April 1984

              "SSS."
                    OF      .
AIR QUALITY MODELS PERTAINING TO PARTICULATE MATTER
    ENVIRONMENTAL SCIENCES RESEARCH LABORATORY
       OFFICE OF RESEARCH AND DEVELOPMENT
      U.S. ENVIRONMENTAL PROTECTION AGENCY
        RESEARCH TRIANGLE PARK, NC 27711

-------
AIR QUALITY MODELS PERTAINING TO PARTICULATE MATTER
                        by
   S.A. Batterman, J.A.  Fay,  D.  Golomb,  J.  Gruhl
                 Energy Laboratory
       Massachusetts Institute of Technology
                Cambridge,  MA 02139
      Cooperative Agreement Number 809229-01
                  Project Officer

                 Jack  H.  Shreffler
        Meteorology and Assessment  Division
    Environmental Sciences Research Laboratory
         Research Triangle Park, NC 27711
    ENVIRONMENTAL SCIENCES RESEARCH LABORATORY
        OFFICE OF RESEARCH AND DEVELOPMENT
       U.S. ENVIRONMENTAL PROTECTION AGENCY
         RESEARCH TRIANGLE PARK, NC 27711

-------
                                DISCLAIMER
    This report has been reviewed by the Environmental  Sciences  Research
Laboratory, U.S. Environmental  Protection Agency,  and approved for
publication.   Approval does not signify that the  contents  necessarily
reflect the views and policies  of the U.S.  Environmental  Protection
Agency, nor does mention of trade names or commercial products constitute
endorsement or recommendation for use.
                                    n

-------
                                 ABSTRACT
    This report describes an evaluation of the  Particle  Episodic Model
(PEM), an urban scale dispersion model  which  incorporates deposition,
gravitational settling and linear transformation  processes  into the
predecessor model, the Texas Episodic Model (TEM-8).   A  sensitivity
analysis of the model was performed,  which included the  effects of
deposition, gravitational settling and  receptor grid  size.
Recommendations are made to improve the performance and  flexibility  of  the
model.

    PEM was applied to a source inventory of  the  Philadelphia  area to
provide a preliminary estimate of source apportionment.  PEM modeling
employed both hypothetical and actual meteorology.   Results indicate that
area source emissions dominate TSP, SO? and sulfate concentrations at
urban receptors.  A large fraction of the inhalable particles  may arrive
from distant sources.

    This report also contains an overview of  receptor models  (RMs) used
for the source apportionment of aerosols.  Some diagnostic  procedures for
RMs are evaluated using a synthetic data set.  Described are RM trade-offs
and protocols and possible hybrid dispersion/receptor models.   Issues
regarding the inter-comparison of source apportionments  from receptor and
dispersion models are highlighted with  reference  to the  1982 Philadelphia
study.

    This report was submitted in fulfillment  of Cooperative Agreement
Number 809229-01 by the M.I.T. Energy Laboratory  under the  sponsorship  of
the U.S. Environmental Protection Agency.  This report covers  a period
from April to October 1983 and work was completed as  of  November 4,  1983.

-------
                            ACKNOWLEDGEMENTS
    The authors appreciate the  guidance  and  assistance offered by the
Project Officers,  Drs.  K.  Demerjian  and  J. Shreffler, and the comments,
discussion and data provided by Drs.  R.  Stevens, T. Dzubay, C. Lewis and
the scientists and engineers of the  environmental  agencies of the City
of Philadelphia and the States  of  New Jersey and Pennsylvania.

-------
                                CONTENTS

Abstract	iii
Figures	,	vi
Tables   	vii

    1. Introduction  	   1

    2. Conclusions and Recommendations 	   3

    3. Evaluation of Particle Episodic Model 	   5
         PEM Verification  	   5
         Transferral and Compilation 	   5
              Comparison to TEM-8  	   6
              Deposition and Settling  	   6
              Transformation 	  10
              Area Sources	11
         PEM Improvements	15
              Deposition, Settling and Transformation  	  16
              Source Specifications  	  18
              Area Sources   	19
              Outputs	18
              Miscellaneous  	  19
         Summary   	21

    4. Source Apportionment Using PEM  	  22
         Philadelphia Source Inventory 	  22
         Meteorology   	29
         Dispersion Modeling 	  29
              Hypothetical Meteorology 	  30
              Actual Meteorology 	  35
         Discussion	40

    5. Receptor Models	43
         Overview of Receptor Models 	  43
              Chemical Mass Balance Receptor Models  	  45
              Multivariate Receptor Models   	  47
              Composite Receptor Models  	  47
         Diagnostic Procedures for CMB Models  	  48
              Collinearity   	  48
              Source Selection 	  51
              Influential Characteristics  	  53
              Robust Regression  	  60
         Hybrid Receptor/Dispersion Models 	  61
         Comparisons Between Source and Receptor  Models   	  64
              Evaluative Criteria  	  64
              Averaging Time	65
         Receptor Model Protocols  	  67
              Protocols for Monitoring and Analysis  	  58
              Modeling Protocols 	  69
              Protocols for Inter-Comparison of Receptor  Models   .  70
         Summary   	70

References   	72

-------
                                  FIGURES
Number                                                                  Page
1.  Ambient downwind concentrations  from test  point  source
    under stability A 	   8
2.  Ambient downwind concentrations  from test  point  source
    under stability D 	   8
3.  Downwind ambient concentrations  for stability B  from  point
    and area sources with transformation	12
4.  Percent secondary sulfate (of sulfur dioxide concentrations)
    for conditions in Figure 3	12
5.  Downwind center line concentrations from 1  km square area  source
    for various receptor grid sizes  (km) under stability  D   	   13
6.  PEM map of SO? concentrations for 1 point  and 3  area  sources
    showing receptors where concentrations  are calculated .......   13
7.  PEM map using 2.5 km receptor grid showing concentrations from
    area sources under stability E and 8 consecutive 45 degree
    wind shifts	15
8.  PEM map modeling area sources as point  sources and stability  D,
    otherwise conditions as in Figure 7	15
9.  Map showing location of point and area  sources in the
    Philadelphia inventory  	   23
10. Map showing emission rates of coarse particles from
    Philadelphia area sources 	   27
11. Map showing emission rates of fine particles from area  sources   .  .   28
12. Map showing emission rates of S02 from  area sources    	     28
13. Map showing receptor grid of Philadelphia  area,  sub-census  tract
    areas, city limits and locations of monitors  	   31
14. Partial regression plots for 8 source classes   	   58

-------
                                   TABLES
Number                                                                  Page
1.  Comparison of ambient participate concentrations and  fraction
    of gaseous concentration for test point source for  several
    deposition and settling velocities  	    9
2.  Summary of TSP, S0£ and $04 emission inventories in the
    greater Philadelphia area 	   24
3.  Summary of TSP and S02 and 504 emission inventories
    in sources in the local Philadelphia area	    25
4.  Primary sulfate emission factors  	   27
5.  TSP source class contributions from PEM using  hypothetical
    meteorology under three stability categories   	   32
6.  S02 and 504 source class contributions from PEM using
    hypothetical meteorology for three stability categories 	   33
7.  Average percent TSP source class contributions from PEM using
    hypothetical meteorology  	   34
8.  Average percent $04 source class contributions from PEM
    modeling using hypothetical meteorology 	   34
9.  Meteorological data used in PEM simulation	36
10. TSP source class contributions using actual meteorology 	   38
11. S02 and $64 source contributions using actual  meteorology ...       39
12. Apportionment of hourly SC>2 and SOa concentrations  at
    selected city receptors for 12 wind directions  	   41
13. Differences between classes of receptor and dispersion  models  ...   46
14. Source signatures from Quail Roost II 	   46
15. Proportionate variance decomposition of source profiles 	   50
16. Simple correlation of source profiles   	   50
17. Variance inflation factors of the source profiles 	   50
18. Results of all possible regression procedure using  synthetic
    data and Mallow's Cp statistic	     54
19. Estimated CMB source apportionment using 8  source classes 	   57
20. DFBETAS row deletion statistics for CMB of  synthetic  data set  ...   57
                                     vii

-------
                                 SECTION 1

                               INTRODUCTION
    This is the final report for  the April-October  1983  phase  of  the
EPA-MIT Cooperative Agreement examining aerosol  models and  their
applications.  The principal objective of this work is to assess  the  two
fundamental approaches to source  apportionment of aerosol concentrations:
source-oriented dispersion modeling (DM) and receptor modeling (RM).   We
aim to define the strengths, limitations, areas  of  applicability  and
possible protocols for the use of these methods  in  the regulatory context.
The specific objectives include:

(1) Assessment of the new Particle Episodic Model (PEM), a  dispersion model
which incorporates deposition, settling and transformation  processes  into
the standard Gaussian plume dispersion algorithm.  PEM  is an extension of
the Texas Episodic Model (TEM-8).  This report contains  a discussion  of the
verification, applicability and limitations of PEM. Suggestions  are  given
to improve the versatility of the model. A validation study,  in  which
predicted and observed concentrations are compared, is the  next logical
step in the model evaluation process.

(2)  Review and general comments  about receptor  models.  Part  of  this
project has involved a review of  the models and  literature  that make  up the
current understanding of receptor techniques, as well as our experience
with receptor models and similar  statistical problems.   The current status
and capabilities of RM are discussed, with  attention to  the potential uses
of RM in the regulatory process and the application of diagnostic tools to
RM which might be part of the modeling protocol.

(3)  Discussion of potential hybrid dispersion-receptor  models.   There are
a number of ways to combine source and receptor  models.  Currently, it is
not clear that hybrid models are  feasible and desirable. However,  it is
clear that contributions from each approach can  be  made  to  improve the
other technique in certain circumstances.  Recommendations  are made

-------
regarding the development and application  of  hybrid models in inter-model
comparisons, such as the Philadelphia Study (below).

(4)  Comparison of DM and RM in  the  Philadelphia Study.   In July and August
of 1982 data were collected in the Philadelphia area,  in  part to compare
source apportionments from DM and RM  on  the same, real  data.  Ground
rules and certain specific tasks are required to make  that comparison as
meaningful as possible.   Guidelines  and a  discussion of issues for this
comparison are presented.  Preliminary modeling results using PEM and a
source inventory compiled for the Philadelphia area are given.

    Data from the Philadelphia Study were  not available for this report.
Consequently, much of the analysis of receptor models  contained herein is
theoretical or hypothetical.  Further analysis using both real and
appropriate synthetic data will  be required to determine  the value and
application of the proposed approaches.

-------
                                  SECTION 2
                       CONCLUSIONS AND RECOMMENDATIONS

    A new dispersion model, the Particle Episodic Model  (PEM),  was  found  to
successfully incorporate deposition,  gravitational settling  and linear
transformation processes to the predecessor model, the Texas Episodic
Model.  Thus, PEM should permit greater realism in urban scale  modeling.
However, some improvements to area source calculations seem  warranted.

    Application of PEM to a source inventory for the Philadelphia area  and a
sensitivity analysis indicate:

(1) Area source emissions may dominate TSP, SOg and sulfate  concentrations
at receptors located within an urban  scale.  Therefore,  source  inventory
data for area sources, including source strengths, operating schedules,
micro-inventories (around receptor sites)  and the degree and method of
aggregation of small sources may be critical to accurate dispersion
modeling.  Distant sources warrant less attention.

(2) A large fraction of the observed  inhalable particle  (IP)  mass,
particularly sulfate, probably comes  from medium and long range transport
and not from local sources.  Preliminary modeling has shown  that sulfate
contributions from local sources are  generally only a few ug/m  , compared
with measured values which average about 24 ug/m .  However,  the modeling
examined a relatively short period,  and the source inventory is known to  be
very approximate.

(3) Particle and gaseous concentrations from middle to far field sources
(greater than 10 km) may be sensitive to deposition velocities, with greater
effects at larger distances.  For these sources, gravitational  settling is
relatively unimportant.  Gravitational  settling may be important only for
sources located in the vicinity of monitors which emit large particles  with
high settling velocities.

-------
(4) Sulfate concentrations  from distant sources are sensitive to
transformation rate,  especially at  low wind speeds.


    Our review of the literature and experience with receptor models
indicate:

(1) The temporal  and  spatial  variation of  source profiles, and the
sensitivity of estimated apportionment to  such variation  is  largely
unknown.   The amount  of data  needed for representative and useful results is
also poorly defined.

(2) Generally, RM studies have been custom designed, including selection of
source signatures, filters  and analysis.   A high degree of subjectivity may
be involved in the interpretation and use  of  data  and models.

(3) Each type of RM is prone  to certain failures and limitations.  Chemical
mass balance RMs are  subject  to problems of collinearity  and influential
points.  Diagnostic procedures can  be used to determine whether these
problems exist; remedial procedures may be able to minimize  their effects.
Standardized protocols may be required to  ensure meaningful  results and
promote appropriate uses of RMs.
    Both dispersion and receptor models have useful  attributes for  the
source apportionment of aerosols.  Dispersion models are  predictive and
diagnostic.  Receptor models are primarily interpretive.   Standardized uses
of receptor approaches are possible;  however, their  applicability and
limitations need to be defined.
    Hybrid models, which combine aspects from both dispersion and receptor
approaches, may have application to many air pollution  problems, including
apportionment of ambient concentrations of criteria  and hazardous
pollutants, visibility impairement and acid deposition.   However, at
present, hybrid models are poorly developed and defined.   The development
and validation of hybrid models  will  require an extensive data set  for
various conditions and localities.
                                   4

-------
                                 SECTION 3
                 EVALUATION OF THE PARTICLE EPISODIC MODEL

    The Particle Episodic Model (PEM) is based on the Texas Episodic Model,
Version-8.  The new model successfully incorporates simplified deposition,
gravitational settling and linear transformation processes into the
predecessor model.  PEM simultaneously calculates concentration and
deposition over an urban scale area in a rectangularly gridded receptor
network.  A discussion of the the underlying concepts and algorithms in PEM
may be found in Rao (1983).
    With respect to deposition, PEM provides greater flexibility and some
increased realism over comparable models.  For example,  the ISC model
simulates gravitational settling by tilting the (normally horizontal)  plume
center!ine; deposition is modeled by the partial reflection of the source
contribution.  Settling and deposition processes alter the distribution of
mass within the plume.  The ISC models's treatment is not altogether
statisfactory since it does not model the modified distribution.
Consequently, the ISC model tends to overpredict concentrations near the
source and underpredict concentrations at large distances.   However,  as
shown later, deposition and gravitational settling only slightly affect
concentrations on the urban scale (less than 60 km), especially for the
expected range of settling and deposition velocities of inhalable particles.

PEM VERIFICATION

Code Transferral and Compilation

    The PEM source code was checked for obvious errors or omissions.   No
mistakes were found.  The source code was then compiled  on the IBM FORTRAN
VS compiler.  Only one "block data" per program is permissible with this
compiler.   The source code was modified accordingly.

Comparison to TEM-8

    PEM was used with simple source and meteorological  conditions in each
stability category.   Results were compared to the TEM-8  model.   This showed
                                    5

-------
good agreement provided that the TEM time step  was  specified  to  be  10
minutes.  Both models use the same P-G-T dispersion coeffients,  but TEM
increases these coefficients depending  on averaging time.  PEM uses these
coefficients for calculating one hour averages.   (The  basic averaging period
in PEM is one hour,  in TEM,  10 minutes.)  Maximum differences between the
two models were under a few  percent. However,  under the most stable
categories at relatively long downwind  distances, maximum  differences could
range up to about 30 percent.  These differences  occur since  TEM uses
"look-up" tables for dispersion coefficients at discrete distances, while
PEM uses piece-wise approximations.  While both methods should be
sufficiently accurate for model applications,  the piece-wise  approximations
in PEM may provide more consistent results.

