oEPA
United States
Environmental Protection
Agency
Office of Air Quality
Planning and Standards
Research Triangle Park NC 27711
EPA-450/4-84-004
December 1983
Air
A Review
Of Recent
Applications of
The SAI Urban
Airshed Model
-------
EPA-450/4-84-004
A Review of Recent Applications
Of the SAI Urban Airshed Model
David E. Lay land
Henry S. Cole*
Presently with
Clean Water Action Project
733 15th Street, N.W.
Washington, D.C. 20005
Monitoring and Data Analysis Division
Office of Air Quality Planning and Standards
U.S. Environmental Protection Agency
Research Triangle Park, North Carolina 27711
December 1983
-------
DISCLAIMER
This report has been reviewed by the Office of Air Quality Planning
and Standards, EPA, and approved for publication. Mention of trade names
or commercial products is not intended to constitute endorsement or
recommendation for use.
ii
-------
TABLE OF CONTENTS
Page
List of Figures iv
1. INTRODUCTION 1
2. DESCRIPTION OF THE MODEL 2
3. DESCRIPTION OF THE STUDIES 5
3.1 Tulsa 5
3.2 Denver 6
3.3 St. Louis 8
3.4 Los Angeles 9
4. EVALUATION OF MODEL PERFORMANCE 12
4.1 Peak Accuracy 14
4.2 Overall Accuracy 18
4.3 Spatial and Temporal Patterns 31
4.4 Conclusions 34
5. MODEL SENSITIVITY 36
5.1 Background Air Quality 36
5.2 Solar Radiation 37
5.3 Vertical Mixing 38
5.4 Emissions Composition 41
5.5 Interaction with Emission Reductions 42
5.6 Conclusions 43
6. ANALYSIS OF CONTROL STRATEGIES 45
6.1 Predicted Effects of Emission Reductions 45
6.2 Estimation of Emission Control Requirements 51
6.3 Veracity of Control Predictions 52
6.4 Conclusions 53
7. SUMMARY AND PERSPECTIVES 55
References 61
111
-------
LIST OF FIGURES
Page
1. Daily peak observed and predicted ozone for St. Louis 16
2. Daily peak observed and predicted ozone for Denver 17
3. Scatter diagram of peak ozone for four cities 19
4. Daily mean bias for St. Louis without autocorrelation 20
5. Daily mean bias for Denver with autocorrelation 22
6. Scatter diagram for July 13, 1976, St. Louis 23
7. Scatter diagram, for August 18, 1975, St. Louis 24
8. Scatter diagram for July 29, 1977, Tulsa 26
9. Scatter diagram for August 3, 1977, Tulsa 27
10. Scatter diagram for June 26, 1974, Los Angeles 29
11. Scatter diagram for June 27, 1974, Los Angeles 30
12. Comparison of 20 day, all station average time series of
observations and predictions for St. Louis 32
13. Ozone response curves, St. Louis 46
14. Ozone response curves, Tulsa 47
15. Ozone reduction vs. (a) maximum observed 03 and (b) maximum
predicted 03 for a change in emissions, Denver 49
IV
-------
1. INTRODUCTION
The U.S. Environmental Protection Agency has sponsored a series
of studies on the use of the SAI Urban Airshed Model in the development of
strategies for ozone control in urban areas. Completed studies include
those in Tulsa, Denver, and St. Louis. Additional work in Los Angeles
has been sponsored largely by groups other than EPA. The four cities are
similar to the extent that all are relatively isolated from the influence
of other urban regions. However, their size and the amounts and types of
data available for modeling vary widely. The purpose of this report is to
summarize and compare the major findings of these studies so that future
users of the model will know what to expect and know best how to use the
resulting information.
The report is organized as follows. Section 2 gives a capsule
description of the SAI Urban Airshed Model. Section 3 provides a brief
synopsis of data bases and methods for each study. The remaining sections
attempt to integrate the results of all the studies. Evaluation of model
performance, a major focus of all four studies, is covered in Section
4. Section 5 examines model sensitivity, emphasizing selected results.
Results of applying the model to control strategy assessment are discussed
in Section 6. Summary remarks and overall perspectives regarding the
model are presented in the final section.
-------
2. DESCRIPTION OF THE MODEL
The SAI Urban Airshed Model simulates the major physical and
chemical processes associated with ozone formation in the polluted
troposphere. These include gas phase chemistry, advective transport,
and turbulent diffusion. The modeling domain is divided into a large
array of grid cells. Horizontally the cells are uniformly sized squares
3 to 5 kilometers on a side. Typically, four or five layers of cells
represent the vertical domain. The depth of the layers is scaled by
the height of the mixed layer and the height of the top of the modeling
domain (region top). The latter typically ranges from 500 meters in
the morning hours to 1000 meters or more in the afternoon. Emissions
are injected into individual cells depending on the location of the
sources, their height of release, and the buoyant rise of individual
stack gas plumes.
The theoretical basis for the SAI Urban Airshed Model rests
on the conservation of mass equation for atmospheric diffusion. Each
reactive species is subject to chemical transformation by the Carbon-Bond
Kinetics Mechanism. The amount of ozone produced by the mechanism
depends on the rate of the nitrogen dioxide photolytic cycle, which in
turn is mediated by the rate of free radical cycling associated with
the oxidation of hydrocarbon compounds. Depending on their molecular
structure, individual hydrocarbon compounds are represented by one or
2
-------
more of five carbon-bond types. Distinguished by their different
reaction rates and pathways, these carbon-bond types serve as surrogates
for the complex mixtures of hydrocarbons found in the urban atmosphere.
Simulations are typically carried out for 12 to 16 hours beginning at
0400 or 0500 LSI, although multi-day simulations have also been
conducted.
Primary inputs to the SAI Urban Airshed Model are point and
area source emission schedules for seven species (NO, N02 and five carbon-
bond types), initial and boundary concentrations both at the surface
and aloft for eight species (seven emitted plus 03), and a variety of
meteorological data. These include a three-dimensional wind field,
mixing depths, solar radiation (expressed as equivalent N02 photolysis
rate), surface temperature, and exposure class, the latter an indicator
of thermal instability.
The SAI Urban Airshed Model was developed by Systems Applications,
Incorporated (SAI) under the sponsorship of EPA's Office of Research and
Development. A more complete description of the model is found elsewhere
(Layland, 1980). Besides the actual simulation program, the model is
provided with a host of preprocessor programs which simplify input
preparation.
Two versions of the simulation program were employed during
these studies. The earlier version was used in the Tulsa and Los Angeles
studies. It contained a numerical technique for advective transport
which upon subsequent testing exhibited some undesirable properties.
For this reason, the numerical technique was replaced with another that
3
-------
is more accurate and generates less "numerical diffusion." This technique,
which tends to give somewhat higher peak concentrations, was incorporated
in a later version that was used in the Denver and St. Louis studies.
The Denver study also compared the later version to the earlier version.
All studies used the CBM II version of the Carbon-Bond Mechanism. However,
substantial differences exist in the preprocessor programs used in the
four studies due, in part, to differences in the availability of input
data.
-------
3. DESCRIPTION OF THE STUDIES
3.1 Tulsa
The Tulsa study was conducted under contract by Systems
Applications, Incorporated. The study is described in a draft report by
Reynolds et al. (1982). The results have been summarized by Layland
et al. (1983).
The air quality and meteorological data required for modeling
were obtained through a special field study conducted by Research Triangle
Institute (Eaton et al., 1979). Monitoring stations were established for
the collection of continuous measurements of ozone, NOX, total nonmethane
hydrocarbons, and winds. Data from ten ozone monitors at both urban and
outlying sites were available for model evaluation. Hydrocarbon species
data were also collected at the surface and aloft. In addition, aircraft
measurements were taken of ozone concentrations aloft.
The emissions data were assembled by Engineering-Science (1980)
from information obtained from the Tulsa City-County Health Department and
the Tulsa Metropolitan Area Planning Commission. The basic emission rates
were developed from stationary source records, transportation modeling of
motor vehicle emissions, and per capita emission factors for miscellaneous
sources.
-------
A total of four days were modeled. The impact of emission
reductions was examined on two of the four days. Measurements aloft were
generally not available on the days modeled. For ozone, morning surface
measurements at rural sites were selected to provide an indication of
ozone aloft on each day. For hydrocarbons, a typical vertical profile
was established for all days based on available aircraft measurements.
Twice daily upper air soundings taken at a National Weather Service sta-
tion in Oklahoma City were used to derive winds aloft and mixing depths.
However, the modeling domain was limited to 1000 meters regardless of the
afternoon mixing depth. Surface winds were interpolated from surface
measurements and were smoothed to reduce divergence. N02 photolysis
rates were derived empirically from hourly, surface-based, total solar
radiation measurements on each day.
3.2 Denver
The Denver study was conducted at the National Center for Atmos-
pheric Research (NCAR) under a cooperative agreement with EPA. The study
grew out of earlier studies by Systems Applications, Incorporated, using
an earlier version of the SAI Urban Airshed Model. The results of the
NCAR study have been presented in a report by Dennis et al. (1983).
