JUNE 1987
ANALYSES OF PEM-2 MODEL EVALUATION RESULTS
FOR SHORT-TERM URBAN PARTICULATE MATTER
ATMOSPHERIC SCIENCES RESEARCH LABORATORY
OFFICE OF RESEARCH AND DEVELOPMENT
U.S. ENVIRONMENTAL PROTECTION AGENCY
RESEARCH TRIANGLE PARK, NORTH CAROLINA 27711
-------
ANALYSES OF PEM-2 MODEL EVALUATION RESULTS
FOR SHORT-TERM URBAN PARTICULATE MATTER
James M. Godowitch
Meteorology and Assessment Division
Atmospheric Sciences Research Laboratory
Research Triangle Park, North Carolina 27711
ATMOSPHERIC SCIENCES RESEARCH LABORATORY
OFFICE OF RESEARCH AND DEVELOPMENT
U.S. ENVIRONMENTAL PROTECTION AGENCY
RESEARCH TRIANGLE PARK, NORTH CAROLINA 27711
-------
NOTICE
The information in this document has been subject to the United States
Environmental Protection Agency's peer and administrative review and it has
been approved for publication as an EPA document. Mention of trade names or
commercial products does not constitute endorsement or recommendation for use,
AFFILIATION
The author is on assignment to the Meteorology and Assessment Division,
Atmospheric Sciences Research Laboratory, from the National Oceanic and
Atmospheric Administration, U. S. Department of Commerce.
-------
ABSTRACT
The Pollution Episodic Model Version 2 (PEM-2), an urban dispersion model,
has been evaluated with measurements from the 1982 Philadelphia Aerosol Field
Study data base in order to investigate its ability to model 12-hour average
concentrations of particulate matter less than 10 micrometers (PMio)« Modeled
fine (< 2.5 pm) and coarse (2.5 - 10 pm) particulate total masses were combined
and then statistically evaluated against corresponding PM^g measurements at six
monitoring sites for a 29-day experimental period.
Modeled results from urban emissions alone underestimated measured concen-
trations by up to a factor of 4, which revealed regional background values were
large and needed to be accurately determined. Regional PMjQ background was
derived from the measured concentration at an upwind site selected as the back-
ground monitor with the modeled PM^Q concentration subtracted because all sites
were in the emissions region. About 70% of the measured PMio at most monitoring
sites was contributed by regional background.
Model performance was determined from statistical measures of difference
and correlation between observed and modeled concentrations paired in time and
location. Statistical results were better for modeled plus background values
versus observed concentrations. Low correlations were found from concentration
pairs composed of modeled concentrations and observed minus background values.
The regional background dominated many of the evaluation statistics since it
represented a large fraction of measured urban PMjQ concentrations. Results
differed based upon the method of accounting for background.
Concentration estimates from PEM-2 and the RAM model are compared from
independent evaluations with this data base. Mean and high-five PM^j concentrations
from the PEM-2 model were about 25% lower than RAM predictions at four sites within
the city limits, however, PEM-2's results were 35-40% lower at two most distant
sites from the urban center. These differences in model results are attributed
to particulate removal by dry deposition and settling processes in PEM-2 and
the different treatments of area source emissions by these models. Results of
statistical measures were still quite similiar for both models. Due to the
dominant role of regional background concentrations in this evaluation study,
it was not possible to conclude which model performed more accurately or that
the PEM-2 provided any clear advantage over the RAM model results.
-------
CONTENTS
Abstract i i i
Fi gures vi
Tables vi 1
Acknowledgement vi i i
1. INTRODUCTION 1
Model Overview 2
2. MODEL EVALUATION DATA BASE 3
Emi ssions 3
Parti oil ate Measurements 5
Meteorological Parameters 8
3. MODEL EVALUATION RESULTS AND DISCUSSION 12
Model Calculation Procedures 12
Regional Background Component , 13
Model Evaluation Statistical Results 19
Comparative Results for the PEM-2 and RAM Models 25
4. SUMMARY AND CONCLUSIONS 32
5. REFERENCES 34
APPENDICES
A Evaluation of Modeled Mixing Heights with PAFS Observations ... 36
B Definitions of Statistical Measures 39
C Observed and PEM-2 Modeled PM Concentrations 41
-------
FIGURES
Number Page
1 Area emissions grid and PEM-2 model calculation grid for the 4
Philadelphia evaluation. Dashed line denotes the city limit
and the numbered locations are the six PAFS sites.
2 Map of the Philadelphia metropolitan area and the locations 6
of the PAFS particulate monitoring sites and Philadelphia
International Airport (PHL) where meteorological observations
were taken.
3 Temperature (solid) and humidity (dashed) profiles for 9
A) before sunrise and B) mid-afternoon on on 3 August 1982
at the Camden (CAM) site.
4 Time series of PM^g measurements (dashed) and modeled (solid) 16
concentrations (without background) averaged over all sites
for each 12-hour period of the PAFS study period. Day 1 to
30 corresponds to 16 July through 15 August 1982.
5 Observed versus model predicted (plus background) PM^g 21
concentrations at all sites over all periods.
The proposed 24-hour standard for PMjg is 150 yg/m .
-------
TABLES
Number Page
1 Averaged day and night total emission rates from area and point ...... 5
sources from the Philadelphia emissions inventory
2 Philadelphia Aerosol Field Study monitoring site information ......... 7
3 Observed mixing heights from the 1982 PAFS period .................... 11
4 Deposition and settling velocities for the PEM-2 evaluation runs ..... 12
concentrations at the PAFS monitoring sites ..................... 14
6 Results of 12-hour averaged wind speed and direction and PMio ........ 17
measured at the upwind site during PAFS.
7 Observed and model predicted PMio at upwind sites .................... 18
8 PEM-2 evaluation statistics for PM^ (Set 1) ......................... 20
9 Statistical results for each PAFS site for PMio (Set 1) .............. 23
10 PEM-2 evaluation statistics for PM^ (Set 2) ......................... 24
11 Statistical results for each PAFS site for PMio (Set 2) .............. 26
12 PEM-2 and RAM model results for PM10 ................................. 28
13 Comparison of PEM-2 and RAM for PMio at the PAFS sites ............... 30
14 Comparison of high-five PMio concentrations for PEM-2 and RAM ........ 31
A-l Statistical results of mixing height model evaluation ................ 38
vii
-------
ACKNOWLEDGEMENT
The author expresses his appreciation to K. Shankar Rao (NOAA/ATDD) for
promptly providing the PEM-2 model results for this evaluation effort.
vm
-------
SECTION 1
INTRODUCTION
The proposed National Ambient Air Quality Standards (NAAQS) regulations
will establish a 24-hour standard for particulate matter in the size range
less than 10 micrometers (PMig). Once this short-term standard is promulgated,
state and local regulatory agencies will be required to develop implementation
plans to attain and maintain the new standards. Air quality dispersion models
are expected to be relied upon for urban regulatory applications and emission
control strategies contained in the state implementation plans. EPA has issued
the Guideline on Air Quality Models (Revised) which contains recommended procedures
and models for various applications (EPA, 1986). The evaluation of a dispersion
model with existing measurements is advocated to determine its applicability and
accuracy in modeling particulates in urban environments.
In support of the Agency's policy to regulate urban particulates, the
Atmospheric Sciences Research Laboratory has sponsored the development and
evaluation effort of the Pollution Episodic Model Version 2 (PEM-2). The PEM-2
is an urban-scale Gaussian plume diffusion-deposition model which has been
designed to compute short-term (up to 24-hours) ground-level concentrations of
one or two species of particulate or gaseous pollutants. The model accounts
for the transport, dispersion, and the deposition and settling processes of
particulates from multiple point and/or area emission sources. The technical
features and complete instructions of the PEM-2 model are contained in Rao (1986).
Two evaluation efforts have been conducted on different versions of this model.
Pendergrass and Rao (1984) present the statistical results when the original
version (PEM) was applied to a selected particulate data set from the St. Louis
Regional Air Pollution Study. The current version of the model, designated as
PEM-2,has been recently evaluated against measurements from the Philadelphia
Aerosol Field Study (Ku and Rao, 1986). In this evaluation, separate model
calculations were made for fine particle (< 2.5 urn ) and coarse particle (2.5 -
10 ym) total masses and statistical measures for observed and calculated concen-
trations were determined for each size range. However, there is a need to know
1
-------
the ability of PEM-2 to estimate PMjo concentrations and how its performance
compares to that of other dispersion models.
This report contains the results of various analyses to evaluate the ability
of PEM-2 to model PM^g concentrations. The model results for fine and coarse
particulates from Ku and Rao (1986) were aggregated and statistically evaluated
against corresponding PMjg measurements derived from observed fine and coarse
particulate concentrations from the Philadelphia data base. Statistical measures
of difference and correlation for modeled and measured concentrations paired in
time and space were used to examine model performance. In addition, comparative
results from PEM-2 and RAM, a model recommended in the EPA guideline for urban
applications, are presented. Results of analyzing the regional background
component are also described in order to assess its contribution and role in
urban particulate model evaluation.
MODEL OVERVIEW
The PEM-2 model is designed for short-term urban scale applications since
it is limited to 24 one-hour scenarios and a 50-km domain. It is capable of
calculating ground-level concentrations and deposition fluxes of one or two
particulate or gaseous species over a uniform model receptor grid from a
maximum of 300 point sources and/or 50 area sources.