Deposition and Settling

    An analysis was performed to verify the model's response  and sensitivity
to changes in settling and deposition velocities.   The settling  velocity W
increases proportionally to the particle density  and the square  of  particle
diameter for particles ranging from about 0.05  to 300  urn in diameter.
Mechanically generated particles, with  diameters  from  about 2 to 50 urn,  have
settling velocities around 1 cm/s.  The settling  velocities of sulfate and
SO. are believed to be about 1 to 3 orders of magnitude smaller  than the
above value (Sehmel, 1980).
    The deposition velocity V is the deposition flux divided  by  the airborne
concentration.  Deposition velocities are functions of micrometeorological
variables (aerodynamic roughness); particle properties (diameter, density,
solubility); surface properties (surface roughness  and moisture); and
measurement height (usually 1-1.5 m).  Deposition is difficult to measure
and the uncertainties are large.  Recommended S02 deposition  velocity
estimates range from 0.04 to 7.5 cm/s (National Academy of Science,  1983);
SO- deposition velocities are generally less than 1 cm/s.  For  large
particles (greater than 50 urn), where gravitational settling  is  the primary
removal mechanism, V=W>0.  In the case  of a gas with perfect  reflection  from
surfaces, V=W=0.  The effect of both settling and deposition  is  proportional
to V/u and W/u, respectively, where u is the wind speed.
    In general, deposition velocities are greater than or  equal  to  the

-------
gravitational settling velocity.  PEM does not permit W>V>0,
representing, for example, re-entrainment of deposited particles  in  a
dust storm or soil erosion during high winds.   Such  situations  can be
modeled as area sources with emission rates which  depend  on wind  speed.
     The effect of varying deposition velocity V and settling velocity W
(in cm/s) is shown in Figures 1 and 2 and Table 1.   These examples use a
test point source with a 1 g/s emission rate,  0.13 m diameter stack  10 m
high, exit velocity of 35 m/s, and exhaust temperature of 400°K,
giving an effective plume release height of 17 m.  Meteorological
conditions are stability A (Figure 1) and D (Figure  2), and 3 m/s
winds.  For small distances (<2 km) and 3 m/s  winds, increasing either
the deposition or settling velocity from 0 to  10 cm/s results in  less
than 10 percent change of ambient concentrations with the test  point
source under stability A (Figure 1) and about  100  percent under
stability D (Figure 2).  Maximum concentrations occur about 0.5 km
downwind under stability category D.
     In contrast to near field circumstances (within 1  to 2 km  away),
concentrations from middle to far field sources (greater  than about  10
km distant) may be very sensitive to settling  and  deposition
velocities.  Table 1 shows the fraction of the gaseous concentration
(i.e., V=W=0) that results with selected deposition  and settling
velocities using the test point source and the meteorological conditions
(Stability D) in Figure 2.  For these conditions, concentrations at
large downwind distances are sensitive to small deposition velocities
(greater than 0.1 cm/s) when the settling velocity is 0.   However,
ambient concentrations may be small at these downwind distances.  For
example, a deposition velocity of 1 cm/s results in  74 percent  of the
gaseous concentration at 3 km, and 33 percent  at 60  km.   For deposition
velocities above 1 cm/s, concentrations are relatively insensitive to
settling velocities from 0 to 1 cm/s.   Settling velocities above 1 cm/s,
(representative of large particles) may have a large effect, producing
considerable near field deposition and low ambient concentrations at
large distances.
     The incorporation of deposition  and settling processes into PEM is

-------
      100'
1
c
-4

§
H



1
w
       10
                                    Q.5

                          DOWN-WIND  DISTANCE IN KILOMETERS
                                                              1.0
Figure  1.  Ambient downwind concentrations from test  point  source under
           stability A.

 3

 C
 O
 H
 I
 u
    1000
     100
      10
      1  -
                      0.5           1.0           1.5

                          DOWN-WIND DISTANCE  IN KILOMETERS
                                                             2.0
ngure 'i. Ambient downwind concentrations from test point source under
          stability D,
                                     8

-------
  Settling
  Velocity
  --cm/s--
Deposition
Velocity
--cm/s--
	Downwind Distance in Kilometers	
 —3..     .-6—     -15—     -30—    -60-
                             Ambient Concentration of Gas in ug/nT
                           8.677     2.978
                                 0.799
                              0.297    0.111
                    Ratio of Aerosol  Concentration to Gaseous  Concentration
0
0
0
0
0.01
0.10
1.00
10.00
0.997
0.969
0.736
0.139
0.996
0.957
0.662
0.086
0.992
0.934
0.534
0.040
0.990
0.909
0.434
0.024
0.982
0.874
0.333
0.009
0.01
0.10
1.00
1.00
1.00
1.00
0.737
0.749
0.877
0.664
0.667
0.819
0.537
0.552
0.708
0.434
0.451
0.606
0.339
0.342
0.477
    2.00
   10.00
 2.00
10.00
0.759
0.125
0.652
0.038
0.474
0.001
0.330
0.000
0.189
0.000
Table 1. Ambient concentrations and concentration ratios (fraction of gaseous
         concentration)  for test point source and specified deposition and
         settling velocities.

-------
expected to produce only small  changes  in  gaseous  and  IP concentrations
since local sources are the primary contributors of ambient concentrations
(excluding regional contributions).  Longer  distances  are required for
significant settling and deposition with  the typical range of velocities,
but concentrations are low at these distances.  This was corroborated using
various meteorological scenarios  and PEM modeling  with the Philadelphia
source inventory (Section 4).  Varying  deposition  and/or gravitational
velocities from 0 to 1 cm/s changed ambient  concentrations by less than  1
percent at most receptors.

Transformation

     Figure 3 shows PEM results for primary  (502)  and  secondary  (SO.)
pollutants from a point source and  an area source  of equal strength, using a
transformation rate of 1 percent  per hour.   No  primary emissions of SO,
were assumed.  Figure 4 shows that  sulfate concentrations, as a  fraction of
S02 concentration, increase proportionally with residence time (or downwind
distance) for the point source.  Sulfate concentrations decline, however,
with residence time due to dilution. For  the area source, the sulfate
fraction increases to a downwind  distance  of 7  km, whereupon it  vanishes due
to a cut-off in the PEM program (see below).
     For low wind speeds and a typical  transformation  rate of 1  percent  per
hour, the transformation of S02 emitted within  an  urban area may
constitute a large fraction of ambient  SO. concentrations.  Since primary
emissions of SO, generally are between  1.5 and  14  percent of primary S02
emissions  (the Philadelphia inventory averaged  4.3 percent), only several
hours residence time is necessary for secondary sulfate to exceed the
primary sulfate.
     Concentrations from point sources  beyond a distance of 60 km are
ignored in PEM.  This may lower the level  of predicted secondary pollutants
since there may be insufficient time for  significant transformation within
60 km of travel.  Concentrations  from area sources are calculated for only
the four downwind grid cells, which explains the sudden drop at  7 km in
Figures 3 and 4.  Consequently, the contribution of secondary pollutants
from area sources is underestimated.
                                   10

-------
3
C
     O
     U
        1QOOO
        1000
         100
          0.1
         0.01
                           5             10             IS
                            DOWN-WIND DISTANCE IN KILOMETERS
Figure 3, Downwind ambient  concentrations for stability B, 0.5 m/s winds.  One
          point source  (1 g/s,  0  plume rise); no secondary emissions; one  area
          source  (1 g/s,  100  m  square); transformation rate of 1 percent/hour.
      (N
      O
      O
      w
      i
      Z
      W
      H
      &
                                 RESIDENCE  TIME IN HOURS
                                    4           6
                                    T
                                                                   10
                          5            10             15
                            DOWN-WIND DISTANCE IN KILOMETERS
                                                              20
Figure 4. Percent secondary  sulfate  (of sulfur dioxide concentrations) for
          conditions  in Figure  3.
                                    11

-------
Area Sources

     Area source predictions  were  compared to  "equivalent" ground level
point sources.   PEM calculates  area  source contributions  in only the four
downwind areas in a pattern selected by the  program and within the area
source itself.   If the receptor grid size chosen  is very  small, a zero
concencentation will result not far  downwind of the area  source.  Figures 3,
4, 5 and 6 show that SC^ and  50^ concentrations vanish at preselected
distances downwind of the area  source.   This treatment may underestimate the
buildup of both primary and secondary pollutants, especially for area
sources which are near or smaller  than the receptor grid  size.  Thus,
receptor grid size should be  choosen so as not to loose important area
source contributions.
     Area source calculations in PEM automatically decrease the stability
category by one, in order to  better  simulate dispersion under urban
condition.  All else being equal,  concentrations  produced by stability
category A and B are the same for  area sources, while they differ for point
sources.  This treatment may  be anomalous since many urban point sources
have low stack heights and undergo similar dispersion conditions as area
sources.
     Area source modeling is  sensitive to the  size of the receptor grid.
Figure 5 shows downwind centerline concentrations from a  1 km square area
source with different receptor  grid  sizes.   The maximum concentration from
the area source is produced by  a 0.5 km receptor  grid.  Receptor grid sizes
larger than the area source (e.g., 2.5 km grid) result in low concentrations
since PEM increases the dimension  of (small) area sources to the grid cell
size and correspondingly reduces the source  intensity  (in g/s/area) of the
area source.  In comparison to  the 1 km receptor  grid, fine grids  (e.g., 0.1
to 0.25 km) produce a somewhat  lower maximum concentrations and rapid
decrease in ambient concentrations with distance.
     The concentration pattern  for the four  receptors downwind of area
sources (selected by the program depending on  the wind direction) may not be
realistic.  Figure 6, a map of  ground level  concentrations generated by PEM,
shows concentrations produced by three area  sources  (0.1, 1, and 5 km on a
side) and a ground level point  source without  plume rise. All sources have
equal source strength.  In contrast  to the area sources,  dispersion from the

-------
      z
      o
      £

      I
      a
      O

      §
      CJ
      Q

      W
         10.
1.0
         0.1
                                              1.0-
0.1'
                                   0.25
                                  0.5
                             1.0        2.0    3.0  4.0   5.0       8.0


                       DOWNWIND DISTANCE FTCM AREA SCUPCE CENTROID IN KM
                                                                10.0
 Figure  5. Downwind center!ine concentrations from 1  km square area source with

            specified  receptor grid  sizes  (km);  stability D;  3 m/s winds;  1 g/s
            emission rate;
                0.00   3 00  * OO   f.OO  • 00  10 OO  11.00  U.OO
                                                     00  U. 00  70.00  22.00  24. OO
                21 6O «7 71 74 47
                0  0  0,37 40S 44* 270 1*0 71 37  2O 1 1   «  4  2  2  1   *0  000000
                0  0  0 0 0  0  0
                            0  0 0 0  0 « I  3 11 12 37  SO 51 «2 «1  S* 51 44 31  3O 29
                                                       2 30  37 42 4^ 49  43 40
                O.OO  2.0O  4.OO  1.00  1.00  10.00 12.0O  14.00 1ft. OO  1B.OO 20.00  22, OO 24. OO
                                                                   24. OO



                                                                   JJ 00



                                                                   32 OO



                                                                   21.09



                                                                   20.00



                                                                   It. 09



                                                                   t«.OO



                                                                   If 00



                                                                   t«.M>



                                                                   t».OO



                                                                   14. OO



                                                                   13.00



                                                                   12.OO



                                                                   11.OO



                                                                   10.OO



                                                                    t 00



                                                                    • -OO '



                                                                    7 09



                                                                    4.00



                                                                    9.0O



                                                                    4.00



                                                                    3.0O



                                                                    2.0O



                                                                    1.0O



                                                                    O.OO
Figure 6. PEM map  of S02  concentrations from (top  to bottom) 3  area and 1 point

           source;  each  source has 1 g/s emission rate; solid lines indicate  area

           sources;  dashed lines  show  receptors where concentrations are calculated,


                                         13

-------
point source affects markedly more downwind receptors.  Point source
concentrations 20 km downwind are still  20 percent of peak concentrations.
Area sources influence  a more restricted set of receptors and concentrations
decrease more quickly with  distance.   (The lower stability category used for
the area source explains only part of  this effect.)  Also, the pattern for
the four downwind cells leaves  large gaps in coverage.
      To examine the above  effects in  urban scale modeling the Philadelphia
area sources were described as  area sources (2.5 km square in size) in one
simulation, and as point sources  (with ground  level release height and no
plume rise) in a second simulation.  Both coarse (10 km) and fine  (2.5 km)
grids were used to obtain eight hour average concentrations under
consecutive 45 degree shifts in wind direction and stability D.  Stability
categories for area sources were  increased by  one to permit comparison with
point sources.  In the  vicinity of the sources, modeling using area sources
produces higher concentrations  than point sources (Figure 7 and 8).
However, since area source  concentrations are  calculated only over 4 grid
cells downwind, receptors located beyond 10 km (4 x 2.5 km) have zero
concentrations using area sources, but non-zero concentrations using point
sources.  Effects of super-position of point source plumes may result in low
(2-3 percent of maximum) concentrations 20 or  30 kilometers downwind (seen,
for example, in the NE  direction  in Figure 7).
      There was generally good  agreement between area and point source
modeling using the coarse grid; concentrations were within 7 ug/m  (of an
average ambient concentration of  50 ug/m ) and perimeter concentrations
                  3                   3
were within 2 ug/m  (average of 3 ug/'m ).
      In summary, area  source modeling is affected by a number of  factors.
The agreement between point and area source modeling depends critically on
the area source dimensions  relative to the mean distance between sources.
Similarly, the agreement between  area  source modeling using various grid
sizes depends on the area source  dimension relative to the receptor grid
size.  Accurate modeling of near  field circumstances may require fine
receptor grids, appropriate to  the micro-inventory scale.  A coarser grid
may be used to esimate  mid  and  far field contributions.  Total
concentrations would be estimated as the sum of the two predictions.
                                   14

-------
                      470.OO  479.00  480. OO 485.00 49O.OO  499.OO  5OO.OO  9O9.0O  510.OO
4490.00
4447 . SO
4449.00
4442.90
AAAfl rtft
**^O. l^tf
4437. SO
4439. OO
4432. SO
4430.00
4437 . SO
4429.00
4422. SO
442O.OO
4417. 9O
4419. OO
44 12. SO
44 1O.OO

Figure 7. PEM map
sources
4491. 2O
4444 . 7O
4446 . 20
4443.70
444 1 . 2O
4438 . 70
4436 . 20
4433.70
4431.20
44 28 . 7O
4428. 2O
4423.70
442 1 . 2O
4418.70
44 1«. 2O
4413. 7O
4411.20
OOOOOO'OOOOOOOOOOO
ooooooooooooooooo
ooooooooooooooooo
00111111111111100
0 1224996868421 1OO
1 2 3 7 12 20 29 29 27 34 32 16 4 2 1 0 0
1 2 4 9 24 47 94 63 70 63 91 36 17 2 1 O 0
1 2 9 9 44 79 82 84 79 66 49 39 17 2 100
1 2 6 29 64 79 81 74 63 98 34 17 4 2 1 0 0
1 3 13 36 71 98 99 86 69 33 11 9 3 1 1 0 0
1 3 14 27 99 99 114 89 37 1O 6 4 2 1 1 00
1 3 9 18 36 62 82 49 11 6 4 2 1 1 O 0 O.
1 2 4 2O 31 42 61 39 8 9 3 1 1 0000
1 2 3 14 16 14 3Q 24 9 3 2 1 00000
01233444321 10 'OOOO
01111121111000000
470. OO 479.0O 48O.OO 489. OO 49O.OO 499.00 9OO.OO 509. OO 91O.OO
using 2.5 km receptor grid showing concentrations from area
; stability E; and 8 consecutive 45 degree wind shifts.
471. 2O 476. 2O 4S1-.20 486. 2O 49I.2O 496. 2O 5O1.2O 9O6.2O 911. 2O
1111222t110OO11t2
11112221111011122
11122221111111222
11122321111122332


3 3 4 611 14 9 1O 16 16 8 8 9 2 1 OO
4 4 9 10 20 26 28 29 29 28 28 17 2 1 1 1 1
3 9 9 14 24 31 43 39 33 36 36 12 1 1 1 1 1
2 4 11 2O 26 38 4O 61 98 30 14 3 3 1 1 1 1
2 4 17 29 99 72 76 69 37 19 3 4 3 3 2 2 2
4 4 13 33 69 102 74 27 1O 6 9 2 3 3 2 2 2
2 6 11 21 94 64 48 31 9 3 4 3 2 2 2 2 1
2 4 19 16 22 94 69 21 4 3 2 3 2 2 2 1 1
3 4 7 4O 49 39 31 9733232 1 1 1 '
344- 4O 412284937222 1 1
                    471.20 476.20  481.2O  486.20  491.2O 496.20 5O1  20  SO6.2O  S11.2O


Figure 8.   PEM map  modeling area sources as  point sources,  stability D,
             otherwise conditions  as in  Figure 7.

                                         15

-------
PEM IMPROVEMENTS

     We found the PEM model  easy to  use  (even without a users guide).
However, the versatility and convenience of  the  PEM  program could be
improved.  Some of the recommendations in this section result from our
experience with the model; others are  proposed in view of expected
applications.  Implementation of the minor changes to the model, e.g.,
book-keeping operations, additional  outputs  and  new  user options, would
facilitate a sensitivity analysis and  source apportionments.  Other changes
might increase the realism of the model.  The somewhat increased complexity,
memory storage and computer  time required by the model would not be
important considerations to  most model users.  Incorporating changes before
the model is released would  be advantageous  since user modifications would
not be necessary thus eliminating further review by  regulatory authorities.
     A brief assessment is given below of the importance, objective, and
ease of implementing the various possible changes.   Many of the suggestions
could be included as options (defaults would use the existing program).