Only routinely collected meteorological and air quality data
were available for the study. No monitoring data existed from which early
morning levels of ozone and precursors transported into the modeling region
or accumulated overnight could be estimated. Five ozone monitoring sites,
located within the urbanized area of Denver, were available for model eval-
uation. No outlying sites had been established to measure afternoon levels
-------
of ozone associated with emissions transported beyond the immediate urban
area. No aircraft measurements were taken.
The emissions data base was assembled by the Colorado Department
of Health from stationary source records and motor vehicle travel records.*
The latter were derived from transportation modeling conducted by the
Denver Regional Council of Governments. Emissions from miscellaneous
area sources were also included. Although the base year inventory for
the study was 1979, a second earlier inventory for 1976 was developed,
using the same techniques, for the purpose of air quality trends analysis.
Simulations were performed for eleven days. Additional simula-
tions were run for a subset of eight days using the earlier inventory.
Since no air quality measurements aloft or at outlying sites were available,
typical urban background values were assumed for all eleven days. Twice
daily upper air soundings routinely made at Stapleton International Airport
and tower data taken at Boulder Atmospheric Observatory provided information
on winds aloft. Surface winds were obtained by interpolation from surface
measurement sites. Mixing heights were derived from the Stapleton sound-
ings. However, the modeling domain was limited to 1300 meters regardless
of the afternoon mixing depth. N02 photolysis rates were based on theo-
retical, clear sky rates rather than solar radiation measurements because
these were unavailable. The photolysis rates were adjusted from sea level
to the ground level elevation of Denver.
*0ne large power plant, Cherokee, was eliminated from the data base due
to problems in treating such a concentrated source of NOX emissions in
the model.
-------
3.3 St. Louis
The St. Louis study was conducted jointly by EPA's Office of
Research and Development and the Office of Air Quality Planning and Stand-
ards. Initial development of the model inputs was done under contract by
Systems Applications, Incorporated. Study results have been described in
two reports (Schere and Shreffler, 1982 and Cole et al., 1983) and a
summary paper (Cole et al., 1982).
The extensive Regional Air Pollution Study (RAPS) in St. Louis
provided the emissions, air quality, and meteorological data needed for
modeling (Strothmann and Schiermeier, 1979). Ozone, NOX, total nonmethane
hydrocarbons, winds, and temperature measurements were taken continuously
at both urban and outlying sites. Data from 25 ozone monitors were avail-
able for model evaluation, mostly within the urbanized area of St. Louis.
An intensive program of upper air soundings was also included as part of
the RAPS program. In addition, an instrumented helicopter provided measure-
ments of ozone concentrations aloft. Although samples were collected, valid
hydrocarbon species data were unavailable, either at surface sites or aloft.
The RAPS point source inventory was developed by EPA directly
from questionnaires, plant visits, stack tests, and other available
information. A detailed transportation data base was developed from
which motor vehicle emissions were derived. Emissions from miscellaneous
area type sources were developed using techniques based on economic
activity and population.
Simulations were performed for twenty days. Additional
simulations for a subset of three days were conducted to examine emission
8
-------
control strategies. Since ozone measurements aloft were not always avail-
able on the days modeled, morning surface measurements at rural sites were
selected on each day to represent ozone aloft. For hydrocarbons and NOX
aloft, a single set of background values was assumed for all twenty days.
The wind field was generated by a simple boundary layer model which used
the surface temperature measurements to perturb the mean boundary layer
flow. Mixing heights were estimated from upper air soundings. N02
photolysis rates were empirically derived from continuous surface-based
total solar radiation measurements for each day. These were then adjusted
using radiative transfer theory to provide layer-averaged rates.
3.4 Los Angeles
During the development of the SAI Urban Airshed Model, numerous
applications were conducted in Los Angeles. The model, as originally
formulated, was tested on six days in September, October, and November,
1969 (Reynolds et al., 1973). Numerous refinements in both physical and
chemical treatments were made subsequently. Later model applications
focused on two days, June 26, 1974 and August 4, 1975. These applications
were originally sponsored by several organizations, including the California
Department of Transportation, the California Air Resources Board, and
Southern California Edison Company. The June 26, 1974 simulation was
extended to include the following day, June 27, in a study sponsored by
the U.S. Department of Transportation. The results from these studies
were later updated in several studies, one sponsored by EPA, and others
by Southern California Edison Company and the Western Oil and Gas
Association. The latter study also included a second two-day episode,
November 7*and 8, 1978. The EPA study focused on the sensitivity of the
9
-------
model predictions to the availability of detailed data for constructing
model inputs (Tesche et al., 1981). Only the results from the most recently
performed studies are discussed in this report.
Historically, ambient monitoring in Los Angeles has been quite
extensive. Considerable information was therefore available to charac-
terize air quality and meteorological conditions for the purpose of
modeling. More than forty monitoring stations provided surface measure-
ments of 03, NMHC, and NOX. Nineteen to twenty-five ozone monitors were
available throughout the Los Angeles basin on different days for use in
model evaluation. An extensive hydrocarbon species monitoring program
had been carried out during the summers of 1974 and 1975 which provided
detailed information on the composition of ambient hydrocarbons. In
addition, special studies had been conducted which involved airborne
sampling of 03 and NOX. Upper air and acoustic soundings were available
for as many as six locations at various times during the day. Numerous
stations provided measurements of surface winds.
Emissions from major stationary sources were derived from
records maintained by the South Coast Air Quality Management District.
Nonhighway area source emissions were developed by the Southern California
Association of Governments based primarily on demographic data. Highway
motor vehicle emissions were prepared using the transportation modeling
approaches developed by the California Department of Transportation and
the California Air Resources Board.
10
-------
As indicated above, recent simulations have been performed on
five days.* In order to establish background concentrations, both at the
surface and aloft, surface measurements at outlying stations and on ele-
vated terrain were used, together with special airborne monitoring data.
Nighttime shear-driven mixing depths were derived from surface wind speeds
and acoustic soundings. Daytime convective depths were based on tempera-
ture soundings. In order to represent the effects of terrain channeling
and the diurnal sea breeze on pollutant transport, a three-dimensional
model was used to develop the wind field. The study used theoretical,
clear sky N02 photolysis rates. These were adjusted to account for aerosol
scattering and altitude above ground.
*A1 though control strategy simulations have been performed for Los
Angeles, those carried out using the more recent versions of the
model have been conducted under private auspices or the results are
otherwise unavailable for inclusion in this report.
11
-------
4. EVALUATION OF MODEL PERFORMANCE
If one could be assured that all model inputs are correct, the
accuracy and precision of a model could be determined by comparing the
model predictions with the observed values. However, the SAI Urban
Airshed Model requires a large number of inputs that are difficult to
specify. Thus, the model performance evaluation tests the application
as well as the model itself. Errors related to model deficiencies and
errors induced by designation of model inputs are difficult to distinguish.
In a broader sense, the evaluation of the performance of a model,
such as the SAI Urban Airshed Model, really tests whether air quality
observations can be explained based on available information and under-
standing of the physical and chemical processes involved. If they
cannot, then one's understanding of the problem is incomplete. Either
the emissions are incorrect, the significance of background pollutant
levels is misjudged, the meteorological conditions are inadequately
portrayed, or turbulent diffusion or chemical transformation is impro-
perly represented. Alternatively, the air quality measurements could
be in error. The integration of information from disparate sources in
a quantitative model can draw attention to needed refinement in any one
or more of these areas.
12
-------
Another limitation is that information available to evaluate
model performance is always incomplete. One of the findings of the
St. Louis study is that a large number of days are required to fully
assess model performance due to a large degree of day-to-day variation
in model prediction errors. Results from the Tulsa and Los Angeles
studies are limited to a small number of days. Only the St. Louis and
Denver studies have a sufficient number of days to characterize the
range of performance likely to be encountered. However, even in these
studies there are limitations in the spatial coverage of air quality
monitors. As indicated in Chapter 6, only in the Denver study was an
attempt made, using limited data, to evaluate the ability of the model
to predict the effect of emission reductions on ambient ozone levels.
It should be noted that evaluation of model performance is
complicated by the absence of a "standard" version of the SAI Urban Air-
shed Model. In Chapter 2, a distinction was drawn between two versions
of the model, one brought about by a major revision to the simulation
program. In addition, different users have tended to make various other
changes to the model, particularly the preprocessor programs. In this
sense, the model has continued to evolve. Thus, comparisons among the
four studies are difficult to assess.
In a similar vein, the approach taken by the investigators
toward simulating the days varied among the four studies. In the Los
Angeles and Tulsa work, a relatively few number of days were simulated.
Attempts were made to optimize the performance of the model, within the
confines of the available data. In the Denver and St. Louis studies,
13
-------
a much larger number of days were simulated and much less attention was
given to any particular day.
4.1 Peak Accuracy
Assessment of peak accuracy often receives the greatest
attention in model evaluation studies since implementation of emission
controls is predicated upon reducing the maximum daily ozone in the
region to the level of the standard. Various measures of peak accuracy
have been proposed. The most useful measures appear to be those which
(1) use predicted values at air quality monitoring stations only and
(2) do not require a rigid pairing in time and space between predicted
and observed concentrations.