Since Rao (1986) provides a comprehensive description of the model, only
the notable features that distinguish PEM-2 from other urban Gaussian plume
models are discussed. The important advancements incorporated into the PEM-2
model are the realistic, yet practical methods of simulating dry removal processes
and chemical transformation. Dry deposition and gravitational settling of
particulates are modeled through analytical solutions of the gradient transfer
equations (Rao, 1984). Deposition and settling velocities are specified by the
user. If these processes are not considered, the model equations reduce to the
familiar Gaussian plume diffusion algorithms. The PEM-2 also contains an
option to consider the new plume rise/ penetration methods of Briggs (1984) for
unstable-neutral conditions, however, the standard Briggs plume rise equations
and the all or no penetration scheme are the default methods. In PEM-2, the
contributions from a particular area source are numerically determined to impact
the concentrations at 9 downwind receptor grid cells, which in many cases may
not extend to the model boundary. Otherwise, PEM-2 is consistent with other
Agency models in its treatment of various technical features and processes.
-------
SECTION 2
MODEL EVALUATION DATA BASE
All the measurements necessary to perform a model evaluation for urban
particulates were obtained during the Philadelphia Aerosol Field Study (PAFS).
The PAFS field program was conducted during an intensive 31-day period from
14 July to 14 August 1982 in the metropolitan region of Philadelphia, Penn-
sylvania. Brief descriptions of the emission inventory, particulate measurements,
and meteorological parameter observations used in the model evaluation are
provided since complete details have been documented in the reports cited herein.
EMISSIONS
A comprehensive inventory of fine (FP) and coarse (CP) particulate total
mass emissions was developed specifically for the experimental period. Some
real-time sampling was performed and an effort was made to determine whether import-
ant sources were operating continuously or off-line at times during the study
period. Nevertheless, it is ackowledged that the hourly emissions were primarily
derived from long-term values and should not be construed to represent actual
emissions measurements (Toothman et al., 1985).
The components of the inventory included 300 major point sources, 289 area
sources, gridded mobile sources, gridded minor point sources, and 25 sources
of Industrial Process Fugitive Particulate Emissions (IPFPE). The area, gridded
mobile, and minor point source emissions were combined into a single file for
input to the model because they were constructed on the same 17 X 17 grid with
individual cell sizes of 2.5 km on a side. The IPFPE sources were also incorp-
orated into this emissions grid where each existed, although their grid cell
sizes were either 0.2 km or 0.5 km. The 42.5 km by 42.5 km area emissions grid
is shown inside the larger PEM-2 model calculation domain in Figure 1.
The model grid was enlarged to 80 km by 80 km to accommodate the 50 out of the
300 major point sources located outside the area source grid. A list of the
major point sources reveals that the top 50 sources contributed about 50% of
the point source particulate emissions (Toothman et al., 1985). The tallest
stack in Philadelphia was 152 m, however, most stack heights were under 100 m.
-------
o
ID-
O
•
O
I
I I..-1 !•••
I-: I
I
t 151 -r i
a I
,r
g
M-.KTFl
I I
I I
I
ITT
T
1 T
I
•m.s
164.S
XUTM (km)
£01.5
S24.S
Figure t. Area emissions grid and PEM-2 model calculation grid
for the Philadelphia evaluation. Dashed line denotes
the city limit and the numbered locations are the six
PAFS sites.
-------
Additionally, the point sources were generally distributed along the Delaware
River which is oriented in a nearly NE-SW line through the middle of the model
domain (Figure 2). Table 1 reveals that point sources represent an important
part of the total particulate emissions in both size ranges in Philadelphia.
TABLE 1. AVERAGED DAY AND NIGHT TOTAL EMISSION RATES FROM AREA AND POINT SOURCES
IN THE PHILADELPHIA METROPOLITAN AREA
POLLUTANT AREA EMISSIONS (kg/s) POINT EMISSIONS (kg/s)
DAY NIGHT DAY NIGHT
FP total mass 7.897 3.021 4.269 3.724
CP total mass 5.133 1.668 1.737 1.536
DAY = 12-hour period starting at 0600 EOT
NIGHT = 12-hour period starting at 1800 EOT
Reference - Ku and Rao (1986)
Diurnal curves in Ku and Rao (1986) indicate that area source particulate emissions
were indeed greater than those from point sources during the daytime period.
FP and CP area emissions displayed a significant increase and decrease during
the early morning and late afternoon periods, respectively, which are associated
with large changes in traffic volume and some industrial activity.
PARTICULATE MEASUREMENTS
There were six PAFS monitoring sites equipped with dichotomous filter samplers
which provided continuous FP and CP measurements during the 31-day experimental
study. Information about the instrumentation, data reduction, and quality
control and assurance procedures are contained in PEDCO (1983). The location
of each site is shown in Figure 2 and detailed information is also provided in
Table 2. All 6 sites were located within the urban emissions region and 4
sites (i.e. BRD, FRB, WTP, ARP) were situated inside the city limits. Site BRD
is centrally located in the downtown area. Sites ARP and CLK are the most
distant monitors from the city at 19.7 km and 16.7 km, respectively, with site
BRD's location as the reference position. Site CLK is situated in the most remote
-------
NORRISTOWN
PENNSYLVANIA
PHILADELPHIA / xs
LIMITS
CHEHST HILL
WOODBUST
CLARKSBORO
N E Vf JERSEY
10
kilometers
0 100
KILOMETERS
Figure 2. Map of the Philadelphia metropolitan area and the locations
of the PAFS participate monitoring sites, and Philadelphia
International Airport (PHL) where meteorological observations
were taken.
-------
TABLE 2. PHILADELPHIA AEROSOL FIELD STUDY MONITORING SITE INFORMATION
SITE ABBREVIATION LOCATION UTM* COORDINATES DISTANCEt
(km) (km)
5 BRD Community Health Service Bldg. 4421.35 N
500 South Broad Street 485.81 E
(Urban/commercial)
7 FRB Fireboat Station 4425.17 N 7.0
Allegheny Ave. & Delaware River 491.67 E
(Suburban/industrial)
8 WTP Water Treatment Plant 4427.83 N 8.0
Ford Road & Belmont Ave. 481.20 E
(Suburban/residential)
12 ARP Northeast Airport 4436.02 N 19.7
Grant & Ashton Roads 498.97 E
(Suburban-rural)
28 CAM Inst. for Medical Research 4419.00 N 6.3
Copewood & Davis Streets 491.70 E
Camden, NJ
(Suburban/residential)
34 CLK Shady Lane Home 4405.40 N 16.7
Cohawkin and County House Roads 480.70 E
Clarksboro, NJ
(Suburban-rural)
t Distance is from Site BRD
* UTM - Universal Transverse Mercator northing (N) and easting (E) in Zone 18
7
-------
setting south of the city with nearby rural surroundings (Figure 2).
The FP and CP total mass measurements consisted of 12-hour average concen-
trations. The two averaging periods of 0600-1800 EOT and 1800-0600 EOT are
essentially representative of daytime and nighttime conditions, respectively.
METEOROLOGICAL PARAMETERS
The model requires hourly values of wind speed and direction, temperature,
mixing height, and stability class. Since there were long periods of missing data
from the meteorological instruments at each site, none of these data was used
in this evaluation. Consequently, hourly surface observations made by the
National Weather Service at the Philadelphia International Airport (PHL) located
in the southwest section of the city were obtained for the model evaluation.
It is recognized that these data do not represent hourly averaged conditions,
and a single point measurement greatly simplifies the complex spatial and time
varying flow patterns expected over an urban region. Nevertheless, in many
model applications, such observations are relied upon when quality-assured
on-site meteorological data are lacking.
High resolution upper air temperature and relative humidity measurements
up to 2000 m were obtained by an airsonde system (PEDCO, 1983) which was launched
three times daily at 0400, 1000, and 1600 EOT from 16 July through 14 August at
the CAM site. The height of the lowest elevated temperature inversion base
defined the mixing height, Z-j. Example morning and afternoon profiles are
shown in Figure 3 to illustrate the large variation in Z-j over the daytime period.
In a few cases when no inversion existed (in some afternoon profiles), Z-j was
determined as the base height of a stable, dry layer which was found where the
product of the temperature and humidity gradients was a maximum negative value.
The near-sunrise temperature profile in Figure 3 reveals that an elevated
nocturnal inversion capped the shallow urban boundary layer (UBL), a common
feature of the temperature structure over cities (Godowitch et al., 1985).
However, there was large variability in the low level morning temperature structure
at this site over the experimental period because of its position outside of the
central urban area. When this site was downwind, the UBL was often situated
between a shallow surface-based inversion and an elevated nocturnal inversion.
8
-------
2500
10
,
080382
U 16 18 20
, TEMPERATURE (C) ,
55
60
65
70 RH 75
80
22
,
85
24
,
90
2500
12.5 15.0 17.5 20.0 22.5 25.0 27.5
TEMPERATURE (C) , ,,,,,,
30.0
40
50
60
70 RH 80
90
100
110
Figure 3. Temperature (solid) and humidity (dashed) profiles
for A) before sunrise and B) mid-afternoon on
3 August 1982 at the Camden site.
-------
Thus, early morning profiles at the CAM site exhibited the same features in the
vertical extent and structure of the UBL as suburban locations described by
Godowitch et al. (1987). Since the CAM site was the only upper air site, some
morning urban Z-j's were defined as the height of the elevated UBL top when a
surface inversion existed at this site. It is recognized that Z-j is spatially
variable in the urban area at this time, but a zero Z-j is unrealistic since
surface-based nocturnal inversions are rare in the central city.
