hr average
PM2.5 concentrations measured at 12 sites in the greater Pittsburgh region.
Mod. = Moderate, USG = Unhealthy for Sensitive Groups, and UH = Unhealthy.
Note that the totals for each month are less than 30 or 31 due to missing data (Dye
et al., 2002a).
4-8
-------
10.0-
8.0-
6.0 -
4.0 -
2.0-
1
f|
1
-
-
"
-H
L
.-I
"L
-
,— 1
,
1
DGood
D Mod
DUSG
• UH
Sun Mon Tue Wed Thu Fri
Sat
Figure 4-2. Summertime day-of-week distribution of the average number of days per year that
PM2.5 concentrations fell into each AQI category from 1999 to 2001 based on the
peak 24-hr average PM2.5 concentrations measured at 12 sites in the greater
Pittsburgh region. Mod. = Moderate, USG = Unhealthy for Sensitive Groups, and
UH = Unhealthy. Note that the totals for each day are different due to missing
data (Dye et al., 2002a).
Climatology in operations
Using climate tables does not require much expertise. The forecaster need only
understand that the tables are tools to guide and bound the air quality forecasts created using
other methods. Consulting climate tables may be useful when other methods predict extreme
events and they may help determine if a forecast of extreme concentrations is warranted. The
forecaster can also use climatological information in the forecast discussion to provide context.
For example, "Tomorrow's predicted peak ozone concentration is 150 ppb; this would be the
first time in two years that ozone has reached this level."
Climatology strengths
• Climatology acts to bound and guide an air quality forecast produced by other methods.
• It is easy to develop.
Climatology limitations
• Climatology is not a stand-alone forecasting method but a tool to complement other
forecast methods.
• It does not account for abrupt changes in emission patterns such as those associated with
the use of reformulated fuel, a large change in population, forest fires, etc.
4-9
-------
4.1.3 Criteria
A criterion is a principle by which something is evaluated. The Criteria method in air
quality forecasting uses threshold values (criteria) of meteorological or air quality variables to
forecast pollutant concentrations. Sometimes called "rules of thumb," the Criteria method is
commonly used in many forecasting programs as a primary forecasting method or combined
with other methods. It serves as a fundamental method on which to start an air quality
forecasting program.
How the Criteria method works
This method is based on the fact that specific values of certain meteorological and air
quality variables are associated with high pollutant concentrations. Once known, forecasters can
look for the occurrence of the criteria in weather forecasts and predict pollutant concentrations
from them. For example, high pollutant concentrations are often associated with hot
temperatures and, thus, temperature can be used as one predictor of pollutant concentration. For
instance, historical analysis may show that a temperature at or above 90°F is required to have an
8-hr ozone concentration greater than 85 ppb in a particular area. Thus, 90°F would be a
threshold value (criterion) for an 8-hr ozone exceedance.
Since ozone and PM2.5 formation is complex, forecasters must use several variables and
associated criteria to accurately forecast ozone or PIVb.s. Table 4-4 shows an example of multi-
parameter criteria used to forecast ozone concentrations in Austin, Texas. This table indicates
the conditions necessary for a 1-hr ozone exceedance. To have an exceedance in Austin in July,
the predicted maximum temperature must be at least 92°F, the temperature difference between
the morning low and afternoon high must be at least 20°F, the average daytime wind speed must
be less than 5 knots, the afternoon wind speed must be less than 7 knots, and the prior day's peak
1-hr ozone concentration must be at least 70 ppb. Note that the meteorological criteria are
predicted values for the next day. If these conditions are not met, then an exceedance is less
likely and, thus, would not be forecasted.
The Criteria method is best suited to help forecast an exceedance, non-exceedance, or
pollution in a particular AQI category range rather than an exact concentration.
Criteria method development
Complete the following steps to develop the Criteria method for air quality forecasting in
a particular region:
1. Determine the important physical and chemical processes that influence ozone or
PM2.5 concentrations in the area. This helps with identification of variables to use for the
criteria. Literature reviews, historical case studies, and climatological analysis (as
discussed in Section 5.2) can help with this step.
2. Select variables that represent the important physical and chemical processes that
influence ozone or PM2.5 concentrations in the area. Useful variables may include:
maximum temperature, morning and afternoon wind speed, cloud cover, relative
humidity, 500-mb height, 850-mb temperature, etc. Statistical software can be used to
4-10
-------
Table 4-4. Criteria for 1-hr ozone exceedances in Austin, Texas, used by the Texas Natural
Resource Conservation Commission (Lambeth, 1998).
"Month
Apr
May
Jun
Jul
Aug
Sep
Oct
Daily
Temp
Max
(above °F)
78
84
84
92
92
87
87
Daily
Temp
Range
(above °F)
20
20
20
20
20
18
18
Daily
Wind
Speed
(below knots)
8.0
8.5
6.0
5.0
5.0
5.0
5.0
Wind
Speed
15-21 UTC
(below knots)
6.0
10.0
9.0
7.0
7.0
7.0
5.0
Prior Day's
Ozone 1-hr
Max
(above ppb)
70
70
70
70
70
75
75
limit the number of variables by identifying the most important and significant ones. A
discussion of variable selection is presented in Section 4.2.
3. Acquire at least four years of recent ozone or PM2.s data and surface and upper-air
meteorological data.
4. Determine the threshold value for each parameter that distinguishes high and low
pollutant concentrations. For example, create scatter plots of ozone or PM2.5 vs.
particular parameters to help determine the thresholds, as shown in Figure 4-3. In this
example, the criterion of 28°C (81°F) helps distinguish higher ozone concentrations from
lower concentrations in Charlotte, North Carolina. With the criterion of 28°C, only two
ozone concentrations greater than or equal to 85 ppb occur when the temperature is less
than the criteria. However, many low ozone concentrations (less than 85 ppb) occur
when the maximum temperature is greater than or equal to 28°C, thus criteria for other
variables (wind speed, cloud cover, etc.) are needed to accurately identify high ozone
days.
5. Use an independent data set (i.e., a data set not used for development) to evaluate the
selected criteria.
6. Keep in mind the following development issues:
• Evaluate threshold values for each month or season to understand how the values
change.
• When emissions compositions change, the peak pollutant concentration associated
with an established criteria may change. When this happens, the criteria method
should be updated or the exceptional event should be noted.
Criteria method operations
The Criteria method is one of the easiest methods to use. Data must only be acquired and
checked against the established criteria to determine the air quality forecast. Although use of this
4-11
-------
to
a.
a.
o
o
o
o
g
CO
E
X
IV
E
15
c
o
£
>.
13U ~
125 -
120 -
115 -
110 -
105 -
100 -
95 -
90 -
85 "
80 -
75 -
70 -
65 -
60 -
55 -
50 -
45 -
*r\ -
Step 1
•
+
* *
V * *•;;
***** **~ '
» »* **
•
^ »
»
»
criteria
»
. *
^ »
• • •
4
* * »* **
«_« A >
* * * »
V >^ *** *
* * *^ * * *
* * * *
« f •
»***
* *
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
Maximum surface temperature (°C) at Charlotte
Figure 4-3. Scatter plot of maximum surface temperature and regional maximum 8-hr ozone concentration in
Charlotte, North Carolina in 1996 (MacDonald et al., 1998).
-------
method does not require an understanding of meteorology and air quality processes, it is
advisable that someone with such knowledge be involved in the development of the method and
check the air quality predictions for physical reasonableness.
Criteria method strengths
• It is easy to operate.
• It is relatively easy to develop, and it can be refined each year as more knowledge is
acquired.
• It is an objective method that alleviates potential biases arising from human subjectivity.
• It complements other forecasting methods. This method can easily be used first to
determine whether or not the situation warrants spending more time on fine-tuning the
forecast or using more sophisticated methods.
Criteria method limitations
• Selection of the variables and their associated thresholds is subjective.
• It is not well suited for predicting exact pollutant concentrations; but better suited for
forecasting pollutant concentrations above or below a certain concentration or AQI
category.
• This is an objective tool that can only predict pollutant concentrations based on
information contained within the observed and forecasted data. Changes in the predicted
weather conditions may not be reflected in the predictor variables and may cause
uncertainty in the air quality predictions.
4.1.4 Classification and Regression Tree (CART)
Classification and Regression Tree (CART) is a statistical procedure designed to classify
data into distinct (or dissimilar) groups. For air quality forecasting, CART enables a forecaster
to develop a decision tree to predict pollutant concentrations based on the values of predictor
variables that are well correlated with pollutant concentrations.
How CART works
CART uses software to develop a decision tree by continuously splitting peak pollutant
concentration data into two groups based on a single value of a selected predictor variable
(Horie, 1988; National Research Council, 1991; Stoeckenius, 1990). The selected predictor
variable and the threshold cutoff value are determined by the CART software. The software
identifies the variables with the highest correlation with the pollutant. It seeks to split the data
set into the two most dissimilar groups. The splitting of the data set and tree development
continues until the data in each group are sufficiently uniform. Predictor variables used in
CART typically include meteorological data (i.e., temperature, wind speed, cloud cover, etc.),
but may also include air quality data or other data such as the day of week or length of day. See
Section 4.2 for a list of common predictor variables.
4-13
-------
Figure 4-4 shows a decision tree for maximum ozone concentrations created using
CART. This decision tree was developed by Horie (1988) for the South Coast Air Basin in
California. As discussed by Horie, of the 73 variables used in the analysis, the temperature at
850 mb describes the greatest amount of the variance in maximum ozone concentration; it was
used as the first data split. This split resulted in the two most dissimilar groups: Group 1 (for
850-mb temperature less than 17.1°C) had an average ozone concentration of 90 ppb, and
Group 2 (for 850-mb temperature greater than 17.1°C) had an average concentration of 230 ppb.
CART was then applied to each group using the same set of 73 predictor variables. The low
ozone Group 1 was split again by 850-mb temperature at 9.9°C, while the high ozone Group 2
was split by 900-mb temperature at 24.3°C. The tree growth continued until there were 10
distinct groups. In this example, the entire decision tree explains 80% of the variance in the
daily maximum ozone concentration.
It is quite simple to forecast pollutant concentrations using the decision tree created by
the CART analysis. For the example shown in Figure 4-4, if the forecasted predictor variables
include an 850-mb temperature of 20°C, 900-mb temperature of 23°C, and southeast morning
winds at Los Angeles International Airport, then the expected ozone concentration would be
182 ppb, as determined by the 1988 decision tree.
Note that slight differences in the predicted variables can produce significant changes in
predicted pollutant levels. For example, if the predicted 900-mb temperature were 25°C instead
of 23°C, the predicted ozone would have been 230 ppb instead of 182 ppb. Careful evaluation of
the accuracy and quality of the predicted weather conditions is needed to ensure an accurate
prediction.
Since this decision tree was developed in 1988, ozone concentrations in the South Coast
Air Basin have dropped dramatically (SCAQMD, 1997) due in part to changes in fuels and
automobile control technologies. To account for changes of vehicle mix and other emissions
changes, a decision tree should be updated frequently.
CART development
Complete the following steps to develop a decision tree using CART:
1. Determine the important physical and chemical processes that influence pollutant
concentrations in a particular area in order to identify the key variables. Literature
reviews, historical case studies, and climatological analysis (as discussed in Section 5.2)
can help with this step.
2. Select variables that properly represent the important physical and chemical processes
that influence pollutant concentrations in the area. A discussion of variable selection is
presented in Section 4.2.
3. Create a multi-year data set of the selected variables. Choose recent years that are
representative of the current emission profile. Reserve a subset of the data for
independent evaluation of the method.
4. Use CART software to create a decision tree on the multi-year data set.
5. Evaluate the decision tree using an independent data set.
6. When emissions compositions change, the pollutant concentrations associated with the
established criteria may change. When this happens, the decision tree should be updated.
4-14
-------
LEGEND
N = Number of Days
Mean O3 = Average Peak Ozone Concentrations (pphm)
O3S.D. = Standard Deviation of O3
LAX7D = Mornina Winds at Los Anaeles International Airport Max °z
DL = Day Length (Mrs) - l 096
850T = 850mb Temperature (°C) _ '
900T = 900mb Temperature (°C) ™ ™ ^ 3 _ ^
TOPT = Top of two inversion temperatures (°C) I3
r 850T
N
Mean 03
03 S.D.
^^^^—^^^
' 850T < 17.1 N
N = 760
Mean 03 = 90
03 S.D. = 46
^^ ' ^^
< 9.9 ^ f 850T > 9.9 ""]
386 N = 374
= 63 Mean 03 = 130
= 23 03 S.D. = 47
( 850T > 17.1 N
N = 336
Mean 03 = 230
03 S.D. = 58
^^---^^ \
r 900T < 24.3 ^1 C 900T > 24.3 A
N = 234 N = 102
Mean 03 = 210 Mean 03 = 280
03 S.D. = 49 03 S.D. = 51
/\ /'\ /^\ /"\
DAYL < 10.8 N
N = 187
Mean 03 = 50
0, S.D. = 14
V )
DAYL > 10 . 8 ^ f DAYL < 10 . 6 \f DAYL > 10 . 6 ] C LAX7D
N = 199 N = 91 N = 283 N
Mean 03 = 75 Mean 03 = 70 Mean 03 = 130 Mean 03
03 S.D. = 24 03 S.D. = 24 03 S.D. = 42 03 S.D.
v J V J v J V
= E to SE V LAX7D = W to NW \ f NZJ7D ± West ^ ( NZJ7D = West s
= 87 N = 147 N = 24 N =78
= 182 Mean 03 = 230 Mean 03 = 230 Mean 03 = 289
43 03 S.D. 44 o S.D. = 46 °3 S.D. = 44
J V J V J V J
850T < 12.7 ^
N = 113
Mean 03 = 111
03 S.D. = 33
V J
' 850T > 12.7
N = 170
Mean 03 = 149
03 S.D. = 41
^ TOPT < 19.9 ^
N =23
Mean 03 = 183
03 S.D. = 40
V J
' TOPT > 19.9 A
N = 124
Mean 03 = 235
0, S.D. = 40
V J
Figure 4-4. Decision tree for daily maximum ozone concentrations in the South Coast Air Basin in the
Los Angeles, California, area (Horie, 1988).
-------
CART operations
The CART method is very easy to use and requires little expertise. Data needed for the
decision tree must be acquired and processed through the tree to determine the ozone forecast.
Use of this method does not require an understanding of meteorology and air quality processes.
However, it is advisable to have someone with meteorological experience evaluate the CART
pollutant predictions for reasonableness.
CART strengths
• Requires little expertise to operate on a daily basis; runs quickly.
• Complements other subjective forecasting methods.
• Allows differentiation between days with similar pollutant concentrations if the pollutant
concentrations are a result of different processes. Since PM can form through multiple
pathways, this advantage of CART can be particularly important to PM forecasting.
CART limitations
• Requires a modest amount of expertise and effort to develop.
• Slight changes in predicted variables may produce large changes in the predicted ozone
or PM2.5 concentrations.
• This is an objective tool that can only predict pollutant concentrations based on
information contained within the observed and forecasted data. Changes in the predicted
weather conditions may not be reflected in the predictor variables and may cause
uncertainty in the pollutant predictions.
• CART may not predict pollutant concentrations during periods of unusual emissions
patterns due to holidays or other events; however, human forecasters can account for
these changes and their potential impact on pollutant concentrations.
4.1.5 Regression Equations
Regression is a statistical method for describing the relationship among variables. For air
quality forecasting, regression equations are developed to describe the relationship between
pollutant concentration (referred to as the predictand, what is being predicted) and other
predictor variables (e.g., temperature, wind speed, etc.). Regression equations have been
successfully used to forecast pollutant concentrations in many areas of the country (Cassmassi,
1987; Dye et al., 1996; Hubbard and Cobourn, 1997; Ryan, 1994).