The rationale for the first criterion is that the alternative,
use of the all-grid daily maximum prediction, tends to bias the evalua-
tion toward overprediction. This is because the daily maximum prediction,
which is selected from the entire modeling region, has a much greater
probability of capturing the actual peak concentration than does a moni-
toring network having a limited number of monitors. The second criterion
recognizes that pairing the observed and predicted concentrations based
on the time and/or location of the observed peak tends to bias the evalu-
ation in the opposite direction, toward underprediction. This is related
to the sparsity of the monitoring network and the nonrandom selection of
the days chosen for modeling. Dennis et al. (1983) offer a further dis-
cussion regarding this aspect of the model evaluation problem. Although
pairing of observed and predicted concentrations is common practice,
this approach to model evaluation, besides its potential for causing a
14
-------
misleading or biased evaluation, fails to recognize that prediction of
the exact time and location of the peak concentration, in a regulatory
context, is of lesser importance than the prediction of the magnitude
of the peak.
Criteria (1) and (2) above have been observed in making the
scatter diagram for St. Louis shown in Figure 1. In this diagram, the
predicted and observed peaks are paired only by day (not by hour or site)
and the predicted peaks are restricted to monitoring sites. Apparent
from this diagram is some tendency for the model to underpredict peak
ozone at the locations of the monitoring sites. However, most of the
predicted maxima fall within ± 30 percent of the observed maxima.
The same presentation is made in Figure 2 for Denver. The
tendency towards underprediction of peak ozone is quite pronounced.
Note that all the predicted daily peaks are less than the corresponding
observed peaks. On average, the daily peak ozone is underpredicted by
about 30 percent.
The relatively small number of days modeled in Los Angeles and
Tulsa limit what can be concluded from these studies. However, the results
from these may be combined with those of Denver and St. Louis in order to
see how well the observed peaks are reproduced for the ensemble of studies.
In principle, the larger the number of studies, the more one can distinguish
errors associated with particular applications and deficiencies that are
inherent in the model itself. However, to the extent that methods for
developing model inputs are common to all studies, systematic errors in the
15
-------
100 200
OBSERVED DAILY PEAK O3 (ppb)
300
Figure I. Daily peak observed and predicted ozone for St. Louis.
16
-------
CO
O
m
o.
160
140
120
100
Q
Q 80
UJ
5 60
UJ
DC
OL
40
20
i I
20 40 60 80 100 120 140 160 180
OBSERVED DAILY PEAK O3 (ppb)
Figure 2. Daily peak observed and predicted ozone for Denver.
(Adapted from Dennis et al., 1983)
17
-------
predictions may simply reveal deficiencies in the methods. Unfortunately,
as was pointed out earlier, neither the models nor the methods were
identical.
Having acknowledged the limitations of direct comparisons among
studies, observed and predicted daily peak 63 for all four cities are
presented together on a scatter diagram in Figure 3. Although the forty
points show an overall tendency toward underprediction, about three-
quarters of the predicted maxima fall within ± 20 percent of the observed
peaks. The results in Figure 3 also indicate the SAI Urban Airshed Model
can give reasonably accurate estimates of peak ozone over a wide range of
concentrations (10 to 35 pphm). The points above the line, while in the
minority, show that either overestimation or underestimation are possible
outcomes. However, the results suggest the Denver application was affected
by unique deficiencies in critical inputs or model treatments.
4.2 Overall Accuracy
While the assessment of peak accuracy is essential, model accuracy
at lower concentrations is also important. Demonstrating accuracy across
the full range of values helps to establish the validity of the model's
treatment of physical and chemical processes and therefore gives greater
confidence in the reliability of the model's predictions when used for
estimating control requirements.
The most extensive analyses of model accuracy are the St. Louis
and Denver studies. A concise summary of the findings for the twenty day
St. Louis study is given in Figure 4. In this figure, model accuracy on
each day is expressed as the daily mean residual (or bias, 0-P) for all
18
-------
- 400
a
a.
ro
O
IU
a.
300
O
£ 20°
O
O
HI
cc
°" 100
ST. LOUIS
A DENVER
O LOS ANGELES
ATULSA
100 200 300 400
OBSERVED DAILY PEAK O3 (ppb)
Figure 3. Scatter diagranv of peak ozone for four cities,
19
-------
cc
UJ
30 -
20 -
f
{
UJ
0 10
z
UJ
Q
LL
I o
O
lO
X -10
h;
(A
5 -20
* -30
T 1 T *
1 i
. . T 1
t f f
-fM i
( i
50
60 70
80
90 100
DAILY MEAN OBSERVED OZONE (ppb)
Figure 4. Daily mean bias for St. Louis without autocorrelation,
20
-------
pairs of observations and predictions.* An estimate of the precision of
the bias statistic is indicated by the 95 percent confidence interval,
which is also shown in Figure 4. These confidence intervals represent
a lower limit, since autocorrelation was not considered in their compu-
tation. Considering the entire set of twenty days, no overall bias is
evident. Moreover, there is no pattern in the daily mean residual with
respect to the daily mean observed ozone.
A similar presentation of the daily mean bias for the eleven
days in Denver is made in Figure 5. The confidence intervals are quite
wide due to the few numbers of observation sites and because autocorre-
lation was considered. Although for most days, individually, the bias
is not significantly different from zero (the confidence interval in-
cludes the zero value for the bias statistic, 0-P), the ensemble of days
clearly shows underprediction.
The overall accuracy of model predictions for any one day is
best ascertained from a scatter diagram of all pairs of observations
and predictions. Two individual days are shown for St. Louis in Figures
6 and 7, August 18, 1975 and July 13, 1976. On both days, there is a
shift toward underprediction at higher concentrations. Of the two,
August 18 shows the greater underprediction at higher concentrations,
yet the daily mean residual, -12.3 ppb, indicates overprediction.
In contrast, the daily mean residual for July 13 is +5.9 ppb. From
Figure 7, the sizeable negative mean residual on August 18 is clearly
*It should be noted that the daily mean residual can be a misleading
indicator of the overall accuracy of the model, as illustrated later on,
21
-------
1 35
£.
H 30
Z
FIDENCE
IS> Ni
O tn
Z
O 15
O
in 10
en
X
t 5
< °
eo
5 '5
<
Q -10
-r
(
1 1
i
{
h
i
t
1
1
1 (
j
1
i
*
i
i
1
<
i
1 J
(
i
i
30
40
50
60
70
DAILY MEAN OBSERVED OZONE (ppb)
Figure 5. Daily mean bias for Denver with autocorrelation,
(Adapted from Dennis et al., 1983).
22
-------
24
21
18
£ 15
CO
O
D
UJ
12
UJ
oc
0- 6
T
A 2 DATA POINTS
O3 DATA POINTS
D4 DATA POINTS
I
I
I
6 9 12 15 18
OBSERVED O3 (pphm)
21
24
Figure 6. Scatter diagram for July 13, 1976, St. Louis
23
-------
o.
Q.
m
O
O
UJ
24
21 -
18.
15
12
= 9
O
5
UJ
oc
a.
A 2 DATA POINTS
O 3 DATA POINTS
D 4 DATA POINTS
A 5 DATA POINTS
6 DATA POINTS
6 9 12 15 18
OBSERVED O3 (pphm)
21
24
Figure 7. Scatter diagram for August 18, 1975, St. Louis.
24
-------
attributable to the large number of overpredictions below 100 ppb observed,
Thus, the daily mean residual, by lumping together all pairs of observa-
tions and predictions, is a relatively poor indicator of the overall
accuracy of the model on any one day.
A tendency toward underprediction at high concentrations is
expected because the sample of days chosen for modeling was not a
random sample. Days were selected because high ozone concentrations
had been observed at monitoring sites. In St. Louis, the spacing
between ozone monitors over a considerable part of the modeling region
is large compared to the size of the urban plume. Thus modest errors
in the spatial alignment of the predicted plume will lower the predicted
concentration at the monitor which had experienced high concentrations
without appreciably raising the predictions at another monitoring site*.
This predisposes the outcome towards underprediction. Indeed, a shift
toward underprediction at high ozone concentrations is apparent on fif-
teen of the twenty test days. On only one day is there a shift toward
overprediction at high concentrations. Though the problem of nonrandom
sampling no doubt exerts an effect on the results, it does not by itself
explain the degree of underprediction on at least eight of the fifteen
days on which the peak ozone concentration is underpredicted in the
St. Louis study. On these days, the all-grid daily maximum prediction
is less than the daily maximum observation.
Scatter diagrams for July 29, 1977 and August 3, 1977 in Tulsa
are shown in Figures 8 and 9. On August 3, 1977, monitoring coverage was
excellent in the area where the urban plume was located. On this day,
25
-------
I I I I I I I I I I I I 1 I I I I I I I
4.0 8.0 12.0 16.0
OBSERVED O3 (pphm)
Figure 8. Scatter diagram for July 29, 1977, Tulsa
(Courtesy of Systems Applications, Inc.),
26
-------
16.0
12.0
CO
O
Q
oi
| 8.0
LU
cc
0.