Table 3 lists the Z-j observations for each time over the study period. There
were large differences in Z-j as values ranged from 100 m to 898 m at 0400 EOT
and 656 m to 2100 m at 1600 EOT. The observed morning and afternoon mixing
heights, and hourly surface observations were input to the RAMMET meteorological
processor (Turner and Novak, 1978) in order to derive hourly values of l-\ and
stability class required by PEM-2.
The observed Z-j's at 1000 EOT provided an interesting data set for comparison
with derived values from RAMMET and with those computed from an empirical urban
mixing height equation (Godowitch et al., 1985). The statistical results are
presented in Appendix A. A notable result was that RAMMET mixing heights were
generally higher and displayed an overall positive bias when compared to observed
values or those from the empirical equation. However, strong conclusions from
such an evaluation of modeled and non-averaged point measurements must be
tempered because the time period around 1000 is transient and often characterized
by rapid growth and large variations in Z-j over very short periods as the
nocturnal inversion is eroded. Additionally, this data set is a relatively
small sample and limited in extent. Nevertheless, the results provide some
evidence about the performance of these mixing height models for this important
time period.
10
-------
TABLE 3. OBSERVED MIXING HEIGHTS FROM THE 1982 PAFS PERIOD
TIME (EOT)
DATE
7/16
7/17
7/18
7/19
7/20
7/21
7/22
7/23
7/24
7/25
7/26
7/27
7/28
7/29
7/30
7/31
8/01
8/02
8/03
8/04
8/05
8/06
8/07
8/08
8/09
8/10
8/11
8/12
8/13
MEAN
STUDY DAY
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
& STD DEV.
0400
301
119
203
296
172
465
167
134
100 E
114
353
100
100 E
898
100
477
305
112
356
172
199
407
487
372
896
544
539
318
333
331 m + 215
1000
551
556
365
749
354
1001
804
186
780
521
1340
500
893
930
164
1195
361
441
685
991
649
723
550
372
517
850
561
638
374
641 m + 283.
1600
1265
1444
2000 E
1987
1556
1710
1719
988
2022
1968
1967
1873
1851
1570
1273
1709
1929
1977
1762
1831
1944
811
1885
1374
1290
1929
656
2100
1850
1651 m + 382
E - estimated value due to missing data; not used to compute means.
11
-------
SECTION 3
MODEL EVALUATION RESULTS AND DISCUSSION
MODEL CALCULATION PROCEDURES
Minor modifications of the program code were needed prior to the model runs.
The area source array size of 50 was increased to 289 in order to include the
entire area emissions inventory. In addition, code changes were made in the
calculations of the concentrations from point sources. Computations were performed
only at the four grid receptors surrounding each PAFS site. Model run time
was significantly reduced since calculations were made at only 24 of 1024 receptor
points in the 32 X 32 model grid.
The technical features and options chosen for the model runs included; urban
wind profile exponents, stack-tip downwash, new plume rise/penetration methods,
and a constant height of 10 m was specified for all area sources (Ku and Rao, 1986).
The deposition (V^) and gravitational settling (W) velocities for the particulate
species are given in Table 4. These values are believed to be representative
TABLE 4. DRY DEPOSITION AND
Particulate
Range
Fine (FP)
Coarse (CP)
SETTLING
V,
Day
0.2
0.5
VELOCITIES FOR THE PEM-2 EVALUATION RUNS
d (cm/s)
Night
0.1
0.5
W
Day
0.0
0.25
(cm/s)
Night
0.0
0.25
for each size range, however, there is uncertainty since there is a lack of
experimental deposition measurements in urban areas. Nevertheless, these
estimates provide for differences between the time periods and for different
size ranges.
PEM-2 was executed to compute hourly concentrations of FP and CP due to
hourly emissions and hourly meteorological parameters for the 29-day period
from 16 July through 14 August 1982. Modeled concentrations were stored on
12
-------
magnetic tape for post-processing into 12-hour averages. The modeled concentration
at each PAFS site was determined as the weighted average of the four surrounding
gridded receptor values because PEM-2 does not allow for input of non-gridded
receptor coordinates. Although background values may be input to the model,
background concentrations were not considered in the evaluation runs. Consequently,
the modeled concentrations represent the urban source contribution only. The
regional background was accounted for in the evaluation analyses.
REGIONAL PARTICULATE BACKGROUND COMPONENT
The determination of the regional particulate component is expected to be an
important factor of urban PM^g modeling because it may be a relatively large
fraction of the total measured concentration at urban receptor sites. This is
particularly pertinent for Philadelphia because it is one in a string of large
urban areas along the northeast corridor where regional pollutant transport is
an important modeling consideration. The regional background measurement must
be representative of the incoming concentration into an urban domain. Hence, it
should be measured at an upwind site or suitable 'remote1 location that is not
impacted by nearby sources or influenced by emissions from the urban area being
modeled. Unfortunately the PAFS sites, as noted earlier, were all located inside
the Philadelphia emissions area. Thus, each site was impacted to a different
extent based on its relative position within the emissions domain. However,
no other PMjg monitoring sites were available to provide a measure of the regional
background. In the evaluation performed by Ku and Rao (1986), the lowest FP
and CP concentrations at one of the 4 outer sites (BRD and WTP were omitted)
were selected for each 12-hour period as representative of the background
values for each species.
Before describing the method to derive regional background, the actual
differences in the PM^Q concentration among the PAFS sites were found from periods
when measurements were available at all sites. The number of samples varied from
only 36 at the BRD site to 56 at the CAM site. Of the 58 cases between 16 July
and 14 August, the results in Table 5 were determined from 29 periods when all
sites were operational. With the exception of site FRB which exhibited the
highest mean values since it was known to be strongly impacted by nearby fugitive
sources, differences in the means were about 25% or less for the other five sites.
13
-------
TABLE
TOTAL
N
MEAN
jfSD
DAY
N
MEAN
+SD
NIGHT
N
MEAN
+SD
5. PM1Q
BRD
29
46.1
16.9
15
48.1
20.5
14
44.0
12.4
CONCENTRATIONS
FRB
29
65.8
25.1
15
73.3
28.4
14
57.8
18.7
(ug/m3) AT
.--• CTTCC „
WTP
29
45.9
16.7
15
45.5
18.7
14
46.2
15.0
THE PAFS
ARP
29
43.8
16.2
15
43.8
19.8
14
43.9
12.0
MONITORING
CAM
29
41.0
16.9
15
43.1
20.1
14
38.9
13.2
SITES
CLK
29
36.5
14.3
15
37.6
14.4
14
35.3
14.6
N = number of concurrent measurements at all six sites
SD = standard deviation
14
-------
The lowest PM^g concentrations occurred at the CLK site, the station furthest south
of the city. The difference in PMjg concentrations between CLK and an urban site
with the 2nd highest values was only about 10 yg/m^. Mean values far the 12-hour
nocturnal period were slightly lower than daytime values at some sites.
The relative similarity in magnitude and correspondence in the day-to-day
variation in PM^g among these sites suggested that concentrations were strongly
influenced by meteorological conditions and the regional background concentration;
two factors impacting the entire urban domain. The time series of observed and
model predicted concentrations averaged over all sites for each period are
displayed in Figure 4. Clearly, observed PM^g concentrations exhibited much
greater variability and their magnitudes were 2 to 4 times larger than model
results. Furthermore, the modeled PM^g series appears to be uncorrelated with
the observations. Recall that model results were computed with emissions,
meteorological parameters and other physical processes; however, the incoming
concentration at the upwind model boundary (i.e. background) was not included
in the model runs. Therefore, regional particulate background must be accurately
accounted for in the model evaluation analysis process since these results
indicate that it is a large contribution to the observed PM^g concentrations.
The hourly surface wind observations were averaged over 12-hour intervals
corresponding to the time period of the measured concentrations in order to
determine which site was upwind of the city. Table 6 contains the 12-hour
averaged wind speed and direction and their standard deviations over each time
interval. It is evident that there were several cases exhibiting significant
wind direction shifts. Nevertheless, the upwind site was selected based on the
mean wind direction for the period. Sites ARP, CAM, CLK, and WTP were considered
upwind under northerly, easterly, southerly, and westerly wind flows, respectively.
The upwind site and its PM^g measurement are also given in Table 6. Both BRD and
FRB sites were omitted as potential upwind sites due to their central locations
within the city.
The results for each upwind site in Table 7 reveal that southerly and
westerly flows were the most numerous during the study period as the WTP or CLK
sites were most frequently upwind. Furthermore, upwind PMjg concentrations
were higher and more variable under these wind flows. The lowest concentrations
occurred under easterly flows. The modeled PMjg concentrations for each site
are also shown in Table 7. As noted earlier, modeled values were expected to
be different at each upwind site. The CAM site exhibited the highest modeled
concentration of all upwind sites. It was furthest from the upwind model
15
-------
lOOi i i i I
E
01
60
o
o
o
s~
Q-
20
i i i r
j i
I I
10
15
20
25
30
Figure 4. Time series of PM,Q measurements (dashed) and modeled (solid)
concentrations (without background) averaged over all sites
for each 12-hour period of the PAFS study period. Day 1 to
30 corresponds to 16 July to 15 August 1982.