Deposition, Settling and Transformation

     The current model requires that deposition, settling and transformation
parameters pertain to all sources and  meteorological conditions in an
averaging period (1 to 24 hours).  The following suggestions provide a
hierarchy of increasing sophistication in handling these processes.

a.  Hourly variations of deposition  and  transformation parameters.
    Importance:  Variable, depending on  importance of deposition and
    transformation.
    Objective:  Portray time varying characteristics, e.g., day/night
    differences of deposition velocities.
    Implementation:  Programming simple.  Little guidance for user.
    Possibly tied to stability and wind  speed inputs.
                                    16

-------
b.  Source-specific deposition,  settling and transformation  parameters.

    Importance:   Moderate for calculations of TSP  where  emission
    characteristics vary widely.   Less important for  gases and  IP with
    similar characteristics.

    Objective:   Increased realism by reflecting different particle  size
    distributions from various sources.

    Implementation:   Programming simple, but specification  possibly
    difficult (especially if parameters vary on hourly basis) due to
    incomplete data.


c.  Particle size stratification of deposition and settling  parameters.

    Importance:   Generally minor, since size (weighted)  average parameters
    may be used in most cases.

    Objective:   More accurate description of particle dispersion by
    reflecting distribution of particle sizes. Also  permits simultaneous
    calculation of TSP, IP, and gaseous pollutants.

    Implementation:  Programming simple.  Handle as additional  pollutant
    (e.g., increase number of pollutants to five and  sum pollutants.)


d.  Site-specific deposition velocities.

    Importance:   Moderate, depending if deposition is important.

    Objective:   Increased realism by modeling surfaces found over large
    areas, e.g., water, urban, grass.

    Implementation:  Programming possibly difficult since present solution
    algorithm may not be suitable.   More complicated  if  deposition  changes
    with time.   User must specify surfaces and deposition rates (possibly
    keyed to surface type.


e.  Insignificant transformation rates.

    Importance:   Minor, depending on transformation rate, source-receptor
    distances and wind speed.

    Objective:   Reduced CPU time by eliminating calculations from (numerous)
    small local  sources and meteorology which produces neglible
    transformation.

    Implementation:  Moderate programming difficulty. Requires look-up
    charts and relative estimates of extent of transformation.  Invisible to
    user.
                                   17

-------
Source Specifications

a.  Hourly variations of source strength  and  other  source characteristics
    (exit velocity,  temperature).
    Importance:   Small for modeling  worst case  conditions,  possibly  large
    for validation and other studies,  especially with  significant daily
    variations of emission sources.
    Objective:  Simulate operating  schedules  of major  emitters.
    Implementation:  Programming simple.  Possibly  large amounts of input
    data.
b.  Determination of source class impacts.
    Importance:  Variable, depending on application.
    Objective:  Expands culpability option,  particularly useful  in  SIP
    revision.  Also defines culpability of  industrial  source  complex with
    multiple emission sources.
    Implementation:  Programming straightfoward,  involving  additional
    specification of source classes and book-keeping of class impacts for
    receptors.  User simply specifies source class.

Area Sources
a.  Make area source calculations independent of  receptor grid size.
    Importance:  Variable, depending on importance of  area  sources  and
    receptor grid size chosen.
    Objective:  Increased realism of area source  modeling.
    Implementation:  Programming possibly difficult.   Invisible to  user.

b.  Concentration and/or other  criteria for determination of  number of
    downwind grid cells.
    Importance:  Variable, depends on importance  and configuration  of area
    sources.
    Objective:  Increased realism of area source  modeling by  using
    meaningful standard rather  than fixed (4 downwind  receptors)
    calculations.
    Implementation:  Programming straightforward.  Invisible  to user.

                                    18

-------
Outputs

a.  Automatic scaling of concentration and deposition flux maps.
    Importance:   Minor, for convenience.
    Objective:   Eliminate useless outputs and model  runs producing  numbers
    either too small or too large for display.
    Implementation:  Programming simple.    Invisible to user.

b.  Culpability list of all (or specified number)  of sources at  specified
    receptors showing major contributors  to receptor concentrations.
    Importance:   Normally minor, since culpability of top 10 point  sources
    is provided.  However, expand to include area  source contributions.
    Objective:  Facilitates source apportionment, especially in validation.
    Implementation:  Programming and use  straightforward.

c.  Hourly concentrations written on tape.
    Importance:   Minor, but convenient.  Present option writes only final  (1
    to 24 hour)  average concentrations on tape.
    Objective:   Permits analysis of critical meteorological conditions,  etc.
    Implementation:  Trivial.

Miscellaneous
a.  Hourly variations of calibration parameters, particularly intercept.
    Importance:   Minor.
    Objective:   Portray changes of background levels.
    Implementation:  Simple.

b.  User specification of wind profile exponents and anemometer  height.
    Importance:   Generally minor.
    Objective:   Tailor model to location, if estimates are available.
    Implementation:  Simple.
                                    19

-------
c.  Incorporate varying terrain.
    Importance: Potentially large with  elevated  sources  and complicated
    terrain.
    Objective:   Model fumigation  of elevated terrain.
    Implementation:   Programming  difficult,  since  algorithms may not  be
    suitable and deposition and settling processes poorly  understood  with
    terrain differences.  Usage simple  by specifying  receptor elevation.
d.  Increase or eliminate cut-off distance for point  sources
    Importance:  Possibly important with certain emission  inventories.
    Objective:   Extends modeling  to tnesoscale (30-300 km).
    Implementation:   Trivial.
e.  Use of updated dispersion  coeffiecients
    Importance:  Important under  some conditions
    Objective:   Standardize models with other UNAMAP  models
    Implementation:   Trivial
e. Long Term Version of PEM
    A long term version of PEM could be very useful.   Such a model  might  be
    similar to the ISC-LT model,  which  calculates  annual concentrations
    using site-specific meteorological  data.  The  joint  probability of  the
    STAR categories, i.e., each wind speed,  wind direction and  stability
    category,  is used to weight ambient concentrations for that category  in
    the determination of annual or seasonal  average concentrations.   This
    would yield spatially varying long-term  concentrations that might be
    more valid than short-term concentrations predictions, which depend on
    accurate source and meteorological  data.
                                   20

-------
SUMMARY

    Of the recommended changes,  we feel  that the  most  Important  improvements
increasing model credibility are those concerning area source modeling.  For
sulfate modeling, hourly variation of transformation rates  and elimination
of cut-off distances for point and area  source  calculations are  the most
relevant.  For TSP modeling, source specific specification  of particle size
or deposition and settling velocity are  important.  We also feel  that the
development of a long term version of PEM for seasonal or annual  particulate
concentrations would complement  existing air quality models.  Most of the
remaining recommendations simply provide greater  convenience for  model users.
                                   21

-------
                                 SECTION 4
                       SOURCE APPORTIONMENT USING PEM

      In this section we describe aerosol  source apportionment in
Philadelphia using the PEM dispersion model.  The source inventory,
meteorology and PEM predictions  are  presented.  The results could be
indicative of the composition of aerosols  obtained by dichotomous samplers
in the Philadelphia study.

SOURCE INVENTORY

      The preliminary source inventory was compiled to obtain the
approximate composition and location of  the major emission sources in the
Philadelphia metropolitan area.  Sources of the data are listed at the end
of this section.  The data do not include  micro-inventories around receptor
sites.  Present data do not allow a  complete classification of sources.
However, this should not seriously handicap the apportionment since the
objective is to obtain only representative inventory and source
apportionment estimates.  More detailed  source and monitor data for the
Philadelphia study should permit more accurate modeling.
      The inventory consists of  a total  of 50 area sources and 104 point
sources.  The location of the sources is shown in Figure 9.  Particle
emissions are broken down into  12 source classes as shown in Table 2.  This
Table also shows the six source  classes  of S0? and (primary) SO, sources.
      The composition of the emissions sources located within the city is
different than shown in Table 2, which  includes a number of large sources
located up to 60 km from the city limits.  Table 3 provides a summary of the
emission sources located within  470  to 500 UTM east, and 4410 to 4440 UTM
north, a region which encompasses Philadelphia city limits with the
exception of the northeast corner (see Figure  13).  In this smaller region,
area sources emit most of the TSP and a  few industrial sources emit most of
the S02 and SO^.
      Area source emissions included both  fine and coarse particles, as well
as S02 and SO^.  Area source emissions were estimated by aggregating
about 600 sub-census tract emission  estimates  into fifty 2.5 km square

                                   22

-------
     4450*
                  0    10   20   30km
                    425.
459.
475,
see.
Figure 9.  Map  showing location of 104 point  sources and 50 area sources used
          in the Philadelphia inventory.   Philadelphia city limits  are shown
          in the center of the figure.   Scales are UTM coordinates.
                                  23

-------
           Category
Number of Sources
           —TSP Emissions--

             g/s      percent
1
2
3
4
5
6
7
8
9
10
n
12

1
2
3
4
5
6
Area sources—coarse
Area source--fine
Refineries
Incinerators
Oil-fired utility
Coal-fired boilers
Metal and steel
NJ County Emissions
Chemical industries
Grain elevator
Misc. ground level
Misc. oil-fired
Totals
Category Number
Area sources
Oil--indus trial
Coal
Oil—miscellaneous
Chemical industries
Refineries
50
50
7
5
8
14
9
6
10
1
3
40
203
of Sources
56
8
14
40
9
7
246.18
17.23
127.30
39.54
299.39
287.10
280.33
266.00
77.89
25.92
24.65
50.76
1742.26
S02 Emissions
g/s percent
317.48 5.0
1531.25 23.8
2335.00 36.4
525.18 8.2
245JO 3.8
1457.30 22.7
14.1
9.9
7.3
2.3
17.2
16.5
16.1
15.3
4.5
1.5
1.4
2.9
100
SO^ Emissions
g/s percent
22.14 8.0
107.19 38.6
35.04 12,6
27.22 9.8
13.89 5.0
72.57 26.1
    Totals
 134
6411.31   100
278.05  100
Table 2. Summary of TSP (top)  and  S02  and  504  (bottom) emission
         inventories in the Philadelphia area.
                                   24

-------
           Category
Number of Sources
           —TSP Emissions--

             g/s      percent
1
2
3
4
5
6
7
8
9
10
11
12



1
2
3
4
5
6
Area sources — coarse
Area source — fine
Refineries
Incinerators
Oil-fired utility
Coal -fired boilers
Metal and steel
NJ County Emissions
Chemical industries
Grain elevator
Misc. ground level
Misc. oil-fired
Totals

Category Number
Area sources
Oil — industrial
Coal
Oil—miscellaneous
Chemical industries
Refineries
50
50
4
5
6
2
4
0
6
1
3
22
153

of Sources
50
6
2
22
6
4
246.18
17.23
95.70
39.54
50.19
66.30
25.11
0.00
19.29
25.92
19.15
36.59
651.20
S02 Emissions
g/s percent
249.15 12.0
1222.96 58.7
502.10 24.1
436.98 21.0
223.70 10.7
548.00 26.3
37.8
2.6
14.7
6.1
9.2
10.2
3.9
0.0
3.0
4.0
2.9
5.6
100
SO^ Emissions
g/s percent
17.23 10.5
85.61 52.3
7.53 4.6
13.10 8.0
12.80 6.5
27.47 16.7
    Totals
 90
2083.19   100
163.68  100
Table 3. Summary of TSP (top)  and S02 and  SO^  (bottom) emission
         inventories in Philadelphia  located within 470 to 500 UTM east
         and 4410 to 4440 UTM  north.
                                   25

-------
areas.  The area source inventory combined emissions  from mobile,  point  and
area source emissions.   However,  coarse TSP emissions are primarily from
road dust; fine TSP emissions are from mobile sources and residential  and
commercial activities;  and SC^ emissions are primarily from residential
oil-fired burners.  Figures 10 to 12  show area source estimates  for the
three pollutants.
    Primary S(L emissions were determined by multiplying  SO^ emissions
by the factors given in Table 4 (National Research  Council, 1983).  Thus,
all S02 sources had primary sulfate emissions.  Area  sources had a
emission weighted sulfate emission factor of 7 percent.  The emission
weighted sulfate emission factor for the entire inventory is 4.3 percent.
    The inventory was compiled from the inventories listed below:

1. City of Philadelphia Air Management Services, Philadelphia,  PA.  B. Glazer,
T. Weir,  (215) NU6-7393
   Comprehensive  inventory of city emission sources listing over 1000  point
   sources including hazardous pollutant emissions and area source emissions of
   fine and coarse particulates, S02, $04, NOx. Limited computer capability.
2.  PEDCO Micro  Inventory, PEDCO Environmental, Cincinnati, OH.   Barb  Siegal,
(513) 782-4700
   Major TSP point sources (over 100 T/yr) in the Philadelphia area.  Completed
   for 1978 monitoring study.  1982-3 update forthcoming.
3. N.J. State Dept. of Environmental Protection, Bureau of Field Operations.
Trenton office:  Dr. Ray Dyba,  (609) 985-3009;  Cherry Hill, Camden County
office:  Terry Juchnowski (609) 984-0616
   Estimates of  criteria pollutant emissions and identification of major
   sources for Burlington, Camden, Glousester and Mercer Counties.  Not
   computerized  (on microfiche).
4. Pennsylvania  Department of Environmental Resources, Inventory section,
Harrrisburg, PA.   John Walker (717) 787-4324
   PA point sources (excluding those located  in the City of Philadelphia)
   including emissions of criteria and other pollutants, operating schedules
   and plant parameters. Sophisticated computerized inventory.
                                   25

-------
         	Fuel	
         Coal
         Residual oil—industrial
                     --residential
         Distillate oil
         Mobile sources
         Miscellaneous
--Percentage Primary S04-
            1.5
            7.0
           13.4
            3.0
            3.0
            5.0
Table 4.  Primary sulfate emission factors (National  Research  Council,  1983).
                                                  Soo
Figure 10. Map showing emission rates (g/s)  of  coarse  particles from
           Philadelphia area sources.  Area  sources  are  2.5  km square.
                                   27

-------
                                                   
-------
5. "Thermal Electric Power Plant Contruction and Annual  Production
Expenses," Energy Information Administration, Washington,  D.C.,  1980.
      Several power plant emissions were estimated using these fuel  records
      and maximum NSPS emissions.

METEOROLOGY

      Surface meteorological data for the PEM simulation were obtained for
the Philadelphia's Northeast airport and the Norristown  meteorological
station.  The Norristown data were received in a preprocessed format,  which
included stability classifications.  No mixing heights were obtained.   PEM
simulations used constant, ficticious mixing heights (as specified).

      These data were obtained from:

         Pennsylvania Department of Environmental  Resources,  Meteorology
         Section, Harrisburg, PA.  Denis Lohman (717)  787-4319

DISPERSION MODELING

     The PEM dispersion model was used to estimate the source apportionment
in the Philadelphia area under several meteorological  scenarios.   PEM  can
simulate dispersion of either two independent pollutants,  or  one  pair  of
coupled pollutants.  Thus, source apportionment of the 12  TSP source classes
was accomplished by 6 model runs, each run modeling two  (independent)  source
classes.  Source apportionment of SOo and SO*  for  the  6  source classes
required 6 model runs, since S02 to SO, transformation was permitted and
each source emitted both S02 and primary sulfate.   A constant 1 percent
per hour transformation rate was used.  Deposition and gravitational
velocities were set to zero.  A tape output of concentrations at  each
receptor was generated by each model  run.   After the 6 runs for each
apportionment (and meterological scenario) were completed,  a  Conversational
Monitoring System (CMS) "macro" program consolidated the concentrations into
one file, which was then transferred to the Time-shared  Reactive  On-Line
laboratory (TROLL), a large, interactive statistical and data analysis
system (Information Processing Center, 1980).   TROLL was used to  generate
the source apportionment tables.  Both CMS and TROLL reside on the MIT
IBM/370 Computer.
                                   29

-------
Hypothetical Meteorology
     First, hypothetical meteorology was  used to portray the average effect
of the sources at each receptor.   Eight hour concentrations were calculated
using consecutive hourly 45 degree wind shifts  (starting at 0 degrees) using
a fine (25 x 25 grid;  60.0 km square)  receptor  network for the following
stability categories and wind speeds:
         stability category   wind speed (m/s)
                A                  2
                D                  4
                E                  3
Due to hourly shifts in the wind  direction  in these scenarios, no single
source contributed longer than one hour to  receptor concentrations (with the
exception of receptors located within  area  sources).
      Based on the results obtained, a coarse (4x4 grid; 30 km square, SW
corner UTM 4410 north, 470 east)  was selected to include both areas of
maximum concentration  and the Philadelphia  monitoring sites.  The receptor
grid and receptor numbers are shown in Figure 13.  This receptor network
encompasses Philadelphia and adjoining suburbs.  The coarse grid was chosen
primarily to reduce computations; it may  not reflect the diversity of
concentrations at the  various monitors in the Philadelphia area.
      Predicted source class estimates are  given in Table 5 for TSP
(excluding sulfate) and Table 6 for S02 and SO,.  The average percent
source class contributions of TSP and  S04 at the 16 receptors are listed
in Tables 7 and 8.  The following results are highlighted:

(1) Receptors show a diverse source apportionment. Receptors 6, 7, 10 and
11, located within Philadelphia,  receive  high loadings from area sources.