How regression equations work
If two variables are correlated, a line or a curve can describe the relationship between
those variables using a mathematical equation. With this equation, pollutant concentrations can
be predicted from other variables. Multi-linear regression is most commonly used to forecast
ozone (Equation 4-1). However, curvilinear regression (Equation 4-2) is useful in ozone
4-16
-------
forecasting because it captures the non-linear relationships of ozone and predictor variables. The
same approach can be applied to predicting PM2 5 concentrations.
O3 = ci Vi + c2 V2 cn Vn + constant (4-1)
O3 = ci Vi+ c2 V22+ c3 V33 , cn Vnn + constant (4-2)
where:
O3 = predictand (could also be PMio, PM2.5, or other pollutants)
c = coefficients (weighting factors)
V = predictor variables
An example of a multi-linear regression equation is shown in Equation 4-3. This model
was developed for forecasting wintertime PM2.5 concentrations for Salt Lake City, Utah (Dye et
al., 2002b). This equation predicts the next-day's average 24-hr PM2.s concentration. Table 4-5
describes the variables used.
PM2.5 = 53.429 + 3.382*Holiday - 0.189*Precip - 0.31*Tmax
- 0.541*SurfaceWS + 1.008*(T@700mb - Tmin)
0.838*(Stability) + 0.183*Td@700mbOOZ - 0.292*WS@850mbOOZ
(4-3)
Table 4-5. Variables used in regression Equation 4-3.
Variable
Holiday
Precip
Tmax
SurfaceWS
T@700mb
Tmin
Stability
Td@700mbOOZ
WS@850mbOOZ
Description
1 for Valentine's Day, Martin Luther King, Jr. Day, Presidents' Day,
Veterans' Day, and Super Bowl Sunday.
2 for Thanksgiving weekend and Christmas Eve through New Year's Day.
1 for weekends immediately preceding or following any of the above
holidays.
0 for all other days.
Forecasted precipitation in inches during the 24-hr forecast period.
Forecasted daytime maximum temperature (°F)
Average resultant wind speed from 12Z to OOZ (0500 to 1700 MST)
Temperature at 700 mb at 12Z (0500 MST) (°C)
Forecasted or observed minimum temperature (°C)
Temperature at 700 mb at OOZ (1700 MST) (°C) minus the forecasted
daytime maximum temperature (°C) at the surface
Dew-point temperature at 700 mb at OOZ (1700 MST) (°C)
Wind speed at 850 mb at OOZ (1700 MST) (m/s)
To use the equation, a forecaster simply inputs the forecasted values into the equation.
Notice that the model uses only weather variables. Thus, a forecaster can use input values from
the 24-hr weather forecasts to make the next-day PM2 5 forecasts.
4-17
-------
Regression equation development
Complete the following steps to develop a regression model for ozone or
PM2.5 concentrations:
1. Determine the important physical and chemical processes that influence ozone or
PM2.5 concentrations in a particular area. Literature reviews, historical case studies, and
climatological analysis (as described in Section 5.2) can help with this step.
2. Select variables that represent the important physical and chemical processes that
influence ozone or PM2.5 concentrations in the area. Statistical software can be used to
limit the number of variables by identifying the most important ones. A discussion of
variable selection is presented in Section 4.2.
3. Create a data set of ozone or PM2.5 values and selected predictor variables. Choose a
minimum of three recent years that are representative of the current emissions profile.
Randomly select about 25% of the data and set them aside for independent evaluation
(Step 5).
4. Use statistical software to calculate the coefficients and a constant for the regression
equation. The process is straightforward and is likely described in the statistical software
manual.
5. Perform an independent evaluation of the regression model using the verification
statistics listed in Section 5.6. Evaluate the performance of the regression equations
using a data set other than the developmental data set.
6. Other development issues to consider include:
• Ozone and PM2.5 are often log-normally distributed; yet regression is best suited for
predicting data that are normally distributed.
• Use the natural log of ozone or PM2.5 concentrations as the predictand to improve
performance.
• Regression tends to predict the mean better than the tails (i.e., high pollutant
concentrations) of the distribution. Creating secondary regression equations to
predict only the high pollutant concentrations may improve forecast accuracy. These
secondary equations can be used when the primary equation reaches a specified
concentration level.
• Be careful not to "over fit" the model by using too many prediction variables. An
"over-fit" model will decrease the forecast accuracy. A reasonable number of
variables to use in predicting ozone or PM2.5 is 5 to 10.
• One variable can likely represent a whole subset of variables. Unique
(i.e., dissimilar) variables should be used to avoid redundancy and co-linearity.
• Stratifying the data set may improve regression performance. Consider dividing the
data set by seasons, weather type, or other meteorological variables. For example,
separate equations might be developed for spring, summer, and fall.
4-18
-------
Regression equation operations
Compared to the relatively extensive effort required to develop regression equations,
operation of the model requires modest expertise. Running the forecast equation only requires
data input into a simple computational program or spreadsheet that contains the regression
equation(s). Although use of the equation(s) does not require an understanding of meteorology
and air quality processes, it is advisable that someone with meteorological experience check the
pollutant prediction for physical reasonableness.
Because the predictor variables are forecasted, they have inherent uncertainty, which
results in an air quality forecast that has a degree of uncertainty. To help quantify this
uncertainty the input values can be altered slightly and the effect this has on the air quality
forecasted can be evaluated.
Regression equation strengths
• Regression analysis is well documented and widely used in a variety of disciplines. It has
been successfully used in ozone forecasting is many areas of the country (Cassmassi,
1987; Dye et al., 1996; Hubbard and Cobourn, 1997; Ryan, 1994).
• Regression software is widely available and runs on a personal computer. It is generally
easy to use.
• Regression is an objective forecasting method that reduces potential biases arising from
human subjectivity.
• Regression can properly weight relationships that are difficult to subjectively quantify.
• Regression analysis can be used in combination with other forecasting methods, or it can
be used as the primary forecasting method.
Regression equation limitations
• Regression equations require a modest amount of expertise and effort to develop.
• Regression equations tend to predict the mean better than the tails (i.e., the highest
pollutant concentrations) of the distribution. They will likely underpredict the high
concentrations and overpredict the low concentrations.
4.1.6 Artificial Neural Networks
Artificial neural networks are computer algorithms designed to simulate biological neural
networks (e.g., the human brain) in terms of learning and pattern recognition. Artificial neural
networks have been under development for many years in a variety of disciplines to derive
meaning from complicated data and to make predictions. In recent years, neural networks have
been investigated for use in pollution forecasting (Comrie, 1997; Gardner and Dorling, 1998;
Ruiz-Suarez et al., 1995). Artificial neural networks can be "trained" to identify patterns and
extract trends in imprecise and complicated non-linear data. Because ozone and PM2.5 formation
are complex non-linear processes, neural networks are well suited for ozone and PM2.5
4-19
-------
forecasting. Note that neural networks require about 50% more effort to develop than regression
equations and provide only a modest improvement in forecast accuracy (Comrie, 1997).
How artificial neural networks work
Artificial neural networks use a complex combination of weights and functions to convert
input variables (such as wind speed and temperature) into an output prediction (such as ozone or
PM2.5 concentration). Figure 4-5 shows the artificial neural network components. A forecaster
supplies the neural network software with meteorological and air quality input data. The
software then weights each datum and sums these values with other weighted datum at each
hidden node. The software then modifies the node data by a non-linear equation (transfer
function). The modified data are again weighted and summed as they pass to the output node.
At the output node, the software modifies the summed data using another transfer function and
then outputs an ozone or PM2.5 prediction. The neural network software offers several choices
for transfer functions.
INPUT LAYER
Meteorological
and Air Quality
input data
HIDDEN LAYER
OUTPUT LAYER
Processing at each hidden node: \ »
1. Weight input variables and sum.
Sj = A1W1+A2W2
2. Transform sum using non-linear
equation.
Pollutant
Prediction
Processing at the output node:
1. Weight the transformed hidden layer "
variables and sum.
S.+l = 8^3+8^4
2. Transform this sum using non-linear
equation and output pollutant predictiorj
P=F(S:+1)
Figure 4-5. A schematic of an artificial neural network (Comrie, 1997).
4-20
-------
Commercial software can be purchased to help forecasters develop and operate a neural
network. Before a prediction can be made, the network software must be "trained" and
developed. Complete the following steps to train artificial neural networks:
1. Supply the software with historical meteorological and air quality data for the input layer.
2. Supply the software with the historical ozone or PM2.5 data.
3. The software establishes nodes within the hidden layer. It then iteratively adjusts the
weights until the error between the output data and the actual data (observed) is
minimized.
4. Neural networks typically use a backpropagation algorithm to adjust the weights to
minimize the error. The error information propagates back through the network. The
software first adjusts the weights between the output layer and the hidden layer and then
adjusts the weights between the hidden layer and the input layer. With each iteration, the
software adjusts the weights to produce the least amount of error in the output data. This
process "trains" the network.
Once the network has been trained (i.e., developed) it can be used to operationally
forecast ozone or PM2.5 concentrations.
Three data sets are required to train a neural network to achieve good generalization on
new data: a developmental set, a validation set, and a test set. The developmental set is used to
develop the neural network. The validation set is used to determine when the network's general
performance is maximized. And the test data set is used to evaluate the trained network.
It is important not to over train the neural network on the developmental data set because
an over trained network will predict ozone or PM2.5 concentrations based on random noise
associated with the developmental data set (Gardner and Dorling, 1998). When presented with a
new data set the network will likely give incorrect output since the new data's random noise will
be different than the random noise of the developmental data set.
Developing artificial neural networks
Complete the following steps to develop neural networks to forecast ozone or PM2.5.
1. Complete historical data analysis and/or literature reviews to establish the air quality and
meteorological phenomena that influence ozone or PIVb.s concentrations in a particular
area. A detailed discussion of this process is contained in Section 5.2.
2. Select parameters that accurately represent these phenomena. Be sure to select
parameters that are readily available on a forecast basis. A detailed discussion of variable
selection can be found in Section 4.2.
3. Confirm the importance of each meteorological and air quality parameter using forward
stepwise regression, for example. See Section 4.2 for more details.
4. Create three data sets: (1) a data set to train the network, (2) a data set to validate the
network's general performance without over fitting the data, and (3) a data set to evaluate
the trained network. The developmental data set should contain at least three years of
data. The validation and evaluation data sets should each contain about one year of data.
4-21
-------
However, with today's changing emissions, a five-year-old data set may have
significantly different characteristics than a current data set.
5. Train the data using neural network software. Be sure not to over train the network as it
must be general enough to work well on new data sets. As the network is trained, use the
validation data set to determine when the network's general performance is maximized.
See Gardner and Dorling (1998) for details.
6. Test the generally trained network on a test data set to evaluate the performance. If the
results are satisfactory, the network is ready to use for forecasting. A discussion of
forecast accuracy and performance is presented in Section 5.6.
Artificial neural networks operations
Compared to the development of the network, the operation of the network is
straightforward and requires little expertise. Data only needs to be acquired and input into the
input layer of the neural network. Although use of the network does not require an
understanding of meteorology and air quality processes, it is advisable that someone with
meteorological experience be involved in the development of the method and evaluate the
pollutant prediction for reasonableness.
Strengths of artificial neural networks
• Ozone and PM2.5 formation are non-linear processes. This method can weight
relationships that are difficult to subjectively quantify and neural networks allow for non-
linear relationships between variables.
• Neural networks should predict extreme values more effectively than regression
equations, provided that the network developmental set contains such outliers.
• Once a neural network is developed, a forecaster does not need specific expertise to
operate it.
• Neural networks can be used in combination with other forecasting methods, or it can be
used as the primary forecasting method.
Limitations of artificial neural networks
• Neural networks are complex and not commonly understood; thus, the method can be
inappropriately applied and difficult to develop.
• Neural networks do not extrapolate data well. Thus, extreme pollutant concentrations not
included in the developmental data set will not be taken into consideration in the
formulation of the neural network prediction.
4.1.7 Deterministic Air Quality Modeling
Deterministic air quality modeling attempts to mathematically represent the important
processes that affect ambient air quality. Air quality modeling actually requires a system of
models that work together to simulate the emission, transport, diffusion, transformation, and
removal of air pollution. These models include:
4-22
-------
• Meteorological models - These models forecast meteorological conditions that determine
transport and mixing; and influence chemistry, emissions, and deposition.
• Emissions models - These models simulate the temporal, spatial, and chemical
distribution of emissions of the pollutant in question, and/or its precursors, from both
anthropogenic and natural sources.
• Air quality models - These models use the forecasts from meteorological and emissions
models to simulate the transport, diffusion, transformation by chemical reaction, and
removal of air pollution. A more detailed discussion of air quality models can be found
in Seinfeld and Pandis (1998).
Historically, deterministic air quality models have been used in air quality planning to
estimate the impact of population growth and emission controls on future air quality. For air
quality planning, these models have been and continue to be used in case study analyses to
understand air pollution processes and to estimate the effects of emissions changes on pollutant
concentrations during episodic conditions. These modeling analyses may take years to set up,
evaluate, and complete. In recent years, high performance computing at low cost has become
available and air quality models have been used with forecasts from prognostic meteorological
models to produce daily air quality forecasts. While it is possible to forecast pollutant
concentrations with simple one-dimensional air quality models, three-dimensional (3-D) air
quality models that simulate the complex interaction of physical and chemical processes are
more suitable for forecasting air quality because they handle multiple processes and allow for
future improvements to the model components.
How 3-D air quality modeling systems work
To predict pollutant concentrations with a 3-D air quality model, meteorological factors,
and emissions must be predicted first. An air quality modeling system links the meteorological,
emissions, and air quality models together to make air pollution forecasts. Figure 4-6 shows
how these components are related in an air quality model system.
Meteorological models
Prognostic meteorological models solve sets of equations that represent fundamental
atmospheric behavior. During the past 10 years, prognostic mesoscale modeling has become an
increasingly common method of developing inputs for air quality modeling. Many newer air
quality models use mesoscale meteorological models as their preferred meteorological driver and
are well suited for air quality forecasting. Currently, the two most widely used meteorological
models for air quality applications are:
• The Penn State/NCAR Mesoscale Model version 5 - MM5 (Grell et al., 1994)
• The Regional Atmospheric Modeling System - RAMS (Pielke et al., 1992)
The most common approach to prognostic meteorological modeling for air quality
applications is to use National Center for Environmental Prediction (NCEP) operational forecast
model (Eta, AVN, NGM, etc.) analysis fields to provide initial and boundary conditions for
MM5. When used for air quality forecasting, the NCEP forecast meteorological fields can be
4-23
-------
Prognostic
Meteorological
Model
Initial Performance
Conditions .'Evaluation
Meteorological
Observations
Topographic and
Landuse Data
Emissions, Industry,
and Human Activity
Data
Air Quality
Observations
Figure 4-6. Schematic showing the component models of an air quality modeling
system.
used for MMS's boundary conditions. In the forecast mode, the MM5 modeling systems can
make weather predictions for use in an air quality model.
Emissions models
Emissions modeling is the process of estimating emissions with the spatial, temporal, and
chemical resolution needed for air quality modeling. The emission inventory includes data for
mobile sources, stationary point sources, area sources, and natural sources. Mobile, biogenic,
and some point/area source emissions can vary substantially with temperature. Mobile source
and some industrial/commercial emission sources also exhibit significant variations by day of the
week. Currently, there are three emissions modeling systems commonly used to provide
Eulerian air quality models with emissions input data.
• Emission Processing System (EPS 2.0) (U.S. Environmental Protection Agency, 1992)
• Emissions Modeling System (EMS-95 - EMS-2002) (Bruckman, 1993)
• Sparse Matrix Operator Kernel Emissions (SMOKE) modeling system (Coats, 1996)
When performing emissions modeling to support air quality forecasting, environmental
(i.e., temperature and solar radiation) and day-of-week effects need to be taken into account.