4.0
4.0 8.0 12.0 16.0
OBSERVED O3 (pphm)
Figure 9. Scatter diagram for August 3, 1977, Tulsa
(Courtesy of Systems Applications, Inc.)
27
-------
little underprediction is apparent in Figure 9. In contrast, underpredic-
tion at high ozone concentrations is seen in Figure 8 for July 29, 1977.
In this case, however, no monitors were located where the high ozone
concentrations were predicted to be. In addition, the highest ozone
concentrations (those greater than 16 pphm) are associated with a monitor
that was calibrated using the NBKI technique, a technique shown in an
intercomparison study to give results that were high by 25 percent.
Of all the studies, Los Angeles has the clear advantage of
having more ozone monitors more evenly distributed. Scatter diagrams
for June 26 and June 27, 1974 are shown in Figures 10 and 11.* The
results for June 26, 1974 show a slight tendency toward overprediction
across the entire range of ozone concentrations. For June 27, 1974, no
bias is apparent in the model predictions up to a concentration level of
40 pphm. However, the very highest concentrations, about 50 pphm, are
underpredicted. This two day episode was recently resimulated using an
improved numerical technique (see Section 2) but there was only modest
improvement for the highest two observations. While underprediction at
these very high concentration levels could have any number of causes,
some material may have been inadvertently lost from the modeling region
when the winds shifted offshore at night, material that would otherwise
have been advected inland again on the following day, June 27.
*It should be noted that the observed ozone concentrations in these
figures have been adjusted to account for differences in calibration tech-
niques for the various monitoring stations. Specifically, ozone concentra-
tions measured at monitors calibrated with the NBKI technique were lowered
by 22 percent.
28
-------
50.00
Q.
Q.
S 40.00 -
cc
z
LU
o
z
O
o
UJ
z
O
N
O
Q
UJ
^ 10.00
O
O
LU
cc
a.
30.00 -
20.00 -
0.00
! -i i q i I i i i i I i i i i I i i i i
0.00 10.00 20.00 30.00 40.00 50.00
OBSERVED OZONE CONCENTRATION (pphm)
Figure 10. Scatter diagram for June 26, 1974, Los Angeles
(Courtesy of Systems Applications, Inc.).
29
-------
a
Q.
z
g
50.00
40.00 -
z
uu
O
Z
O
O
UJ
Z
O
N
O
O
ui
O
5
UJ
DC
a.
30.00 -
20.00 -
10.00
0.00 10.00 20.00 30.00 40.00 50.00
OBSERVED OZONE CONCENTRATION (pphm)
Figure 11. Scatter diagram for June 27, 1974, Los Angeles
(Courtesy of Systems Applications, Inc.).
30
-------
4.3 Spatial and Temporal Patterns
In addition to peak accuracy, in an unpaired sense, and overall
accuracy, another aspect of model accuracy is the replication of spatial and
temporal patterns. The success of the model at replicating these patterns
hinges on its ability to predict ozone concentrations at a particular time
and location. This is a rigorous requirement for any model.
Very generally speaking, spatial patterns tend to be controlled
by the wind field and temporal patterns tend to be controlled by sunlight
intensity, in the absence of major perturbations in other model parameters.
However, a distinction between spatial and temporal patterns is partly an
artificial one. Pollutant transport and reaction kinetics are interconnected.
A significant error in the timing of the ozone build-up necessarily carries
over and influences the location of the peak ozone concentrations. Never-
theless, examining spatial and temporal patterns can be useful for further
evaluating model performance.
The ability of the model to reproduce the spatial and temporal
patterns of ozone is difficult to judge from calculated spatial and tem-
poral correlation coefficients. The temporal pattern of ozone production
is so highly correlated with sunlight intensity that most any reasonable
diurnal pattern of sunlight intensity leads to high temporal correlation
coefficients, on average. Although time series of predictions do not
necessarily closely follow the observations, the two are found to be in
phase when the results for all days are averaged, as seen from the results
of the St. Louis study in Figure 12. At any particular station, the tem-
poral pattern can be strongly influenced by the wind field, which can move
31
-------
150
PREDICTED
RESIDUAL
-25
14 16 18 20
Figure 12. Comparison of 20 day, all station average time series of
observations and predictions for St. Louis.
32
-------
the urban ozone plume toward or away from the monitoring station, or by
individual NOX plumes, which can rapidly deplete ozone levels. By con-
trast, the spatial correlation coefficient is sufficiently sensitive to
minor displacements in the location of the ozone plume that the resulting
correlations are uniformly low. Low correlations are found even when
visual inspection of spatial isopleths indicates fair to good agreement
with the observation network.
For these reasons, direct visual inspection of time series plots
and spatial isopleth maps offers the best opportunity for evaluating the
ability of the model to reproduce temporal and spatial patterns. The two
types of presentations are interrelated and should be examined in tandem
to avoid misdiagnosis. Frequently, a problem in the temporal pattern at a
particular monitoring station, as shown in a time series plot, can be
caused by a simple displacement of the ozone plume. The latter may be
apparent from a series of spatial isopleth maps. This kind of problem may
be unrelated to either the chemical kinetics mechanism, the parameteri-
zation of turbulent diffusion, or the treatment of emission sources.
Instead, it may be attributable entirely to errors in the wind field.
For control strategy purposes, the location and timing of the
ozone plume is less important than the magnitude of the peak concentra-
tions. As discussed previously, errors in the location of the plume may
cause the model to appear to underpredict in relation to the observations
at the monitoring locations. Visual inspection of spatial isopleth maps
can help to identify if this is the major cause of the underprediction
33
-------
or whether other factors may be involved on any particular day. Spatial
isopleth maps can also help to identify possible causes of overprediction.
Despite the important effects of errors in the wind field, the
results of the four studies indicate a general correspondence between the
fields of observed and predicted ozone concentrations. However, results
from both the Tulsa and St. Louis studies exhibit a tendency to under-
estimate the spatial extent of the ozone plume, i.e., the area of ozone
concentrations elevated above regional background levels. This appears
to be true even when the magnitude of the peak prediction agrees well with
that of the peak observation. Evidence of this is necessarily qualitative,
as exact comparisons are impossible, given a sparse observation network.
4.4 Conclusions
From the results of the four studies, a number of general
conclusions can be reached regarding how the SAI Urban Airshed Model
can be expected to perform and the methods by which it can be evaluated.
These are as follows:
o The accuracy of the model predictions can be expected to
vary greatly from one simulation day to the next, even
on days with similar levels of detail in input informa-
tion. This is evident from each of the four studies.
o The accuracy of the model predictions can also be expected
to vary considerably from one study (i.e., city) to the
next, depending in part on the availability of input
information and in part on the methods by which it is
prepared for input to the model. For example, model pre-
dictions in Denver show significantly less accuracy than
those in Tulsa, St. Louis, or Los Angeles.
34
-------
o Overall, the model results have been shown to exhibit a
tendency toward underprediction. However, this is judged
to be a serious problem in only one of the four studies (i.e.,
Denver). It is unclear whether the cause is related to
inadequacies in the input information or in the model
itself. There is no strong evidence that other applications
with other data bases would necessarily exhibit the same
tendency.
o Determining the accuracy of the model predictions is an
inexact science. Simple statistical representations of
model performance (e.g., the mean residual, the root
mean square error, or the correlation coefficient) can be
expected to provide limited insight into the ability of
the model to predict the maximum ozone concentration, the
concentration used as a design value for control strategy
analysis.
o In gauging the accuracy of the model at predicting peak
ozone concentrations, rigorous pairing of the observations
and predictions in time and space has the effect of causing
the model to appear to underpredict. This is related to the
sparsity of the network of observations and the nonrandom
selection of days for model evaluation. In contrast,
direct comparison of the daily maximum prediction (all
grid maximum) to the daily maximum observation (all station
maximum) has the effect of causing the model to appear to
overpredict. This is related to the different sample
sizes from which the maximum concentrations are drawn.
Therefore, undue reliance on either of these comparisons
is not recommended, although together they do serve to
establish probable bounds on the true accuracy of the
model predictions.
o Due to problems brought about by rigorous pairing of obser-
vations and predictions, statistical approaches to model
evaluation require subjective interpretation in order to
judge the accuracy of model predictions of peak concentra-
tions on any given simulation day. Users may find that
spatial isopleth maps and time series plots are useful tools
for making these interpretations.
o Scatter diagrams of all pairs of observations and predic-
tions serve to convey quickly the major features of the
overall predictive capability of the model on any given
simulation day. Scatter diagrams are highly recommended
for this purpose.
35
-------
5. MODEL SENSITIVITY
Given that the predictions of the SAI Urban Airshed Model, like
those of other models, have less than perfect accuracy, it is useful to
consider what factors contribute significantly to errors in the model pre-
dictions. During the development and application of the model, a variety
of testing has been conducted to evaluate the importance of various model
inputs and to identify those to which the model predictions are most sensi-
tive. Only the results of the studies discussed in Section 3 are considered
here. Results from other studies using earlier versions of the model are
considered less reliable. Presumably, refinements in the specification of
these inputs, or greater availability of information upon which they are
based, will lead to more accurate predictions, at least in the absence of
serious deficiencies in the model itself.