(2) Point sources contribute less than 1  ug/m3  under stability A compared
                3
to up to 60 ug/m  from area sources.   Point sources have larger
contributions under other stabilities. However, area sources dominate both
TSP and S02 concentrations.

(3) Low sulfate concentrations, generally about 5 percent of TSP, reflecting
the relatively high wind speeds which  advect the pollutants outside the area
resulting in little SO^ to SO,, transformation,  as well as the small
proportion of primary sulfate.
                                   30

-------
            470
U T M  COORDINATES EAST

   480               490
        4450
       4440
    1
    z
    0)
    a
    8
    u
       4430
       4420
       4410  *


Figure 13. Map showing 5 km receptor grid of  Philadelphia area.  City  limits shown
           with dashed line.   Small  numbered  regions are sub-census  tract areas
           Large numbers indicate receptor number for coarse  (10 km) grid used in
             ^11.^'  .Squares indicate  location of five monitors of  the
             iladelphia study;  the sixth  is located 7 km south of map
                                   31

-------An error occurred while trying to OCR this image.

-------An error occurred while trying to OCR this image.

-------
Stability  Wind  	--Percent  Contribution	•	

Category  Speed  area-coarse  area-fine  oil-ind   NO-misc  grain   oil-misc  other
A
D
E
2 ra/s
4
3
84.1
60.9
70.4
12.6
9.1
10.5
0.8
2.6
1.5
0.5
18.2
13.0
0.1
4.5
2.2
0.3
1.1
0.5
1.5
3.6
1.9
Table 7. Average (16 receptor) percent TSP source contributions  from PEM
         modeling using hypothetical meteorology.  Source classes contributing
         less than 1 percent of total TSP are labeled "other," and include
         refineries, coal, steel, chemical and soil.
Stability  Wind      	Percent Contribution	
Category  Speed      area      oil-ind    coal   oil-misc  chemical  refineries
A
D
E
2 m/s
4
3
60.4
39.0
56.1
17.7
21.2
14.2
5.4
3.6
4.9
6.0
22.0
13.6
2.0
3.8
2.2
8.3
10.0
9.0
Table 8. Average (16 receptor) percent 504 source class contributions from
         PEM modeling using hypothetical meteorology.
                                    34

-------
Actual Meteorology

     Actual meteorological data were  used to  simulate  three  consecutive  12
hour periods, starting with 6 AM,  August  22,  1982  (Table  9).   These  periods
were selected since they showed diversity:  transition  from unstable  daytime
conditions to stability E (period  1); transition to  stability  A with some
moderate and fairly persistent winds  in categories 3 and  C (period 2);
neutral stability and low winds (period 3).   Wind  speed,  direction,
stability, and temperature data were  complete for  these periods.
     Source class apportionments for  TSP, S02 and  S04  are shown in
Tables 10 and 11.  It is seen that area sources dominate  as  before,  but
other source classes may be significant contributors to both TSP  and SO^
concentrations.  According to the source class inventory  (Tables  2 and 3),
area sources emit nearly one-half of  the TSP  within  the city area.   This is
reflected in the source apportionment of Table 10.
     Coal burning is the largest S02  emission class, and  industrial  oil
burning produces most of the primary  sulfate.  In  the  vicinity of the city,
however, most S02 and SO^ is emitted  by 6 industrial oil  users and 4
refineries located near the city.   The highest S02 and SO^
concentrations from these sources (refineries) were  11.4  and 1.4  ug/m ,
respectively, for receptor 9 and period 1.   These  point sources have
elevated stacks ranging from 40 to 84 m which generally result in low ground
level concentrations.  Consequently,  area source emissions also dominated
the source apportionment of S02 and SO*.
     The highest predicted sulfate concentration  in  the three  periods was
6.3 ug/m  (receptor 11, period 1). A search  for conditions  that  produced
the maximum sulfate concentrations was performed  by  calculating hourly
concentrations using the coarse grid  with incremental  30  degrees  shifts  of
wind direction under stability D for  the receptors located in  the city
(Table 12).  The maximum concentrations from point sources are considerably
smaller than the average area source  contributions.  Only in one
circumstance was the point source contribution roughly comparable to the
area source contribution (receptor 10, wind direction  310 degrees, with  an
                                      3
area source contribution of 1.016 ug/m  and an industrial oil  source
                          i
contribution of 0.588 ug/m ).  These  results imply that most point sources
contribute little to sulfate levels in the city under  most conditions.
                                   35

-------
SCENARIO
NUMBER

1


2


3


4


5


6


7


8


9


to


It


12


SCENARIO
NUMBER
1


2


3


4


9


6


7


8


9


10


11


12


STABILITY WIND SPEED
CLASS CLASS

A


a


c


00


00


ON


E


£


ON


E


E


E


STABILITY WIND SPEED
CLASS CLASS
ON


00


DO


C


a


C


B


B


A


A


A


A


HIND
SPEED
 on


88. OO


49.OO


313. OO


77 JM


109. OO
AMBIENT
TEMPERATURE
(DEO C)


28.00


26. OO


23. OO


25. OO


24. OO


23. OO


22.OO


22.OO


21. CO


21.OO


19. OO


19. OO
AMBIENT
TEMPERATURE
(DEC C)


19.00


21.0O


24. OO


25. OO


26. OO


27. OO


78. OO


27.00


28. OO


28. OO


28. OO


28. OO
Table 9. Meteorological data used for PEM simulation.   All mixing  heights  set
         to 2000 m and inversion penetration  factor  set to unity.   Top:  period
         1; bottom; period 2.
                                   36

-------
MARIO
STABILITY WIND
SPEED WIND WIND VINO
AMBIENT
M8ER CLASS CLASS SPEED SECTOR DIRECTION TEHPERATUt

1


2


3


4


S


a


7


8


9


10


11


12



A


8


C


00


DO


00


00


00


00


00


ON


E


(M/S)

1.30O


1. 10O


1.2CO


1.200


1.40O


1.00O


1.3OO


1.000


LOCO


1.00O


t.ooo


1.0OO

(DEO)


334.00


323. CO


17. OO


108.00


167.00


173. CO


2O9.00


219. OO


3O9.0O


284.00


147.00


93.0O
(oca c)


28. CO


26. OO


24. OO


23. OO


23. OO


22.0O


22.OO


21. CO


20.0O


19. OO


18.00


17.00
Table 9. (continued)  Meteorological data; period 3,
                                   37

-------
         AREA-C  AREA-F  REFIN.
                              - -  -  SOURCE CLASS CONTRIBUTION  IN UG/M3  	


                              INCIN. OIL-INO   COAL  STEEL   NJ-MIS CHEM.   GRAIN
                                                - - -          TOTAL


                                                SOIL  OIL-MIS CONCENT
RECEPTOI
NUM8EI
1. ...
2. . . .
3 ...
4. . .
g
6
7
8
9
1O. . .
11. ..
13...
13. ..
14. . .
15...
1«. . .
*
9
12. MI
8.577
o.ooo
o.ooo
21 581
214 9IO
153 2SO
0 OOO
15 197
. 185. 170
. 298.450
o-.ooo
. 1O. 639
. 47.240
. 111. 480
O.OOO


1.769
0.826
O.OOO
O.OOO
3. 344
33. 1O8
19.257
O.OOO
2. 198
27.876
47.321
O.OOO
1.829
7. 188
16.378
0.000


0.018
O.OOO
O.OOO
O.OOO
0.084
O.OO3
O.OOO
O.OOO
0. 119
0.066
O.OO4
0.001
0.313
0. 169
0.007
O.O03


O.OOO
0.034
O.OOO
O.OOO
0 016
O.OO2
O.OO8
O.OOO
O.OO3
O.OO3
0.017
0.003
O.O34
O.OO3
0.218
O.OO7


1.070
0.001
o.ooo
0.247
0 276
O.044
0.013
1.451
0.060
O.OOS
0.033
0.079
0.073
0.334
0.298
O. 113


0. 144
0.04S
O.447
O. 181
0.340
0.048
1.O03
1.537
0.263
0.063
0.540
2.253
0.083
0. 176
0.294
0.993


0.026
O.OOS
0 502
0.016
0.376
0.08O
0.492
0.070
O.OOO
0.022
0.036
3.648
0.010
O.OU
0. 103
0.054


O.OOO
O.OOO
o.ooo
o.ooo
o.ooo
O.026
O.OOO
o.ooo
1 .977
2.735
0.224
0.000
O.OO1
6.076
0.037
3.493


0. 108
O.OOO
O.OOO
O.OOS
0.065
0.005
0.027
0.050
0.097
0.013
0.063
0.322
O.O12
0. 107
0.262
1.013


O.OOO
O.OOO
O.OOO
O.OOO
0.003
O.OOO
o.ooo
o.ooo
4.833
0.22O
O.OOO
O.OOO
O.OO7
11.832
0.003
O.OO1


O.OOO
0.000
o.ooo
o.ooo
3. 4 1O
o.ooo
o.ooo
o.ooo
2.863
O.346
O.OOO
0.083
3. 139
O.629
0.066
O.OOO


O.O74
0.013
0.014
O.OOO
0.251
0.031
0.043
O. 154
O.21S
0.246
0. 1 16
0.023
0.07S
O.429
0.262
0.033


15.3SO
7.501
O.963
0.453
29.745
248.257
174.093
3.263
27.825
216.766
348. 8O4
6.413
16.019
74. 197
129.404
5.711
MEAN	  67.29O  10.O58   0.049   0.022  O.2S6  0.528   0.341   0.911   0.135   1.056   0.859   0.124  81.422
           AREA-C  AREA-F  REFIN.
-  -  - SOURCE CLASS CONTRIBUTION IN UG/M3 - -  •

INCIN. OIL-INO   COAL   STEEL   NJ-MIS  CHEM.
                                                                           GRAIN
-  -  -         TOTAL


SOIL  OIL-MIS CONCENT
RECEPTOR
NUMBER
1 . . 4 Rjn
2. ...
3
4. ...
5 ...
6 . .
7
8 ...
9 .
10. . .
11. ..
12. . .
13. . .
14. . .
15. . .
16. . .
4 963
2 276
0.816
6 75O
69 404
5 1 098
0 088
5 594
. SO. 321
. 88.269
O.OOO
1 . 39 1
3.332
19.372
O.OOO
0.711
0. 716
0.313*
0. 124
O. 989
1O. 664
6. 509
0.012
0. 870
7.565
14.211
O.OOO
O. 191
O.502
2.749
O.OOO
0. 505
0.1(9
0. OO3
O.OO2
0. 547
0. 04 1
0 OO5
0. OO3
0 t49
0.019
0.012
O.OOO
O.OO1
O.OOO
O.014
O.OOO
0. 082
0.076
O 049
O.O07
0. OO6
0 155
0 06O
O OOO
0. OO2
O.OO9
O.OOO
O.OOO
O.O07
O.O25
O.OOO
O.OOO
0 205
0. 068
0 316
0. O0 1
0 . 03O
0 642
0 032
0. OOO
0 067
0. 124
O.OO1
O.OOO
0.073
0.01O
0.482
O.OOO
O O95
O.O37
0 O38
0 . OO6
O O9O
0 030
0 O57
0 016
0 O73
O.O22
0.022
O.O69
O.OOS
O.O44
0.03O
0.058
O/"> * K
• \> ' 3
0 OO6
Onns
UWD
0 O05
0 O24
O 089
O O30
0 OOO
O OO2
0.016
0.001
O 004
O.O09
o.ooe
O.OOS
o.ooo
Onfi3
• vo<
0. 137
O 177
0 1 1 3
O 442
1 047
O 1 6O
0 226
2T7H
. J ' a
0.483
1.852
O. 169
0. 149
0.427
0.096
3.311
Orvc t
• UO I
0 128

0 . 040
0 0 1 3
0 O25
01 SO
. 1 09
Ortcs7
. L/3 /
OOOfl
U\JO
OO 1 1
. \J i I
0.076
0.028
0.013
O.OO7
0.007
O. 199
0.026
Oe*7Q
.3/9
Omo
. U Jtf
O. OOO
OfW>
. \J\J\J
Onn f
. Uv T
OO 1 1
. U 1 1
0 • OOO
O . OOO
OA17
• U J /
19.737
O.OOO
O.OOO
O.031
O.OOO
O.OO1
o.ooo

0 . (02
OO9T
• U* J
0.013
Of\fi*
• UU*
1 1 tt7
T . 1 0 /
01*7 "3
. J f >J
O. f!26
0 . isOO
0 . OOO
0.021
O.OO6
0.001
o.oot
O.O69
13.685
o.ooo

0. 193 7 . 25 1
01 
-------An error occurred while trying to OCR this image.

-------
Table 12 also shows that area source contributions remain constant under
varying wind directions.  According  to  this modeling, only the area source
in which the receptor resides is an  important  contributor to sulfate.  Note
however, that the relatively large receptor  grid  size of 10 km encompasses
16 of the 2.5 area sources.   This results in  less variation than might be
found with smaller grid sizes.
     The PEM results support the observation  in Section 3 that secondary
sulfate may compose a large  fraction of the  sulfate mass, especially from
distant sources.  Sulfate may attain about 10  percent of the S02 mass,
although only 4.3 percent primary sulfate is emitted by all sources.
However, it should be recalled that  the S04  emission factor for area
sources was 7 percent. (A molecular  weight ratio  of  1.5 (S0^/S02) was
used for mass conversion.)

DISCUSSION

     PEM modeling of the Philadelphia area is  constrained by the  limited
accuracy of the source inventory.  In particular, estimates of area source
emissions, which dominated the apportionment,  are very  crude.  More
accurate modeling may require better inventories, especially around
receptor sites and more detailed meteorological  data.   The  Norristown
station does not necessarily provide representative  surface observations
for the Philadelphia area.  Also, the omission of mixing  height data will
bias results.  Mixing height was simply set to 2000  m.  Lower mixing
heights were found to alter point source contributions  by  up to 40
percent.  Very low mixing heights (e.g., 200 m)  eliminated  point  source
contributions since elevated sources had plumes that penetrated the ceiling
layer.
     Using 9 monitoring sites in Philadelphia from May  to  September  1979,
Suggs and Barton  (1983) obtained 24 hour concentrations of  fine particles
ranging from 2 to 54 with a mean of 29 ug/m .   TSP  concentrations at  the
                                     3                      3
same sites ranged from  15 to 172 ug/m  with a mean  of 56  ug/m  .
Predicted sulfate concentrations at the four receptors  located within
Philadelphia (numbers 6,  7,  10, 11)  ranged from 0.5  to  6.3  ug/m   for  the
three 12 hour periods.  The corresponding TSP predictions  ranged  from  58  to
346 ug/m3.
                                    40

-------An error occurred while trying to OCR this image.

-------
Thus, PEM predictions have the correct order of magnitude,  although  TSP
predictions seem high, and sulfate seem low.  However,  observed particle
concentrations include non-sulfate species as well  as background regional
levels.  Background concentrations were not included in the PEM modeling.
                                    42

-------
                                 SECTION 4
                              RECEPTOR MODELS
     The principal objectives of this section include  the  following:

(1) Brief review of receptor models (RMs);

(2) Application and evaluation of several  diagnostic procedures  designed to
minimize the spurious results obtained by  chemical  mass  balance  RMs, and
suggestions for protocols which  incorporate these diagnostic  tools;

(3) Suggestions for the development and application of hybrid models,  i.e.,
models which combine aspects of  dispersion and receptor  models;

(4) Description of issues concerning inter-comparisons between results of
receptor and dispersion models;

(5)  Discussion of issues concerning RM protocols.