-------
TABLE 6. RESULTS FOR 12-HOUR AVERAGED WIND SPEED AND DIRECTION AND PM10 CONCENTRATIONS
AT THE UPWIND SITE DURING PAFS
PRECIP.
CASES
DATE
Mon/Day
7/16
7/16
7/17
7/17
7/18
7/18
7/19
7/19
7/20
7/20
7/21
7/21
7/22
7/22
7/23
7/23
7/2U
7/21.
7/25
7/25
7/26
7/26
7/27
7/27
7/28
7/28
7/29
7/29
7/30
7/30
7/31
7/31
8/01
8/01
8/02
8/02
8/03
8/03
8/OU
8/01.
8/05
8/05
8/06
8/06
8/07
8/07
8/08
8/08
8/09
8/09
8/10
8/10
8/11
8/11
8/12
8/12
8/13
8/13
START
TIME
(EOT)
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
ws
(m/s)
1.8+0.6
2.U +0.8
3.U+1.3
3.0+0.7
3.9+0.8
3.0+0.7
3.5+0.6
2.7+0.8
I*. 0+1. 2
3.2+0.6
1*.2+0.7
3.2+0.5
I*. 1+0. 9
2.0+0.8
2.7+0.7
3.2+0.7
1*. 5+1.1*
2.9+0.9
U.2+1.5
1*. 0+0.6
3.7+1.1
3.0+0.8
3.0+1.0
3.0+1.1
I*. 6+1.8
3.8+1.0
U.2+0.7
1.8+0.7
3.8+1.1
1.8+0.9
3.1+1.1*
3.2+0.9
3.9+0.6
2.U+0.8
2.7+0.9
3.0+0.7
1*.2+0.5
2.5+1.0
2.9+1.0
3.1+0.8
I*. 0+0. 7
3.1+1.5
5.3+0.5
3.5+0.8
3.5+1.1
3.3+0.7
1*. 1+0.8
3.6+1.0
U.7+0.8
3.3+1.0
1*. 2+1.1*
3.1+1.2
U.8+0.9
3.7+0.8
U.U+.10
2.2+0.6
3.0+0.7
2.6+0.6
WD
(DBG)
225. +2 U.
218 .+19.
231*. +11*.
218. +16.
252. +20.
233. +10.
268. +29.
262. +18.
320. +71*.
3. +23.
9. +2U.
319. +20.
316. +18.
252. +103.
157. +86.
318. +63.
9S.+63.
2U3 .+1*8.
2U6.+19.
226. +1U.
282. +23.
328 .+1*1*.
131*. +92.
169. +59.
181*. +1*9.
313. +10.
313. +20.
281*.+1*5.
186. +2l*.
170. +20.
255. +76.
250. +56.
328. +13.
289. +23.
281*.+1*2.
331.+36.
3. +17.
3UO. +17.
258. +55.
209. +16.
239. +15.
326.+79.
93. +37.
86. +35.
172.+75.
225. +18.
188. +9.
192 .+1*8.
238. +1U.
225. +20.
257. +23.
331*. +3U.
31. +19.
32. +23.
38. +37.
l.+ll*.
332. +33.
288. +21.
SITE
CLK
CLK
CLK
CLK
WTP
CLK
WTP
WTP
WTP
ARP
ARP
WTP
WTP
WTP
CLK
WTP
CAM
CLK
CLK
CLK
WTP
WTP
CAM
CLK
CLK
WTP
WTP
WTP
CLK
CAM
CLK
CLK
WTP
WTP
WTP
WTP
ARP
ARP
ARP
CLK
CLK
WTP
CAM
CAM
CAM
CLK
CLK
CAM
CLK
CLK
WTP
WTP
WTP
CAM
ARP
ARP
ARP
WTP
PM-in
(ug/nr
3U. 1
M
35.2
28.3
57.1
6l.5
55.6
1*3.3
1*0.3
33.2
29.0
25.8
28.2
28.2
39.0
37.2
M
20.7
3U.1
1*5.2
70.5
68.0
50.9
53.1
76.7
31.1
25-5
1*1.2
25.8
32.1
1.7.1
39.9
'1*1.5
38.3
1*6.6
1*7.2
29.8
1*0.1
1*6.1
1*9.6
66.5
51.5
36.6
23.7
2U. 3
16.1*
26.5
3U.2
27.5
27.5
1*1*. 0
1*6.1*
25.3
18.9
29.6
35.6
26.9
31.2
.5
17
-------
boundary when upwind and also impacted by more emissions. In an attempt to
account for these contributions of urban emissions at the upwind site, the
regional background (PM^) was determined by subtracting the predicted concentration
(PMp) from the observed concentration at the upwind site (PMUp) for each period.
Table 7 also reveals that the resultant mean regional background derived with this
method varied from 21.3 yg/m3 at the CAM site to 37.2 yg/m3 at the CLK site.
The overall regional background of 32.4 yg/m3 from all periods compares closely
with 34 yg/m3 determined by Batterman et al. (1986) from the lowest concentration
for each period without consideration of the wind flow. The data base was sufficient
in number to show that the regional particulate background is directionally
dependent. This finding suggests that monitoring sites, strategically-located
around an urban area, may be needed to accurately resolve the background particulate
component for various flow regimes.
SITE
UPWIND
WTP
ARP
CAM
CLK
N -
TABLE 7.
H
22
8
7
19
number of
OBSERVED AND MODEL PREDICTED PM10 AT UPWIND SITES
PMUp
(yg/m3)
41.5+12.6
33.8+6.5
31.5+10.7
39.7+16.2
cases
PMp
(ug/m3)
8.5+3.2
4.4+1.7
10.2+3.1
2.5^3.0
Background (PMb)
(yg/m3)
32.9
29.4
21.3
37.2
32.4
WIND DIR.
(Deg.)
286+62
348+41
126+53
218+28
AND SPEED
(m/s)
3.3
3.3
3.5
3.4
18
-------
MODEL EVALUATION STATISTICAL RESULTS
The statistical measures of difference and correlation between observed and
modeled concentrations paired in time and location described by Fox (1981) were
computed to evaluate model performance. The definitions for these measures and
other statistical parameters are given in Appendix B. Two separate sets of
statistical results were determined in this model evaluation. For Set 1,
observations (Oj) were compared to the corresponding sum of model predictions and
background (i.e. Pi + PMt>). For Set 2, results were obtained for model predictions
(Pi) and corresponding resultant observed values derived by subtracting the
background for each period from the measured concentration (i.e. Oi - PM^).
The evaluation statistics in Table 8 contain the results from all sites for
all periods (total), and separately for day and night cases. In this analysis
where PM& was included with predictions, paired values from the upwind site
were omitted in the sample data set. Although this criterion reduced the sample
size (N), the statistical measures would be artifically improved if upwind
values were included due to the method of determining PM^. Table 8 shows that
PEM-2 slightly underestimated PMjQ concentrations with an overall positive bias
(d) of 5.3 yg/m3. A value of 0.96+0.3 for the ratio of P/0 is favorable.
However, measures of correlation from linear regression analysis depart from
desired values. A relatively large intercept (A) of 23.7, a slope (b) of only
0.41, and a correlation coefficient (R) of 0.56 were determined from all avail-
able paired concentrations. Figure 5 shows that the few cases of high observed
concentrations being greatly underpredicted had a definitive influence on the
linear correlation measures and also effected the difference measures, such as
bias. Nevertheless, a large majority of concentration pairs in Figure 5 are
within a factor of 2 of each other. Several of the statistical results in
Table 8 indicate that PEM-2 performed slightly better for the daytime periods
as revealed by the lower bias, smaller mean squared error (MSE,j), and higher R.
Interestingly, the correlation coefficients for PMjg are ^n between the higher
values for FP and low values for CP in Ku and Rao (1986). However, it must be
remembered that PM^ is a large factor in this set of concentrations pairs and
its strong influence is reflected in these evaluation results.
The statistical results in Table 9 are presented to show how PEM-2 performed
at each site for all periods. As in the previous analysis, concentration pairs
at the upwind site were also excluded in these calculations. A notable outcome
19
-------
TABLE 8. PEM-2 EVALUATION STATISTICS FOR PMio (SET 1)
t
STATISTIC
N
0
P*
Bias (d)
Ml
P*/0
sd
A
b
R
MSEU
MSES
MSEd
MFE
IA
TOTAL
244
49.2+_22.3
43.9+16.4
5.3
12.0
0.96+0.3
19.0
23.7
0.41
0.56
185.1
201.2
386.3
0.7
0.7
P* = Pj + Background
t - nomenclature and equations
Units are yg/m^ for appropriate
DAY
123
51.7+23.9
47.3+18.0
4.4
13.0
1.0+0.3
18.8
22.8
0.47
0.63
193.0
176.1
369.1
0.5
0.8
of statistical measures
measures
NIGHT
121
46.7^20.3
40.4+13.9
6.3
11.0
0.92^0.3
19.2
26.9
0.29
0.42
156.7
247.0
403.8
0.9
0.6
defined in Appendix B
20
-------
f\>
100
80
60
Ou
O
40
o
Ul
az
o.
20
0
i i I
1 I ' '' ' ' b1 ' ' ' I '
— Q
I ' ' ' ' I
I I 1 I
I I I I
o -
111 l_ I I I I I I I I I I I I I I I I I I I I I I I I
I I
25
50
75 100
3M,n Observed
125 150
175
Figure 5. Observed versus model predicted (plus background) PMln concentrations
at all sites over all periods. The proposed 24-hour standard
for PM1Q is 150 ug/m .