5.1 Background Air Quality
Assumptions regarding background air quality appear to have
significant effects on ozone predictions. These assumptions are reflected
in the model inputs for initial and boundary concentrations, both aloft
and at the surface. When reactive hydrocarbons aloft were reduced by a
a factor of two in the Tulsa study, from 84 to 42 ppb C at 500 meters alti-
tude, the maximum predicted ozone concentration was reduced by 17 percent.
Oxygenated organics were also reduced in this test from 12 to 4 ppb C.
Similar tests were conducted on two days in the Denver study. On both
36
-------
days, reactive hydrocarbons aloft and at the boundaries of the modeling
region were reduced from 52 to 19 ppb C and oxygenated organics from
5.0 to 1.2 ppb C. For the day on which the maximum ozone was observed
within the urban area, the effect was negligible. However, on the day
when the urban ozone plume was transported outside the urban area, the
maximum predicted concentration was lowered by 12 percent. In the Los
Angeles study, use of total NOX measurements and an assumed HC/NOX
ratio of 7 for establishing initial concentrations for hydrocarbons,
instead of actual hydrocarbon species measurements, was found to lower
ozone concentrations above 12 pphm by an average of 40 percent.
The effect of background ozone has been examined in the Denver
study. A change from 2 pphm to 9 pphm increased the maximum ozone pre-
diction by an equal amount on several days. Even on the day having the
highest ozone concentrations, the maximum prediction was increased nearly
4 pphm.
The importance of background depends on the wind field and
the mixing which ensues with higher wind speeds and greater vertical
wind shear, as well as on the rate of growth of the mixed layer. Of
course, background air quality becomes more important, in a relative
sense, as emissions are reduced. Better measurements of background
concentrations of ozone and the whole array of ozone precursors are
necessary to improve the accuracy of model predictions.
5.2 Solar Radiation
The accumulation of ozone in the polluted troposphere is caused
.»
by the photodissociation of nitrogen dioxide which in turn is driven by
37
-------
solar radiation. The accurate specification of N02 photolysis rates is
therefore essential to the accurate prediction of ozone concentrations.
At equilibrium, ozone concentrations should be directly proportional to
the N02 photolysis rate, in the absence of hydrocarbons. Tests conducted
in the St. Louis study show that although the maximum ozone prediction is
slightly less sensitive than this, the relationship between N02 photolysis
rate and peak ozone is only marginally less than one-to-one.
The use of theoretical, clear sky, ground level photolytic
rates may not accurately reflect the true photolytic rates in urban
atmospheres at a particular time and location. Ground level rates under-
state the true photolytic rates aloft due to the attenuation, near the
surface, of short wave radiation by ozone absorption and aerosol scat-
tering, resulting in lowered ozone predictions at the surface. Scattering
by clouds can significantly alter both the intensity and spectral distri-
bution of sunlight, thereby either enhancing or depressing photolytic
rates, depending on cloud cover and solar angle. The inability to explic-
itly consider radiative transfer in the model, or simply to provide
vertically varying photolysis rates on input, consequently places a limi-
tation on the accuracy with which ozone concentrations can be predicted.
5.3 Vertical Mixing
Vertical mixing has a strong effect on model predictions. These
effects are partly associated with the impact which background concentra-
tions aloft and elevated emissions, above the morning mixed layer, have on
surface concentrations as the afternoon mixed layer develops. The extent
of vertical mixing is controlled jointly by three principal user-specified
38
-------
inputs: (a) vertical grid spacing, (b) mixing height/region top, and
(c) exposure class. Vertical grid spacing is important because mixing is
assumed to be instantaneous within any grid cell on a time scale comparable
to the time step used for numerical integration. Specification of the mix-
ing height/region top places a lid on the mixed layer. The exposure class,
in combination with wind speed, determines the vertical eddy diffusivity at
the grid mesh points. Although vertical eddy diffusivity is a complex func-
tion involving several parameters, exposure class is the most important,
independently controlling parameter. Given the importance of vertical mix-
ing, and the interrelationships between the various user-specified inputs,
careful attention must be given to the way in which these inputs are estab-
lished, within the context of the theoretical framework of the model.
Tests conducted in the Denver and Los Angeles studies on vertical
grid spacing produce conflicting results. This is likely associated with
interactions with other model inputs which differ for the two studies. In
the Denver study, when the grid spacing in the mixed layer was increased
from 250 to 330 meters at midday, ozone concentrations were reduced up to
24 percent at one station, while the daily maximum was reduced 8 percent.
Little effect from vertical grid spacing was seen for the Los Angeles study.
However, the Los Angeles work demonstrates very clearly the importance of
pollutants which reside in the layers aloft, below the region top but above
the morning mixed layer- When the aloft layers were eliminated, large
increases or decreases were seen at individual stations, although the daily
maximum ozone concentration was reduced only 5 percent. In some cases,
surface ozone concentrations more than doubled, due to the removal of
NOX emissions emitted into the layers aloft, thereby eliminating ozone
39
-------
scavenging which would otherwise have taken place when the emissions
were mixed to the surface later in the day.
The specification of mixing height/region top is also tied to
the entrainment of pollutants from aloft. Results from both the Denver and
St. Louis studies show that an increase in afternoon mixing height/region
top, in the range of 25 to 30 percent, actually caused a small increase in
surface ozone concentrations, indicating that entrainment was more impor-
tant than dilution. At the same time, a decrease of 25 percent in the
mixing height in the St. Louis study also increased the daily maximum ozone
prediction, by an amount that was much less than, or equal to, the increase
expected from a simple box model treatment of dilution in the absence of
chemical reaction.
Exposure class is used in the model, as indicated previously,
to categorize the thermal instability of the atmosphere. Tests conducted
during the Denver study reveal that the accuracy of model predictions is
closely tied to the exposure class parameter. A modification to the
exposure class that corresponds to a change from unstable to neutral
stability increased peak ozone concentrations by as much as 38 percent
at individual stations. The daily maximum ozone was increased similarly.
Unfortunately, exposure class is a rather vaguely defined parameter,
one that depends on solar radiation as well as surface heating, or
temperature. Perhaps these would better be input to the model directly
and, together with surface winds, be used to calculate the vertical
eddy diffusivity, thereby avoiding the need for an exposure class
parameter altogether.
40
-------
It is interesting to note that the change in the exposure class
improved the accuracy of the CO predictions in the Denver study. The CO
predictions were low compared to the observations. However, CO measure-
ments may not be a good yardstick for judging the adequacy of the treatment
of vertical mixing. This is because high CO concentrations occur in local-
ized "hot spots" which, on a spatial scale, are much smaller than the size
of a typical grid cell.
5.4 Emissions Composition
A number of tests have been carried out to examine the effect on
ozone predictions of changes in the chemical composition of emitted hydro-
carbons and of emissions of oxygenated organic compounds. When a chemical
composition typical of motor vehicle exhaust was used in place of a less
reactive urban mixture, the daily maximum ozone prediction was increased
in the St. Louis study by as much as 30 percent. A similar test in the
Los Angeles study shows the daily maximum ozone prediction was little
affected but ozone concentrations above 12 pphm were increased by as much
as 14 percent, on average. At individual stations, increases ranged as
high as 50 percent. In the St. Louis study, ozone predictions are par-
ticularly sensitive to emissions of organic oxygenates. An increase in
the fraction of total emitted carbon which is oxygenated, from 0.015 to
0.059 (as carbon), increased the daily maximum ozone concentration by 12
percent, even though the increase in total carbon was offset by decreases
in the amount of aromatic and olefinic compounds. Results in Tulsa show
an increase in the fraction of oxygenated organics, from 0.008 to
0.036, has considerably less effect, however, than a reduction of 12
»
percent in total hydrocarbon emissions. In this test, the net effect
41
-------
was to lower the daily maximum ozone prediction by 9 percent. About
half the reduction in total hydrocarbon emissions was accounted for by
aromatic compounds, the other half by paraffins, on a per carbon basis.
Taken together, the results from these studies indicate that accurate
model predictions require that attention be given to the composition of
emitted hydrocarbons and especially to emissions of oxygenated organic
compounds.
5.5 Interaction with Emission Reductions
Of special interest is the interaction between emission reduc-
tions and uncertainties in other model input parameters. Changes in ozone
concentrations which are predicted to occur as emissions are reduced may
be influenced significantly by other model inputs. However, very limited
analyses have been done in this area. In the Tulsa study, simulations of
the effect of a 30 percent reduction in hydrocarbon emissions were made for
two different levels of background hydrocarbon concentrations aloft. Back-
ground hydrocarbons differed by a factor of two (42 vs. 84 ppb C at 500
meters altitude) while organic oxygenates differed by a factor of three (4
vs. 12 ppb C at 500 m) for the two cases. At the lower background level,
the daily maximum ozone was reduced by 27 percent. However, at the higher
background level , ozone was reduced only half as much, just 14 percent.
The Denver study examined the effect of vertical mixing on the
relative change in ozone concentrations when emissions are increased.