OVERVIEW OF RECEPTOR MODELS

     Receptor models are procedures which  use observed aerosol
characteristics to identify and  quantify the sources of  ambient  air
pollutants.  The aerosol characteristics most frequently measured  include
the chemical and elemental mass, optical properties and  particle size
distribution.  Less frequently measured characteristics  include  isotope
ratios, organic and inorganic compounds, and crystalline structure.  Unlike
dispersion models, RMs do not require meteorological or  source data.   Source
profiles may be used, but are not necessary.
     In general, RMs make use of the following assumptions:

Non-reactive (conservative) aerosols.  The total  aerosol mass at a receptor
is a linear sum of the aerosol contributions from individual  sources.  In
general, there can not be any transformation of the aerosol from time of
emissions at the source through  atmospheric transport  to collection and
ultimate analysis.
                                   43

-------
Stable aerosol characteristics.   Characteristics of the aerosols are a
linear sum of the aerosol  characteristics  from  individual sources, e.g.,
elemental ratios are assumed to  be  constant  between source and receptor.

Identifying aerosol characteristics for  source  classes.  Source
apportionment is possible  only for  those source classes that have
identifying characteristics.  No unique  characteristic is needed: just a
unique combination of characteristics.

Description of source profiles.   Chemical  mass  balance models require
quantification of source characteristics,  e.g., elemental ratios.
Multivariate models require less precise descriptions, but patterns of
characteristics must be recognizable.  Sources  with unknown or highly
variable characteristics may not be distinguishable.

Miscellaneous statistical  assumptions.   For  regression models, the number of
characteristics must exceed the  number of  source classes, and errors
(residuals) should be normally distributed and  uncorrelated.  For
multivariate models, the number  of  filters must be sufficient for the
degrees of freedom required for  the number of filter  characteristics and
sources used.

     Deviations from these assumptions will  degrade the validity of the
receptor model.  The magnitude of the deviations that typically occur  in RM
applications is not known at the present time.  Also  unknown  is the
susceptibility of RMs to such deviations.  There are  obvious situations
where the assumptions are violated  and RM  applications may not be useful.
For example, RM assumptions do not  hold  for  secondary pollutants, e.g., smog
or sulfate, which undergo extensive transformation and scavenging.  Also,
RMs can not apportion sources within a source category, e.g., determining
which of two power plants is culpable,  since these sources have similar
aerosol characteristics.
     There are many approaches to receptor modeling based on the above
assumptions.  Cooper (1980) has  classified receptor models by the
interpretive approach used to associate  aerosol characteristics with
emission sources.  These approaches include  1)  regression analysis of

                                   44

-------
aerosol characteristics, which has been labeled the chemical mass  balance
(CMB) approach;  2) multivariate analyses of the variability of  aerosol
characteristics;  3) composite receptor models which combine CMB and
multivariate methods; and 4)  a miscellaneous category including  enrichment
or depletion processes, trajectory analyses and, in general, hybrid
dispersion/receptor models.   Table 13 highlights some of  the major
differences between model approaches.  ERT (1981)  and Thurston  (1983)
provide reviews of current models.  A compilation  of recent work in receptor
modeling is found in APCA (1982).   Some aspects of the different models are
discussed below.

Chemical Mass Balance Receptor Models

     Most chemical mass balance (CMB) models minimize the sum of squared
residuals (SSR) between observed and predicted aerosol  characteristics:

    Minimize SSR = (A S_ - CJT (A ^ - £)

where:  A is the vector of estimated source apportionment  (a. is  the
       —«.                                                   I
       contribution of source class i to total pollutant  mass).

       S is the source profile matrix (s. . is a measure  of aerosol
       ~~                                i»J
       characteristic j emitted by source class i).

       C_ is the vector of observed ambient characteristic (c^ is a measure
       of characteristic i).

CMB models require measured characteristics for one or  more time periods and
source profiles.  Table 14 shows the source profiles for  sources used  in the
Quail Roost II study of RMs.   Ambient sampling and analysis procedures are
quite sophisticated.  However, source sampling involves greater  practical
difficulties and less accuracy.  Moreover,  source  profiles often are
unavailable or highly variable within source classes (e.g., from one oil
burner to another), as well  as dependent on operating conditions (e.g., fuel
burned).  Consequently, the matrix of source profiles,  S^  may be very
approximate.  CMB models using inappropriate source profiles may produce
                                   45

-------
                                Dispersion
                                Models
         CMB
Identification of culpability
     Individual sources
     Source classes
     Combined source classes

Source apportionment
     Quantitative
     Qualitative

Uncertainty estimates
     Quantitative
     Qualitative

Data Requirements
     Source strength,  location
     Source profiles
     Meteorology
-Receptor Models	

 Multivariate  Composite
                                 X


                                 X
X

X
Table 13.  Differences  between  major classes of receptor and dispersion models,
           STEEL* STEELS OIL*
                          	 SOURCE CLASS 	


                          INCIN OILS  COAL* 4GGREG  GL*SS BASALT SOIL
                                                            *UTO  WOOD
c. . . .
NA. . . .
AL . . .
SI . .
s . . .
CL, . . .
K
CA
TI . . .
V
CR. . . .
MN . . .
FE. . . .
NI . .
CD . . . .
ZN . . ,
AS. . .
BH. . .
P8. . . .
. . 149196
9383
13196
8995
34948
33141
1O20O
634 1O
1448
189
3848
15216
. 123383
3O72
231O
2831 1
82
91
7876
214629
14438
1OSC9
1 1653
46829
36707
1281O
52372
1 143
149
3OS9
22968
154932
2797
397O
42555
114
146
13OO8
4O617
32536
6947
195793
135376
824
746
836
757
9289
227
1 17
13315
5297
655
356
12
27
811
39991
133280
4855
7424
132 183
2974O7
2484 1O
1 1O73
1013
25
89O
36O
5721
93
4248
2 1OBO
129
831
16663
198625
2O676
6O88
76426
123929
13O71
1329
10614
1 1OO
691
431
118
133OO
50
595
1 136
339
399
4133
30450
8648
39387
142O63
15868
1O46O
1O4O7
36O48
284S
14O
348
28O
18909
187
1313
3756
287
ISO
3O39
985
5335
597 13
235756
6589
8O8
13520
4 1693
2793
65
60
906
4993O
33
580
490
41
13
119
2036
158353
3632
1989
178681
3079
32591
4428
336
161 1
3537
253
9104
1**71
432
S54
651
17
507
150
20 158
85596
166164
226
194
8580
45793
7910
223
118
1633
81767
85
96
83
12
1
4
68593
4271


92

12629
28O
3 1O
203O
996OS
1O3
161
391
20
1O
31O
518497 1
74649
832O
93O9
8782
8717
27 14
1 1288
7
1O6
89
598
7O43
6 1
3
3248
12
387O8
133622
575515
3327
82 16
798 1
866
1OO3O
9426
7769
19
2
3
20
1O32
s
3
617
12
9«
10O
7O45
4459
1 640 1 8
155397
1 33869
18979
8O47
35386
9OI1
797
330
348
71849
635
38O
688
162
11
1197
 Table 14.  Source signatures from Quail Roost  II  in ug/g.
                                     46

-------
        dB[uaAO  aq;  ;eqi ipns (XLU;BIU oi^sue^oBueqo aounos p3AU3sqo aq; pue
saunpsooud SLSA~IBUB  UO:PBJ. 6ui.sn p3ALuap) SLXB  UO^OBJ. aqq. saiBos PUB SS^BIOU
 SISA"IBUB UO^OBJ. UOI.;BIUUOJ.SUB^ }36uB:j.  'souB^suL  uoj  'XLUiBiu 3[Lj.oud aounos
    3SLoaud B }noqi.LM luauiuoiauoddB aounos j.o sa^BuiL^sa aAL;Bli.^LiBnb apuojd
  o^ A"^iiiqB ai)^ aABi) siapoui uo^daoaj a^BLJBAL^inui/uoj.ssau6au
                                                            uo^oaoay
                 *SJ.SA"|_BUB ^uauodmoD [Bdpuiud  PUB SLSA"IBUB wia^snp  *SLSA"[BUB
                'uoLSsa-iBaj a^BLUBAi^iniu apntoui.  saunpaooud a^eiaeAi^inw
                                                               •saounos  g o^  9
sdBijjad 6uiqsj.n6iuisip 'siapoiu gwo uBqq. uoi^niosau uawoi aABq o; pua^
                    uuoi.^uoddB [BH^OB ai|^ ;ou inq 'qoBa A"q
      j.o ^unouiB  aij^  PUB sadA"^ ao-inos aq; a^Boipuj. siapoiu asaq^.
                   ^snqou apuoud o^ pa^oaQoo aq ^snui saiduiBS ^uaiqiuB (£861
              09 °^  (£861 'd^d) Ot ^noqB aouis  ^as^j.o ^BqMauios si a6B^UBApB
siqi  *UMOu>|un SL  saounos A"|.a>|j.i j.o aiij.0w»d uo A"^L^uapL aq^ uaqM pasn aq A"BUI
        'snqi   'XL^BIU ai.Lj.oud aounos q.LOL[dxa UB BuLuinbau }ou j.o
                     Bq s[apoiu a^etJBAL^inw  'xaidmoo A"ua  "
  A"aqi  •soiisj.ja^OBJBqD [oso-iaB 6uouiB sdi.qsuoi.^B[au ^ua^sisuoo uoj. >|00[ PUB
 suoi^BAuasqo ^uaiquiB  j.o satuas auu} B asn siapoui uo^daoau
                                                  s[apow uo^daoay

                                                                         a6pij
  PUB ^snqou  'sanbiuqaa^ aouBLUBA aAL^Daj.j.a  'sa-iBnbs  ^SB3[ pa^t)6i.aM  '
        }SBai A"jBuipuo   isaunpsooud BULMOQOJ.  sq^  j.o  A"UB asn siapooi
                                                  •sua^ndiuoD A*UBIU uo aj.qBij.BAB
    A"|.uoiuiuoo  auB  sa6B>|DBd uoi.ss3u63-i *A"I^SBI   *(S3SS8io 3ounos L °^  9 i
    uoi4nios3u q6i.q A"IULBJ. 3ABq os[B A~3qj.  *p3p3au SL a|.diuBS US^I.LJ.  auo
    PUB 'paALuap  aue  saiBuiL^sa aounos aAnB^L^UBnb 'a[duiLS A"uaA A"n.Bindaouoo
            aou.LS  swy a;BLUBAL^[nui -IBAO sa6B^uBApB auios aABq s|.apoiu
  paunsBaui aq; JL  suoaua ao saBusqo UBIUS o; A";LAL;LSuas PCIB sa[i.j.oud aounos
      auios uaawq.aq ^LUBILUILS aq; o; anp  (sq.[nsau  sno.Lunds o; psai A"BIU
 *siua[qoud IBOL;SL;B;S  [BuaAas o; auoud auB OS^B siapoui asaqi  *suouua

-------
estimated source profiles is  maximized.   Then, conventional CMB procedures
are used to calculate mass contributions.  This  procedure may provide useful
results when source profiles  are  not  known a  priori.
     In addition to target transformation factor analysis, composite
receptor models include multiple  linear  regression/factor analysis and
effective variance CMB/factor analysis  procedures.

DIAGNOSTIC PROCEDURES FOR CMB MODELS

     This section discusses several  problems  which  are  frequently
encountered in CMB receptor models.   These problems are collinearity,
influential characteristics and outliers.  In part, the reliability and
quantitative performance of CMB receptor modeling depend on resolving these
problems.  Some relatively new diagnostic procedures, which have not been
applied previously (at least  not  in  our  search of the literature) are
discussed and demonstrated.  These  procedures are developed in depth by
Belsley, Kuh and Welsch (1980)  and  are  implemented  in the TROLL statistical
package  (Information Processing Service, 1980).

Collinearity

     Collinearity exists when source profiles, e.g., elemental ratios, are
similar or linearly dependent.   In  CMB  models, collinearity may result from
1) similar source signatures; 2)  too many source classes, and 3) too few
observations, e.g., not enough elemental measures.  In  general, collinearity
may lead to unstable, negative and/or too large  estimates of source
contributions from one or more source classes.  Ordinary least squares (OLS)
in CMB models can not be rationally applied  before  such dependencies have
been established.
     There are many diagnostic tools which may detect the presence of
collinearity.  These include:
     -eigensystem analysis and condition number
     -singular value decomposition  and  variance  decomposition proportions
     -examination of the correlation matrix  of the  source signatures
     -variance inflation factor
These methods, which use only the source profile S^  (not the observations of

                                   48

-------
aerosol characteristics), are illustrated using the (contaminated)  source
profile matrix given to the Quail Roost II participants.   This matrix (Table
14) contains 13 source profiles, each consisting of 19  elements.   (The
matrix was contaminated in Quail Roost II by adding random measurement
errors to each element of the true source matrix.)
     The singular value decomposition (SVD) and variance  decomposition
proportions (VDP) analyses are performed using the  procedures of  Belsley,
et.al. (1980).  The SVD consists of the first three columns of Table  15,
where "rows" (1st column) correspond to eigenvectors of the source  profile
matrix, ranked by decreasing "singular values" (2nd column); and  "condition
indices" (3rd column)  are the ratio of the largest  singular value to  the
singular value of each row.  (This analysis is similar  to eigensystem
analysis: singular values are the square roots of the eigenvalues,  but  are
computed in a way which decomposes the source matrix S^, rather than S S_.)
The remaining columns  show the VDP, where "coef"  represents the 13  source
profiles.  The VDP, which range from 0 to 1 and sum to  1  in each  column,
show the fraction of variance associated with each  source class and each
eigenvector.
     Col linearity which may be serious enough to degrade  source
apportionments is indicated by singular values with high  condition  indices
(greater than about 100) and two or more high VDPs  (greater than  about  0.8
or 0.9).  Table 15 clearly shows the dependence between oil-A and oil-B
sources (condition index 122, coefficients 3 and 5)  and a somewhat
surprising multiple dependence between steel-A, steel-B,  glass, basalt, and
soil sources (condition index 300, coefficients 1,  2, 8,  9, and 10).  The
coefficients for these source classes may be poorly estimated by  OLS
regressions.
     Somewhat similar  dependencies for the Quail  Roost  II profiles  are  shown
by the correlation matrix (Table 16).  Very high  or low correlations  (near +
1) indicate strong dependencies between two sources.  There are six pairs of
profiles that have correlations above 0.9, namely,  steel-A and steel-B
(arguments 1 and 2); coal-A and aggregate (6 and 7);  aggregate and  glass (7
and 8); aggregate and  basalt (7 and 9); basalt and  soil (9 and 10)  and  auto
and wood (11 and 12).   Note that correlation coefficients do not  show
multiple dependencies.
     The variance inflation factor (VIF) does indicate  both single  and

                                   49

-------
                                  VARIANCE ptco»»o*nion MATRIX
IW
1
1
3
4
S
i
7
a
9
10
It
I]
13
S1NO.VM..
I S]t«
1. ill 14
1 29317
0. §36041
0.730SM
O 4467OI
O. 413176
0. 1(1581
O 13SO3S
0 044149
0 0276 13
O 020733
0 008438
COMB
1
1
3
3
3
B
8
13
18
S7
91
113
300
.INDEX