These effects may be included in pre-computed, model-ready emissions inputs for various cases,
or may be calculated at run-time based on the predictions of the prognostic meteorological
model. Of the currently available emissions models, only SMOKE has been used to calculate
emissions at run-time in a real-time forecasting system (McHenry et al., 1999).
4-24
-------
Air quality models
Three-dimensional air quality models are classified as being either Lagrangian or
Eulerian depending on the method used to simulate the time-varying distribution of pollution
concentrations. Lagrangian (trajectory) models follow individual air parcels over time using the
meteorological data to transport and diffuse the pollutants; they may also include chemical
transformations. This approach is computationally efficient when treating a limited number of
emission sources. However, it is difficult to properly characterize the interaction of a large
number of individual sources when nonlinear chemistry is involved and these models have
limited usefulness in forecasting secondary pollutants. Eulerian models use a grid of cells
(vertical and horizontal) where the chemical transformation equations are solved in each cell and
pollutants are exchanged between cells. These models can produce three-dimensional
concentration fields for several pollutants but require significant computational power.
Typically, the computational requirements are reduced through the use of nested grids, with a
coarse grid used over rural areas and a finer grid used over urban areas where concentration
gradients tend to be more pronounced.
The Hybrid Single-Particle Lagrangian Integrated Trajectories with a generalized non-
linear Chemistry Module (HY-SPLIT CheM) model is an example of a Lagrangian model used
to forecast air quality on a regional scale (Stein et al., 2000). However, because of the limited
ability of Lagrangian models to handle the large number of emission sources needed for urban-
scale forecasting, Eulerian models are often used to forecast air quality.
The typical Eulerian air quality model is bounded on the bottom by the ground, on the top
at some specified height, and at some distance on all four sides, depending on the size of the
meteorological modeling domain and the area of interest. The volume of the modeling domain is
divided into grid cells. For regional air quality models, the grid cells are typically on the order of
tens of kilometers in length and width (4 to 36 km are common) with 5 to 50 vertical layers. The
grid cell size is chosen to maximize the resolution for a given computational budget. Smaller
grid cells will result in higher resolution and generally greater model accuracy, but at a higher
computational cost. Within each grid cell a series of physical and chemical processes are
simulated as shown in Figure 4-7.
Currently available Eulerian air quality models include.
• The California Grid Model: CALGRTD (Yamartino et al., 1992).
• The Comprehensive Air Quality Model with Extensions: CAMx (Environ, 1998).
• The Community Multiscale Air Quality model: CMAQ (U.S. Environmental Protection
Agency, 1998).
• The Multiscale Air Quality Simulation Platform: MAQSIP (Odman and Ingram, 1996).
• The SARMAP Air Quality Model: SAQM (Chang et al., 1996).
• The Urban Airshed Model with Aerosols: UAM-AERO (Lurmann, 2000; Lurmann et al.,
1997).
• The Urban Airshed Model with Carbon Bond IV Chemistry: UAM-IV (Morris et al., 1990).
• The Variable Grid Urban Airshed Model: UAM-V (Systems Applications International,
1999).
4-25
-------
RISING MIXING
Figure 4-7. Schematic illustration of the processes in an Eulerian photochemical model cell
(Schere andDemerjian, 1984).
Setting up a 3-D air quality model
Substantial staff and computer resources are needed to establish a scientifically sound and
automated air quality forecast system based on a 3-D air quality model. Even when using
existing meteorological, emissions, and air quality models, the effort to integrate and refine the
entire system enough to produce reliable forecasts may be large.
To implement a 3-D air quality modeling system for predicting air quality in a region, the
following steps are suggested.
1. Design and plan the system
• Decide on which pollutants to forecast.
• Define modeling domains (meteorology and air quality) considering geography and
emission sources. The definition should include horizontal and vertical resolution of
the models.
• Select component models considering forecast pollutants, domains, component model
compatibility, availability of interface programs, and available resources.
• Determine hardware and software requirements for the system.
• Identify sources of needed meteorological, emissions, and air quality data, and
methods for acquiring these data.
4-26
-------
• Prepare a detailed plan for acquiring and integrating data acquisition, modeling, and
analysis software.
• Identify what the final products will be.
• Plan for continuous real-time evaluation of the modeling system.
• Prepare a reasonable implementation schedule that plans for problems.
2. Identify and allocate the resources needed
• Staff for system implementation and operations.
• Computing and storage consistent with the selection of domains and models.
• Communications for data transfer into and out of the modeling system. Sufficient
network bandwidth must be provided for downloading external data from the Internet
and transferring data within the local network to ensure other operations are not
affected.
3. Acquire required geophysical data
• Topographical data
• Land use data
4. Implement the data acquisition and processing tools, component models (emissions,
meteorological, and air quality), and analysis programs.
• Implement each program individually.
• Use standard test cases to verify correct implementation.
5. Develop the emission inventory
• Acquire needed emission inventory related data.
• Review the emissions data for accuracy. Errors in the emissions data can result in
large errors in the air quality model output.
• Be sure that the emissions data reflect the most recent emissions data available unless
there are known problems that indicate earlier data is more suitable.
• Project emissions to the current year using growth and control factors.
• Update the base emission inventory annually unless annual projections are to be
included in the run-time processing of emissions. If projections are to be included in
the run-time processing, growth and control factors should be reviewed and updated
annually.
6. Test the data acquisition and processing tools, component models, and analysis programs
• Test the operation of all data acquisition programs, preprocessor programs,
component models, and analysis programs as a system using real data sources.
4-27
-------
• Review the prognostic meteorological forecast data for accuracy over several weeks
under various weather patterns. Errors in the meteorological input field can result in
large errors in the air quality output. If the meteorological model shows persistent
and significant errors (particularly winds) it may not be reasonable to continue system
implementation without first resolving the source of errors.
• Run the combined meteorological/emissions/air quality modeling system in a
prognostic mode using a variety of meteorological and air quality conditions.
Evaluate the performance of the modeling system by comparing it with observations.
Refine the model application procedures (i.e., the methods of selecting boundary
conditions or initial concentration fields, the number of spin-up days, the grid
boundaries, etc.) to improve performance. It may be necessary to refine the options
or inputs for any one of the component models that make up the system.
7. Integrate data acquisition and processing tools, component models, and analysis
programs into an operational system
programs into an operational system
• After achieving satisfactory results in the testing phase, implement automated
processes for data acquisition, the daily data exchange from the prognostic
meteorological model and the emissions model to the 3-D air quality model and
analysis programs, and forecast product production.
• Implement automated processes by using scripting and scheduling tools. Since the
entire modeling process will take a significant amount of time to complete, there
should be some method of tracking the progress of the modeling.
• Verify that the forecast products reflect the actual model predictions.
8. Test, evaluate, and improve the integrated system
• Run the model in real-time test mode for an extended period. Compare output to
observed data and note when there are model failures.
• After obtaining satisfactory results on a consistent basis, use the modeling system to
forecast pollutant concentrations.
• Document the modeling system.
• Continuously evaluate the system's performance by comparing observations and
predictions.
• Implement improvements as needed based on performance evaluations and new
information.
Three-dimensional air quality modeling system operations
Operation of the 3-D air quality modeling system should be completely automated.
Ideally, the forecaster should only need to review the model forecast for reasonableness.
However, in reality there will be times when parts of the system will fail. Therefore, operational
procedures should be established for monitoring of the models as they execute and for recovery
from failures.
4-28
-------
The forecaster should continuously monitor model performance statistics and graphics to
identify persistent errors or biases. Problems with the model predictions that are significant
should be discussed with the system developer.
Strengths of 3-D air quality models
• Three-dimensional air quality forecast models are phenomenological based, simulating
the physical and chemical processes that result in the formation and destruction of air
pollutants.
• They can forecast for a large geographic area.
• They can predict air pollution in areas where there are no air quality measurements.
• The model forecasts can be presented as maps of air quality to show how predicted air
quality varies over a region hour by hour. The maps can be animated to show where air
pollution is expected to form and how it will evolve over the course of a day.
• Three-dimensional air quality forecast models can be used to further understand the
processes that control air pollution in a specific area. For example, they can be used to
assess the importance of local emission sources or long-range transport.
Limitations of 3-D air quality models
• Inaccuracies in the prognostic model forecasts of wind speeds, wind directions, extent of
vertical mixing, and solar insulation may limit 3-D air quality model performance. Small
discrepancies in winds over 24-hr to 48-hr periods can produce significant shifts in the
spatial pattern of predicted ozone concentrations over a region.
• Emission inventories used in current models are often out of date and based on uncertain
emission factors and activity levels. Three-dimensional air quality forecast model
accuracy depends on accurate emission inventory modeling.
• Site-by-site ozone concentrations predicted by 3-D air quality forecast models may not be
accurate due to small-scale weather and emission features that are not captured in the
model.
4.1.8 The Phenomenological/Intuition Method
Phenomenological/intuition forecasting involves analyzing and conceptually processing
air quality and meteorological information to formulate an air quality prediction.
Phenomenological/intuition forecasting can be used alone or with other forecasting methods such
as regression or criteria. Although intuition is commonly defined as "the perception of truth or
fact, independent of any reasoning," for air quality forecasting intuition is the perception of truth
or fact (the prediction) derived from reason (the conceptual processing of meteorological and air
quality data).
This method is heavily based on the experience provided by a meteorologist or air quality
scientist who understands the phenomena that influence ozone or PlV^.s. This method balances
4-29
-------
some of the limitations of objective prediction methods (i.e., criteria, regression, CART, and
neural networks).
How the Phenomenological/Intuition method works
This method depends on an individual's capabilities and/or experience in three major
areas:
1. Understanding the processes that influence ozone or PM2.5. The basic component to
phenomenological/intuition forecasting is developing a robust and accurate conceptual
understanding of the important phenomena that control pollutant concentrations. This
conceptual understanding should include information on synoptic, regional, and local
meteorological conditions, plus air quality characteristics in the forecast area.
2. Synthesizing information. Vast amounts of data are needed to forecast air pollution.
Forecasters will analyze both observed and forecasted weather charts, satellite
information, air quality observations, and pollutant predictions from other methods. Each
piece of information or prediction from other methods must be evaluated and given a
relative weight.
3. Developing a consensus. Some information or data will likely be contradictory and
should be dismissed. For example, available weather data may be conducive to high
ozone concentrations (light winds, clear skies, and high temperatures) and forecasting
criteria may suggest high ozone concentrations, yet the regression equation may predict
only modest ozone concentrations. The forecaster must take into account the historical
performance of each method/data source, accept some, reject others, and issue the ozone
forecast based on a general agreement of the forecasts.
Phenomenological/Intuition method development
The fundamental step in developing a Phenomenological/Intuition forecasting method is
acquiring a conceptual understanding of the important physical and chemical processes that
influence ozone or PM2.5 concentrations in an area. Literature reviews, historical case studies,
and climatological analysis (as discussed in Section 5.2) can help with this. Although much
knowledge can be gleaned from these sources, the greatest benefit to the method is the
development of intuition, which only comes from forecasting experience.
Phenomenological/Intuition method operations
Compared to other forecasting methods, phenomenological/intuition forecasting requires
a high level of expertise. The forecaster needs to have a strong understanding of the processes
that influence ozone or PM2.5 concentrations and needs to apply this understanding on a daily
basis. Typically, the forecaster will evaluate meteorological forecast models and use pattern
recognition that equates the meteorological fields to ozone or PIVb.s concentrations. For
example, the forecaster may observe a high-pressure ridge building into the forecast area and
equate this with high PM2.5 concentrations. The forecaster will repeat this process for several
other meteorological and air quality data fields, weigh the combined influence of these fields,
and output a forecast. For example, some predictor variables may indicate high pollutant
4-30
-------
concentrations, while others indicate moderate and low pollutant concentrations. By processing
all of this information in the conceptual model, the forecaster develops a pollutant prediction.
Strengths of the Phenomenological/Intuition method
• The Phenomenological/Intuition method allows for easy integration of new data sources.
For example, if a new wind monitor is installed, the forecaster can quickly make use of
this additional data. Whereas, other objective methods, such as regression, may require
re-creation of forecasting algorithms.
• The Phenomenological/Intuition method allows for the integration and selective
processing of large amounts of data in a relatively short period of time.
• This method can be immediately adjusted as new truths are learned about the processes
that influence ozone or PM2.5.
• The effect of unusual emissions patterns associated with holidays and other events can
easily be taken into account.
• Extreme or rare events may be more accurately forecasted. Generally, objective methods
such as regression or neural networks do not capture extreme or rare events.
• The Phenomenological/Intuition method is a good complement to other more objective
forecasting methods because it tempers their results with common sense and experience.
Limitations of the Phenomenological/Intuition method
• The Phenomenological/Intuition method requires a high level of expertise. The
forecaster needs to have a strong understanding of the processes that influence ozone or
PM2.5 concentration and needs to apply this understanding in both the developmental and
operational processes of this method.
• Since the Phenomenological/Intuition method is subjective, forecaster bias is likely to
occur. Using an objective method as a complement to this method can alleviate these
biases.
4.2 SELECTING PREDICTOR VARIABLES
Many of the methods discussed in Section 4.1 use predictor variables to forecast ozone or
PM2.5. This section provides guidelines for selecting the candidate predictor variables to use in
ozone or PM forecasting efforts. Tables 4-6 and 4-7 list common predictor variables that
influence ozone and PM2.5 concentrations. Consider the following issues when selecting
predictor variables.
• Understand the phenomena. Before selecting particular variables it is important that
forecasters understand the phenomena that affect ozone or PIVb.s concentrations in their
region. This understanding can be gained through review of past air quality studies in the
area, conducting a historical analysis of meteorology and ozone or PM2.5, and/or doing a
literature review as described in Section 5.2.
4-31
-------
Table 4-6. Common predictor variables used to forecast ozone.
Variable
Maximum temperature
Morning wind speed
Afternoon wind speed
Cloud cover
Relative humidity
500-mb height
850-mb temperature
Pressure gradients
Length of day
Day of week
Morning NOX
concentration
Previous day's peak
ozone concentration
Aloft wind speed and
direction
Usefulness
Highly correlated with ozone and ozone
formation
Associated with dispersion and dilution of
ozone precursor pollutants
Associated with transport of ozone
Controls solar radiation, which influences
photochemistry
Surrogate for cloud cover
Indicator of the synoptic-scale weather
pattern
Surrogate for vertical mixing
Causes winds/ventilation
Amount of solar radiation
Emissions differences
Ozone precursor levels
Persistence, carry-over
Transport from upwind region
Condition for
High Ozone3
High
Low
-
Few
Low
High
High
Low
Longer
-
High
High
-
Relative condition is location- and season-dependent.
Table 4-7. Common predictor variables used to forecast PM2.5.
Variable
500-mb height
Surface wind speed
Surface wind direction
Pressure gradient
Previous day's peak
PM2.5 concentration
850-mb temperature
Precipitation
Relative Humidity
Holiday
Day of week
Usefulness
Indicator of the synoptic-scale weather
pattern
Associated with dispersion and dilution of
pollutants
Associated with transport of pollutants
Causes wind/ventilation
Persistence, carry-over
Surrogate for vertical mixing
Associated with clean-out
Affects secondary reactions
Additional emissions
Emissions differences
Condition for
High PM2.5a
High
Low
-
Low
High
High
None or light
High
-
-
Relative condition is location- and season-dependent.
4-32
-------
Capture the important phenomena. The variables selected should capture the important
phenomena that affect pollutant concentrations in the region. For example, research may
show that high background PM2.5 concentrations are needed to produce high
PM2.5 concentrations in the forecast area. Thus, using yesterday's PM2.5 concentration as
a surrogate for background PM2.s concentration may improve forecast accuracy.
Select observed and forecasted variables. Predictor variables can consist of observed
variables (e.g., yesterday's ozone or PIVb.s concentration) and forecasted variables
(e.g., tomorrow's maximum temperature). Using forecasted predictor variables is critical
since tomorrow's pollutant concentrations are more strongly related to tomorrow's
weather conditions than to today's or yesterday's pollutant concentrations.