-------
is the poor model performance for site FRB, which is highlighted by the large
positive bias and low correlation coefficient, although other statistical
results also deviate strongly from those at the other sites. The fugitive
dust sources are believed to have caused the high CP concentrations at site FRB.
While mean CP concentrations were 50% or less of the mean FP values at the other
sites for the study period, mean CP was slightly greater than FP at site FRB.
It appears that the fugitive emissions may have been greatly underestimated in
the inventory. Additionally, PEM-2 was not designed to treat subgrid scale
sources of varying areal extent that differ from the larger uniform area grid size.
It is evident the high PM^g observations in Figure 5 occurred at site FRB.
When all concentration pairs from site FRB were removed from the data set,
several statistical measures improved dramatically. For example, model bias
was reduced from 5.3 pg/m^ to 0.7 yg/m^, and R increased from 0.56 to 0.74.
Nevertheless, the dominant role of the background concentrations on this
set of evaluation statistics has emerged. The percentages in the bottom line in
Table 9 indicate that the magnitudes of the derived PM^ were about 70% or more
of the measured PM^Q concentrations, except at site FRB. Thus, the magnitude and
variations in PMjo are primarily composed of a large regional background
component at 5 of 6 PAFS sites with a smaller contribution due to urban sources.
Due to the large influence of PM^ on the previous evaluation statistics,
the Set 2 of statistical results were obtained with concentration pairs composed
of model predictions and resultant observed values minus the regional backgrounds.
Several of the statistical measures in Table 10 reveal poor model performance
as results differ largely from their values in Table 8. Although bias and
absolute error d are nearly equivalent in both sets of results, their magnitudes
represent a much greater fraction of the observed and predicted means in Table
10. Furthermore, correlation measures in Table 10 reveal a problem often
encountered when evaluating a data set composed of relatively small values.
Differences between small numbers are inherently magnified. The relatively
large absolute error, near zero slope, and small R in Table 10 also indicate
considerable scatter and little correlation between these paired concentrations.
In fact, the standard deviations are comparable to the mean resultant observed
values. These results also demonstrate the powerful role of the regional particulate
background in this evaluation process. Consequently, model accuracy is difficult
to discern as some statistical results differ based on how the background value
is applied.
22
-------
TABLE
STATISTIC
N
0
P*
Bias (d)
d
P*/0
sd
A
b
R
MSEU
MSEg
MSEd
MFE
IA
PMb/0
P* - PI +
9. STATISTICAL RESULTS FOR EACH PAFS SITE FOR PMjQ (SET 1)
_____ O TT17
BRD
36
45.7+16
50.1+15
-4.5
7.9
1.1+0.
9.2
13.4
0.80
0.83
73.3
28.5
101.8
-0.7
0.9
72%
background
FRB
50
.0 67.4+31.0
.5 44.2+16.4
23.2
24.7
2 0.7+0.3
28.7
29.8
0.21
0.40
222.0
1121.1
1343.1
2.9
0.6
48%
PMh
WTP
31
46. 6+1 8. 6
43.0+18.2
3.6
9.2
0.9+0.2
11.6
6.4
0.78
0.80
114.5
28.6
143.1
0.6
0.9
70%
ARP
44
47.3+15.8
40.1+14.9
7.2
9.8
0.9+0.2
11.0
6.8
0.70
0.74
97.3
72.8
170.2
1.1
0.8
69%
CAM
48
45.4+17.4
48.9+15.8
-3.5
9.1
1.1+0.3
11.7
17.6
0.69
0.76
104.6
40.8
145.4
-0.6
0.8
76%
CLK
35
36.8+13.5
35.7+11.5
1.0
7.5
1.0+0.3
11.5
9.9
0.70
0.66
112.7
16.8
129.5
0.2
0.8
85%
23
-------
PARAMETER
N
0*
P
d
d
P/0*
Sd
A
b
R
MSEU
MSES
MSEd
MFE
IA
where 0* =
TABLE 10. PEM-2 EVALUATION
TOTAL
240
16.8+18.5
10.9+6.7
5.9
11.8
1.4+7.4
18.5
9.8
0.06
0.18
43.3
332.7
376.0
0.5
0.4
0 - PMh
STATISTICS FOR
DAY
120
18.6+18.1
13.2+7.3
5.4
12.5
1.6+8.9
17.9
11.5
0.09
0.23
50.0
297.1
347.1
-0.01
0.4
PM10 (SET 2)
NIGHT
120
15.0+18.7
8.5+5.1
6.4
11.0
1.3+5.6
19.1
8.3
0.01
0.06
25.6
379.4
404.9
1.0
0.3
24
-------
An interesting feature evident in these sets of results is the reversal in
magnitude of the unsystematic (MSEU) and systematic (MSES) mean square errors.
The large MSES values relative to MSEU in Table 10 suggest model modifications
are in order if improvement in model performance is desired. On the other hand,
the large MSEU in Table 8 reveals that errors are also unsystematic, which suggests
modeled results may be the best that can be expected.
The statistical results for each site over all periods in Table 11
substantiate the poor model performance found earlier at site FRB. In particular,
correlation coefficients are considerably lower, and even negative for site CLK,
compared with those in Table 9. Distinctive features that appeared in both sets of
results are overpredictions at sites BRD and CAM, and relatively large under-
predictions at sites ARP and CLK, the most distant monitors from the city.
These particular results are examined further in the next section where PEM-2
results are compared to RAM model predictions.
COMPARATIVE RESULTS FOR THE PEM-2 AND RAM MODELS
Both PEM-2 and RAM are Gaussian plume models with many of the same technical
features and options. However, there are a few important differences between
these models, which may yield variations in estimated concentrations. A relevant
factor is that PEM-2 accounts for dry deposition and gravitational settling,
while these processes cannot be considered by the RAM model. Another diff-
erence between these models is in the treatment of area emissions and their
source heights. PEM-2 computes contributions from no more than eight upwind area
cells to the concentration at a given receptor, while RAM can consider the
impact at a particular receptor from all upwind area sources. Additionally,
a single area source height can be specified in the PEM-2 runs. In contrast,
RAM allows for input of different area source heights. Both models deal with
area sources emissions according to the widely used narrow plume hypothesis.
The details of the RAM evaluation effort and results are documented in
Anderson et al. (1986). Both models were executed with the same emissions
data base, and hourly meteorological parameters, although there may be small
differences in Z-j since Anderson et al. (1986) obtained observed values from
a nearby rawinsonde site instead of using the airsonde data. Options
for the RAM runs were set based on regulatory recommendations (EPA,1986).
25
-------
TABLE 11.
STATISTIC
N
0*
P
d
Ml
P/O*
Sd
A
b
R
MSEU
MSES
MSEd
MFE
IA
where 0* =
STATISTICAL RESULTS FOR
RRD
36
12.9+9.
17.2+6.
-4.3
7.9
3.4+9.
9.2
14.1
0.24
0.38
31.4
70.5
101.8
-1.4
0.6
Oi - PMh
FRB
50
6 35.0+28.6
2 11.8+5.8'
23.2
24.7
6 0.6^0.8
28.7
11.2
0.02
0.08
32.2
1310.9
1343.1
4.0
0.4
EACH PAFS
CTTF
WTP
30
14.4+12.7
9.9+4.9
4.5
8.8
0.5+3.4
10.8
6.8
0.22
0.56
16.2
115.7
131.9
0.8
0.6
SITE FOR
ARP
44
14.5+11
7.3+4.
7.2
9.8
-.1+4.
11.0
6.2
0.08
0.21
16.8
153.3
170.1
1.5
0.5
PMio (SET 2)
CAM
47
.1 11.0+9.7
2 13.8+6.6
-2.8
8.5
1 3.8^13.
10.6
12.4
0.13
0.19
41.0
77.7
118.7
-1.3
0.4
CLK
33
6.9^7.4
4.0+_3.0
2.9
6.2
3 0.0_+1.8
8.2
4.2
-0.03
-0.07
8.6
64.2
72.8
-1.8
0.3
26
-------
A single area source height of 10 m was input in the PEM-2 model runs. In
contrast, one of three possible source heights was specified for area sources
in RAM; namely, 13.7 m, 9.1 m, or 4.6 m, were assigned to area sources based on
emission rate (Anderson et al., 1986). The grid size for both models was 2.5
km. This means the contributions from area sources beyond 20 km from a site
were not considered by PEM-2 in the concentration calculations for a given
site. All area sources upwind of a site were included in RAM.
The statistical results from evaluations of both models are presented
in Table 12. The RAM results are taken from Anderson et al. (1986). Their
background values were added to the PEM-2 model results in this phase of the
analysis to make the direct comparison with RAM possible. The sample sizes
differ slightly between the model results because a 31-day period starting on
14 July was considered in the RAM evaluation. The common feature found from
both sets of statistics in Table 12 is that PEM-2 predictions were consistently
lower than RAM's results in the mean and peak concentrations. In the set where
modeled predictions are considered alone, the PEM-2 mean value of 10.1 yg/m3
is 3.1 yg/m3 less, or about 75% of RAM's mean of 13.2 yg/m3.
The model differences described earlier are believed responsible for the
different predictions. Modeled concentrations by PEM-2 are reduced due to loss
by deposition processes. RAM's consideration of more upwind areas sources and
lower source heights for some area sources, particularly those grid cells outside
the city, compared a uniform 10 m height for PEM-2 also contributed to higher
concentrations since PEM-2 computations are limited to eight upwind area grids.