.3141
.019O4
.93734
.51335
. <«733
. 13711
.9419
.7491
2177
.5114
092
.029
COEF.1
O.
O.
0.
O.
0.
0.
0.
0.
O.OOl
0
O.O14
0 044
0 931
coir. a
0.
0.
o.
0.
0.
0.
0.
0.
O.003
0.014
0.034
0 O4
0 111
COCF.3
o.
0.
o.
o.
0.
o.
O.OO3
O.OO3
O.OO9
0.08
0 021
0 837
O.O67
COEF . 4
0-
0.
0.032
O.OS1
0. 123
O.OO8
O.OO7
0.
0.039
0.647
0 014
O.O13
0.074
COtf .5
o.
O.
0.
o.
o.
0.
o.
o.
o.oot
O.O19
0 139
0.836
O.019
COEF.«
0,
0.
o.
g.
o.
o.
o.
o.
o.ooc
o. 101
0.07ft
O.O3*
0.711
COEF. 7
0.
O.
0.
0.
0.
0.
0.
O.
0.
0.
0.399
0 111
0.631
COtf .1
0.
0.
0.003
0.
0.003
O.002
O.OH
O.01
O.
0. 101
0.013
O.OOl
0.849
COJf 9
0.
0.
0.
0.
0.
0.
0
0.
0.
O.O69
0.014
0.
O.919
COtf . 1O
O.
O.
0.
0.
0.
0.
0.
0
0 OO»
0.00*
0.041
0 003
0.93
COtf . 1 1
O.
6.003
0.
O.OO2
O.OO3
O.OO3
O.C47
0.63
0.003
O 089
0 O53
0 0»1
0. 141
COtf . 12
O.
O.
O
O.
o.
o.
0.
O 02S
0
0 OO1
0 11
0 634
0 117
cotf. i:
0.
o.
o.
0
0 OO3
O.O34
O.OOl
O.OO6
O.019
0 14
0 42
0 321
O.OS1
Table 15. Proportionate variance decomposition of the contaminated Quail Roost  II
          source profiles. "Coef correspond to 13 source class  profiles  (columns
          of Table 10. See Table 17 for  identification of profile numbers.)
ARG1
AKG2
ARG3
M)G4
ARG3
ARC*
ARG7
AftGB
ARC?
ARC 10
ARC 11
ARG12
MG13
ARC I
1.000
0.983
0.048
0.014
0.407
0.136
0.031
-0.060
0.150
0.104
0.441
0.712
0.09O
ARG2
l.GOO
0.097
0.024
0.468
0.124
-0 . 002
-0.004
O.OY3
0.092
0.728
0.774
O.078
ARG3
1.000
0.040
0.400
0.822
0.74?
0.38?
0.404
0.717
0.043
0.088
0.703
AR64
1.000
0.108
-0.054
-0.141
0.404
-0.182
-0.177
-0.013
-0.004
-0.00?
AR09
1.000
0.403
0.176
0.324
0.117
0.2B4
6.764
0.802
0.327
ARG*
l.OOO
0.96?
-0.036
0.877
0.931
0.084
0.123
0.681
AR07
1.000
-0.078
0.746
0.745
-O.I IS
-0.07?
0.70*
ARGB

1.000
-O.071
-0.141
-O.O24
-O.088
0.23?
AR07


1.000
0.740
-0.13?
-0.106
0.76?
ARG 10



l.OOO
0.026
0.068
0.747
AR011




l.OOO
0.937
-O.I 44
ARG12





l.OOO
-0.076
                                                                            ARG13
                                                                            l.OOO
 Table  16.  Simple  correlation  for  the  Quail  Roost II  source profiles. "Arg"
           numbers refer to source profiles.  (See Table 17 for identification.)
  •Number-   --Source—     --VIF-
     1        Steel-A       2963
     2        Steel-B       2258
     3        Oil-A          728
     4        Incinerator      5
     5        Oil-B         1541
     6        Coal-A        1646
     7        Aggregate     2449
•Number-
    8
    9
    10
    11
    12
    13
•Source—
  Glass
  Basalt
  Soil
  Automobile
  Wood
  Coal-B
-VIF--
    84
  2320
  2319
    23
   652
   108
 Table 17. Variance inflation factors for the Quail Roost II source profiles,

                                    50

-------
multiple dependencies.   The VIFs for the same profiles  are  shown  in Table
17.  A high VIF (above about 30) may indicate strong  single or multiple
dependencies.  Table 17 suggests that a majority of the sources are  involved
in strong collinear relationships.   However,  the number of  dependencies  and
the sources which contribute to them are not  indicated  by the VIFs.
     We believe that the use of correlation coefficients and the  singular
value decomposition and variance decomposition procedure of the source
profiles will provide a comprehensive identification  of single and multiple
dependencies.  These procedures indicate the  source classes which will be
poorly estimated as well as those which are not adversely affected by the
collinearity.
     Remedial procedures for alleviating the  problems caused by collinearity
include:
     -eliminating or combining sources
     -using additional filter characteristics, e.g.,  more elemental analysis
     -using ridge regression
     -using principal components regression
Only the first procedure, eliminating (or combining)  sources, is  described
here.
     Eliminating sources from the CMB model is, perhaps, the simplest method
of handling collinearity.  The subset of sources to be  used can be selected,
in part, from the collinearity analysis.  For the Quail Roost profiles,  four
or five of the collinear sources (e.g., oil-B, steel-A, glass and basalt)
might be eliminated, based on the SVD and VDP analysis.  However,  it  is
necessary to account for the possibility that an actual source is eliminated
in the CMB model.  This is the problem of source selection, described below.

Source Selection

     In Quail Roost II, the subset  of sources for CMB modeling was selected
as follows:  1) Initially, all sources were included  in CMB modeling.  2)
Sources with negative source contribution estimates were removed  one  at  a
time, discarding sources with the largest c/./a. (uncertainty/estimated
source contribution), and repeating CMB calculations  for each removal.   3)
When all source contribution estimates were positive, sources with the
largest s'./a. were removed, and the CMB calculations  were repeated after

                                   51

-------
each removal.  4) After all  cj'./a. were less  than  1, discarded sources
were added experimentally.   This  last step  is  necessary  since this procedure
does not necessarily guarantee the  best model.
     The "all possible regressions"  (APR) procedure may  provide a more rapid
and systematic procedure of  determining sources  in CMB models.  These
computerized routines efficiently generate  all possible  regressions, that
is, all possible combinations of  sources (Montgomery  and Peck,  1982).  The
procedure calculates regressions  with 1 to  p sources, where  p may be
specified by the user (and is less  than or  equal  to n).   For each level of
p, the APR procedure calculates all  regressions  and determines the models
with the best fits, then prints out the sources  included in  a specified
number of the best models.   This  procedure  is  as  fast as stepwise
regression, and is less arbitrary in the sense that no combination of
sources is omitted.
     For n sources there are 2n-l  possible  regression models, a
potentially large number.  Practically, however,  if strong collinearity
involving k sources exists,  then  only a maximum  of n-k sources  need be
included in CMB calculations, and 2n~ -1 regressions  need be calculated.
With the Quail Roost II data set, for example, several of the 13 sources had
strong collinear relationships and  four or  five  fewer sources need be
included in any CMB model.
     The APR procedure is demonstrated using a synthetic data set created
for the Philadelphia area.   Synthetic data  sets  have  the advantage that
complete control is maintained over the data:  the "true" apportionment and
the errors are known.  However, the error structure and  composition of the
synthetic data sets may not be realistic.  The synthetic data set created
here was designed only to illustrate the diagnostic procedures.  The data
set was constructed using the PEM dispersion model, source  inventory and
meteorology from Philadelphia (Section 3) and the same  procedure used in
Quail Roost II.  In brief, the 13 profiles  from  Quail Roost  II  were assigned
to the  12 source classes used in PEM modeling, multiplied by their  predicted
contributions to receptor concentrations, and contaminated  with random
errors.  The synthetic data set used for this evaluation was derived from
period 2 meteorology at receptor 10 (see Section 4),  and had major
                                                  3                  3
contributions from three sources.  Soil: 50.72 ug/m  ; wood:  19.74 ug/m  ;
                         3
and automobile: 7.80 ug/m .   It also included minor contributions from

                                   52

-------
other sources,  as shown in Table 19.   Ordinary  least  squares with 8  profiles
and a constant  were used in a  simple  CMB  approach.  Four  profiles, as
indicated in the col linearity  analysis, were  eliminated from the CMB model.
     Mallow's C  statistic, which is  related  to the mean  square error of
the estimated source contribution,  was the  criterion  used for  source
selection in the APR procedure (Table 18).  For p  sources included in a
model (p ranging from 1 to 13),  the five  best models  were generated.  For
example, the best one variable model  used Coal-A;  the second best used
Oil-B, (sources 6 and 5 respectively). The best two  variable  models used
soil and wood sources (sources 10 and 12).  In  general, models with  low
values of C  are of interest.   This corresponds to a  9 variable model.
However, Table  18 shows that a number of  models with  various combinations
and numbers of  variables have  about the same  fit,  i.e., there  may be no
clearly superior model.  The Table also shows that the models  selected tend
to be quite similar, i.e., include or exclude similar sources.  Interesting,
while all models composed of five or  more sources  included  the major
contributing sources of the synthetic data  set  (sources  10, 11, and  12),
models of fewer variables may  not include these sources.  Models with a
large number of sources tended to include sources  that were identified as
being collinear, e.g., coal-A  and aggregate (sources  5 and  7).  This
indicates that  a model with fewer than 9  variables may be preferable since
collinear source profiles will not be accurately estimated.
     Since the  all possible regressions procedure  tests all combinations of
sources it may be more comprehensive  than the procedure used in Quail Roost
II.  However, unlike the Quail Roost  II procedure, it does  not focus on
removing negative source contributions.  This may  not be  a  serious drawback
since the procedure generates  many "near  optimum"  models.  If  the 2nd or 3rd
best model of p sources contains negative source estimates, then models
containing p-1  sources may be  examined.  The  likelihood of  significant and
negative source estimates should decrease as  models contain fewer sources.

Influential Characteristics

     Influential characteristics (or  observations) are filter
characteristics that exert a disproportionate influence on  the fitted
model.  For example, lead is an influential characteristic  for automobiles,

                                   53

-------
          REGRESSIONS WITH 1  VARIABLES (CP)

      CRITERION      VARIABLES
       109293.
      113790.
      116458.
      125723.
      142717.
 6
5
10
12
11
          REGRESSIONS WITH 2  VARIABLES (CP)

      CRITERION      VARIABLES
         1390.26
        7533.54
       13928.
       49429.3
       50748.8
 10  12
6  12
6  1 1
5  6
?  6
          REGRESSIONS WITH 3 VARIABLES (CP)

      CRITERION      VARIABLES
          711.606
        1094.24
        3933.7
        5O33.32
        5745.85
 6  10  12
1O  11  12
6  7  12
6912
4  6  12
                   REGRESSIONS WITH 7 VARIABLES  (CP)

               CRITERION      VARIABLES

                    25.4399     2  4  5  1O  11   12  13
                   27.2665     3  6  9  10  11   12  13
                   28.0219     5  6  9  1O  11   12  13
                   49.3224     4  5  6  10  11   12  13
                   91.4441     6" 3  9  1O  11   12  13

                   REGRESSIONS WITH 8 VARIABLES  (CP)

               CRITERION      VARIABLES

                     7.7S109    4  5  6  7  10   11  12  13
                    7.86788    3  4  6  7  10  11   12  13
                   11.6995     2  4  5  6  10  11   12  13
                   11.9773     1  4  5  6  10  11   12  13
                   13.9464     1  2  4  5  10  11   12  13

                   REGRESSIONS WITH 9 VARIABLES  (CP)

               CRITERION      VARIABLES

                     6.64801    2  4  5  6  7  10   11  12  13
                    6.99388    1  4  5  6  7  10 11  12  13
                    8.83725    3  4  5  6  7  10 11  12  13
                    9.32423    3  4  6  7  9  10 11  12  13
                    9.39112    4  5  6  7  8  10 11  12  13
          REGRESSIONS WITH 4 VARIABLES (CP)

      CRITERION      VARIABLES
          446.823
         485.597
         51 1 .658
         519.943
         552.52 I
 6  10  11   12
6  10  12   13
      10
      10
      to
12
12
12
          REGRESSIONS WITH 5 VARIABLES (CP)
      CRITERION
                    VARIABLES
           61.0221     5  10  11   12   13
         243.687      6  9  1O  11   12
         272.C88      6  10  11   12   13
         296.008      1  6  10  11   12
         341.835      2  6  10  11   12

          REGRESSIONS WITH 6 VARIABLES (CP)

      CRITERION     VARIABLES

           52.6796     5  6  10  11   12   13
          62.2877     4  5  10  11   12   13
          62.7126     5  7  10  11   12   13
         135.011      6  9  10  H   12   13
         173.831      4  6  9  10  11  12
                             REGRESSIONS WITH 1O VARIABLES (CP)

                         CRITERION      VARIABLES
8. 1 1S24
8.29817
8.43179
S. 61338
8.66667
REGRESSIONS
CRITERION
10.026
10.2101
10.2512
1O. 3361
10.5527
REGRESSIONS
CRITERION
12.0252
12.026
12.1117
12. 1339
12.28O8
REGRESSIONS
CRITERION
14.
1 2
2 3
2 4
2 4
1 3
WITH
4
4
5
5
4
1 1
5
S
6
6
5
6
6
7
7
6
7
7
9
8
7
VARIABLES
10 11 12 13
10
10
10
10
11
1 1
1 1
11
12
12
12
12
13
13
13
13








(CP)
VARIABLES
1 2
2 4
2 3
1 4
1 3
WITH
3
5
4
5
4
12
4
6
5
6
5
5
7
6
7
6
6
a
7
B
7
VARIABLES
7
9
9
9
9
10
10
10
10
10
11
1 1
1 1
t 1
1 t
12
12
12
12
12
13
13
13
13
13





(CP)
VARIABLES
1 2
1 2
1 2
2 3
1 3
WITH
3
3
4
4
4
13
4
4
5
5
5
5
5
6
5
6
6
S
7
7
7
VARIABLES
7
7
8
8
a
9
8
9
9
9
10
10
10
10
10
11
1 1
1 1
t 1
1 1
12
12
12
12
12
13
13
13
13
13
(CP)
VARIABLES
1 2
3
4
5
6
7
8
9
1O
11
12
Table  18.  Results  of  all  possible  regression  procedure using  synthetic data  set
             and  Mallow's Cp  statistic.   Table  17 shows  the  identification  of the
             sources  (shown under  variables).
                                            54

-------
since the estimated automobile contribution  is  roughly  proportional to the
measured lead concentration,  which  is  a  fairly  unique tracer for this
source.  In general,  estimated source  contributions may depend more on the
influential points than the rest of the  data.
     Influential characteristics may increase the  accuracy and decrease the
uncertainty of the estimate of the  source contribution  if the
characteristics are correct.   Such  (legitimate)  extreme characteristics are
useful as indicators of a particular sources in RM, and also serve to show
which elements or characteristics should be  analyzed  in the filter sample.
However, the source apportionment estimate may  be  misleading if the
influential characteristics result from  improperly recorded data,
measurement errors, or inappropriate source  profiles.
     Many diagnostic procedures exist to detect influential
characteristics.  These procedures may be classified  into three general
families (Belsley, et. al.  1980):

1. Single and multiple row deletions:  Essentially, one or more filter
characteristics is removed from the analysis, and  the source apportionment
is re-estimated.  Large changes between  the  source apportionment estimated
with missing characteristics and with  the complete set  of characteristics
may indicate critical (i.e.,  possibly unique) filter  characteristics and
possibly uncertain source estimates.  This process is repeated for each
filter characteristic.  Single and multiple  row deletion measures include:
       change of estimated source class  coefficients—DFBETA(S)
       change of fit--DFFIT(S), PRESS

2. Examination of the source profile (X) matrix:   These measures examine
only the source profile matrix and look  for  single or multiple
characteristics which are unique to each profile.  If few characteristics
distinguish a profile, the source apportionment may be  sensitive to these
characteristics.  Common measures include:
       diagonals of the projection matrix H  = X(XTX)  x'
       covariance matrix—COVRATIO

3.  Examination of residuals:  Residuals (the difference between observed
and fitted filter characteristics in CMS models) may  be examined to

                                   55

-------
determine if least squares procedures  are  appropriate and tests of
significance are valid,  i.e.,  by  examining whether the residuals are
correlated and normally distributed.   In addition, plots of residuals may
provide a great deal of information concerning  influential points.  Residual
diagnostics include:
       normal probability plots
       standardized and studentized residuals
       partial residual plots
       partial regression leverage plots

     Two types of diagnostics  are illustrated using  the synthetic data set
and ordinary least squares procedures.  Five of the  profiles that showed
col linearity problems were removed from the model.   Actual and estimated
source contributions are shown in Table 19. This apportionment is meant
only to illustrate the use of  diagnostics: no attempts were made to improve
the apportionment, e.g., by omitting  sources with negative contributions.
     DFBETAS are single row deletion  statistics, which quantify the effect
of an individual filter characteristic on  the estimated source class
contribution.  DFBETAS indicate the change of the estimated source
contribution in terms of its standard error. A reasonable cut-off for
determining influential characteristics is +_ 1, where omitting a
characteristic, e.g., Vn concentration, might change the estimated source
contribution by one standard deviation (of its  uncertainty).
     Table 20 shows DFBETAS for the  synthetic data  set.  Each row of  the
Table shows the DFBETAS for the removal of one  characteristic from the CMB
model.  For example, removing  characteristic 1, the  mass of carbon, from the
regression would change the estimated oil-A contribution by 3.83 times the
standard error of this coefficient,  a large amount.  Five characteristics
(1=C, 3=AL, 7=K, 13=Fe, and 19=Pb) have DFBETAS which exceed the cut-off for
a several source classes.   As expected,  Pb is  an  influential characteristic
for the auto source class.  The four  other influential characteristics
affect steel-B, oil-A, incinerator,  aggregate,  coal-A, soil, coal-B and wood
classes.
     Partial residual plots permit a  visual indication of  influential
subsets.  Each plot reveals the relationship between particular source
profiles, adjusted for its collinearity with the other profiles, and  the