• Ensure data availability and reliability. Make sure that data is easily obtainable from
reliable source(s). Ensure that data will be available by a specified time every day, so
that forecasts can be issued in a timely manner. For example, if a forecast needs to be
issued for tomorrow's ozone or PM2.5 concentration by 1 100 LST, all predictor variables
and data must be available before 1 100 LST.
Using the above guidelines and understanding the mechanisms and phenomena that
influence ozone or PM2.5 concentrations, a forecaster might select as many as 50 to 100 variables
for consideration. These variables are the starting point for the statistical analysis, but will need
to be reduced to a smaller number of the most useful variables.
Statistical analysis techniques can be used to identify the most significant variables.
Following is a list of the types of statistical analyses that can be performed. For further details
on statistical methods, see Wilks (1995).
• Cluster analysis is a method used to partition data into similar and dissimilar subsets.
Many of the variables may be somewhat similar (e.g., maximum surface temperature and
900-mb temperature), and cluster analysis can be used to identify these similarities. One
variable can likely represent a whole set of similar variables. Unique (i.e., dissimilar)
variables should be used to avoid redundancy. Statistical software can be purchased and
used for performing cluster analysis.
• Correlation analysis is used to evaluate the relationship between the predictand
(i.e., pollutant levels) and various predictor variables. Correlations range from +1
(high-positive relationship) to 0 (no relationship) to -1 (high-negative relationship).
Variables used for this type of analysis should have a high-positive or high-negative
correlation. A high-positive correlation indicates that increases in the variable are
associated with increases in the next day's pollutant concentration. Some of the variables
may be both similar and highly correlated, for example, maximum surface and 850-mb
temperatures. In this case, one or two variables would suffice. Spreadsheet programs
(Excel, Lotus, etc.) or statistical software can be used to calculate correlation.
• Step-wise regression is an automatic procedure that allows the statistical software
(SAS, Statgraphics, Systat, etc.) to select the most important variables and generate the
best regression equation. When using this approach, it is important to question and
evaluate the results. A common problem with this technique is that the resulting
4-33
-------
regression equations may contain too many variables that cause them to over fit the data,
producing inaccurate predictions.
• Human selection is another means of selecting the most important predictor variables. A
forecaster can visually evaluate the relationship among variables using scatter plot
matrices, for example.
This selection process results in a series of key variables that can be used with the
forecast methods described in Section 4.1 to predict pollutant concentrations in a forecast region.
4-34
-------
5. STEPS FOR DEVELOPING AN AIR QUALITY FORECASTING PROGRAM
This section describes the major steps for setting up and operating an air quality
forecasting program. For each step, major issues are identified and suggestions are provided for
resolving them. Understanding the users' needs (Section 5.1) and understanding the processes
that control air quality (Section 5.2) are the first steps to developing a forecasting program.
Information to help forecasters choose one or more forecasting methods is presented in
Section 5.3. Section 5.4 identifies the types and sources of air quality and meteorological data.
Section 5.5 explains the importance of having a forecasting protocol. Section 5.6 explains how
to evaluate the quality of a pollutant forecast.
5.1 UNDERSTANDING FORECAST USERS' NEEDS
The success of an air quality forecasting program depends partly on accurate predictions,
but also on meeting the needs and objectives of forecast users. As discussed in Section 3, air
quality forecasts are used for three major purposes: public health notification, episodic control
programs (Action Days), and scheduling specialized monitoring programs. The questions
provided below are designed to help identify forecast users' needs.
• Who will use the forecast? Forecasters who understand their audience will have more
insight into potential ways to improve the forecast.
• For how many months are forecasts needed? Understanding how long the ozone or
PM2.5 season lasts helps forecasters plan the resources (labor and data) needed for air
quality forecasting. The analysis techniques described in Section 5.2 can help determine
the length of the season for each pollutant.
• What periods should a forecast cover? Typically, air quality forecasts are made for the
current- and next-day periods; however, they can be extended to include two- to five-day
predictions. Note that longer-range predictions will likely be less accurate.
• Are three-day forecasts needed for weekend/holiday periods? During weekends and
holidays staff may be unavailable to produce daily forecasts. In this case, two- and/or
three-day forecasts may be needed to cover these periods. A plan should be in place to
handle the situation if conditions change appreciably from initial forecasts.
• When should forecasts be issued to ensure meeting public outreach deadlines?
Preparatory work is needed to communicate forecast information to the public,
particularly during air quality episodes. Issuing forecasts as early in the day as possible
helps ensure that they can be effectively communicated to the public.
• Should forecasts be re-issued? If so, under what conditions? Sometimes weather
conditions change rapidly after a forecast has been issued. Re-issuing an ozone or
PM2.5 forecast may improve the forecast accuracy, but could lead to public confusion and
questioning of the forecast and possibly jeopardize the credibility of the outreach
program.
5-1
-------
• What are the accuracy requirements? For example, is an error of±20ppb acceptable?
What about missing a forecast by two AQI categories? It is important to understand the
error tolerance of forecast users. Exceeding this threshold can lead to reduced credibility.
• Are air quality forecasts issued for maximum regional concentrations or for site-
specific maximums? Forecasting difficulty and uncertainly is greater for smaller
forecasting regions. It is more difficult to make a site-specific forecast than a regional
one. The user's forecast needs and tolerance for accuracy must be balanced with the
resources available to produce the forecast. Most air quality agencies issue regional
ozone and PM2.5 forecasts.
• Are PM2.5 forecasts issued for 24-hr average PM2.s concentrations or for sub-daily
averages? Since the AQI for PM2.s and PMio is based on a 24-hr average, forecasts are
generally made for the AQI standard. However, when conditions are rapidly changing,
forecasting PM2.5 on a shorter averaging interval may be more desirable for the end-users
of the forecast.
• Should forecasts be made for specific concentrations or concentration ranges
(e.g., AQI categories)? Generally, forecasts used for public health notification are
provided in concentration ranges or AQI categories, which allow for easier forecast
interpretation. However, agency personnel who make decisions about specialized
sampling (e.g., collecting VOC measurements) may benefit from specific concentration
forecasts.
• Should a forecaster hedge high or low? Ideally all forecasts would be accurate. In
reality, all forecasts contain uncertainty. Hedging a forecast either high or low allows the
forecaster to account for conditions that could influence pollutant concentrations such as
fronts or winds. It is important to identify whether forecast users want to minimize "false
negatives" (forecast for low ozone, but actually observe high ozone) or "false positives"
(forecast for high ozone that doesn't occur).
• What types of interactions with forecast users are needed? In addition to receiving
forecasts, forecast users may benefit from a brief discussion with the forecaster. This
discussion allows the forecaster to pass on verbally any details or uncertainty about the
forecast and any possible scenarios that might result in changes to the forecasted values.
• Are written forecast discussions of predicted weather and air quality conditions
required? These discussions provide additional information to help users interpret the
predicted pollutant values. Written explanations can convey fine points and uncertainties
about the forecast.
• How should forecasts be disseminated? Many methods exist for disseminating forecasts
(fax, phone, e-mail, Internet, pager, etc.). Identify appropriate primary and secondary
(backup) methods to disseminate forecasts to users.
• How should missed forecasts be handled? Missed forecasts, particularly large misses,
should be examined and discussed with forecast users. By identifying and explaining the
causes of error, forecasters can learn from past mistakes, and users can better understand
the forecast process and its limitations.
5-2
-------
5.2 UNDERSTANDING THE PROCESSES THAT CONTROL AIR QUALITY
The next step in developing an air quality forecasting program is understanding how and
why pollution forms in an area. Section 2 provides a general discussion of the chemical
processes and weather phenomena that influence ozone and PM2.5 concentrations. This section
presents methods and examples to help with identification and understanding of the processes
and phenomena that influence ozone and PM2.5. Understanding these processes and phenomena
will improve a forecaster's capabilities. Common methods for developing this understanding
include reviewing literature from past research and conducting data analyses.
5.2.1 Literature Reviews
The most efficient and generally the easiest way to start understanding pollution in a
particular area is by reviewing existing literature on the topic. Ozone pollution has been studied
for three decades, and scientists have produced a plethora of papers and reports for most areas of
the country. Although PM2.5 pollution has also been studied for many years, few areas are
forecasting for PM2.5.
Articles published in the Journal of Applied Meteorology and Atmospheric Environment
often contain very pertinent information. Other literature sources, such as reports from
local/regional ozone studies that may be available through government agencies, should also be
considered. Broadening the literature review to include nearby regions may provide important
information that is directly applicable to air quality processes in the forecast area.
Some good general reference sources include:
• National Research Council (1991) - Explains how tropospheric ozone forms and
provides details about ozone chemistry.
• Seinfeld and Pandis (1998) - Provides a basic overview of atmospheric chemistry (ozone
and PM2.s) and describes how meteorology affects atmospheric chemistry.
• Wallace and Hobbs (1977) - Provides general meteorological information about weather
maps, atmospheric stability, and atmospheric motions from the synoptic-scale to the
local-scale.
• Wilks (1995) - Describes statistical techniques and how these can be applied to
meteorological data. Many of the techniques discussed can also be applied to air quality.
5.2.2 Data Analyses
Once a literature search is completed, data analysis can help forecasters learn more about
the processes that control ozone or PM2.5 concentrations in their area. Data analysis is the
process of exploring data to answer questions. It can be performed in three steps: developing
questions (i.e., hypotheses), acquiring data, and using analytical methods to answer the
questions. Depending on the air quality agency's resources, data analysis efforts can range from
simple statistical analyses to large field studies with subsequent research and computer
modeling. Following is a discussion of some basic analysis procedures to help explain the
processes that control ozone or PM2.5 concentrations in a forecast area.
5-3
-------
The first step in performing data analysis is having clearly defined questions; this will
increase the effectiveness of the research. Types of questions to ask include:
Temporal distribution of ozone or
• During what weeks/months do ozone or PM2.5 episodes occur?
• At what time of day do the highest ozone or PM2.s concentrations occur? For how many
hours do high ozone or PM^.5 concentrations typically last?
• For how many consecutive days do high ozone or PM2.s episodes typically last?
• Do maximum ozone or PM2.5 concentrations vary by day of week?
• During what time of year do PM2.5 episodes occur?
• Do ozone and PM 2.5 episodes coincide with one another? When do PM2.s-only episodes
occur?
Spatial distribution of ozone or PMj.g
• Where do the highest ozone or PM2.5 concentrations occur? Do the highest
concentrations occur at different times for different sites?
• Have emissions patterns changed in recent years?
Monitoring issues
• Has the monitoring network changed recently?
• What are the different PM2.5 monitoring methods and how do they compare to one
another?
• Are adjustments made to the continuous PM2.5 data to make them better match the
FRM standards? If so, how accurate are these adjustments?
Meteorological and air quality processes
• What types of synoptic weather patterns are associated with high ozone or
PM2.5 concentrations?
• Does local carryover contribute to ozone or PM2.5 concentrations?
• Do surface or aloft transport pollutants from other areas contribute to ozone or PM2.5 in
the forecast area?
• How do local flow patterns influence ozone or PM2.5 concentrations?
• How does the aloft temperature structure influence peak ozone or PM2.5 concentrations?
• What types of weather patterns are associated with cloud cover?
• Are there rare emission events such as forest fires, agricultural burning and tilling,
windblown dust, etc. that cause PM2.5 episodes?
5-4
-------
The remainder of this section discusses these questions in more detail and explains why
each question is important. Also included is an example analysis technique. These examples are
intended as a starting place for an understanding of the important processes that produce ozone
or PM2.5 in a forecast area.
Temporal distribution of ozone or PM
2.5
Question: During what weeks/months do ozone or PM2.5 episodes occur?
Why: Helps define the forecasting period.
Technique: Create frequency plots of the number of exceedances by month (or week) for
several years. Figure 5-1 shows that in the New Jersey and New York City
metropolitan region, a forecasting season for ozone would last from May
through September, since most of the 1-hr and 8-hr exceedances are confined to
these months.
16
14
12
re
Q
HI
2 10
TO
*
HI
u
X
LU 8
6
D)
D1 -hr exceedance of 125 ppb
• 8-hr exceedance of 85 ppb
I
Figure 5-1. Distribution of the average number of days with 8-hr and 1-hr exceedances by
month for the New Jersey and New York City region from 1993-1997
(NESCAUM, 1998).
5-5
-------
Question: At what time of day do the highest ozone or PM2.5 concentrations occur? For
how many hours do high concentrations typically last?
Why: Knowing the typical time and duration of high ozone and PM2.5 concentrations
can help public outreach personnel properly notify the public so they can take
appropriate action to minimize exposure. Understanding the diurnal cycle of
ozone or PM2.s can help in producing a more accurate forecast.
Technique: Create frequency plots of the time of peak ozone or PM2.5 concentrations. For
example, Figure 5-2 shows that the highest occurrence of 1-hr ozone
exceedances is at 1400 EST in the New Jersey and New York City region, but
ranges from 1000 to 1700 EST.
25
20
ro
•o
01
u
i
10
5 --
1000
1100
1200 1300 1400
Time of Maximum Ozone (EST)
1500
1600
1700
Figure 5-2. Distribution by hour of daily maximum 1-hr ozone concentration on days that
exceeded 125 ppb in the New Jersey and New York City region from
1993-1997 (NESCAUM, 1998).
5-6
-------
Question: For how many consecutive days do high ozone or PM2.5 episodes typically last?
Why: Knowing the typical duration of episodes can help guide the forecast. For
example, if episodes never last more than two days in a particular area, the
occurrence of a three-day episode in the future is unlikely; therefore, the
forecaster would be cautious to forecast high pollution for three straight days.
Technique: Create a frequency plot (such as the one shown in Figure 5-3) of the number of
continuous days with high ozone or PIVb.s concentrations. Figure 5-3 indicates
that a typical ozone episode of 125 ppb exceedances lasts one to two days and is
never longer than four days in the New Jersey and New York City region.
50
45 -
40 -
35 -
VI
a 30
ui
"5 25
d 1 -hrexceedance of 125 ppb
• 8-hr exceedance of 85 ppb
20--
15 --
10 --
5 --
Tli
6 7 8 9 10
Length of Episode (days)
12 13
14
15
Figure 5-3. Average annual frequency of ozone episode length for the 8-hr and 1-hr standards in
the New Jersey and New York City region from 1993-1997 (NESCAUM, 1998).
5-7
-------
Question: Do maximum ozone or PM2.5 concentrations vary by day of week?
Why: Weekday and weekend differences in commute traffic and some industrial
processes can lead to a variation in ozone and PM2.5 concentrations given similar
weather conditions.
Technique: Create frequency plots of the number of exceedances by day of week. For
example, Figure 5-4 shows that in the New Jersey and New York City region
1-hr ozone exceedances are more likely to occur on Tuesdays and Wednesdays
and somewhat less likely to occur on the weekends. Notice that the 8-hr
exceedances show no day-of-week dependence. Thus, given similar
meteorological conditions, a 1-hr ozone forecast on Saturday through Monday
should be lower than one for Tuesday through Friday in this region. Notice that
weekend 8-hr exceedance frequency is higher than all days except Tuesday.
D1-hrexceedance of 125 ppb
• 8-hr exceedance of 85 ppb
Sun
Mon
Tue
Wed
Thu
Fri
Sat
Figure 5-4. Distribution of the average number of 8-hr and 1-hr exceedances by day of
week for the New Jersey and New York City region (NESCAUM, 1998).
5-8
-------
Question: During what time of year do PM2.5 episodes occur?
Why: Unlike ozone, for many areas of the United States PM2.s episodes can occur at
any time of the year. Understanding what time of year PIVb.s episodes occur in
the forecast area will help define the forecast season or seasons, so resources can
be effectively allocated. Also, understanding when PM2.5 episodes occur helps
the forecaster determine what meteorological phenomena may be influencing air
quality.