A modification was made to the exposure class which corresponds to a
change from unstable to neutral stability. The daily maximum ozone
prediction was increased nearly twice as much for the neutral case as
42
-------
for the unstable case. It appears from these results, together with
those from Tulsa, that emission control requirements could be quite
sensitive to the extent to which urban emissions are dispersed and
become mixed with background air.* Thus, the accuracy of estimates of
needed emission controls may depend significantly on the availability
and reliability of background air quality measurements and on the treat-
ments of winds and vertical mixing, all of which act in complementary
fashion.
5.6 Conclusions
Tests of model sensitivity in the four studies allow the follow-
ing conclusions to be reached. Users are cautioned that these are based
on limited sensitivity analyses which cannot be generalized to apply to
all situations.
o Background air quality levels for ozone and organic compounds
can be expected to significantly influence ozone predictions,
particularly when emissions are reduced. Proper characteriza-
tion of these is needed to better ensure the accuracy of model
predictions, particularly the characterization of hydrocarbons
and oxygenated organic compounds.
o Vertical mixing of emissions with background air exerts a
strong effect on model ozone predictions. Users should
realize the extent of mixing depends on a complex interplay
involving a number of model input parameters. The specifi-
cation of these must be coordinated to achieve good results.
Given the significance of vertical mixing and the difficulty
of representing turbulent diffusion in atmospheric models,
independent evaluation of this component of the model is
suggested in order to better ensure the accuracy of the
model predictions.
*This is also consistent with the significant difference found in the
Denver study with respect to the two versions of the model discussed in
Section 2. .The version with less "numerical diffusion" gave a relative
reduction in the peak ozone which was one third greater, for the same
emission reduction.
43
-------
Users should be aware that N02 photolysis rates have a
strongly controlling effect on model ozone predictions.
In the present formulation of the model, only a relatively
simplistic approach to the specification of the N02 photo-
lysis rate is possible. In the interest of greater accuracy
a more realistic representation of these rates is needed, one
based on physical mechanisms for radiative transfer in urban
atmospheres.
Alternative distributions of hydrocarbon emissions in
terms of chemical composition can be expected to produce
significant changes in ozone predictions under certain
circumstances. These changes are related to differences in
the reaction rates and pathways of different hydrocarbon
species. Relatively large perturbations are required to
produce major effects. However, organic oxygenates appear
to have an effect on ozone concentrations which is out of
proportion to their relative rate of emission compared to
hydrocarbons. This reflects the high degree of efficiency
with which these compounds promote ozone formation in the
Carbon-Bond Mechanism. Particular attention to these
compounds in the preparation of emission inventories appears
necessary to improve the accuracy of model predictions.
44
-------
6. ANALYSIS OF CONTROL STRATEGIES
The studies in all four cities have analyzed the effect of
emission controls on model ozone predictions. Specific control strategies,
involving emission reductions for individual sources or source types, and
uniform emission reductions, involving across-the-board emission reductions
for all sources, have both been examined. Since the studies in St. Louis,
Denver, and Tulsa were conducted or sponsored by EPA, emphasis is given to
the results from these studies. Control strategy work in Los Angeles was
done under private auspices and is not included. Control strategy simula-
tions have been performed on three days in St. Louis, two days in Tulsa,
and eight days in Denver. Analysis of the results of these simulations
has focused on the daily maximum ozone predictions.
6.1 Predicted Effects of Emission Reductions
Control strategy simulations for St. Louis and Tulsa show that
for a given reduction in hydrocarbon emissions at constant NOX, the daily
maximum ozone concentration is reduced by a percentage which is less than
or equal to the percentage reduction in emissions. The St. Louis and
Tulsa results are shown in Figures 13 and 14 in the form of ozone response
curves. These give the relative change in the daily maximum ozone pre-
diction versus the relative change in hydrocarbon emissions. Immediately
evident is the large disparity in the results from one simulation day
45
-------
Figure 13. Ozone response curves, St. Louis.
46
-------
0)
o
I
Z
g
o
D
O
UJ
DC
60
40
20
T
T
T
OJULY 29
AUGUSTS
20
40
60
80
HC EMISSIONS REDUCTION, percent
100
Figure 14. Ozone response curves, Tulsa.
47
-------
to the next. This is a reflection of the strong dependence which ozone
concentrations have on both meteorology and background air quality. The
shapes of the curves appear to be somewhat different for the two cities,
the Tulsa response curves flattening much faster as hydrocarbons are reduced.
However, this merely reflects the lower ozone concentrations which prevail
on the Tulsa days. As emissions are reduced, the relative change in ozone
is lessened as background levels of ozone and its precursors are approached.
Results from the St. Louis and Tulsa studies suggest there is
little relationship between the maximum observed or predicted ozone con-
centration and the sensitivity of model ozone predictions when emissions
are changed. This is illustrated more clearly in the results from the
Denver study shown in Figure 15. From this diagram, there appears to be
no correlation between the percentage reduction in ozone due to a change
in emissions and either the maximum ozone observation or the maximum
ozone prediction. Different days exhibit different ozone reductions for
the same emission reduction.
A potentially significant finding of the St. Louis study is the
tendency of the daily maximum ozone prediction to migrate downwind as
hydrocarbon emissions are reduced. This suggests that rather than basing
required emission controls on ozone predictions at a particular monitoring
site, controls should be based on the daily maximum prediction, regardless
of location. Otherwise, the ozone problem may simply be shifted downwind,
to locations further from the immediate urbanized areas of the city.
Looking at individual control measures, results for both Tulsa
and St. Louis show that controls on mobile source hydrocarbon emissions
48
-------
a)
OBSERVED
.0
a
CO
O
180
170
160
150
140
130
120
110
100
I I I 1 1 I I I 1 1
_
D D
D D
D
D
_
QD -
i i i i i i i i i i
2 4 6 810121416182022
% CHANGE (ALL-GRID MAXIMUM)
b)
PREDICTED
. ,
J3
a
a.
CO
O
^
X
2
1 /U
160
150
140
130
120
110
100
90
80
-7r>
1 1 I 1 I 1 (*j 1 I 1
_ ^
I D D
DD
_
D
DD -
- _
i i I i i i i i I i
2 4 6 810121416182022
% CHANGE (ALL-GRID MAXIMUM)
Figure 15. Ozone reduction vs. (a) maximum observed 0- and (b) maximum
predicted 0_ for a change in emissions, Denver.
(Adapted from Dennis et al. 1983).
49
-------
are more effective than controls on stationary source hydrocarbon emissions
on a pound-for-pound basis. This is because mobile source emissions have
larger olefin, aromatic, and aldehyde components than do stationary sources.
These compounds contribute to the formation of ozone with greater efficiency
than do paraffins.
The St. Louis simulations also indicate that controls on
hydrocarbon emissions from both stationary and mobile sources combined
cause ozone to be reduced more than indicated by the sum of the reductions
for the two source types individually, as if having a synergistic effect.
The synergism is a reflection of the nonlinearity of the ozone response
curves. This can be seen from Figure 13, where the curves are shown to
be slightly S-shaped when emissions are reduced.
Very little systematic study has been done on the effect of
changes in NOX emissions, either increases or decreases, on daily maximum
ozone concentrations. Results from the St. Louis study indicate that
decreases in NOX emissions tend to offset ozone reductions achieved by
controlling hydrocarbon emissions. However, studies done in Los Angeles
reportedly show the opposite. Ozone formation is limited in part by rapid
NOX removal mechanisms under Los Angeles type conditions. Substantial
reductions in both hydrocarbons and NOX appear to be necessary to cause
large reductions in ozone in the eastern part of the Los Angeles basin.
A similar effect may occur downwind of other urbanized regions in rural
areas. Thus, although increases in NOX emissions may decrease urban ozone
levels, ozone levels in rural areas may be increased. However, the role
of urban NOX emissions in rural ozone formation remains speculative.
50
-------
6.2 Estimation of Emission Control Requirements
Model simulation results show clearly that ozone formation is
a complex process in which emissions, meteorology, and background air
quality interact. The day on which the highest ozone concentration is
observed is not necessarily the day that will require the greatest emis-
sion controls. Results from both the Denver and Tul sa studies suggest
ozone reductions may be more difficult to achieve on days when the daily
maximum ozone prediction is located in outlying areas, away from the
immediate urban environs. Considerations such as these, combined with
the fact that the monitoring network may not detect the highest ozone
concentrations on any particular day, dictate that several days, at a
minimum, be simulated for the purpose of control strategy development.
Ozone response curves, such as those in Figures 13 and 14, are
a useful tool for estimating emission control requirements. Given the
percentage reduction needed to lower the maximum prediction on a partic-
ular day to the level of the standard, the curves provide an estimate
of the required percentage reduction in hydrocarbon emissions. In this
way, the model results are used in a relative sense. The advantage to
this approach is that model errors are implicitly controlled. For
example, if the relative error (percent over or underprediction) in the
maximum prediction is the same in the pre- and post-control simulations,
then the relative change in the maximum ozone prediction will remain free
of error. Using the model in an absolute sense, such that the maximum
prediction for the post-control scenario is not allowed to exceed the
level of the air quality standard, could lead to either an over or
51
-------
underestimation of the degree of emission control required depending on
whether the model is over or underpredicting the maximum concentration.