                                   56

-------
-Source Class-
Steel a
Steelb
01 la
Incin
Oilb
Coala
Aggreg
Glass
Basalt
Soil
Auto
Wood
Coalb
          	Source Apportionment in ug/m	
          -Actual--      -OLS Estimate-     -Robust Estimate-
            0.016
            0.000
            0.121
            0.009
            0.275
            0.022
            0.021
            0.076
            0.019
            50.321
            7.795
            19.737
            0.000
 3.6
-1.0

13.7
-9.6
44.0
 4.7
20.7
-6.4
13.8
-9.7
44.1
 4.7
20.6
-6.5
Table 19. Estimated CMB  source apportionment  using  8 source classes and
          synthetic data set; actual apportionments are PEM predictions; only
          coefficients with 95 percent  confidence are shown.
OBSER.
STEELB  OILA    INCIN  COALA  AGGREG   SOIL    AUTO
                                                         WOOD
                        COALB  CONST
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
1. 170
O.02S '
0.938
O.O39
0.021
O.453
1 .597
-O. 155
0.045
0. 155
O.OO2
O.012
3.227'
O.OJJ
O.O17
•O.50G
-0.007
-O.631
O.OS6
3.831
0.066
1.867
-O.422
-0.514
-O.316
-0.557
-1.555
-O.O61 ,
-0.201
-O.OO1 '
O.O14
2.354
-0.056
O.OO4
O. 163
O.O03
0.227
0.313
' -3.258)
0. 136
0. 122
O.O96
O.OO3
0.089
-2.369
-0.728
-0. 159
-O.O27
0.002
-0.019
1 . 1 14
0.011
0.047
O.426
-O.OO3
0.060
0.207
-O.54S
O.OO1
-O. 117
-O. 197
0.055
0.538
r.797)
0.447
0. 130
0. 175
0.001
O.O03
- 1.834
O.O43
-0.015
-O.5O7
-O.OO4
-0.504
-0.051
-O.539
-O.023
0. 192
O. 111
-O.037
-0.551
-1.878
"O.287
-O. 142
-0. 140
-O.OO1
-0.010
0.982
-O.O31
0.016
O.545
O.OO4
0.5O2
-O.01Z

0.
-0.
-O.
0.
0.
1 .
- 1 .
0.
O.
-0.
0.
1.
0.
-O.
-0.
-O.
-O.
0.
1.688
065
346
314
213
434
322
62O
1 17
041
OOO
018
090
OO4
009
441
OO2
353
104

0
-O
O
O
-O
0
O
-O
0
0
-0
0
O
0
0
-0
0
-3
0.219
.314
. 171
.087
.029
. 190
.010
.029
. 136
. 141
.004
.030
.275
.064
.065
.241
.CO9
.-603
.850
.43.510
-O.208
-O. 109
0. 145
-0.055
-0.416
-1 .742-1
0. 198
-0.044
-O.206
-O.OO3
0.004
-0.41Q
-O.063
-0.025
0.406
O.OO9
O. 158
f. 136
' -3.015
-O. 1O8
-2.517
0.720
-O.439
-O.223
-O.437
1.472
-0.073
0.097
0.002
-O.03O
- 17702
O.041
0.026
0.331
-O.OO2
0.050
-O.238
-O.31S
-O.010
-O. 198
0.07O
0.016
-0.2O7
-O.659
0.075
0.236
-0.440
-0.010
0.062
-0.48S
-O. 173
-O. 145
-0. 159
O.024
1.069
0.312
Table 20. DFBETAS row deletion statistics for CMB of  synthetic  data  set,
          Observations refer to the elemental measures  in Table 13.
                                    57

-------
residuals (the difference between  fitted and  observed filter
characteristics), adjusted for  the other sources  (i.e.,  the source
contributions of the other sources have  been  removed from the residuals).
Thus, each plot illustrates the fit between source  profiles and filter
characteristics where the effect of the  other sources has been removed.
     Partial residual plots for 8  source classes  in the  CMB model using the
synthetic data set are shown in Figure  14.  For example, the top  left plot
is the partial residual plot for the soil class.  The estimated soil
contribution is related to the  slope of  the regression line (shown with
asterisks); the significance or fit of  the data to  the soil profile is
indicated by the scatter around the regression line.  The plot for the
incinerator class shows a negative slope (and thus  negative source estimate)
and considerable scatter, both  of  which  imply that  results for this source
may not be meaningful.
     Influential points or subsets typically  form isolated clusters of
point(s) located some distance  from the  majority  of points on the partial
residual plots.  Influential points heavily influence the regression line,
i.e., the source estimate, since the regression line minimizes the sum of
squared differences.  For example, point D (indicated with an arrow)
(characteristic 7=K) is an influential  point  for  the oil-A, incinerator,
coal-A, aggregate and soil classes.  This point alone heavily weights the
regression line for oil-A, incinerator,  coal-A and  aggregate source classes,
which were poorly estimated by  OLS.  The validity of the K measurement is
thus highly important for these source  estimates.   Similarly, point K
(19=Pb) is influential for the  auto and wood  sources.  Error in the lead
measurement will highly bias the auto estimate.   However, the wood source
estimate may be more resistant  since other characteristics lie on the
regression line.  This information is similar to  that provided by the
DFBETAS.  The partial regression plots  also show  an influential subset—not
definitively indicated by the DFBETAS--in the case  of coal-B, which has two
influential points, B (3=A1) and D (7=K).
     DFBETAS and partial residual  plots  may be effective ways to determine
the effect of both single and multiple  influential  characteristics in CMB
models.  Influential characteristics may require  further analysis to
determine the legitimacy of the characteristics.
                                   58

-------
no««-»
•••
t.as



O.3S



•O.7S


•1.7S
9 •*•• °-« '•? -o. us -0.078 -o.oas o.oas o.or»
0.2S
SOIL .H'c
.
la

OILA J . *
A «0
 	 " 	 '• 	 """"1 ••«! AGGREG
HOC
t* FA
» I J
02J O.O4
I A.
.
K *
a.
• -O.Oi

.
COALB •' i<0 . -«•'<>
* u
• 1

H > A
•Fl
•HA3J
a oa
K .
1 *

.J


   -o.oso -o.oas  o.ooo , o.oai  o.oso  O.OTS  o. too
Figure 14.  Partial  regression plots for 8  source classes.  Adjusted  elemental
            mass  on  ordinate; adjusted source profile (as indicated)  on
            abscissa.    Fitted least squares  line is shown with asterisks.
                                     59

-------
Robust Regression

     Robust regression procedures  are  designed to diminish the effect of
outliers, such as erroneously measured filter characteristics, data entry
mistakes, etc.  In ordinary least  squares  (OLS), outliers tend to have a
disproportionate effect on estimated source contributions.   In robust
regression, the influence of outliers  is constrained by downweighting
outliers.  Robust procedures are especially useful when the  normality
assumptions of OLS are in question.  The principal value of  robust
regression in CMB RM, perhaps, is  to provide a check on OLS  results.
Agreement between the OLS and robust regression results may  help insure that
the estimated source class contributions are meaningful.  Robust regression
may be viewed as an automated way  of determining the influence of outliers
(as well as influential points).
     Most robust procedures employ iteratively reweighted least squares.
First, source estimates and residuals  are  calculated using ordinary OLS.
Then, a specified influence function (Hampel, 1974)  is used  to downweight
large residuals in a weighted least squares regression.  An  iterative
process using weighted residuals and the influence function  continues until
convergence.  Watson's "effective  variance" procedure also uses iteratively
reweighted least squares procedures.   However, the effective variance
procedure gives more weight to ambient source measurements which are
measured with precision, as well as source class estimates which are
certain.  It does not diminish the effect  of outliers which  tend to pull the
OLS solution.
     Robust procedures were used on the  synthetic data sets  using 8 source
signatures in TROLL.  The standard convex  Huber influence function was used
to weight the squared residuals (Peters, et. al., 1981).  Robust results
were very similar to OLS results,  i.e, few characteristics were
significantly downweighted (Table  19).  This probably occurred because the
contamination method did not generate  outliers  (only normal  errors were
used.)  In practice, however, results  using OLS and  robust procedures may
vary.
     It is also possible to determine  which characteristics  were
downweighted in the robust regression.  Examination  of such  points may
identify possible errors.  We note that  it is possible to combine robust

                                   60

-------
procedures with ridge regression and the effective variance  technique.
Lastly, a procedure similar to robust regression,  called bounded-influence
regression, designed to limit the bias caused by contamination  of  data   such
as that caused by keypunch mistakes (Peters,  et. al.,  1981 )f may have
application to CMB apportionments.

HYBRID RECEPTOR/DISPERSION MODELS

     This section discusses several possibilities for  hybrid
dispersion/receptor models.  Hybrid models use a receptor model approach in
conjunction with source emission rates, source locations and/or dispersion
information.  Such models conceivably might provide a  more accurate and
flexible means of source apportionment.  They may also help  reconcile
differences between source and receptor models.
     Some potential uses of hybrid RM/DM approaches include:

Operational validation.  Receptor and dispersion models are  used separately.
Inter-model comparisons may help identify emission inventory deficiencies,
i.e., missing or mis-estimated emission sources. Inter-comparisons can help
"confirm" apportionment estimates by employing a "consensus"  standard.

Diagnostic validation.  Model comparisons are used to  refine and calibrate
model parameters, e.g., dispersion coefficients  and source profiles.  Such
validation may be appropriate for model development.

Complementary use. Outputs from receptor and  dispersion models  are
combined.  For example, DM might be used to apportion  regional  or  background
concentrations, and RM might estimate local concentrations.   Other examples
include using RM with filter samples collected according to  the observed
frequency of meteorological conditions, to determine long term  averages;
collecting filter samples during meteorological  patterns which  result in the
highest DM predictions; restricting the profiles in CMB models  to  upwind
sources; and calculating back-trajectories which incorporate transformation
and deposition for receptor determined mass source apportionments  (e.g.,
Thurston, 1933).  Some of these approaches may be more practical with
short-term (e.g., hourly) filter samples.

                                   61

-------
Coupled models.   All  data,  including meteorology, source emissions and
profiles and filter characteristics are  used  in coupled models.  Coupled
models intimately combine receptor and dispersion approaches, and thus may
be more complicated than other  hybrid approaches.

     Coupled models might be classified  by their primary orientation, around
either dispersion or receptor approaches.   Some speculations concerning
possible coupled hybrid dispersion/receptor models  are given below.
     For receptor oriented hybrid models,  outputs of  DMs might  be treated
stochastically.   CMB receptor models might then be  extended by  either
including DM predictions as priors  in Bayesian optimizations, or by  using DM
predictions of source contributions as coefficient  weights in weighted
optimizations.  Analogous methods are possible for  multivariate RMs.  Here,
preprocessed meteorological data might become new variables in  a factor
analysis or multiple regression model.   Association of particular
meteorological patterns with source class  signatures  may identify the likely
direction and distance of contributing sources.  Alternately, source
contributions predicted by DMs  may become  variables in factor analyses,
along with the usual filter characteristics.   Associations of predicted
source contributions with the respective source profiles may indicate areas
of agreement between the models.
     The second general type of coupled  models is based around  dispersion
modeling.  In some ways, this approach  is  similar to  the diagnostic
validation discussed earlier applied to  DMs.   The (deterministic) inputs or
parameters of DMs are considered as stochastic variables  (with  estimated
distributions).   Optimization is used to match DM predictions to observed
concentrations or receptor determined apportionments, at the same time
maximizing the probability of the stochastic  variables  in the DM.  Yamartino
(1982) and Cooper (1982) have used  a time  series of observed concentrations
to estimate average source strength by region and sector, respectively.  The
identification of large emissions in areas without  known sources may help
quantify the regional (background)  contribution.  These studies did  not use
aerosol characteristics.
     More sophisticated coupled models may be possible  if short-term (CMB)
receptor model results are available and of suitable  accuracy.  This may

                                   62

-------
permit the estimation of short-term parameters in  dispersion  models.   For
example, source inventories generally include only annual  average emission
rates.  However, hourly and seasonal emission rates are  highly variable.
The distribution of hourly emissions may be estimated.   The optimization
model may be used to select emission rates, within the assigned
distribution, that produce results closest to the  source apportionment
determined previously using RMs.  DM using the revised emission rates then
might provide the hybrid source apportionment.  Similar  approaches may be
possible for other variables in dispersion models, e.g., dispersion
coefficients and wind direction.

     There has been relatively little work in the  development of hybrid
models.  Some of these models may be used to help  apportion both local and
regional pollutants in the Philadelphia study.  Data from this study, as
well as from realistic synthetic data sets, may provide  a means to test the
feasibility of the various hybrid models.
                                   63

-------
COMPARISONS BETWEEN SOURCE AND  RECEPTOR MODELS

      Some aspects of model validation and  model  inter-comparisons are
discussed in this section with  reference  to the Philadelphia study.  The
evaluation of similarities and  differences  between dispersion and receptor
models may help explain strengths  and weaknesses  of  the models, and possibly
indicate the certainty of the source apportionment.
      One approach for identifying and quantifying the likely causes of
model agreement (or discrepancy) might use  regression or factor analyses.
These procedures may be used to associate measures of agreement between RM
and DM results with meteorological data,  averaging time, source class, and
other information.

Evaluation Criteria

      This section addresses the quantitative "performance" measures used to
evaluate model results.  In the Philadelphia study,  both RM and DM results
will be compared, and the true  source  apportionment  is not known.  Such
inter-model comparisons involve somewhat  different concerns than model
validation studies where predicted concentrations are compared to observed
concentrations, or the Quail Roost exercise, which inter-compared only
receptor models.
      There are several useful  strategies for inter-model comparisons:

(1) Comparison between all models.  Provides the  most information and makes
the fewest assumptions, but (n-l)l comparisons must  be made for n models, a
potentially large number.

(2) Comparison of each model's  results to the mean or median result.  This
is useful since it highlights deviations  from the mean or median
prediction.  However, there is no  assurance that  the mean or median
prediction is the best estimate of source apportionment.

(3) Comparison to an estimate of true  source apportionment, based on
considering and weighing all available evidence.  Joint and independent
application of RM and DM as suggested  by  Core, et.al.,  (1982) might be the

                                   64

-------
means to obtain the source apportionment estimate.   Additional ways of
combining receptor and source models were discussed  previously.

     While the first two strategies permit direct comparisons of  model
results, it is felt that the third strategy may be the most comprehensive
approach to model inter-comparison.  However,  as discussed previously under
the section on hybrid receptor/dispersion modeling,  there are not yet
objective and practical approaches which appropriately combine all evidence.
     A second issue of RM and DM comparisons concerns the quantity
compared.  In general, it would seem that the  quantities compared should
relate to modeling objectives.  Comparison of  (possibly size fractionated)
source class contributions of the particulate  mass might be appropriate, as
used in Quail Roost II.  It is important to realize  that neither  DMs nor RMs
apportion regional sulfate, which may contribute much of the fine particu-
late mass.  DMs, such as PEM, do not model regional  sources (beyond 60 km
distance).  RMs may not accurately apportion sulfate since sulfate does not
have identifiable tracers.  Consequently, apportionment of the sulfate in
the Philadelphia study will involve large uncertainties.  Apportionment of
sulfate might be broken out separately from the apportionment of  local
sources, to help clarify the contribution from regional sources.
     Third, the different nature of the outputs produced by DM and RM should
be recognized in inter-comparisons.  DMs provide deterministic point
estimates;  RMs provide statistical or qualitative estimates.  While it is
possible to weight comparisons between RM results with estimates  of their
statistical confidence, this can not be accomplished easily with  DM.  One
possible solution is to compare only the statistically significant RM
results to DM results.  Statistically insignificant  results might be set to
zero.  One or several confidence levels might  be selected for this purpose.
(Confidence intervals should be explicity incorporated in the calculation of
the best estimate of source apportionment, as  discussed later.)