The relative similarity in results in Table 12 for both models make it
difficult to state which model is superior or is more acccurate. While the
correlation measures for PEM-2 are slightly better than the RAM results, the
large positive biases reveal both models greatly underpredicted observed values.
This indicates that attention should be focused on the background values derived
by Anderson et al. (1986) because of the large contribution of PM^ to the total
measured concentration described herein and by Batterman et al. (1986) with this
PAFS data base. A mean PMb of 20.5 yg/m3 in Anderson et al. (1986) compares
with 32.4 yg/m3 presented in Table 7. The difference between these values is
large enough to account for the significant model bias found in Table 12.
Anderson et al. (1986) derived the regional particulate component from the
PAFS data after analyses of the lowest concentrations obtained every sixth day
27
-------
TABLE
PARAMETER
N
0
P
d
A
b
R
MAX 0
MAX P
2nd High
0
P
(1 ) 0^ versus
(2) Pi versus
12. PEM-2 AND
PEM-2
(D
305
47.4
30.6
16.8
17.6
0.27
0.56
161.2
58.3
101.8
57.9
P! + PMjj from
Oi - PM>, from
RAM MODEL RESULTS FOR
RAMT
(D
318
47.5
33.8
13.7
21.6
0.26
0.52
161.2
71.3
101.8
63.1
Anderson et
Anderson et
PEM-2
(2)
305
26.9
10.1
16.8
8.2
0.07
0.19
144.5
35.4
76.1
35.4
al. (1986)
al. (1986)
PM10
RAMt
(2)
318
26.9
13.2
13.7
11.7
0.06
0.13
144.5
49.1
76.1
43.0
t Results from Anderson et al. (1986)
28
-------
for 1982 at one of 12 designated IP sites in a network surrounding Philadelphia.
Measurements from this network were not utilized in the PEM-2 evaluations.
However, different methods produced the large differences in PM^, not the avail-
ability of additional measurement sites.
Separate statistical results are given in Table 13 to show model differences
for each PAFS site. Both models performed poorly for site FRB. There were
also large positive biases for both models at individual sites. PEM-2 results
were lower than RAM predictions at each site. On the other hand, correlation
coefficients and linear regression measures were slightly better for PEM-2
at most sites.
An interesting feature to be explored is the difference in model predictions
for the various sites. Part 2 of Table 13 shows the model results differed by
5.2 pg/m3 at site BRD to 2.6 ug/m3 at site CLK. However, a more revealing
result when comparing these model predictions is the ratio of this difference
to the RAM prediction (i.e. PR/W - PPEM / PRAM) at eacn site. Interestingly,
the percentage of this ratio was about 20% for BRD site and at the 3 sites closest
to it, but results jumped to 35% and 40% at sites ARP and CLK, respectively.
These latter two sites are most distant from site BRD. It appears that differences
between the models were accentuated when approaching the boundary of the model
domain, which in this case is farthest from the city and important emissions.
Further investigation of the models is required to determine the relative
importance of the different factors in producing these differences. Nevertheless,
it is clear that PEM-2 predicts lower PM^g concentrations than RAM, and the
differences increased away from the city and also farther from the major emission
sources.
Of particular relevance in regulatory applications is how a model simulates
the highest concentrations since these are the values which may exceed a pollutant
standard and provide the basis for the design concentration upon which a control
strategy is implemented. Table 14 contains the high-5 modeled concentrations
for both PEM-2 and RAM. The results are for predicted concentrations plus back-
grounds from Anderson et al. and predictions without background. Results for
these peak concentrations are similar to those obtained from the mean concentration
results. PEM-2 predictions were almost always lower than RAM's results. The
values of the ratio PEM-2/RAM were lowest at the 2 outermost sites, ARP and
CLK. Both models significantly underpredicted the peak observed concentrations
even when the regional background values of Anderson et al. (1986) were used.
The PMb values derived herein were generally much larger than those determined
by Anderson et al., and their use would have eliminated the large positive bias.
29
-------
TABLE 13. COMPARISION OF PEM-2 AND RAM FOR PM AT THE PAFS SITES
SET (1) P + PMb from Anderson et al . (1986)
----------------------------- s ITE --------------------------------
BRD FRB WTP ARP CAM CLK
PARAMETER PEM-2 RAM PEM-2 RAM PEM-2 RAM PEM-2 RAM PEM-2 RAM PEM-2 RAM
N 36 37 51 53 55 57 53 55 56 59 54 57
0 45.7 45.7 66.8 66.8 44.9 45.4 45.8 46.3 44.0 43.7 37.8 37.9
P 37.3 42.4 31.9 34.8 30.2 32.5 27.4 31.2 34.4 37.8 24.4 27.1
d 8.4 3.3 34.9 32.0 14.7 12.9 18.5 15.1 9.5 5.9 13.4 10.8
A 16.3 21.4 24.7 28.8 8.5 14.8 7.3 8.6 16.0 19.6 7.6 11.7
b .46 .46 .11 .09 .48 .39 .44 .49 .42 .42 .45 .40
R .75 .68 .33 .30 .84 .73 .74 .76 .68 .68 .83 .72
SET (2) 0 - PMb from Anderson et al. (1986)
0 25.6 25.7 46.6 46.3 24.1 24.5 25.3 25.6 23.2 22.9 17.2 17.4
P 17.2 22.4 11.8 14.4 9.4 11.6 6.9 10.6 13.7 17.0 3.9 6.5
P Diff 5.2 2.6 2.2 3.7 3.3 2.6
A 14.4 19.2 11.4 14.4 4.9 10.7 4.8 5.9 13.4 15.9 5.6 8.3
b 0.110.12 0.01-0.00 0.190.04 0.080.18 0.01 0.05 -. 10 -. 10
R 0.18 0.16 0.04 -0.00 0.49 0.10 0.22 0.37 0.02 0.08 -.21 -.14
P Diff = RAM - PEM-2 prediction
30
-------
TABLE 14. COMPARISON OF THE HIGH-FIVE PM10 CONCENTRATIONS FOR PEM-2 AND
BRD FRB WTP ARP CAM
MAXIMUM
PEM-2
RAM
2ND HIGH
PEM-2
RAM
3RD HIGH
PEM-2
RAM
4TH HIGH
PEM-2
RAM
5TH HIGH
PEM-2
RAM
HIGH-5 AVE.
PEM-2
RAM
RATIO
PEM-2/RAM
HIGH-5 Obs.
(1)
55.2
71.3
53.1
64.6
50.6
63.4
49.5
61.7
48.9
60.3
(2)
30.9
49.1
29.8
37.1
27.3
33.4
26.3
32.8
25.0
31.6
(1) (2)
58.3 28.4
63.1 40.9
51.0 27.8
60.9 26.1
49.4 24.2
59.9 25.3
48.2 21.8
56.1 24.5
46.8 19.8
50.9 24.3
51.5 27.9 50.7 24.4
64.3 36.8 58.2 28.2
0.80 0.76 0.87 0.86
AVE. 73.7 44.1 140.3 114.9
(1) P + background
(2) P without background added
(1)
57.9
57.8
51.8
51.0
48.5
49.9
45.1
49.4
44.7
48.1
49.6
51.2
0.97
80.5
(2)
24.7
27.7
19.4
27.0
17.0
22.4
16.0
22.2
15.5
21.9
18.5
24.2
0.76
51.3
(1)
51.8
62.2
48.7
56.4
46.2
56.1
45.7
51.2
41.2
47.5
46.7
54.7
0.85
79.4
(2)
20.8
29.0
16.1
26.8
15.7
23.4
14.1
21.1
13.3
20.8
16.0
24.2
0.66
50.8
(1) (2)
57.0 35.4
63.1 43.0
54.1 28.0
61.9 33.6
52.6 25.9
55.8 32.1
52.2 25.3
54.8 29.6
50.4 24.9
53.5 27.5
53.3 27.9
57.8 33.2
0.92 0.84
80.0 47.5
RAM
CLK
(ll (2)
40.6 15.0
45.7 23.5
39.4 14.7
44.2 21.0
37.8 13.7
41.4 20.4
37.3 11.5
39.9 20.6
36.8 10.3
39.4 18.3
38.4
42.1
0.91
70.4
13.0
20.8
0.62
36.9
6-SITE
AVE.
(2)
25.9
35.5
22.6
28.6
20.6
26.2
19.2
25.1
18.1
24.1
21.3
27.9
0.76
-------
SECTION 4
SUMMARY AND CONCLUSIONS
An evaluation of the PEM-2 urban dispersion-deposition model has been performed
with measurements from the 1982 Philadelphia Aerosol Field Study data base.
Observed and model predicted 12-hour PMig concentration pairs were statistically
evaluated to determine model performance over a 29-day study period. Analysis
of the observed concentrations from 6 monitoring sites indicated the regional
particulate background was a large contributor to urban PMjg concentrations in
Philadelphia, as the mean PMig concentrations from concurrent measurements at
these sites differed by only 10 iig/m^. Consequently, the method to derive regional
background values consisted of consideration of the wind flow and contribution
from urban emissions at the upwind site. Regional background component was
found to be about 70% or more of the measured PMig concentration at 5 of 6
sites. The remaining urban site was strongly impacted by local fugitive emissions.