Technique: Create histograms of the number of days with PM2.5 concentration above a
certain threshold by month. The threshold may be the NAAQS 24-hr
PM2.5 standard.
Question: Do ozone andPM2.s episodes coincide with one another? When do PM2.s-only
episodes occur?
Why: In many areas of the country, PM2.5 and ozone episodes coincide during the
"summertime" (May through September). If this is the case for the forecast area,
then understanding the processes that influence ozone may help forecasters
understand the processes that influence PM2.5 during the summertime.
PM2.s-only episodes occur in the wintertime under weather conditions conducive
to high pollutant concentrations and when there is not enough sunlight (due to
the low sun angle) to drive the photochemical reactions that produce ozone.
PM2.s-only episodes can also occur in the summertime, for example, when there
are unusual emissions that produce PM2.5 or if conditions are conducive to
pollution buildup, but clouds block sunlight thus keeping ozone concentrations
low while PM2.5 concentrations are high.
Technique: Create a scatter plot of the daily AQI based on ozone versus the daily AQI based
on PM2.5. For example, Figure 5-5 shows a scatter plot of the summertime daily
AQI based on ozone and PM2.5 for Washington D.C. for 1999 through 2001. As
noted on this plot the AQIs are not well correlated meaning that ozone and PIVb.s
episodes don't coincide very often in Washington D.C. However, there are days
on which high AQIs do agree. These days should be investigated to understand
what makes them different from the others.
Spatial distribution of ozone or PMj.5
Question: Where do the highest ozone or PM2.5 concentrations occur? Do the highest
concentrations occur at different times for different sites?
Why: Different areas in the forecast region may have very different ozone or
PM2.5 characteristics due to spatial variation in emissions and meteorology.
When forecasting for large areas it may be necessary to sub-divide the region to
account for differences in ozone or PM2.5 concentrations based on differences in
emissions and weather across the region.
Technique: Plot a map of the average peak ozone or PM2.5 concentration and the time of the
peak for each site on exceedance days.
5-9
-------
20 40 60
Figure 5-5. Scatter plot of the summertime daily AQI based on peak 8-hr ozone concentrations
versus the AQI based on 24-hr average PM2.5 concentrations for Washington D.C.
for 1999 through 2001. The one-to-one line (thin line) and linear best fit (thick
line) are also shown.
Question: Have emissions patterns changed in recent years?
Why: If emissions patterns have changed, weather conditions that have historically
produced ozone or PM2.5 exceedances, may now result in lower concentrations.
Technique: Determine if significant emissions changes (e.g., the use of reformulated fuel) or
shifts in population have occurred in the forecast region. For example, less
reactive emissions may result in peak ozone concentrations occurring farther
downwind and/or in lower ozone concentrations. Changes in 862, NOX, or
ammonia emissions can influence the production of secondary PIVb.s.
Monitoring issues
Question: Has the monitoring network changed recently?
Why: Changes in a monitoring network can cause significant differences between
historic and currently observed ozone or PM2.5 concentrations. If a new monitor
was recently installed downwind of a major emission source area, then the
observed time and peak ozone or PM2.5 concentrations for the entire area may
change significantly due to this new site. These types of monitoring network
changes must be taken into account when analyzing historic and current ozone
or PM2.5 concentration data.
5-10
-------
Technique: Create a plot of the historic monitoring network and compare it to a plot of the
current network. If new sites have been added in recent years, determine if the
new sites have caused an increase in the number of exceedance days in the
forecast region.
Question: What are the different PM2.5 monitoring methods and how do they compare to
one another?
Why: Different monitoring methods (FRM, BAM, TEOM) can report very different
concentrations, even when sites are collocated. The differences are often
weather condition dependent (temperature and humidity). If forecasting tools
are developed using historic FRM data and then real-time TEOM or BAM data
are used for daily forecasting, then these differences need to be considered.
Technique: Create a scatter plot of daily 24-hr average PM2.5 concentrations from collocated
FRM and continuous monitors. For example, Figure 5-6 shows a comparison of
1999-2001 FRM and TEOM 24-hr average PM2.5 data collected at a site in
Forsyth County, North Carolina. The blue line is the 1:1 line. The regression
equation shows a slope near 1 and the r is high (.96), which indicates a good
relationship among the data.
0-
a
50
45
40
FRM=1.055(TEOM)-0.426
R2=0.9577
TEOM 2.5 [ug/m3]
Figure 5-6. Comparison of FRM and TEOM 24-hr average PM2.5 data collected in
Forsyth County, North Carolina, from 1999 to 2001 (Courtesy of Lewis
Weinstock).
5-11
-------
Question: Are adjustments made to the continuous PM2.5 data to make them better match
the FRM standards? If so, how accurate are these adjustments?
Why: Forecast tools are usually developed from FRM data. In real-time, only
continuous data are available to verify the forecast. Since collocated FRM and
continuous monitors often report different PM2.5 concentrations, adjustments to
the continuous data need to be made to properly verify the forecast.
Technique: Create scatter plots of collocated FRM and adjusted continuous monitor data
(similar to that shown in Figure 5-6). If the data points fall within roughly 15%
of the 1:1 line, then the adjusted data represent the FRM data; otherwise,
evaluate why the data are different and consider developing new techniques to
adjust the continuous data to match the FRM data. For example, developing
equations for each season may produce better results.
Meteorological and air quality processes
Question: What types of synoptic weather patterns are associated with high ozone or
PM2.5 concentrations?
Why: Synoptic-scale weather features are large (1000 km or more) weather
circulations that influence regional weather conditions that, in turn, strongly
influence the production and transport of ozone and PM2.s and their precursors.
By reviewing weather forecast charts, forecasters can identify historical weather
patterns associated with particular pollutant concentrations in the forecast
region.
Technique: Analyze historical weather charts that depict synoptic features. Classify the
surface and aloft synoptic weather patterns and create frequency plots showing
the synoptic pattern versus the number of high ozone or PM2.5 concentration
days, moderate ozone or PM2.5 concentration days, and low ozone or
PM2.5 concentration days. EPA's AIRNow program provides historical ozone
maps since 1998 (www.epa.gov/airnow/).
For example, Figure 5-7 shows a surface synoptic pattern associated with high
ozone in Pittsburgh, Pennsylvania (Comrie and Yarnal, 1992). Historic daily
weather map sources include the Daily Weather Map series issued daily by the
National Oceanic and Atmospheric Administration (NOAA)3 and the archive
analysis of the National Center for Environmental Prediction weather models
available on the Internet at www.arl.noaa.gov/ready/arlplota.html.
Question: Does local carryover contribute to ozone or PM2.5 concentrations?
Why: When ozone or PM2.5 episodes occur over several days, day-to-day pollution
buildup can contribute to the daily ozone and PM2.5 concentrations. That is,
today's pollution (if not dispersed, deposited, or permanently reacted away) will
contribute to tomorrow's pollution.
3 Daily Weather Maps, Climate Prediction Center, Room 811, World Weather Building, Washington, DC 20233
5-12
-------
Figure 5-7. A surface synoptic pattern associated with high ozone in Pittsburgh, Pennsylvania
(Comrie and Yarnal, 1992).
Technique: Carryover ozone can be investigated by examining ozone data from surface sites
at which the ozone data show no overnight titration by NO. Since there is no
titration for PM2.5 most sites that are not significantly influenced by local
sources can be analyzed. Create scatter plots of overnight ozone or
PM2.5 concentrations vs. the next-day concentrations. Examine the plots to see if
there is a relationship between overnight concentrations and the next-day
concentrations.
For example, Figure 5-8 shows the relationship between 0200 EST ozone
concentrations at a mountainous site in western North Carolina (a site that is
representative of regional carryover) and North Carolina daytime peak ozone
concentrations. The influence of background ozone or PM2.5 can also be accessed
by analyzing aloft data collected by aircraft, on a tower, or on a nearby mountain
instead of or in addition to the surface data.
5-13
-------
160
S" 140
120
100
80
60
*\ \ s
.*• - *» ^
• *• '•"
• ***** **
2 40
• *
1
20
10 20 30 40 50 60 70 80
0200 ESJ ozone concentration (ppb) at Fry Pan site in Haywood County, North Carolina
90
100
Figure 5-8. Scatter plot of 0200 EST ozone concentrations at a mountainous site (Fry Pan) in
Haywood County, North Carolina, versus North Carolina daily regional
maximum ozone concentrations for June to September, 1996 (MacDonald et al.,
1998).
Question: Do surface or aloft transport of pollutants from other areas contribute to ozone
or PM2.5 in the forecast area?
Why: Long-range transport of pollutants can contribute significantly to local ozone or
PM2.5 concentrations. It is important for a forecaster to understand if and when
this occurs in order to accurately forecast ozone or PM2.5.
Technique: Computing back trajectories is a useful way to examine the potential for
long-range transport (Figure 5-9). For selected days, create 12-hr, one-day, and
two-day back trajectories at several levels. Determine if these trajectories
originate in areas with high ozone or PM2.5 concentrations. An excellent tool for
computing back trajectories is interactively available on the Internet at
www.arl.noaa.gov/ready/hysplit4.html (Draxler and Hess, 1997).
5-14
-------
0)
T3
'•a UN
79* 78* 77* ISO 75* 74*
Longitude
Figure 5-9. Back trajectories during an ozone episode in the northeastern
United States showing possible transport of pollutants from regions to
the west (Ryan et al., 1998).
Question: How do local flow patterns influence ozone or PM2.5 concentrations?
Why: Local flow patterns such as land-sea breezes, up/down slope flows, and terrain
guided flows can play a large role in transporting ozone and PM2.s. Such flows
may locally transport pollutants from upwind sources to downwind cities or
recirculate pollutants within metropolitan areas. Understanding the flow
processes in the forecast area will greatly improve the forecasts.
Technique: Compute back trajectories on high, moderate, and low ozone or
PM2.5 concentration days. A good resource for computed trajectories is
www. arl. noaa.gov/ready/hysplit4. html.
An example of a 24-hr back trajectory for a monitoring site in Crittenden County,
Arkansas, (near Memphis, Tennessee) on a high ozone day is shown in
Figure 5-10. This simple trajectory shows both surface and aloft flow from the
northeast portion of the domain.
5-15
-------
4300
4200 - -
4100 -•
4000 - -
3900 -
3800 -
3700
Memphis Airport
700
800
900
1000
UTME(km)
1100
1200
1300
Figure 5-10. A 24-hr back trajectory from Crittenden County, Arkansas, starting at 1400 EST
on August 25, 1995, and ending at 1300 EST on August 26, 1995. Trajectories
were computed using surface wind data from the National Weather Service's site
at the Memphis airport and upper-air data from a radar wind profiler located at the
airport (Chinkin et al., 1998).
Question: How does the aloft temperature structure influence peak ozone or
PM2.5 concentrations?
Why: Aloft temperature structure strongly influences vertical mixing and dilution of
pollutants. A stable atmosphere produces less vertical mixing and dilution of
pollutants which leads to higher ozone or PM2.5 concentrations.
Technique: In many areas of the country 850-mb temperature is a good indicator of aloft
stability and inversion strength. Forecasted 850-mb temperatures can therefore
be used to estimate the amount of mixing and dilution of ozone and PM2.s. For
example, New York State ozone forecasters use an 850-mb temperature greater
than 15°C as one criterion for forecasting high ozone concentrations (Taylor,
1998), while Sacramento, California, forecasters set the 850-mb temperature
criterion at a temperature greater than 18°C (Dye et al., 1996).
Question: What types of weather patterns are associated with cloud cover ?
Why: Cloud coverage limits the photodissociation of NO2; this is a key step in ozone
formation. Accurately predicting cloud coverage will improve ozone forecast
accuracy. Clouds can also increase the conversion of 862 to sulfate.
Technique: The National Weather Service' (NWS') computer forecast models predict
relative humidity at several altitudes with reasonable accuracy. Analyzing these
predictions along with satellite images can help to forecast cloud cover. Model
5-16
-------
output statistics (MOS) predict the amount of cloud cover but are not always
accurate. For example, anvils from distant thunderstorms are often not
accurately predicted. Performing case studies of days when a model's cloud
predictions are wrong and understanding the types of weather patterns that cause
inaccurate cloud forecasts, allow forecasters to identify such conditions in the
future.
Question: Are there exceptional emission events such as forest fires, agricultural burning
and tilling, windblown dust, etc. that cause PM2.5 episodes?
Why: Rare emission events often cause PM2.5 episodes. Since these events are rare,
they are difficult to predict using objective forecasting tools or techniques alone.
By understanding when such events are likely to occur (for example, wildland
fires occur most often in the summer and fall), forecasters can watch for such
events and account for them in the daily forecasting when needed.
Technique: Perform case study analysis of past PIVb.s episodes to determine the cause of the
high concentrations. See Section 2.2.5 for more information on rare events.
Once forecasters understand the chemical and meteorological processes that influence
ozone and PM2.5 concentrations, they can start selecting methods to forecast these pollutants, as
discussed in the next section.
5.3 CHOOSING FORECASTING METHODS
Once the needs of the forecasting program have been identified, forecasters will need to
choose a forecasting method or combination of methods to predict ozone or PM2.5. The
method(s) chosen will primarily depend on the available resources and experience. Listed below
are a number of issues to consider when selecting a forecasting method.
Resources Cost may be the major factor guiding method selection. When
determining the overall cost of a particular method, the costs
associated with both developing and operating the method should
be considered. Development costs versus operating costs can
vary greatly between methods. For example, the development of
a regression model may be fairly expensive compared to the
development of a criteria method; however, the operational costs
may only be slightly different.
Severity of The severity of the air quality problem and the frequency of
problem high concentrations in the forecast region will also guide the
choice of method. For example, a region with very few ozone
episodes may only need a simple and inexpensive method to
forecast a few high ozone days. On the other hand, a region that
experiences many days with high ozone during the summer and
numerous days with high PM2.5 in the winter may benefit from
developing several forecasting methods that can be utilized
year-round.
5-17
-------
Balancing Balancing resources between multiple forecasting methods may
methods minimize the limitations of the methods while compounding
their strengths. Also, balancing objective and subjective
methods may increase forecast accuracy.
Adding Once a forecasting method has been selected, the program is not
methods limited to retaining this single method. Building a program from
one simple method in the first year to multiple methods in future
years is a cost-effective approach to increase the accuracy of a
forecasting program.
Expertise Some methods require a high level of meteorological experience
and forecasting expertise. Working with a university or other
agency to develop a forecasting method may be beneficial if
in-house resources are not available.
5.4 DATA TYPES, SOURCES, AND ISSUES
Once a forecasting method or methods has been selected, data needs should be addressed.
Air quality and meteorological data are needed for both developing the method(s) to predict air
quality and for operationally forecasting. This section identifies the types and sources of
meteorological and air quality data as well as issues to consider when acquiring and using data.
A variety of data types, both meteorological and air quality, are available for developing
prediction methods and forecasting air quality. The general data requirements of each method
are listed in Table 4-1. Table 5-1 summarizes data types and typical parameters. These data
types include surface and upper-air meteorological data, both observed and forecasted. Data
needs for a particular forecast region will depend on the specific meteorological and air quality
phenomena to be predicted.
Locating a data source is often a major part of developing a forecasting method.
Table 5-2 lists many of the major sources of air quality and meteorological data. Historical data
can also be found in air quality studies that were conducted in the forecast region.
Most forecasting programs require many types of data from difference sources to fulfill
all of the forecaster's needs. Each of these data sources provide data at different costs, in
different file formats, and with varying degrees of reliability and quality. When acquiring data,
consider the following issues:
5-18
-------
Table 5-1. Data products for developing forecasting methods and for forecasting weather and ozone.