Averaging Time

     Bias, distributional effects and extrapolation  errors are three effects
associated with averaging time which may influence model inter-comparisons.
     DMs usually can predict long-term, i.e.,  seasonal and annual,

                                   65

-------
concentrations with greater accuracy than  short-term, i.e, hourly to daily,
concentrations.  Many of the errors  of  short-term DMs result from inaccurate
or incomplete data from the source  inventory, meteorology and ambient
monitoring.  Many of these short-term errors are random, and tend to be
averaged out in long-term concentration averages.   (These generalizations
may not apply to the modeling of  regional  pollutants.)
     RMs also may be susceptible  to  short-term errors since source class
profiles are usually assumed to be  time and source  averages of measured or
estimated profiles.  The errors in  the  source profile may be small if many
individual emission sources compose  a source class  (and  if there are no
systematic biases).  For example, if in a  particular source class there are
n individual sources which equally contribute to observed concentrations,
then the uncertainties of the source profile may be expressed as:
                               standard deviation  of individual profiles
Standard deviation of profile =	-—;	

since variances are additive. This  suggests that source profile errors will
be small for source classes composed of many sources, even in short time
periods.  Thus, the accuracy of RM based apportionments  of automotive, oil,
soil, etc., source classes may be largely  insensitive to averaging time.  In
contrast, the accuracy of apportionments for some source classes composed of
only a few sources may be more sensitive to averaging time.  Thus, long-term
apportionments from both RMs and  DMs may be the most valid, although some
short-term RM results may be equally valid.  This may be tested by
contrasting measures of agreement between  RM and DM results for different
source classes at different averaging times.
     It may be worthwhile to evaluate model agreement in short time periods
since, in many circumstances, it  may not be possible to  sample and analyze a
large number of samples.  This might exclude multivariate models, since
these models require many observations. However, such comparisons may yield
information of practical value regarding the accuracy and reliability of CMB
models.   (The Quail Roost II exercise primarily used multivariate models,
and thus only the average apportionment over 40  12-hour  averaging periods
was evaluated.)
     The confidence level or uncertainty in short-term  (and  long-term)
source apportionments should be  included in inter-model  analyses.  The
uncertainty of single and multiple  short-term apportionments, e.g., by

                                   66

-------
applications of CMB models, might be estimated using  statistics  such  as  the
mean coefficient of variation of estimated contributions  in  particular
source categories.
     The second effect of averaging time is its influence on pollutant
distributions.  It is widely recognized that pollutant  concentrations at
short averaging times have approximately lognormal  distributions and  large
standard deviations.  Concentrations at long averaging  times have much
smaller standard deviations.  Similarly, the standard deviation  of  source
apportionments will depend on averaging time.  Comparisons between  DM and RM
should account for the these distributions by employing appropriate
performance measures, as discussed by the AMS (1979)  and  Fox (1981).  This
also suggests that hybrid RM-DM or RM models which  use  short averaging times
(e.g., several hours) may be subject to non-normal  errors.   Consequently,
such models should employ robust procedures which are resistant  to  such
"contaminated" data, such as robust regression discussed  earlier.
     The third effect of averaging time concerns extrapolation errors.   In
general, seasonal  or annual source apportionments using RM and (necessarily)
a limited number of filters will require extrapolation  of data.   The  number
of filter samples  required to produce representative  longer  term source
apportionments is  unknown, but is likely to be dependent  on  the  site  and
circumstance.

RECEPTOR MODEL PROTOCOLS

     Pace (1983) has discussed possible roles for receptor models in  the
regulatory framework.  There are a number of important  and unresolved
questions concerning the practical use of receptor  models in the regulatory
setting.  Protocols represent well defined and justified  methods which help
assure quality results.   Protocols for receptor models  may be classified
into three areas:   1) protocols for the physical aspects  of  the  study,
including the design, siting, operation,  and quality  control  and assurance
of monitoring and  filter analyses;  2) modeling protocols, which include
model selection, treatment of uncertainty and diagnostics; and 3) protocols
for model inter-comparisons, to determined preferred  approaches.  This
section discusses  aspects of RM protocols using this  classification.
                                   67

-------
Protocols for Monitoring and  Analysis

       There are a number of  general  issues related to the amount and
quality of the information required  in  RM  studies.  These are important
issues since most costs of RMs  are associated with sampling and analysis of
filters, rather than modeling.   These  issues and trade-offs include:

(1) Number of analytic tests  on the  filter samples:  There is almost an
unlimited amount of analytical  testing  that can be done on a set of
samples.  Obviously, it is important to arrange testing such that the most
cost effective methods are conducted first.  Analytical tests that can break
col linear relationships between sources can be determined by col linearity
diagnostics.  Site-specific cut-offs for analyses might be established,
beyond which further analysis is not cost-effective.  This trade-off between
more and fewer analytical tests could  be defined prior to regulatory use.

(2) Number of samples:  A minimum number of filter samples is required for
representativeness; a larger  number  of  samples is required for multivariate
receptor approaches.  Here again it  may not be worthwhile to collect
additional samples.  Documentation of  RM performance for specific sites and
situations might help define  the necessary number of samples.

(3) Source sampling:   More extensive  source sampling may increase the
accuracy and selectivity of RM.  Such  analysis is expensive, time consuming,
and may not yield representative signatures.  Thus, as in filter analysis,
the costs of source sampling  can be  balanced against its decreasing marginal
value.

(4) Length of sampling period:   At this point most sampling periods are 12
hours in length.  This allows a representative sample of seasonal conditions
with a reasonable number of filters.  However,  12 hour intervals make  it
difficult to use the information contained in wind direction and other
meteorological data.  It is now possible to obtain hourly samples with
sufficient loadings for the analytical  procedures.  With hourly samples, the
source apportionment problem  might well decompose  into different problems
for each wind direction and thus yield finer source resolution.  The cost,

                                   68

-------
however, of running 20 to 25 days of hourly samples  through  the  analytical
procedures would be prohibitive.   Thus,  a trade-off  exists between  the
representativeness of the data and the ability to utilize meteorological
information.

Modeling Protocols

      The diagnostic procedures discussed previously are a subset of
modeling protocol issues.  The main objectives of using diagnostic  tools are
to increase the speed and effectiveness of receptor  modeling and to decrease
the chance of spurious results.  Diagnostics might be incorporated  into
model protocols to 1) detect the  presence of problems such as collinearity,
outliers and influential observations, 2) assess  the extent  to which  the
source apportionment is degraded, and 3)  determine whether corrective action
is necessary.  The usefulness of  the various diagnostic procedures  might be
subject to the same sort of evaluation as model comparison,  in order  to
determine which procedures are the most appropriate.
      A receptor modeling protocol, which incorporates diagnostics, might
include:

      1.  Use of source inventory, dispersion modeling, factor analysis, or
      microscopy to identify major source contributors.

      2.  Use of measured or estimated source profiles in an effective
      variance CMB or multivariate analysis to quantify impacts  and possibly
      detect additional sources.

      3.  Estimate the effect of  collinearity in  the source  profiles  (e.g.,
      using singular value decomposition).

         If necessary, re-estirnate apportionment  (e.g., using ridge or
         principal components regression, or deletion/combination of  source
         classes using all possible regressions).  Obtain positive  and
         significant source contributions.
                                   69

-------
      4.  Determination of the  sensitivity  to outliers and influential
      points using DFBETAS,  partial  residual plots, and/or effective
      variance techniques.

      5.  Confirm CMB results using  robust  regression techniques, or "fresh"
      data.

Clearly, receptor model protocols will  depend on the circumstances and the
effort of the study.

Protocols for Inter-comparison  of Receptor  Models

      Model  inter-comparison serves  the purposes of documenting model
performance  and uncertainty, and determining model applicability, i.e.,
matching models and uses.   Inter-comparison studies of receptor models have
not stressed both objectives, largely due to limited data.  Model
inter-comparisons may include measures of model adequacy for different
source and background compositions  (e.g., rural western vs. urban eastern
sites); varying averaging  times; and various levels of study effort  (e.g.,
with or without source sampling).   More effort  is needed to estimate
confidence levels which are representative.

SUMMARY

      Some of the major issues  in the use of receptor models are:

(1) Art versus science in  the application of RM.  Currently there is a
substantial  amount of imposed "intelligence" that is  introduced  into RM
procedures by experienced  practitioners. Present RM  studies are generally
custom designed, including selection of source  signatures, filters and
analysis.  In the regulatory context, some  of the discretionary elements of
RM might have to be removed to  reduce the chances for misuses  of the
models.  Regulatory protocols must  balance  the  need to ensure  robust and
representative results with the interpretative  and case-specific nature of
RMs.  A middle ground might ensure  that monitoring, filter analysis, quality
assurance and control procedures are adequate,  and RM results  are not

                                   70

-------
spurious.  Also,  appropriate measures  of  uncertainty must be  included.

(2) Development of representative  source  profiles  and  synthetic data sets.
The variation in  source profiles  (over time  and  space) and  the deviation
from the RM assumptions can be determined from source  or near field
measurements, or  perhaps derived using wind  trajectory analysis.  Synthetic
data sets with realistic error structures and compositions  for various
aerosol regimes,  e.g.,  rural western vs.  urban eastern, may be used to
construct RM protocols.  Such data sets may  help identify the resolution and
uncertainties of  the various receptor  models as  well as the necessary study
effort (sampling  and analysis).  Inter-model comparisons using this data can
be used to select preferred models and diagnostic  approaches.

(3) Hybrid receptor/dispersion approaches.   Hybrid models require more data
than either approach alone.  Such  models  may be  cost-effective if source
inventories and meteorological data are available.   If hybrid models permit
significantly improved  performance in  terms  of accuracy and flexibility of
source apportionment, their expense and complexity may be justified in other
circumstances.  At present, hybrid models have not been demonstrated and
critically evaluated.
                                   71

-------
                                 REFERENCES


Air Pollution Control Association (1982):  Specialty Conference  on  Receptor
    Models Applied to Contemporary Pollution Problems.Danvers, MA,  APCA
    Publication. 388 pp.

Cooper, D.W., Receptor-Oriented Source-Receptor Analysis,  loc.  cit.

DeCesar, R., Evaluation of Multivariate and Chemical Mass  Balance  Approaches
    to Aerosol Source Apportionment using  Synthetic Data and an Expanded
    Kacs Data set, loc. cit.

Yamartino, R.Y., Formulation and Application of a Hybrid Source-Receptor
    Mode 1. loc. crT.

American Meteorological Society (1979): Air Quality Modeling and the  Clean
    Air Act:  Recommendations to EPA on Dispersion Modeling for Regulatory
    Applications. Boston, MA. 268 pp.
Belsley, D.A., E. Kuh, R.E. Welsch (1980):  Regression Diagnostics:
    Identifvino Inf—•«---•' r>_j._ —i  •«	~ ?•*/*.-,-,?	1±...   mm
    Sons, New York.
Identifying Influential  Data  and  Sources  of  Col linearity. John Wiley and
          Yc
Cooper, Y.A., J.G. Watson: Receptor Oriented Methods of Air Particulate
    Source Apportionment. J. Air Poll. Control Assoc. 30 1116-1125.

Core, J.E. et. al. (1982): Particulate Dispersion Model Evaluation:  A New
    Approach Using Receptor Models. J. Air Pol. Control Assoc.  32_ 1143-1147.

Energy Information Administration (1980):  Thermal Electric Power Plant
    Construction and Annual Production Expenses. Washington, D.C.

Environmental Research and Technology, Inc. (1981):  The State of the Art of
    Receptor Models Relating Ambient Suspended Particulate Matter to
    Sources. Report P-A42Z to US Environmental Protection Agency, Research
    Triangle Park, NC.

Fox, D.G. (1981): Judging Air Quality Model Performance, Review of the Woods
    Hole Workshop. American Meteorological Society,  Boston, MA.

Hampel, F.R. (1974): The Influence Curve and its R^le in Robust Estimation.
    J. Amer. Statist. Assoc. j)9 383-394.

Huber, P.J. (1981): Robust Statistics. John Wiley and Sons, New York.

Information Processing Center (1980): TROLL Users Guide. Massachusetts
    Institute of Technology, Cambridge, MA.

Montgomery, D, E.A. Peck (1982): Introduction to Linear Regression Analysis.
    John Wiley and Sons, New York.
                                   72

-------
                           REFERENCES (continued)


National Academy of Science, (1983): Acid Deposition:  Atmospheric Processes
    in Eastern North America. National Academy Press,  Washington, D.C. 275pp.

Pace, T.G. (1983): Models to Develop Controls Strategies for PM-10, in
    Proceedings of the 76th Annual Meeting of the Air  Pollution Control
    Assn. Atlanta, GA, June 19-24.

PEDCo Environmental (1979): TSP Source Inventory Around Monitoring Sites in
    Selected Urban Areas, Philadelphia. Report to US Environmental
    Protection Agency, Monitoring and Data Analysis Division, Research
    Triangle Park, NC. 92 pp.

Rao, Shankar K. (1983): Plume Concentration Algorithms with DepositIon,
    Sedimentation and Chemical Transformation IAG-AD-13-F-1-70/-Q.Rational
    Oceanographic and Atmospheric Administration, Oak  Ridge, TN. 87 pp.

Sehmel, George (1980): Particle and Gas Dry Deposition:  A Review. Atmos.
    Environ. 14 983-1011

Suggs, J.C., R.M. Burton (1983):  Spatial Characteristics of Inhalable
    Particles in the Philadelphia Metropolitan Area. J.  Air Pol. Control
    Assoc., 33_ 686-91.

Texas Air Control Board (1979): Texas Episodic Model,  User's Guide
    PB80-227572. National Technical Information Service, Springfield, VA.

Thurston, G.D. (1983): A Source Apportionment of Particulate Air Pollution
    In Metropolitan Boston. Ph.D. Thesis, Harvard School of Public Health,
    boston, MA.
                                   73

-------
                                   TECHNICAL REPORT DATA
                            (Please read Instructions on the reverse before completing)
1. REPORT NO.
                             2.
                                                           3. RECIPIENT'S ACCESSION NO,
4. TITLE AND SUBTITLE
                                                           5. REPORT DATE
 Air Quality Models Pertaining  to  Particulate Matter
                                                           6. PERFORMING ORGANIZATION CODE
r. AUTHOR(S)
S.A.  Batterman, J.A. Fay, D. Golomb,  J.  Gruhl
                                                           8. PERFORMING ORGANIZATION REPORT NO.
9. PERFORMING ORGANIZATION NAME AND ADDRESS
 Energy Laboratory
 Massachusetts Institute of Technology
 Cambridge, MA  02139
                                                                     = LEMENT NO.
                         - 3156  (FY-84)
             11. CONTRACT/GRANT NO.

              Cooperative  Agreement
              809229-01
12. SPONSORING AGENCY NAME AND ADDRESS
Environmental  Sciences Research  Laboratory - RTP, NC
Office of Research and Development
U.S.  Environmental Protection Agency
Research Triangle Park, NC 27711
             13. TYPE OF REPORT AND PERIOD COVERED
              Final 04/83  -  10/88	
             14. SPONSORING AGENCY CODE
              EPA/600/09
15. SUPPLEMENTARY NOTES
16.ABSTRACT This rep()rt describes an  evaluation of the Particle  Episodic Model (PEM), an
urban  scale dispersion model which  incorporates deposition, gravitational  settling and
linear transformation processes into  the predecessor model, the  Texas  Episodic Model
(TEM-8).   A sensitivity analysis  of the  model was performed, which  included the effects
of deposition, gravitational settling and receptor grid size.  Recommendations are made
to improve the performance and flexibility of the model.
     PEM  was applied to a source  inventory of the Philadelphia area to provide a pre-
liminary  estimate of source apportionment.  PEM modeling employed both hypothetical and
actual  meteorology.  Results indicate that area source emissions dominate  TSP, S0? and
sulfate concentrations at urban receptors.  A large fraction of  the inhalable particles
may arrive from distant sources.
     This report also contains an overview of receptor models  (RMs) used for the source
apportionment of aerosols.  Some  diagnostic procedures for RMs are  evaluated using a
synthetic data set.  Described are  RM trade-offs and protocols and  possible hybrid
dispersion/receptor models.  Issues regarding the inter-comparison  of  source apportion-
ments  from receptor and dispersion  models are highlighted with reference to the 1982
Philadelphia study.
17.
                                KEY WORDS AND DOCUMENT ANALYSIS
                  DESCRIPTORS
                                              b.lDENTIFIERS/OPEN ENDED TERMS
                             COSATl Field/Group
18. DISTRIBUTION STATEMENT
   RELEASE TO PUBLIC
19. SECURITY CLASS (This Report I

 UNCLASSIFIED
                                                                         21. NO. OF PAGES
20. SECURITY CLASS (This page)
 UNCLASSIFIED
                                                                         22. PRICE
 EPA Form 2220-1 (R»v. 4-77)   PREVIOUS EDITION is OBSOLETE

-------