In many situations, however, the relative error would not be
expected to remain constant, in part due to the strong nonlinearity of
the model. In addition, the inputs themselves may readily contribute to
a nonlinear type of behavior. For example, in a typical control strategy
simulation, background concentrations, presumably associated mainly with
uncontrolled anthropogenic, biogenic, or other sources, are held constant
while emissions are reduced. In this way, uncertainties in the specifi-
cation of background concentrations can have a large effect on the rela-
tive change in the peak ozone prediction in response to emission
reductions.
6.3 Veracity of Control Predictions
In studies employing the SAI Urban Airshed Model, and indeed
in most air pollution modeling studies, the assumption is made that if
it can be demonstrated a model adequately reproduces air quality obser-
vations for one emissions level, then the predictions will be equally
valid at another, generally lower, emissions level. This may frequently
be a reasonable assumption for nonreactive pollutants, but is more likely
to be a matter of faith for photochemical pollutants such as ozone. The
strongly nonlinear character of atmospheric chemical reactions can nullify
such assumptions. Dennis et al. (1983) argue it is necessary to test the
veracity of control predictions directly.
An attempt to do this was made in the Denver study. Model
simulations for an ensemble of eight days were conducted with emission
52
-------
inventories for two different historical periods. The average change in
the daily maximum ozone prediction was compared to the trend in observed
ozone concentrations during an overlapping time period. Although trend
analysis is necessarily imprecise, due largely to the confounding effects
of meteorology, the results are encouraging. The average change in the
daily maximum ozone prediction (10.9%) agrees favorably with the observed
trend (13.6% change) despite the fact that the peak ozone concentrations
are themselves significantly underpredicted in the Denver study. The
reliability of these results is difficult to judge, however, due in part
to inconsistencies between observed and predicted trends in nonmethane
hydrocarbon and NOX concentrations. Only by means of ongoing modeling
efforts and continuation of monitoring programs, both for ozone and its
precursors, will it be possible to demonstrate conclusively that photo-
chemical models are capable of predicting changes in ozone concentrations
which occur due to the implementation of emission control strategies.
6.4 Conclusions
Applications of the SAI Urban Airshed Model to date allow the
following conclusions to be reached. Users are cautioned that these are
based on limited analyses which may not be applicable to every situation
in which the model might be applied.
o Uniform changes in hydrocarbon emissions will result in
nonlinear changes in peak ozone predictions which are in
the same direction. However, the magnitude of the relative
change in peak ozone is likely to be less than the relative
change in emissions.
53
-------
The degree of hydrocarbon emissions reductions needed to
lower ozone concentrations to the level of the standard
can be expected to vary strongly from day to day. Days
with the highest observed peak ozone are not necessarily
those needing the greatest emission control. Thus, it is
essential that control requirements be based on a sample
comprised of a number of days.
Simultaneous reductions in both hydrocarbon and NOX emissions
are likely to be less effective than hydrocarbon emission
reductions alone in controlling peak ozone concentrations in
or near urban areas. However, the extent to which urban NOX
emissions may contribute to rural ozone is unknown.
Reductions in mobile source emissions can be expected to be
more effective than equal reductions in stationary source
emissions. Thus, consideration of differences in the com-
position of hydrocarbon emissions among source types is
recommended for the purpose of control strategy development.
In some cases, the daily maximum ozone prediction migrates
downwind as emissions are reduced. Thus, emission control
requirements should not be based on ozone predictions con-
strained to a particular monitoring site.
Reductions in background precursor concentrations which
parallel emission reductions can be expected to have a sub-
stantial effect on peak ozone concentrations for days with
moderate to strong wind speed and a lesser effect on
stagnation days. Such reductions cannot be recommended,
however, as there can be no assurance that background
levels will be favorably influenced by emission control
efforts.*
The accuracy of model predictions of the ambient air impact
of emission controls is difficult to ensure, even if the
usual model performance evaluation indicates good results.
This is related to the strongly nonlinear character of atmos-
pheric chemistry and diffusion. Assurance of good control
predictions would require the model to be evaluated with data
bases which encompass a change in emissions. Unfortunately,
suitable data bases of this sort are only rarely available.
In this regard, the Denver study provides preliminary indica-
tions which are encouraging, although the data available are
inconclusive.
*Background is to be distinguished from interurban transport of freshly
emitted ozone precursors between cities closely situated, as in some parts
of the Northeastern U.S. Such transport has not been addressed in the four
studies.
54
-------
7. SUMMARY AND PERSPECTIVES
This report has reviewed recent applications of the SAI Urban
Airshed Model involving studies in Tulsa, St. Louis, Denver, and Los
Angeles. All four cities are relatively isolated from other urban areas.
Thus, interurban transport was not a factor in these studies. Both Los
Angeles and St. Louis had extensive data bases available, including air
quality and meteorological data. In Denver, only a relatively sparse data
set was available, derived nearly entirely from routine monitoring activi-
ties. Tulsa was intermediate with respect to data availability. All
studies used similar techniques in the development of the emission inven-
tories. Neither the simulation program nor the preprocessor programs were
identical in the four studies. These similarities and differences must be
considered when attempting to generalize from the study results.
The results of the four studies have been discussed in regard to
evaluation of model performance, model sensitivity, and analysis of control
strategies. It is hoped that potential users of the SAI Urban Airshed
Model may benefit from the experience described here. Clearly, however,
the reliability of the model results depends heavily on the amount and
quality of the data available for modeling and the ability and experience
of those persons who must of necessity make various judgments regarding
the data, its interpretation, and its suitability and representativeness
55
-------
with respect to the requirements of the model. Other users may be more
or less successful than the results of these studies would suggest.
Conclusions regarding model performance, model sensitivity, and
control strategies are given at the end of each section and will not be
repeated here. However, some general impressions to leave with the reader
are summarized as follows. One is that the accuracy of the model, in
terms of agreement between the model predictions and observed ozone con-
centrations, is judged to be quite good, with the exception of the Denver
study where less data were available. When the peak ozone predictions
are compared to the peak ozone observations for the ensemble of days
modeled in the four studies, the accuracy of the model predictions is
on the order of 30 percent. On any given day, the model is more likely
to underpredict than to overpredict. It should be recognized that deter-
mining the accuracy of the model in a quantitative sense is not a simple
exercise; no single statistic is able to satisfactorily characterize the
ability of the model to predict the peak 1-hour ozone concentration
used for control strategy analyses. Although the studies do suggest a
tendency to underpredict the peak 1-hour ozone concentration, sensitivity
analyses indicate the tendency to underpredict is well within uncertainties
attributable to the model inputs. This does not mean, of course, that the
model inputs are responsible for all errors in the model predictions or
that the model cannot be improved.
A second impression is that background concentrations, particu-
larly hydrocarbon and oxygenated organic compounds as well as ozone,
exert a considerable influence on the model predictions. Although effects
56
-------
are seen in the base case, the impact of background concentrations is far
greater when emissions are reduced. The implications for control strategy
analyses are obvious. The importance of background can be assumed to be
further heightened the more the plume of urban emissions becomes mixed with
background air. This in turn is controlled by the particular set of
meteorological conditions on any given day as well as by peculiarities
associated with the grid model itself.
Substantial time and effort have been expended to gain perspective
on the use of the SAI Urban Airshed Model as a predictive tool for ozone
in urban areas. The four studies reviewed in this report underscore the
level of resources necessary to mount an application of the model in a
particular locale. At present, a total cost of 50,000 dollars can be
expected per day simulated. This includes the cost of analyzing the air
quality and meteorological data and preparing it for input to the model.
Also included are various preliminary simulations and sensitivity analyses
which are necessary on any given day to optimize and refine the modeling
approach. Special expertise is required to both exercise the model, in a
scientific sense, and execute the model, in an operational sense. The
per day cost could be higher depending on the complexity of the meteor-
ological regime, for example if terrain or coastal effects are present,
and the manner by which these are treated. Paradoxically, the less data
that are available for consideration in modeling, the lower the costs
are likely to be. Not included is the cost of preparing a spatially,
temporally, and chemically resolved emission inventory or the cost of
collecting ambient data. Application of a three-dimensional grid model
can place additional, more stringent, informational requirements on the
57
-------
emissions and ambient data bases than are routinely made at the State and
local levels depending, in part, on the perceived needs of a particular
application. However, the major costs associated with the collection and
maintenance of high quality data bases are borne by other air pollution
abatement activities and are not unique to modeling.
Due to resource demands, casual or routine application of the
model is not feasible. This must be considered in any regulatory usage
of the model. Another impediment to routine usage is the lack of complete
documentation with respect to operation of the model and adequate guidance
concerning the role and significance of various input parameters.
In addition, there are difficulties in the interpretation of
model predictions for setting emission limitations. As pointed out
earlier, the day on which the highest ozone concentration is observed
at a monitoring site is not necessarily the day that will require the
greatest emission controls. This is related not only to the nonlinear
behavior which occurs when chemical and physical processes interact in
the atmosphere, but also to the limited size of monitoring networks.