Data Type
Surface Meteorological
Surface Air Quality
Upper-air Meteorological
(rawimondes, radar profilers,
andsodars)
Aloft Air Quality Observations
(towers, mountains, and aircraft)
Weather Charts
Weather Radar
Satellite
Meteorological Model Forecasts
Text Weather Forecasts
Variables
WS, WD, T, RH, Solar Rad.,
Cloud Cover, Vis, P
Ozone, PMio, PM2.s, Oxides of
Nitrogen, Carbon Monoxide,
VOCs
Vertical Profiles of WS, WD, T,
RH
Ozone, PMio, PM2.5, Oxides of
Nitrogen, Carbon Monoxide,
VOCs
Surface (WS, WD, T, RH, P)
850 mb (WS, WD, T, Height)
700 mb (WS, WD, T, Height)
500 mb (WS, WD, T, Height),
Others
Precipitation
Cloud Cover (visible and
infrared)
T, RH, WS, WD, Cloud Cover,
Vis, P, and others at many levels
Discussions
Forecasted/
Observed
Observed
Forecasted and
Observed
Observed
Observed
Forecasted and
Observed
Observed
Observed
Forecasted
Forecasted and
Observed
Frequency
Hourly
Hourly
Twice per day to
hourly
Variable
Typically twice per day
Sub -hourly
Hourly and/or
Sub -hourly
Twice per day
Four or more
times per day
WS = wind speed
WD = wind direction
T= temperature
RH = relative humidity
Vis = visibility
P = pressure
-------
Table 5-2. Major sources of air quality and meteorological data.
Data Source
U.S. EPA Aerometric
Information Retrieval System
(AIRS)
National Climate Data Center
(NCDC)
NOAA National Data Centers
(NNDC)
Regional Climate Centers
Purdue University
Commercial Weather Service
Providers (WSP)
Type of Data Source
Historical
Historical
Historical
Historical
Historical and
Real-time
Real-time
Types of Data
Surface air quality
Surface Meteorological, Upper-air
Meteorological, Weather Charts,
Radar, Satellite
Surface Meteorological, Upper- Air
Meteorological, Weather Charts,
Satellite, Radar, Climate
Surface Meteorological, Upper-air
Meteorological, Climate
Information
Surface Meteorological, Upper-air
Meteorological, Satellite, Radar,
Model Forecast, Text Weather
Forecast
Surface Meteorological, Upper- Air
Meteorological, Weather Charts,
Satellite, Radar, Model Forecast,
Text Weather Forecast
Phone Number
(916) 541-5586
(828)271-4800
(828)271-4800
Western Regional Climate Center
(775)674-7010
High Plains Climate Center
(402) 472-6706
Midwestern Climate Center
(217)244-8226
Northeast Regional Climate Center
(607)255-1751
Southeast Regional Climate Center
(803) 734-9560
Southern Regional Climate Center
(225) 388-5021
-
-
Web Site
www.epa.gov/ttn/airs
www.ncdc.noaa.gov
www.nndc.noaa.gov
www.wrcc.dri.edu
www.hpccsun.unl.edu
mcc.sws.uiuc.edu
met-www.nrcc.cornell.edu
www.sercc.com
www.srcc.lsu.edu
www.arl.noaa.gov/ready/arlplota
.html
Comprehensive list of WSPs at:
www.ugems.psu.edu/~owens/
WWW Virtual Library/commer
cial.html
to
o
-------
Cost
Reliability
Quality control
Data formats
Ingest methods
Hardware
requirements
Redundancy
A significant amount of data is available for free on the Internet
from the NWS, Government Laboratories, and Universities. The
reliability of these data is reasonable for forecasting, but access
to it may suffer from Internet outages. Weather Service
Providers (WSPs) supply weather data to television stations,
private industry, and government agencies. WSPs typically
charge a startup fee for display and data acquisition software.
Most charge a monthly/yearly data subscription fee and
automatically send the data to a subscriber's computer.
Reliability for this type of service is generally very high.
Knowing that data will always be available when needed it is
critical to a forecasting program's success. Unreliable data will
reduce forecast effectiveness and may lower forecast accuracy.
Receiving higher quality data (data with fewer errors and
inconsistencies) decreases the necessity for personnel to
thoroughly review the data before it is used. All historical data
should be reviewed for quality prior to developing a forecasting
method.
The number of data formats used should be minimized in order
for data to be decoded and processed more efficiently. Different
time standards, reporting units, quality control codes, etc., can
produce additional decoding/processing effort.
Determining how data will flow from the source to the
forecasting location is important. Ingest methods are typically
the Internet, telephone, and satellite. Internet and telephone
telemetry are cost effective. Satellite delivery systems are very
reliable, yet are typically more expensive. Seek to automate as
many of the data ingest tasks as possible, so the forecaster can
spend more time on the nuances of the prediction.
Hardware needs are a function of the amount of data and data
processing required. The greatest convenience of WSPs is that
they can supply multiple data types through one software
package and computer. Combining this service with Internet use
on the same computer not only improves resource efficiency, but
also provides additional data types and redundancy at little extra
cost.
Having a backup or secondary data source is a prudent practice.
Consider the risks of not having data available for forecasting.
For key information and data, identify several sources from
which the data can be obtained.
Using high quality meteorological and air quality data is important to develop accurate
forecasting methods. In addition, obtaining reliable real-time data is a key component for
operationally forecasting.
5-21
-------
5.5 FORECASTING PROTOCOL
A forecasting protocol describes the daily operating procedures from data acquisition to
forecast production and dissemination. A protocol helps guide personnel through the forecasting
process. It ensures that all activities are performed on time without the need for last minute
decisions and helps maintain consistency from one forecaster to the next. This section explains
what to include in a forecasting protocol.
Preparing a forecast often requires that various personnel complete numerous steps. To
standardize this process, written procedures should be prepared and tested that the forecast team
can follow on a regular basis. A forecasting protocol will likely include:
• Descriptions of the meteorological conditions that produce high ozone and PM2.5
concentrations in the area.
• A schedule of daily tasks and personnel responsibilities. The easiest way to create this
schedule is to work backwards from the time that the forecast is due to the time initial
procedures need to begin. It is likely that the schedule will differ for high, moderate, and
low ozone days. An example of a basic schedule is shown in Table 5-3.
• Steps to take to arrive at a forecast, including key decision points that help forecasters to
quickly identify low ozone and PM2.5 days, thus allowing time for the more difficult
forecasts.
• Forms and worksheets for documenting data, forecast information, forecast rationale, and
comments that forecasters can analyze and evaluate later.
• Phone and fax numbers and e-mail addresses of key personnel.
• Names, fax and phone numbers, and e-mail addresses of forecast recipients.
• Troubleshooting and backup procedures for the key components necessary to produce
and issue the pollutant forecasts such as: backup forecasters, redundant data acquisition
methods, and forecast dissemination.
These written procedures save time and effort and should be an integral part of any
forecasting program.
5.6 FORECAST VERIFICATION
Verification is the process of evaluating the quality of a forecast by comparing the
predicted air quality to the observed air quality. As part of a forecasting program, forecasters
should regularly evaluate the forecast quality. The benefits of verifying ozone or PM2.5 forecasts
include:
• Quantifying the performance of forecasters and/or the forecast program,
• Identifying trends in forecast performance over time,
• Quantifying improvements from new (or changes in) forecasting methods/tools,
5-22
-------
Table 5-3. Example of a forecasting protocol schedule.
Time
0900
0930
0945
1015
1030
1100
1125
-0930
-0945
-1015
-1030
-1100
- 1125
-1130
Activity
Run air quality data acquisition programs. Review data for
and accuracy.
completeness
Acquire observed and forecasted meteorological data from the Internet.
Review forecast weather maps.
Run regression model to forecast ozone.
Evaluate forecasted weather conditions and air quality using the
Phenomenological/Intuition method.
Produce the final forecast; write forecast discussion.
Fax forecast to air district officials and place
forecast on the
Internet.
• Comparing verification statistics to those from other agencies that forecast air pollution.
• Demonstrating the performance of forecasts to program participants, stakeholders, and
the media.
The verification process can be complex since there are many ways to evaluate a forecast
including accuracy, bias, and skill. Since no one statistic can fully reflect the performance of a
program, many verification statistics must be completed in order to completely evaluate the
quality of a forecast program.
Two basic types of forecasts exist: discrete forecasts of specific concentrations and AQI
category forecasts (e.g., good, moderate, etc.). Verification statistics differ for these two types of
forecasts. This section explains how verification statistics can be computed and interpreted for
both types of forecasts and provides a schedule for verifying forecasts (Section 5.6.1).
Section 5.6.2 describes verification statistics for discrete forecasts. Section 5.6.3 describes
verification statistics for category forecasts. An on-line glossary of verification terms can be
found at www.sec.noaa.gov/forecastjverification.
5.6.1 Forecast Verification Schedule
Forecasts should be evaluated frequently to identify any problems or downward
performance trends. A schedule of verification tasks follows:
Daily
If forecasts are significantly missed (off by more than 30 ppb or two
AQI categories), the causes of the missed forecasts should be
examined. Write a forecast retrospective, which is a several page
document that details what went wrong and includes recommended
changes to forecast methods or procedures. Figure 5-11 shows an
example outline for a forecast retrospective.
5-23
-------
Monthly
Annually
Compute the forecast verification statistics described in this section.
Compare these with statistics from previous months and review
statistics with forecasters.
Compute the forecast verification statistics described in this section.
Compare these statistics from previous years, review statistics with
forecasters, and compare them to other forecasting programs.
Forecast Retrospective
Date
1. Summary of event
Provide a brief synopsis of what happened.
2. Forecast rationale
Explain the steps and thought processes used to make the
forecast.
3. Actual weather and air quality conditions
Discuss all aspects of the weather that occurred. Use
weather maps, satellite images, observations. Review the
relevant air quality conditions.
4. Revision to forecasting guidelines
Recommend any changes to forecasting procedures.
Figure 5-11. Example outline of a forecast retrospective.
5.6.2 Verification Statistics for Discrete Forecasts
Several verification statistics can be computed for forecasting programs that predict
discrete pollutant concentration values. Table 5-4 lists four statistics commonly used to verify
discrete forecasts and explains how to compute and interpret these statistics. The four statistics
are:
Accuracy Average "closeness" between the forecast and observed values.
Bias Indicates, on average, if the forecasts are underpredicted or
overpredicted.
Skill score Percentage improvement of a forecast with respect to a reference
forecast (typically a climatology or persistence forecast).
Correlation A measure of the relationship between forecasts and
observations and if the two sets of data change together.
The first step in the process is to pair the forecast and observation data for each forecast
issued. Then use the equations listed in Table 5-4 to calculate the verification statistics.
5-24
-------
Table 5-4. Verification statistics computed on discrete concentration forecasts.
Statistic Name
Accuracy
(mean absolute error)
Bias
(mean error)
Skill score
Correlation
What it Measures
Average "closeness" between
the forecast and observed
values. Summarizes the
overall quality of the
forecasts.
It indicates, on average, if the
forecasts are underpredicted
or overpredicted.
Percentage improvement of a
forecast with respect to a
reference forecast (typically a
climatology or persistence
forecast).
A measure of the relationship
between forecasts and
observations. It measures if
two sets of data change
together.
How to Compute it
1 . Take the absolute difference
between forecast (/) and
observation (o) for all forecasts
(N).
2. Sum the differences and divide by
N.
1 . Take the difference between
forecast (/) and observation (o) for
all forecasts (N).
2. Sum the differences and divide by
N.
1 . Compute accuracy for a reference
forecast (Aref), either climatology
or persistence (see Section 4.1 for
details).
2. Compute accuracy (A) for the
forecast.
3. Divide accuracy by the reference
accuracy and subtract from 1 .
Use correlation functions in
spreadsheet or statistics software
(Excel, Lotus) or compute using the
following:
1 . Compute the co-variance
(Cov(f,o)) using the equation.
2. Compute the standard deviation
for the forecasts (sf) and
observations (so).
3. Divide the co- variance by the
product of the standard deviations
to compute the correlation (Cfo).
Equation
1 ( " ~\
"-M?"-)
1 (N \
5=ly(/_0)
N(^ ')
(A \
SS=\ 1*100
(AM )
n (Co^f,o)}
P ( rf-so }
where: -1 > Cfo < 1
i N '\
Cov(/,o)=— ^(/j-M/XOj-M,,)
where (Jo = mean observed value
l^f = mean forecasted value
„ ( N \2
^•z/;- z/J
, , j-i v^-i ;
"J 1 JV(AT-l)
In ( ii '
kz°Hz°,
•o-l -"' '•'"'
~° \ N(N-l)
Units
Units of
variable (ppm,
ppb, ug/m3,
etc.)
Units of
variable (ppm,
ppb, ug/m3
etc.)
%
How to Interpret it
• Lower numbers are best.
• Values indicate the
uncertainty in any single
forecast.
• Values near 0 are best.
• Values <0 indicate
underpredicting (i.e.,
forecasts are too low).
• Values >0 indicate
overpredicting.
• 0% indicates no
improvement (or skill) over
the reference forecast.
• 50% or more indicates a
significant improvement in
skill.
• Values close to 1 are best.
• Positive correlation
indicates that large forecast
values are associated with
large observed values.
• Negative correlation occurs
when small values of one
set are associated with large
values of the other.
• No correlation occurs when
values in both sets are
unrelated.
• High correlation does not
necessarily denote high
accuracy.
to
-------
To illustrate how to compute and interpret these statistics, Table 5-5 shows hypothetical
forecasts and verification statistics. In this case, a forecaster made hypothetical forecasts (F) for
11 days. For reference purposes, forecasts were also made using the Persistence method (FperS)
discussed in Section 4.1. A reference forecast is a baseline against which to compare a forecast.
Any other forecast method can be used, but typically persistence, climatology, and random
chance are used as reference forecasts.
Table 5-5. Hypothetical forecasts for an 11-day period showing a human forecast (F),
observed values (O), and forecasts using the Persistence method (FPers).
Accuracy (A), accuracy for persistence forecast (Apers), bias (B), skill-score
(SS), and correlation (C) were computed using the equations provided in
Table 5-4.
Date
1-Jul
2-Jul
3-Jul
4-Jul
5-Jul
6-Jul
7-Jul
8-Jul
9-Jul
10-Jul
11-Jul
77
(ppb)
80
90
70
80
90
120
120
100
130
100
60
O
(ppb)
80
100
80
80
90
90
120
110
100
100
60
Fpers
(ppb)
70
80
100
80
80
90
90
120
110
100
100
A = 8 ppb
AperS = 14 ppb
B = 3 ppb
SS = 40%
C = 0.77
For the 11-day period, the forecaster's accuracy was 8 ppb, meaning that, on average, the
forecasts were within ±8 ppb of the observed maximum. A slight positive bias of 3 ppb indicates
that the forecasts are slightly higher than the observed values. The skill score is 40%, which
indicates that the forecasts are 40% better than the reference forecast using the Persistence
method. The last statistic, correlation, had a value of 0.77 indicating that most day-to-day
changes in the forecasts are also reflected in the observations.
More detailed information on forecast verification can be found in Murphy (1991; 1993),
Murphy and Winkler (1987), and Wilks (1995).
5-26
-------
5.6.3 Verification Statistics for Category Forecasts
This section describes verification statistics that can be computed for category forecasts
and provides several examples of how to interpret the verification statistics. Verification
methods differ slightly based on the number of forecast categories. At the simplest level, a
forecast can be issued for two categories (e.g., forecast an Ozone Action Day to occur or not to
occur). A two-category forecast is discussed below. Three- or more-category forecasts are
discussed later in this section.
Two-category forecast
Creating a contingency table (also called a frequency table) is the first step in evaluating
a category forecast. Figure 5-12 shows a contingency table of forecasted and observed events.
This table is the basis for calculating all verification statistics for category forecasts. It is
constructed by counting the frequency of occurrence of each event and assigning it to the
appropriate cell.
Forecasted
Observed
Exceedance
1 §
no yes
a
c
b
d
Figure 5-12. Contingency table for a two-category forecast.