Statistical results'were computed from two sets of concentration pairs;
modeled plus background versus observed, and modeled only versus observed minus
background. Better performance was obtained when background values were added
to model predictions; however, these results were greatly influenced by the
large regional background values. For the other set, large errors and little
correlation was found from the modeled and resultant observed values minus
background since magnitudes were relatively small. Therefore, model performance
and accuracy was difficult to judge because of the dominant role of background
for this data set.
A comparison of PEM-2 and RAM model results was conducted by applying the
regional background values derived for the RAM evaluation. Results revealed
that PEM-2 predictions were lower than RAM concentrations both for mean and
peak values. The factors responsible for lower PEM-2 concentrations include;
loss of mass by dry deposition and gravitational settling processes simulated
by PEM-2 which were not treated in RAM, and different methods to handle area
source emissions and area heights. PEM-2 considers the contributions from up
32
-------
to 8 upwind area sources, while all upwind area sources were included in
computations by RAM. The relative importance of these factors was not assessed,
but deserves further investigation.
Large positive biases in the comparision of PEM-2 and RAM results may
be explained by regional backgrounds too low due to the method used by Anderson
et al. (1986). The PM^g background concentrations derived for and used in
the PEM-2 evaluation were considerably larger than those computed by Anderson
et al. (1986). The latter were used in the RAM model evaluation and were also
added to the PEM-2 results for the purpose of the model comparison. Otherwise,
the strong similarity in the statistical results revealed little evidence to
distinguish between the performances of these models. Model accuracy was also
difficult to assess from the evaluation statistics due to the dominant role of
the regional background component. Future evaluations are necessary to adequately
assess the impact of deposition processes for urban particulate applications.
33
-------
REFERENCES
Anderson, M. K., R.T. DeCesar, E.T. Brookman, J. A. Foster, R. J. Londergan,
1986: Example modeling to illustrate SIP development for the new part-
iculate matter NAAQS. Draft Report to Office of Air Quality Planning and
Standards, U. S. Enviromental Protection Agency. Contract No. 68-02-3886.
Batterman, S., J. A. Fay, and D. Golumb, 1986: Local and regional contributions
to urban particulate matter. U. S. Environmental Protection Agency,
EPA/600/3-86/052. Available from NTIS, Springfield, VA 22161. NTIS No.
PB86-236965.
Briggs, G. A., 1984: Plume rise and bouyancy effects. Atmospheric Science
and Power Production. D. Randerson, Ed. DOE/TIC-27601, Tech. Info. Center,
Oak Ridge, TN, Chapter 8, 327-366.
Environmental Protection Agency, 1986: Guideline on air quality models (Revised).
EPA Publication No. EPA/450/2-78/027R. Available from NTIS, Springfield, VA
22161, NTIS No. PB86-245248.
Fox, D. G., 1981: Judging air quality model performance. Bull. Amer. Meteorol.
Soc.. 62(5), 599-609.
Godowitch, J. M., J.K.S. Ching, J.F. Clarke, 1985: Evolution of the nocturnal
inversion layer at an urban and nonurban site. J. dim, and Appl. Meteorol.,
24, 791-804.
Godowitch, J. M., J.K.S. Ching, J.F. Clarke, 1987: Spatial variation of the
evolution and structure of the urban boundary layer. Boundary-Layer
Meteorol.,38. 249-272.
Ku, J. Y., and K. S. Rao, 1986: Evaluation of the PEM-2 using the Philadelphia
Aerosol Field Study data base. U.S. Environmental Protection Agency,
EPA/600/8-86/016. Available from NTIS, Springfield, VA 22161, NTIS No.
PB86-167921.
PEDCO, 1983: The 1982 Philadelphia Aerosol Field Study; Data Collection Report.
Atmospheric Sciences Research Laboratory, U.S. Environmental Protection
Agency, Contract No. 68-02-3496. Research Triangle Park, NC 27711.
Pendergrass, W. R., and K. S. Rao, 1984: Evaluation of the Pollution Episodic
Model using the RAPS data. U. S. Environmental Protection Agency Publication
EPA/600/8-84/087. Available from NTIS, Springfield, VA 22161. NTIS No.
PB84-232537.
34
-------
Rao, K. S., 1984: Plume concentration algorithms with deposition, sedimentation
and chemical transformation. U. S. Environmental Protection Agency,
EPA/600/8-84/042. Available from NTIS, Springfield, VA 22161. NTIS No.
PB84-138742.
Rao, K. S., 1986: User's guide for PEM-2: Pollution Episodic Model (Version 2).
U.S. Environmental Protection Agency. EPA/600/8-86/040. Available from
NTIS, Springfield, VA 22161. NTIS No. PB87-132098.
Toothman, D. A., J.C. Thames, J.C. Yates, R.R. Segall, and J.N. Bolstad, 1985:
An emission inventory for urban particle validation in the Philadelphia
AQCR. U. S. Environmental Protection Agency. EPA/600/3-85/041. Available
from NTIS, Springfield, VA 22161. NTIS No. PB85-207611.
Turner, D. B., J. N. Novak, 1978: User's guide for RAM. U. S. Environmental
Protection Agency. EPA/600/8-78/016a. Available from NTIS, Springfield,
VA 22161. NTIS No. PB29491 and PB294792.
35
-------
APPENDIX A
EVALUATION OF MODELED MIXING HEIGHTS WITH PAFS OBSERVATIONS
Mixing heights derived from two different methods were statistically
evaluated against observed values at 1000 LOT. It is acknowledged that the
sample size of 29 is relatively small and the evaluation is limited in extent
to a particular hour. However, this effort was undertaken with the intentions
of obtaining some valuable evidence to distinguish between the two methods
and to determine how accurately Z-j can be simulated for this rapidly changing
time period.
Two sets of urban Z-j's were computed; from the RAMMET pre-processor program
and by an empirical equation developed by Godowitch et al. (1985). The hourly
mixing heights for the PEM-2 evaluation were provided by RAMMET. It's technique
is to derive I\ by linear interpolation between local sunrise and 1400 LST with
observed morning minimum and afternoon maximum values. The other method is an
empirical mixing heights growth equation derived from a second order least square
regression fit to a large data set from St. Louis, Missouri. The empirical urban
Z-j growth equation, in a slightly modified form from Godowitch et al. (1985),
is given by
Z^t) = Z^o) - 7.0t + 24.3 t2 (A-l)
where Zi(o) is the urban mixing height at sunrise for a particular day and t is
in hours after sunrise. Equation (A-l) is only applicable between sunrise and
the time when the nocturnal inversion is destroyed, which usually occurs around
mid-morning. According to Equation A-l, urban mixing height growth is distinctly
nonlinear during this time period. The destruction of the nocturnal inversion
layer marks the end of the morning transition period. It is followed by a
relatively short period of more rapid, even explosive, growth that often occurs
through a thicker overlying layer of much weaker stability until the mixing
height of the previous afternoon is reached. It is likely that the observations
at 1000 LOT were obtained under both stages, as the mean time of complete
inversion destruction in St. Louis was around this hour.
The results of statistical analysis for both methods are shown in Table A-l.
A large negative bias was found for the RAMMET model. In fact, it's values
overpredicted observations in 24 of 29 cases. The results from Equation (A-l)
showed more balance with a small negative bias. There were 17 overpredictions
36
-------
and 12 underpredicitons with this model. In addition, the results for the
average absolute error and mean P/0 were more favorable with Equation (A-l).
On the otherhand, the linear regression measures of slope and intercept suggested
a slightly better correlation between RAMMET and observed pairs. However, the
total mean square error (MSEj) in the RAMMET results is dominated by a large
systematic component (MSES), which suggests model revision is necessary if
performance is to be improved.
Thus, even though a mixed picture appears when considering all the
statistical measures, the results for Equation (A-l) are better than RAMMET's
for most of the statistical parameters. The large standard deviations for the
observations demonstrates the amount of variability in Z-j during this time of
rapid change. Nevertheless, a tentative conclusion from this limited evaluation
data set is that RAMMET often overpredicts urban mixing heights at this time period.
Additional evaluations with larger data sets are needed to confirm these results.
37
-------
TABLE A-l. STATISTICAL RESULTS OF MIXING HEIGHT MODEL EVALUATIONS
PARAMETER
METHOD
RAMMET
EQUATION (A-l)
N
0
P
d
|d|
P/0
A
b
R
MSEu/MSEd
MSEs/MSEd
29
641.4 +_ 282.8
917.4 +_ 182.3
-276.0
313.8
1.7 +_ 0.7
711.6
0.32
0.50
17.8%
82.2%
29
641.4 +_ 282.8
675.9 +_ 215.5
-34.5
240.6
1.3 +_ 0.6
538.0
0.22
0.28
45.8%
54.2%
38
-------
APPENDIX 8
DEFINITIONS OF STATISTICAL MEASURES
The nomenclature and equations for the various statistical measures used to
determine model performance are given. The measures of difference and correlation
are computed from N samples of individual observations (0-j) and model predictions
(P-j). Means and standard deviations for observed and predicted values were
computed from standard formulas. Measures of difference are based on residuals
between observed and predicted concentrations where;
Bias di = oi - Pn- (B-l)
d" = - Z di (B-2)
. N 1
average absolute gross error |d"| = - Z |dj| (B-3)
N
standard deviation of bias Sd = [- z (dj-d)2]1'2 (B-4)
Measures of correlation between 0^ and Pi from linear regression analysis
include the intercept (A), slope of a line (b) and correlation coefficient (R).