Thus, it is not possible to pick the day a priori upon which emission
control requirements should be based. Given the current form of the
ozone NAAQS and assuming sufficient data were available, one would need
to model every day in a three-year period in order to be sure of select-
ing the day upon which emission control requirements should be based. In
practice, some subset of days during the three-year period, namely those
having relatively high observed ozone levels, would likely suffice. This
subset of days is anticipated to include a relatively large number of
58
-------
days that would require modeling. However, modeling of even a modest
number of such days requires the expenditure of considerable resources.
Further, it is difficult to establish valid criteria for selecting a few
days to be modeled. The predictions of the SAI Urban Airshed Model for a
small set of days on which high ozone concentrations have been observed
simply do not allow any statistically based finding to be made regarding
the probability of a violation. The problems discussed here are posed by
all complex models, either physicochemical models, such as the SAI Urban
Airshed Model, or complicated Gaussian models, where a large number of
sources are involved.
Results from the studies reviewed in this report indicate the SAI
Urban Airshed Model is capable of reproducing observed ozone concentrations
with an accuracy that is comparable to other numerical models in the field
of science and engineering and is frequently better than that obtainable
from Gaussian models for other pollutants. However, the ability of the
model to adequately predict changes in ozone concentrations brought about
by emission reduction programs is still an open question. A long-term
ambient monitoring and emissions tracking program is required to settle
this issue. As a partial interim step, sensitivity studies could be under-
taken with the data bases currently available to examine the uncertainties
associated with model predictions of the air quality impact of emission
controls.
The SAI Urban Airshed Model appears to be best suited as a
long-term planning tool. The model is also useful as a benchmark for com-
parison with ^impler photochemical models or as a tool for the development
59
-------
of State Implementation Plans where resources and data bases permit. As a
planning tool, the model is particularly useful for answering "what if?"
type questions. If alternative motor vehicle fuels, such as methanol,
come into widespread use in the future, how will ozone concentrations be
affected? If assumptions which have been made concerning levels of oxy-
genated organics aloft are off by a factor of two, how will predictions
of the air quality benefit of emission controls be influenced? If avail-
able data are unrepresentative of winds aloft, how will estimates of needed
emission reductions be modified if an alternative wind treatment is used
instead? In order that determining the answers to these questions might
be other than an exercise in model sensitivity, devoid of reality, the
model must have a firm scientific footing.
In order to serve in these capacities, the model must be
accepted as "state of the art" by the air pollution community. Periodic
improvements in the formulation of the model to incorporate research devel-
opments in the fields of photochemistry and turbulence and diffusion will
likely be required if regulatory needs call for the retention of such a
role into the future. Areas where such developments may prove fruitful
lie with (a) the further refinement of the chemical kinetics mechanism,
particularly the chemistry of aromatics and the treatment of background
reactivity, (b) the coupling of photolytic rates with atmospheric radia-
tive transfer, (c) the treatment of vertical mixing in the planetary
boundary layer, (d) the explicit treatment of sub-grid scale point source
diffusion, and (e) the creation of three-dimensional wind fields from
sparse data sets.
60
-------
REFERENCES
1. Cole, H. S., C. F. Newberry, W. Cox, G. K. Moss, D. Layland (1982)
"Application of the Airshed Model for Ozone Control in St. Louis,"
Paper 02-20.1, 75th Annual Meeting of the Air Pollution Control
Association, New Orleans, Louisiana, June 20-25, 1982.
2. Cole, H. S., D. E. Layland, G. K. Moss, C. F. Newberry (1983)
"The St. Louis Ozone Modeling Project," EPA-450/4-83-019, U.S.
Environmental Protection Agency, Research Triangle Park, North
Carolina, 207 p.
3. Dennis, R. L., M. W. Downton, R. S. Keil (1983) "Evaluation of Per-
formance Measures for an Urban Photochemical Model," EPA-450/4-83-021,
U.S. Environmental Protection Agency, Research Triangle Park, North
Carolina, 219 p.
4. Eaton, W. C., C. E. Decker, J. B. Tommerdahl , F- E. Dimmock (1979)
"Study of the Nature of Ozone, Oxides of Nitrogen, and Nonmethane
Hydrocarbons in Tulsa, Oklahoma Volume 1. Project Description and
Data Summaries," EPA-450/4-79-008a, U.S. Environmental Protection
Agency, Research Triangle Park, North Carolina, 219 p.
5. Engineering-Science (1980) "Emission Inventories for Urban Airshed
Model Application in Tulsa, Oklahoma," EPA-450/4-80-021, U.S. Envi-
ronmental Protection Agency, Research Triangle Park, North Carolina,
280 p.
6. Layland, D. E. (1980) "Guideline for Applying the Airshed Model to
Urban Areas," EPA-450/4-80-020, U.S. Environmental Protection Agency,
Research Triangle Park, North Carolina, 184 p.
7. Layland, D. E., S. D. Reynolds, H. Hogo, W. R. Oliver (1983) "Demon-
stration of Photochemical Grid Model Usage for Ozone Control Assess-
ment," Paper 83-31.6, 76th Annual Meeting of the Air Pollution Control
Association, Atlanta, Georgia, June 19-24, 1983.
8. Reynolds, S. D., M. K. Lui, T. A. Hecht, P. M. Roth, J. H. Seinfeld
(1973) "Urban Air Shed Photochemical Simulation Model Study Volume I.
Development and Evaluation," EPA-R4-73-030a, U.S. Environmental
Protection Agency, Research Triangle Park, N.C., 150 p.
9. Reynolds, S. D., H. Hogo, W. R. Oliver, L. E. Reid (1982) "Application
of the SAI Airshed Model to the Tulsa Metropolitan Area," U.S. Environ-
mental Protection Agency Contract No. 68-02-3370, Systems Applications,
Inc., San Rafael, California, 392 p.
61
-------
10. Schere, K. L., J. H. Shreffler (1982) "Final Evaluation of Urban-
Scale Photochemical Air Quality Simulation Models," draft report,
Environmental Sciences Research Laboratory. U.S. Environmental
Protection Agency, Research Triangle Park, North Carolina.
11. Strothman, J. A., F. A. Schiermeier (1979) "Documentation of the
Regional Air Pollution Study (RAPS) and Related Investigations in
the St. Louis Air Quality Control Region," EPA-600/4-79-076,
U.S. Environmental Protection Agency, Research Triangle Park,
North Carolina, 715 p.
12. Tesche, T. W., C. Seigneur, L. E. Reid, P. M. Roth, W. R. Oliver,
J. C. Cassmassi (1981) "The Sensitivity of Complex Photochemical
Model Estimates to Detail in Input Information," EPA-450/4-81-031a,
U.S. Environmental Protection Agency, Research Triangle Park,
North Carolina, 186 p.
13. Ibid. "Appendix A--A Compilation of Simulation Results,"
EPA-450/4-81-031b, 245 p.
62
-------
TECHNICAL REPORT DATA
(Please read Instructions on the reverse before completing}
1. REPORT NO.
EPA-450/4-84-004
3. RECIPIENT'S ACCESSION NO.
4. TITLE AND SUBTITLE
5. REPORT DATE
A Review of Recent Applications of the SAI Urban
Airshed Model
December 1983
6. PERFORMING ORGANIZATION CODE
7. AUTHOR(S)
8. PERFORMING ORGANIZATION REPORT NO
David E. Layland, Henry S. Cole
9. PERFORMING ORGANIZATION NAME AND ADDRESS
U.S. Environmental Protection Agency
Office of Air Quality Planning and Standards
Monitoring and Data Analysis Division (MD-14)
Research Triangle Park, North Carolina 27711
10. PROGRAM ELEMENT NO.
A13A2A
11. CONTRACT/GRANT NO
12. SPONSORING AGENCY NAME AND ADDRESS
13. TYPE OF REPORT AND PERIOD COVERED
14. SPONSORING AGENCY CODE
15. SUPPLEMENTARY NOTES
16. ABSTRACT
Reviewed are studies in St. Louis, Denver, Los Angeles, and Tulsa on the use
of the SAI Urban Airshed Model for the development of strategies for ozone control.
The SAI Urban Airshed Model is a three-dimensional photochemical grid model incor-
porating the Carbon-Bond Kinetics Mechanism. The report discusses the availability
of aerometric data in the four studies and the sensitivity of model predictions to
selected input parameters. An analysis is presented of the performance of the model
with respect to ambient ozone observations. Issues and results pertaining to use of
the model for control strategy analyses are identified and discussed. It is antici-
pated that potential users may benefit from the experiences in applying the model
which are described here.
17.
KEY WORDS AND DOCUMENT ANALYSIS
DESCRIPTORS
Air pollution
Atmospheric models
Photochemical reactions
Smog
Ozone
Nitrogen oxides
Hydrocarbons
b.IDENTIFIERS/OPEN ENDED TERMS
SAI Urban Airshed Model
Carbon-Bond Mechanism
St. Louis
Denver
Los Angeles
Tulsa
c. COSATI I iold/Group
18. DISTRIBUTION STATEMENT.
19. SECURITY CLASS (Tin's Report)
21 NO. OF PAGFS
67
20. SECURITY CLASS (Thin
22. PRICE
EPA Form 22201 (Rev. 4 77) PREVIOUS EDITION is OBSOLETE
------- |