Using the contingency table shown in Figure 5-12, a perfect forecasting program would
have values in cells "a" and "d" indicating correct "hits". While both "a" and "d" represent
correct forecasts, events forecasted in cell "a" are generally more frequent and easier to predict
than those in cell "d". In the real world, imperfect forecasts result in values in cells "b" (false
alarms) and "c" (misses). The verification statistics listed in Table 5-6 are used to evaluate the
quality of two-event categorical forecasts. The statistics include:
Accuracy
Bias
False alarm
rate
Critical
success index
Probability of
detection
Skill score
Percent of forecasts that correctly predicted the event or non-event.
Indicates, on average, if the forecasts are underpredicted (false
negatives) or overpredicted (false positives).
Percent of times a forecast of high pollution did not actually occur.
How well the high-pollution events were predicted; it is unaffected by
a large number of correctly forecasted, low-pollution events.
Ability to predict high-pollution events.
Percentage improvement of a forecast with respect to a reference
forecast, typically a climatology or persistence forecast.
5-27
-------
Table 5-6. Verification statistics used to evaluate two-category forecasts. Lower case letters in the equations correspond
to those in Figure 5-12.
Statistic name
Accuracy (A)
(percent correct)
Bias (B)
False Alarm Rate (FAR)
Critical Success Index
(CSI), also called Threat
Score
Probability of Detection
(POD)
Skill Score (SS)
What it measures
Percent of forecasts that
correctly predicted the event or
non-event.
Indicates, on average, if the
forecasts are underpredicted
(false negatives) or
overpredicted (false positives).
The percent of times a forecast
of high pollution did not
actually occur.
How well the high-pollution
events were predicted. Useful
for evaluating rarer events like
high-pollution days. It is not
affected by a large number of
correctly forecasted, low-
pollution events.
Ability to predict high-
pollution events (i.e., the
percentage of forecasted high-
pollution events that actually
occurred).
Percentage improvement of a
forecast with respect to a
reference forecast, typically a
climatology or persistence
forecast.
How to compute it
Divide the number of "hits" (cells a
plus d) by the total number of forecasts
issued.
Divide the number of forecasted high
ozone events (cells b plus d) by
observed high ozone events (cells c
plus d).
Divide the high ozone forecasts that
were missed (cell b) by the total
number of high ozone forecasts (cells b
plus d).
Divide the number of high-pollution
"hits" (cell d) by the total number of
forecasts plus the number of misses
(cells b, c, and d).
Divide the correct forecasts of high
pollution (cell d) by the total number
of observed high-pollution events
(cells c plus d).
1 . Compute accuracy for a reference
forecast (Aref), such as
climatology or persistence (see
Section 4. 1 for details).
2. Compute accuracy (A) for the
forecast.
3 . Divide accuracy by the reference
accuracy and subtract from 1 .
Equation
A=(a+d)/N *100
B = b + d
c + d
FAR=(b)/(b+d)*100
CSI = d/(b+c+d)*100
POD=d/(c+d)*100
SS = (A/Aref-l)
*100
Units
%
%
%
%
%
How to interpret
• Higher numbers are better.
• For example, 65 means that 65% of the
forecasts were correct in predicting
ozone or PM2 5 above or below a given
threshold and 35% of the forecasts
missed.
• Values closer to 1 are best.
• Values <1 indicate underpredicting
(i.e., the event occurred more often
than it was forecasted).
• Values >1 indicate overpredicting.
• Smaller values are best.
• 0=no false alarms (perfect forecast of
high events).
• 50 means that half of the forecasts for
high events did not materialize.
• Higher numbers are best.
• For example, 66% indicates that two-
thirds of the forecasts, or actual
exceedances, were correctly predicted.
• Higher numbers are best.
• For example, 70% indicates that, of the
10 high-pollution events, 7 were
correctly predicted.
• 0% indicates no improvement (or skill)
over the reference forecast.
• 50% indicates a significant
improvement in skill.
to
oo
-------
The most important statistics for evaluating the success of a program are Accuracy, False
Alarm Rate, and Critical Success Index.
To help understand these verification measures, statistics were computed for two
hypothetical forecasting programs as shown in Figure 5-13. This example evaluates the forecast
performance for a two-category forecast: a prediction of 8-hr ozone concentrations at or above
85 ppb and below 85 ppb. "Program LM" is typical of a large metropolitan area with many
8-hr exceedances, whereas "Program SC" represents a smaller city with few exceedances. Both
programs were evaluated for a 180-day period.
•o
0)
0)
I/)
.a
O
No
Yes
A
B
FAR
CSI
POD
SS
Aref
Program LM
Forecasted
No Yes
130 8
20 22
84
0.71
27
44
52
78
50
Program SC
Forecasted
•n ^^^^^^^
0)
8! NO
n Yes
A
B
FAR
CSI
POD
SS
Aref
No Y<
160 1
6 :
91
1.56
79
15
33
87
50
3S
1
J
Figure 5-13. Hypothetical verification statistics for a two-category forecast for Program LM
that has many ozone exceedances and Program SC with fewer exceedances.
The accuracy (A) of Program SC is slightly higher than that for Program LM mostly due
to correctly predicting the non-event (i.e., below 85 ppb), which can skew accuracy to be higher
in areas with few high ozone days. Accuracy by itself does not fully describe the performance
differences between the two programs. The second statistic, bias (B), measures the tendency to
underpredict an event (false-negative) or overpredict an event (false-positive). The two
programs exhibit opposite values of bias with Program LM forecasting over twice as many false-
negatives as false-positives and Program SC forecasting nearly twice as many false-positives as
false-negatives.
False alarm rates (FAR) for the two programs are significantly different; Program SC has
nearly three times the FAR of Program LM, 79% versus 27%, respectively. This means that
79% of the high ozone forecasts from Program SC missed. "Crying wolf almost eight out often
times may decrease credibility of the ozone-forecasting program. Program LM's FAR is typical
of many ozone forecasting programs. The critical success index (CSI) measures the forecaster's
5-29
-------
ability to predict the high ozone events, while excluding the large occurrence of correctly
forecasted low ozone days. Program SC has a very low CSI of 15%, meaning that only 15% of
the high ozone events were forecasted correctly even though the accuracy was higher for
Program SC. Program LM does a much better job of predicting the high ozone events and has a
CSI of 44%. Program SC also has a lower probability of detection (POD) than Program LM,
meaning that forecasters in Program SC have a difficult time predicting the high ozone event
when it actually does happen.
The last of the statistics is the skill score (SS), which measures the forecaster's
performance relative to mere chance or another reference method. In the example, the accuracy,
for the reference method (i.e., random chance) is 50%. As with accuracy, Program SC has the
higher skill score; it represents an 87% improvement over chance. But again, Program SC's
results are skewed by a few high ozone events and the frequent (easier to forecast) low ozone
events. Overall, Program SC has a higher accuracy and skill score due to the high number of
correctly forecast low ozone events. Yet, Program LM does a much better job at predicting the
high ozone events as measured by the FAR, CSI, and POD statistics.
A contingency table can be constructed with random forecasts and can be used as a
reference forecast in the skill score computation. Figure 5-14 shows the expected number of
events for each cell. The marginal totals of both the forecasts and the observations are the same
as the marginal totals in the contingency table shown in Figure 5-12 (Stephenson, 2000).
Contingency tables and statistics were computed for Program LM and its random forecast as
shown in Figure 5-15.
0)
o
c
re
•c
0)
0)
o
X
LU
0)
0)
Forecasted Exceedance
no yes
no
yes
(b+a)(c+a)
/n
(c+d)(a+c)
/n
(b+d)(a+b)
/n
(b+d)(c+d)
/n
Figure 5-14. Contingency table for random forecast.
The accuracy, critical success index (CSI), and probability of detection (POD) are all
higher for Program LM than those of its random forecast's. This is reflected in the skill score
which represents a 25% improvement over the random forecast.
5-30
-------
Program LM
"S
t
ft
o
No
Yes
A
B
FAR
CS\
POD
SS
Aref
Forecasted
No Yes
130 8
20 22
84
0.71
27
44
52
25
68
•o
0) ^^_^^_^
j> No
0 Yes
A
B
FAR
CSI
POD
Random
Forecasted
No Y
115 2
35
68
0.71
77
11
17
9S
3
7
Figure 5-15. Hypothetical verification statistics for a two-category forecast for Program LM
and its random forecast.
Three- or more-category forecasts
Categorical forecasts can have three or more categories. In this case, computing the
verification statistics becomes more complicated. For the four-category table shown in
Figure 5-16, the accuracy (A) of the entire forecasting program is computed using the following
equation:
A=[(k+p+u+z)/N]*100
(5-2)
where:
N = total number of events in the table
Computing the verification statistics listed in Table 5-6 first involves collapsing a four-
category table to a two-category table. To collapse the table, complete the following steps:
1. Pick an ozone or PM2.5 concentration that separates two categories (such as 85 ppb,
which separates the Good and Moderate categories from the Unhealthy for Sensitive
Groups and Unhealthy categories in Figure 5-16). The ability to forecast above or below
this value will be evaluated, not the performance of each category in the table.
2. Each cell of the four-category table must be assigned to one of the four cells (a, b, c, or d)
of the two-category table shown in Figure 5-12. Each cell should be assigned according
to the following criteria for the four possible scenarios:
Cell a = event not forecasted and not observed.
Cell b = event forecasted but not observed.
5-31
-------
-Q Good
CD
••- Moderate
-Q Unhealthy for
O Sensitive Groups
Unhealthy
Forecasted
& .^ s^^r xr
(y ^ c^& \j^
k
0
s
w
1
p
t
X
m
q
u
y
n
r
V
z
FA
Figure 5-16. Contingency table for a four-category forecast.
Cell c = event not forecasted but observed.
Cell d = event forecasted and observed.
3. Once the assignments are made, total all of the values corresponding to each letter and
place them in the respective cell of the two-category table.
4. Calculate the verification statistics described in Table 5-6.
These forecast evaluation measures enable a forecaster to objectively quantify how well
pollution is forecasted. They should become a regular part of any forecasting program.
5.6.4 Methods to Further Evaluate Forecast Performance
Statistics provide useful values but do not indicate the reasons why forecast performance
may be poor. The following questions can help determine the cause of performance problems.
1. Are there any modifications in the monitoring network that could possibly skew the
observations, for example, outage problems and new or decommissioned monitoring
sites? Have the emissions patterns changed around the monitoring site (e.g., is a rural site
now surrounded by buildings and roadways?) that might make the data collected at that
site unrepresentative?
2. Are there any systematic forecast biases that are not evident in annual summary statistics,
but are shown by evaluating the forecast performance on a finer time scale? For
example, Figure 5-17 shows the forecast bias each day and with a 2-week running
average that reveals a positive bias starting around July 1 that persists throughout the
season.
5-32
-------
3. Are there any personnel changes in the forecasting team? Should there be more cross-
training to transfer knowledge and experiences to the new forecasters?
4. Did the forecasters interpret meteorological information incorrectly? Were all forecast
and observation data available to make an accurate forecast?
5. How reliable is the forecast guidance, including meteorological observations, forecast
model output, and other meteorological forecast products? How accurate were the key
weather forecasts (e.g., maximum temperature, wind speed, cloud cover, etc.) used to
formulate the air quality forecast?
Date
Figure 5-17. An example of forecast bias for a 24-hr ozone forecast. The dashed line shows
the bias each day, and the solid line shows the 2-week moving average of the
bias.
5-33
-------
6. REFERENCES
Allen G., Sioutas C., Koutrakis P., Reiss R., Lurmann F.W., Roberts P.T., and Burton R.M.
(1997) Evaluation of the TEOM method for measurement of ambient particulate mass in
urban areas. J. Air & Waste Manag. Assoc. 47, pp. 682-689.
Altshuller A.P. and Lefohn A.S. (1996) Background ozone in the planetary boundary layer over
the United States. J. Air & Waste Manag. Assoc. 46, pp. 134-141.
Bruckman L. (1993) Overview of the Enhanced Geocoded Emissions Modeling and Projection
(Enhanced GEMAP) System. In proceeding of the Air & Waste Management
Association's Regional Photochemical Measurements and Modeling Studies Conference,
p 562, San Diego, CA.
Cassmassi J.C. (1987) Development of an objective ozone forecast model for the South Coast
Air Basin. Presented at the Air Pollution Control Association 80th Annual Meeting, New
York, NY, June 21-26.
Chang J.S., Jin S., Li Y., Beauharnois M., Chang K.H., Huang H.C., Lu C.H., and Wojcik G.
(1996) The SARMAP Air Quality Model, SARMAP Final Report Part 1.
Chinkin L.R., Main H.H., Anderson C.B., Coe D.L., Haste T.L., Hurwitt S.B., and Kumar N.
(1998) Study of air quality conditions including ozone formation, emission inventory
evaluation, and mitigation measures for Crittenden County, Arkansas. Report prepared
for the Arkansas Department of Pollution Control and Ecology, Little Rock, AR by
Sonoma Technology, Inc., Petaluma, CA, STI-998310-1837-DFR, November.
Claiborn C.S., Finn D., Larson T.V., and Koeing J.Q. (2000) Windblown dust contributes to high
PM2.5 concentrations. J. Air & Waste Manag. Assoc. 50, pp. 1440-1455.
Coats CJ. (1996) High performance algorithms in the sparse matrix operator kernel emissions
modeling system. Proceedings of the Ninth Joint Conference on Applications of Air
Pollution Meteorology of the American Meteorological Society and the Air and Waste
Management Association, Atlanta, GA.
Comrie A.C. (1997) Comparing neural networks and regression models for ozone forecasting. J.
Air & Waste Manag. Assoc. 47, pp. 653-663.
Comrie A.C. and Yarnal B. (1992) Relationships between synoptic-scale atmospheric circulation
and ozone concentrations in metropolitan Pittsburgh, Pennsylvania. Atmos. Environ. 26B,
pp. 301-312.
Cooperative Program for Operational Meteorology, Education and Training (COMET) (2002).
Available on the Internet at http://www.comet.ucar.edu.
Dockery D.W. and Pope C.A., III (1994) Acute respiratory effects of particulate air pollution.
Annu Rev Public Health 15, pp. 107-132.
Draxler R.R. and Hess G.D. (1997) Description of the Hysplit 4 modeling system. Technical
memorandum by NOAA, ERL ARL-224, December 24.
Dye T.S., MacDonald C.P., and Miller D.S. (2002a) Summary of PM2.5 forecasting tool
development for Pittsburgh, Pennsylvania. Technical memorandum prepared for the
6-1
-------
Pennsylvania Department of Environmental Protection and Allegheny County Health
Department, Pittsburgh, PA, by Sonoma Technology, Inc., Petaluma, CA, STI-901491-
2259-TM, September.
Dye T.S., Miller D.S., and MacDonald C.P. (2002b) Summary of PM2.5 forecasting program
development and operations for Salt Lake City, Utah during winter 2002. Technical
memorandum prepared for Ms. Cheryl Keying, Utah Department of Environmental
Quality by Sonoma Technology, Inc., Petaluma, CA 94954, 901491-2243-TM, August.
Dye T.S., Ray S.E., Lindsey C.G., Arthur M., and ChinkinL.R. (1996) Summary of ozone
forecasting and equation development for the air districts of Sacramento, Yolo-Solano,
and Placer. Vol. I: ozone forecasting. Vol. II: equation development. Final report
prepared for Sacramento Metropolitan Air Quality Management District, Sacramento,
CAby Sonoma Technology, Inc., Santa Rosa, CA, STI-996210-1701-FR, December.
Environ (1998) User's Guide - Comprehensive Air Quality Model with Extensions (CAMx).
Version 2.0. By Environ International Corporation, Novato, CA, December.
Falke S. (1999) Draft PM2.5 topic summary. Available on the Internet at
http://capita.wustl.edu/PMFineAVorkgroup/Status&Trends/Reports/In-
progress/PM25Maps/PM25Maps/sld001.htm.