These measures are computed by;
I (OJX P'i)
b ' —2 - (B-5)
where 0'^ =0^ - 0 and P'^ = P1 - P, the differences between observed and
predicted values and their means.
a = P" - bO" (B-6)
\ X P\)
R = - (B-7)
39
-------
The estimate of a prediction PE] is given by
PEj = a + b 01 (B-8)
The unsystematic (MSE ^ and systematic (MSE 5) components of the mean square
error are given by
MSEU = - KPi - PEi)2 and (B-9)
MSES = - ^(PEi - Oi)2, respectively. (B-10)
The mean square error of the differences (MSE,j) is computed from;
MSEd = - Edi2 (B-ll)
N
Mean fraction error (MFE) is an indicator of the model's overall bias to
underpredict or overpredict.
MFE = - I - (B-12)
N
The index of agreement (IA) is given by;
1 - Z dn-2
IA = - (B-13)
2 C|Pi - oT+ |o^|]2
40
-------
APPENDIX C
OBSERVED AND PEM-2 MODEL PM10 CONCENTRATIONS
A listing of corresponding 12-hour average observed (0) and predicted (P)
concentrations at each site is given for the 29-day period from the PAFS program.
The model evaluation period for the PEM-2 study was from 16 July through 13 August
1982. Missing values are designated by 999. The time (local daylight time, LOT)
represents the starting hour of the averaging period.
41
-------
DATE TIME BRD FRB WTP ARP CAM CLK
(M/D) (LOT) PO P 0 P 0 PO PO PO
7/16 06 28.6 999.0 17.4 999.0 19.4 999.0 6.9 999.0 14.2 57.3 1.2 34.1
7/16 18 27.6 999.0 18.6 999.0 16.4 72.6 11.6 999.0 21.6 61.3 2.5 999.0
7/17 06 26.5 999.0 27.8 57.2 24.7 96.5 20.8 60.8 28.0 52.3 1.5 35.2
7/17 18 16.2 52.8 13.2 38.7 6.5 45.9 8.1 40.7 15.6 29.8 1.5 28.3
7/18 06 20.1 999.0 18.0 75.4 15.5 57.1 15.7 70.2 19.6 71.4 1.1 65.5
7/18 18 8.4 999.0 6.4 999.0 4.4 71.7 4.7 88.3 11.4 62.0 1.2 61.5
7/19 06 23.5 51.2 19.2 999.0 12.6 55.6 16.1 65.7 22.6 56.8 1.6 56.7
7/19 18 13.3 51.0 7.9 999.0 6.3 43.3 5.0 55.7 9.7 45.5 2.4 40.5
7/20 06 23.3 46.7 21.8 56.0 8.8 40.3 14.0 44.6 23.1 43.3 4.0 38.4
7/20 18 21.6 31.4 12.8 33.9 10.4 39.1 6.5 33.2 14.0 24.4 4.8 26.0
7/21 06 16.5 27.8 10.3 39.1 9.5 28.2 3.6 29.0 14.0 17.7 8.0 23.1
7/21 18 13.0 999.0 5.8 40.9 5.8 25.8 2.9 69.3 7.6 47.6 2.7 22.4
7/22 06 14.7 999.0 12.4 82.9 8.0 28.2 4.2 35.9 15.6 36.8 10.3 27.2
7/22 18 13.5 999.0 7.2 60.1 6.7 28.2 3.5 48.8 9.6 59.4 2.9 51.4
7/23 06 26.3 62.0 24.2 67.8 15.6 999.0 7.3 50.5 25.3 47.7 14.7 39.0
7/23 18 27.3 47.4 14.1 43.1 16.0 37.2 9.5 40.3 14.7 999.0 2.8 999.0
7/24 06 16.7 999.0 12.6 34.6 7.8 28.9 5.8 45.6 15.4 999.0 9.6 999.0
7/24 18 9.0 999.0 5.0 39.7 5.4 34.4 2.9 41.8 5.3 28.5 1.5 20.7
7/25 06 16.7 999.0 16.4 51.1 15.4 35.5 12.4 63.1 16.8 46.9 9.6 34.1
-------
DATE TIKE BRD FRB HTP ARP CAN CLK
(tVD) (LOT) PO PO PO PO PO PO
7/25 IB 7.1 53.7 5.7 70.4 3.2 59.1 3.3 65.7 8.9 50.Z 1.1 45.2
7/26 06 12.8 67.6 12.7 139.5 8.0 70.5 9.5 72.6 13.7 80.4 1.2 63.5
7/26 18 13.2 61.9 7.1 63.3 6.3 68.0 4.8 58.8 9.1 59.1 1.5 69.6
7/27 06 16.7 54.6 11.0 71.8 9.0 46.3 5.0 43.7 14.3 50.9 7.0 44.2
7/27 18 16.2 999.0 8.2 55.8 8.2 53.4 5.1 52.9 8.3 59.6 1.8 53.1
7/28 06 11.0 999.0 7.3 132.7 7.4 61.3 3.8 999.0 6.8 101.8 .9 76.7
7/28 18 19.1 999.0 11.3 161.2 13.0 31.1 8.0 34.2 11.9 36.6 1.5 28.8
7/29 06 15.4 27.8 11.4 38.2 7.5 25.5 3.7 22.7 19.7 22.2 6,4 16.5
-P*
00 7/29 18 14.2 32.3 7.3 99.0 6.8 41.2 3.5 37.2 10.8 31.9 2.5 20.6
7/30 06 30.9 46.8 28.4 78.4 17.0 45.7 12.5 35.9 35.4 29.5 11.5 25.8
7/30 18 17.3 40.2 11.6 999.0 11.3 36.8 9.0 24.9 10.2 32.1 .7 30.4
7/31 06 19.0 999.0 10.5 64.9 10.8 47.3 5.5 53.9 10.1 50.4 1.1 47.1
7/31 18 20.2 54.9 11.7 51.4 10.1 62.9 6.5 55.3 14.0 48.9 4.6 39.9
8/01 06 10.5 - 44.8 10.7 53.3 6.8 41.5 5.7 33.3 11.8 36.7 4.5 38.1
8/01 18 5.8 41.3 3.9 46.8 3.1 38.3 1.6 43.4 5.2 44.4 2.0 33.4
8/02 06 22.9 48.1 19.8 75.2 10.4 46.6 9.8 42.9 24.9 54.5 8.0 44.9
8/02 18 18.0 999.0 9.1 141.5 8.4 47.2 6.2 37.5 12.8 47.3 3.6 43.4
8/03 06 16.0 999.0 10.9 126.7 7.8 39.6 5.0 29.8 14.7 37.3 7.5 39.3
8/03 18 13.1 999.0 5.9 64.0 6.2 41.8 3.0 40.1 7.5 58.4 2.5 43.5
8/04 06 25.0 76.6 16.1 82.0 14.8 70.9 5.3 77.7 23.1 58.1 15.0 44.2
-------
DATE
(M/D)
8/04
8/05
8/05
8/06
8/06
8/07
8/07
8/08
8/08
8/09
8/09
8/10
8/10
8/11
8/11
8/12
8/12
8/13
8/13
TIME
(LOT)
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
06
18
BRD
P 0
19.3
16.7
20.0
13.0
19.1
21.9
9.0
14.2
5.2
8.5
14.8
11.8
12.5
13.7
17.3
8.4
16.5
29.8
17.9
58.7
100.2
57.0
47.0
25.8
999.0
31.2
26.3
999.0
999.0
999.0
46.2
45.2
29.7
28.3
33.8
37.9
27.3
31.3
FRB
P
11.0
19.8
15.5
10.2
8.6
11.2
5.4
15.7
4.8
5.8
10.7
12.1
7.9
9.2
7.9
4.8
7.3
19.2
8.5
0
64.1
106.5
57.4
72.0
31.1
35.9
85.0
29.7
999.0
41.4
41.9
91.9
53.3
53.8
23.6
72.9
53.0
92.5
61.6
HTP
P
10.8
19.4
10.2
8.1
12.7
12.3
5.2
10.7
4.1
5.1
8.7
7.1
6.1
6.4
8.3
3.7
7.9
14.9
8.7
0
72.6
89.2
51.5
49.9
30.6
38.1
21.6
31.6
36.3
40.7
999.0
44.0
46.4
25.3
26.8
27.6
38.6
25.1
31.2
ARP
P
7.3
13.3
10.1
7.6
4.4
3.7
3.5
12.0
2.8
3.4
8.7
8.9
5.5
4.0
4.1
1.7
3.5
6.5
4.8
0
58.7
88.1
48.8
42.6
28.9
26.6
26.3
30.1
31.8
49.7
40.4
36.5
44.6
999.0
999.0
29.6
35.6
26.9
37.9
CAM
P
11.9
15.6
16.2
12.7
10.0
10.2
6.0
15.1
4.2
6.2
13.2
13.3
9.7
13.0
9.8
5.7
7.9
25.9
11.2
0
59.7
84.5
44.2
36.6
23.7
24.3
16.0
33.5
34.2
34.9
38.3
48.0
41.4
24.8
18.9
28.0
35.4
22.2
35.2
CLK
P
2.6
.9
1.6
3.4
2.5
5.4
.8
1.2
0
49.6
66.5
48.3
39.3
20.9
27.8
16.4
26.5
.3 999.0
.9
.9
.9
1.2
6.1
4.7
3.3
2.8
13.7
3.2
27.5
27.5
43.1
40.8
28.0
21.7
24.7
28.8
25.5
25.8
------- |