Falke S.R., Husar R.B., and Schichtel B.A. (2001) Fusion of SeaWiFS and TOMS satellite data
with surface observations and topographic data during extreme aerosol events. J. Air &
Waste Manag. Assoc. 51, pp. 1579-1585.
Gardner M.W. and Dorling S.R. (1998) Artificial neural networks (the multilayer perceptron) - a
review of applications in the atmospheric sciences. Atmos. Environ. 32, pp. 2627-2636.
Grell G.A., Dudhia J., and Stauffer D.R. (1994) A description of the fifth-generation Penn
State/NCAR mesocale model (MM5). Prepared by National Center for Atmospheric
Research, Boulder, CO, NCAR Technical Note-398.
Health Effects Institute (2002) Prespectives: understanding the health effects of components of
the paniculate matter mix: progress and next steps. Available on the Internet at
www.healtheffects.org.
Hering S. and Cass G. (1999) The magnitude of bias in the measurement of PM2.5 arising from
volatilization of particulate nitrate from Teflon filters. J. Air & Waste Manag. Assoc. 50,
pp.613-632.
Horie Y. (1988) Air Quality Management Plan 1988 Revision, Appendix V-P: Ozone episode
representativeness study for the South Coast Air Basin. Report prepared for the South
Coast Air Quality Management District, El Monte, CA by Valley Research Corporation,
Van Nuys, CA, VRC Project Number 057, March.
Hubbard M.C. and Cobourn W.G. (1997) Development of a regression model to forecast ground-
level ozone concentration in Louisville, KY. Atmos. Environ, (submitted).
Husar R.B. (1998) Spatial pattern of 1-hour and 8-hour daily maximum ozone over the OTAG
region. Presented at the Air & Waste Management Association's 91st Annual Meeting &
Exhibition, San Diego, CA, June 14-18.
6-2
-------
Jorquera M.E. (1998) The use of episodic controls to reduce the frequency and severity of air
pollution events. Submitted to the Transportation Research Board, Transportation and Air
Quality Committee, Federal Highway Administration, Baltimore, MD, February.
Lambeth B. (1998) Texas Natural Resource Conservation Commission, Austin, TX. Personal
communication.
Lurmann F.W. (2000) Simplification of the UAMAERO Model for seasonal and annual
modeling: the UAMAERO-LT Model. Report prepared for South Coast Air Quality
Management District, Diamond Bar, CA by Sonoma Technology, Inc., Petaluma, CA,
STI-999420-1996-FR, August.
Lurmann F.W., Wexler A.S., Pandis S., Musarra S., Kumar N., Seinfeld J.H., and Hering S.V.
(1997) Development of an acid deposition model (UAM-AERO) for the South Coast Air
Basin. Final report prepared for California Air Resources Board, Sacramento, CA by
California Institute of Technology, Pasadena, CA, Sonoma Technology, Inc., Santa Rosa,
CA, and Aerosol Dynamics, Inc., Berkeley, CA, ARE Contract No. 92-311.
MacDonald C.P., Roberts P.T., Main H.H., Kumar N., Haste T.L., Chinkin L.R., and Lurmann
F.W. (1998) Analysis of meteorological and air quality data for North Carolina in support
of modeling. Report prepared for North Carolina Department of Environment and Natural
Resources, Division of Air Quality, Raleigh, NC by Sonoma Technology, Inc., Petaluma,
CA, STI-997420-1818-DFR, October.
Main H.H. and Roberts P.T. (2001) PM2.5 Data Analysis Workbook. Draft workbook prepared
for the U.S. Environmental Protection Agency, Office of Air Quality Planning and
Standards, Research Triangle Park, NC, by Sonoma Technology, Inc., Petaluma, CA,
STI-900242-1988-DWB, February.
McHenry J., Seaman N.L., Coats C.J., Lario-Gibbs A., Vukovich J., Wheeler N., and Hayes E.
(1999) Real-time nested mesoscale forecast of lower tropospheric ozone using a highly
optimized coupled model numerical prediction system. Preprints in AMS Symposium on
Interdisciplinary Issues in Atmospheric Chemistry, Dallas, TX, January 10-15, American
Meteorological Society.
Morris R.E., Myers T.C., Carr E.L., Causley M.C., and Douglas S.G. (1990) User's Guide for the
Urban Airshed Model. Volume II: User's Guide for the UAM (CB-IV) Modeling
System. Prepared for the Office of Air Quality Planning and Standards, U.S.
Environmental Protection Agency, Research Triangle Park, NC, EPA-450/4-90-007B.
Murphy A.H. (1991) Forecast verification: its complexity and dimensionality. Mon. Wea. Rev.
119, pp. 1590-1601.
Murphy A.H. (1993) What is a good forecast? An essay on the nature of goodness in weather
forecasting. Weather and Forecasting 8, pp. 281-293.
Murphy A.H. and Winkler R.L. (1987) A general framework for forecast verification. Mon. Wea.
Rev. 115, pp. 1330-1338.
National Research Council (1991) Rethinking the Ozone Problem in Urban and Regional Air
Pollution, National Academy of Sciences/National Research Council, National Academy
Press, Washington, DC.
6-3
-------
NESCAUM (1998) 8-hr and 1-hr ozone exceedances in the NESCAUM Region (1993-1997).
Report prepared by the Northeast States for Coordinated Air Use Management, Boston,
MA.
Odman T. and Ingram C.L. (1996) Multiscale Air Quality Simulation Platform (MAQSIP):
source code documentation and validation. MCNC Technical Report, ENV-96TR002-
vl.O.
Paul R.A., Biller W.F., and McCurdy T. (1987) National estimates of population exposure to
ozone. Paper no. 87-42.7 presented at the Air Pollution Control Association 80th Annual
Meeting and Exhibition, Pittsburgh, PA.
Pielke R.A., Cotton W.R., Walko R.L., Tremback C.J., Lyons W.A., Grasso L., Nicholls M.E.,
Moran M.D., Wesley D.A., Lee T.J., and Copeland J.H. (1992) A comprehensive
meteorological modeling system - RAMS. Meteor. Atmos. Phys. 49, pp. 69-91.
Prospero J.M. (1999) Long-range transport of mineral dust in the global atmosphere: Impact of
African dust on the environment of the southeastern United States. Proc. Natl. Acad. Sci.
96, pp. 3396-3404.
Ruiz-Suarez J.C., Mayora-Ibarra O.A., Torres-Jimenez J., and Ruiz-Suarez L.G. (1995) Short-
term ozone forecasting by artificial neural networks. Advances in Engineering Software
23, pp. 143-149.
Ryan W.F. (1994) Forecasting severe ozone episodes in the Baltimore metropolitan area. Atmos.
Environ. 29, pp. 2387-2398.
Ryan W.F., Dickerson R.R., Doddridge E.G., Morales R.M., and Piety C.A. (1998) Transport
and meteorological regimes during high ozone episodes in the mid-Atlantic region:
observations and regional modeling. Preprints of the 10 Joint Conference of the
Applications of Air Pollution Meteorology with the Air and Waste Management
Association, January 11-16, Phoenix, AZ., pp. 168-172, American Meteorological
Society, Boston, MA.
Saxton K., Chandler D., Stetler L., Lamb B., Claiborn C., and Lee B.H. (2000) Wind erosion and
fugitive dust fluxes on agricultural lands in the Pacific northwest. Transactions of the
ASAE43, pp. 623-630.
SCAQMD (1997) 1997 revision to the Air Quality Management Plan - Appendix II: current air
quality. South Coast Air Quality Management District, Diamond Bar, CA.
Schauer J.J., Rogge W.F., Hildemann L.M., Mazurek M.A., Cass G.R., and Simoneit B.R.T.
(1996) Source apportionment of airborne particulate matter using organic compounds as
tracers. Atmos. Environ. 30, pp. 3837-3855.
Schere K.L. and Demerjian K.L. (1984) User's guide for the photochemical box model (PBM).
U.S. Environmental Protection Agency, Office of Research and Development, Research
Triangle Park, NC 27711, EPA-600/8-84-022a.
Schwartz J. (1994) What are people dying of on high air pollution days? Environ Res 64 (19), pp.
26-35.
Seinfeld J.H. and Pandis S.N. (1998) Atmospheric chemistry and physics: from air pollution to
global change., J. Wiley, New York.
6-4
-------
Stein A.F., Lamb D., and Draxler R.R. (2000) Incorporation of detailed chemistry into a three-
dimensional Lagrangian-Eulerian hybrid model: Application to regional tropospheric
ozone. Atmos. Environ. 34, pp. 4361-4372.
Stephenson D.B. (2000) Use of the "odds ratio" for diagnosing forecast skill. Weather and
Forecasting 15 (2), pp. 221-232.
Stoeckenius T. (1990) Adjustment of ozone trends for meteorological variation. Presented at the
Air and Waste Management Association's Specialty Conference, Tropospheric Ozone and
the Environment, Los Angeles, CA, March 19-22.
Systems Applications International (1999) User's Guide to the Variable-Grid Urban Airshed
Model (UAM-V). By Systems Applications International, Inc./ICF Consulting, 101
Lucas Valley Road, Suite 160, San Rafael, California 94903.
Taylor R. (1998) New York State Department of Environmental Conservation, Albany, NY.
Personal communication.
Turpin B.J., Huntzicker J.J., and Hering S.V. (1994) Investigation of organic aerosol sampling
artifacts in the Los Angeles basin. Atmos. Environ. 28, pp. 3061-3071.
Turpin B.J., Huntzicker J.J., Larson S.M., and Cass G.R. (1991) Los Angeles summer midday
particulate carbon-primary and secondary aerosol. Environ. Sci. Technol. 25, pp. 1788-
1793.
U.S. Environmental Protection Agency (1984) Quality assurance handbook for air pollution
measurement systems: Volume II. Ambient air specific methods. Sections 2.1, 2.2, 2.6,
and 2.9. Report prepared by the Environmental Monitoring Systems Laboratory, U.S.
Environmental Protection Agency, Research Triangle Park, NC, EPA-600/4-77-027a,
July.
U.S. Environmental Protection Agency (1986) Guideline on the identification and use of air
quality data affected by exceptional events. EPA-450/4-86-007, July. Available on the
Internet at www.epa.gov/ttn/amtic/files/ambient/criteria/reldocs/4-86-007.pdf.
U.S. Environmental Protection Agency (1992) User's guide for the urban airshed model.
Volume IV: User's manual for the emissions preprocessor system 2.0. Part A: Core
FORTRAN system. Report prepared by U.S. Environmental Protection Agency, Office of
Air Quality Planning and Standards, Research Triangle Park, NC, EPA-450/4-90-
007D(R), June.
U.S. Environmental Protection Agency (1996) Clearinghouse for inventories and emission
factors. On U.S. Environmental Protection Agency electronic bulletin board.
U.S. Environmental Protection Agency (1997a) National air pollutant emission trends, 1990-
1996. Report prepared by the Office of Air Quality Planning and Standards, Research
Triangle Park, NC, EPA-454/R-97-011, December.
U.S. Environmental Protection Agency (1997b) Survey and review of episodic control programs
in the United States. EPA 420-R-97-003, September.
U.S. Environmental Protection Agency (1998) EPA third-generation air quality modeling
system, Models-3, Volume 9B: user manual. Report prepared by the National Exposure
6-5
-------
Research Laboratory, Office of Research and Development, U.S. Environmental
Protection Agency, Research Triangle Park, NC, EPA-600/R-98/069(a), June.
U.S. Environmental Protection Agency (2001) Data quality objectives (DQOs) and model
development for relating federal reference method (FRM) and continuous PM2.5
measurements to report an air quality index (AQI). Available on the Internet at
www.epa.gov/ttn/amtic/files/ambient/monitorstrat/aqidqorept.pdf. EPA-454/R-01-002,
February.
University of Illinois Urbana-Champaign, Department of Atmospheric Sciences (2002) Weather
World 2010 educational web site. Available on the Internet at
http://ww2010.atmos.uiuc.edu/(Gh)/guides/mtr/home.rxml.
Wallace J.M. and Hobbs P.V. (1977) Atmospheric Science, Academic Press, New York.
Ward D.E. (1999) Smoke from wildland fires. Health Guidelines for Vegetation Fire Events,
Lima, Peru, 6-9 October 1998, Background papers, WHO.
Wilks D.S. (1995) Statistical methods in the atmospheric sciences, Academic Press, San Diego,
CA, p 467.
Wilson W.E. and Suh H.H. (1997) Fine particles and coarse particles: concentration relationships
relevant to epidemiologic studies. J. Air & Waste Manag. Assoc. 47 (33), pp. 1238-1249.
Yamartino R.J., Scire J.S., Carmichael G.R., and Chang Y.S. (1992) The CALGRID mesoscale
photochemical grid model -1. Model formulation. Atmos. Environ. 26A, pp. 1493-1512.
6-6
-------
This page is intentionally blank.
-------
TECHNICAL REPORT DATA
(Please read Instructions on reverse before completing)
1. REPORT NO.
EPA-456/R-03-002
3. RECIPIENT1 S ACCESSION NO.
4. TITLE AND SUBTITLE
Guidelines for Developing an Air Quality (Ozone and PM2.5)
Forecasting Program
5. REPORT DATE
June 2003
6. PERFORMING ORGANIZATION CODE
7. AUTHOR(S)
Timothy S. Dye, Clinton P. MacDonald, Craig B. Anderson, Hilary R. Hafner
Neil J. M. Wheeler, Alan C. Chan (Sonoma Technology, Inc.)
8. PERFORMING ORGANIZATION REPORT NO.
STI-902461-2295-FR
9. PERFORMING ORGANIZATION NAME AND ADDRESS
Information Transfer Group-AIRNow Program (E143-03)
U. S. Environmental Protection Agency
Office of Air Quality Planning and Standards
Research Triangle Park, NC 27711
10. PROGRAM ELEMENT NO.
27/53/E
11. CONTRACT/GRANT NO.
Contract No. GS-10F-0181K
Task Order No. 1303
12. SPONSORING AGENCY NAME AND ADDRESS
13. TYPE OF REPORT AND PERIOD COVERED
Office of Air Quality Planning and Standards
Information Transfer and Program Integration Division
U.S. Environmental Protection Agency
Research Triangle Park, NC 27711
Final
14. SPONSORING AGENCY CODE
EPA/200/04
15. SUPPLEMENTARY NOTES
Document is available electronically at "http://www.epa.gov/airnow/"
16. ABSTRACT
This report provides technical guidance to help air quality agencies develop, operate, and
evaluate ozone and PM2.5 forecasting programs. This document provides:
1) Background information about ozone and PM2.5 and the weather's effect on these
pollutants.
2) A list of how air quality forecasts are currently used.
3) A summary and evaluation of methods currently used to forecast ozone and PM2.5.
4) Steps to develop and operate an air quality forecasting program.
5) Information on the level of effort needed to set up and operate a forecasting program.
17.
KEY WORDS AND DOCUMENT ANALYSIS
DESCRIPTORS
b. IDENTIFIERS/OPEN ENDED TERMS
c. COSATI Field/Group
Ozone, Air Quality Forecasting, AQI
PM2.5, Ambient Monitoring
Air Quality Data and Forecasting,
AQI Reporting
18. DISTRIBUTION STATEMENT
Release Unlimited
19. SECURITY CLASS (Report)
Unclassified
21. NO. OF PAGES
125
20. SECURITY CLASS (Page)
Unclassified
22. PRICE
EPA Form 2220-1 (Rev. 4-77) PREVIOUS EDITION IS OBSOLETE
-------
United States
Environmental Protection
Agency
Office of Air Quality Planning and Standards
Information Transfer and Program Integration Division
Research Triangle Park, NC
Publication No. EPA-456/R-03-002
June 2003
Information Transfer Group - AIRNow Program (E143-03)
EPA-456/R-03-002
www.epa.gov/airnow
June 2003
------- |