TSD-1e

CMAQ Model Performance and Assessment

8-Hr OTC Ozone Modeling 

Bureau of Air Quality Analysis and Research

Division of Air Resources

New York State Department of Environmental Conservation

Albany, NY 12233

February 23, 2006

Air quality model evaluation and assessment 

One of the tasks that is required as part of demonstrating attainment
for the 8-hr ozone NAAQS is the evaluation and assessment of the air
quality modeling system that has been utilized to predict future air
quality over the region of interest. As part of the attainment
demonstration, the SMOKE/CMAQ modeling system was applied to simulate
the pollutant concentration fields for the base year 2002 emissions with
the corresponding meteorological information. The modeling databases for
meteorology using MM5 (TSD-1a), the emissions using SMOKE (TSD-1b and
TSD-1c), and application of CMAQ (TSD-1d) provides simulated pollutant
fields that are compared to measurements, in order to establish the
credibility of the simulation. In the following sections a comparison
between the measured and predicted concentrations is performed and
results are presented, demonstrating on an overall basis the utility of
the modeling system in this application.

The results presented here should serve as an illustration of some of
the evaluation and assessment performed on the base 2002 CMAQ
simulation.  Additional information can be made available by request
from the New York State Department of Environmental Conservation.

Summary of measured data

The ambient air quality data, both gaseous and aerosol species, for the
simulation period of May through September 2002 were obtained from the
following sources:

EPA Air Quality System (AQS)

EPA fine particulate Speciation Trends Network (STN)

EPA Clean Air Status & Trends Network (CASTNet)

Interagency Monitoring of PROtected Visual Environments (IMPROVE) 

Pinnacle State Park, NY operated by Atmospheric Science Research Center,
University at Albany, Albany, NY

Harvard Forest, Petersham, MA operated by Harvard University, Boston, MA

Atmospheric Investigation, Regional Modeling, Analysis and Prediction
(AIRMAP) operated by University of New Hampshire, Durham, NH

NorthEast Ozone & Fine Particle Study (NE-OPS), led by Penn State
University and other research groups in Philadelphia, PA

Aircraft data obtained by the University of Maryland, College Park MD

Wet deposition data from the National Atmospheric Deposition
Program/National Trends Network (NADP/NTN), Atmospheric Integrated
Research Monitoring Network (AIRMoN), and the New York State Department
of Environmental Conservation (NYSDEC)

Measured data from sites within the Ozone Transport Region (OTR) plus
the rest of Virginia were included here.  The model-based data were
obtained at the grid-cell corresponding to the monitor location; no
interpolation was performed. 

Ozone (O3)

Hourly O3 is measured at a large number of State, Local, and National
Air Monitoring Stations (SLAMS/NAMS) across the US on a routine basis,
and the data from 208 sites were extracted from the AQS database ( 
HYPERLINK "http://www.epa.gov/ttn/airs/airsaqs/aqsweb/aqswebhome.html" 
http://www.epa.gov/ttn/airs/airsaqs/aqsweb/aqswebhome.html ).  Hourly O3
concentrations from the Harvard Forest Environmental Management Site in
Petersham, MA (  HYPERLINK
"http://www.as.harvard.edu/data/nigec-data.html" 
http://www.as.harvard.edu/data/nigec-data.html ); Pinnacle State Park in
Addison, NY (  HYPERLINK "http://www.asrc.cestm.albany.edu" 
http://www.asrc.cestm.albany.edu ); and the four University of New
Hampshire AIRMAP sites (http://airmap.unh.edu) were also included in
this database.  The EPA CASTNet program collects hourly O3 at generally
rural locations across the US (  HYPERLINK "http://www.epa.gov/castnet" 
http://www.epa.gov/castnet ); data from 22 sites, including two from
West Virginia, were used in the model evaluation.

Fine particulate matter (PM2.5)

The 24-hour average Federal Reference Method (FRM) PM2.5 mass data
collected routinely at SLAMS/NAMS sites across the US were extracted
from AQS (257 sites).  Hourly PM2.5 mass was also included in this
database, primarily extracted from AQS (54 sites).  Hourly PM2.5 mass
were also taken from the Thompson Farm, NH AIRMAP site, Pinnacle State
Park, and the NE-OPS site in Philadelphia, PA (  HYPERLINK
"http://lidar1.ee.psu.edu"  http://lidar1.ee.psu.edu ).

Fine particulate speciation

The 24-hour average PM2.5 and fine particulate speciation (sulfate
(SO4), nitrate (NO3), elemental carbon (EC), organic carbon/organic mass
(OC/OM), and soil/crustal matter) from Class I areas across the US,
collected every 3rd day, were obtained from the IMPROVE web site ( 
HYPERLINK "http://vista.cira.colostate.edu/IMPROVE/Default.htm" 
http://vista.cira.colostate.edu/IMPROVE/Default.htm ).  In addition to
these parameters, the EPA STN (  HYPERLINK
"http://www.epa.gov/ttn/amtic/speciepg.html" 
http://www.epa.gov/ttn/amtic/speciepg.html ) also reports ammonium (NH4)
to AQS; data from this network are collected every 3rd or 6th day. Data
from 49 STN sites, generally in urban areas and often collocated with
FRM monitors, and 21 IMPROVE sites (including Dolly Sods, WV) were used
in this analysis.  Organic mass is assumed to equal 1.8×OC, and
soil/crustal matter is assumed to consist of oxides of Al, Ca, Fe, Si,
and Ti.  The STN OC data are blank-corrected by removing a
monitor-specific, constant blank, and these values are available from  
HYPERLINK
"http://www.epa.gov/airtrends/aqtrnd03/pdfs/2_chemspec0fpm25.pdf" 
http://www.epa.gov/airtrends/aqtrnd03/pdfs/2_chemspec0fpm25.pdf ; the
IMPROVE OC blanks are assumed to equal zero.

Criteria gaseous pollutants

Hourly carbon monoxide (CO; 97 sites), nitric oxide (NO; 75 sites),
nitrogen dioxide (NO2; 97 sites) and sulfur dioxide (SO2; 134 sites) are
also included in this model evaluation database.  A large majority of
these sites are SLAMS/NAMS monitors located primarily in urban in
suburban areas, but data from the Harvard Forest, Pinnacle State Park,
and AIRMAP sites are also included here.

Non-methane hydrocarbons

	

While there are several dozen hydrocarbon species measured routinely,
for this model evaluation database the focus was on Carbon Bond IV
species groups that consist of a single primary species.  For this
reason only ethene (C2H4), isoprene (C5H8), and formaldehyde (HCHO)
concentrations were extracted from AQS.  Hourly C2H4 and C5H8 data from
19 Photochemical Assessment Monitoring Stations (PAMS) sites and 24-hour
average HCHO from 18 air toxics sites are included in this database.

University of Maryland aircraft data

The University of Maryland performed 144 aircraft spirals at 41 regional
airport locations over 26 days from May-August 2002 (  HYPERLINK
"http://www.atmos.umd.edu/~RAMMPP"  http://www.atmos.umd.edu/~RAMMPP ). 
Spirals are approximately 20-45 minutes in duration, over which time the
atmosphere from about 0-3 km is sampled.  The concentrations of O3, CO,
and SO2 from these spirals were included in this database, and help
provide a semi-quantitative evaluation of CMAQ performance above the
ground surface.  Minute average aircraft data were compared to the
nearest instantaneous 3-dimensional CMAQ output.

Wet deposition 

The NADP (http://nadp.sws.uiuc.edu) collects wet deposition samples
across the US, through the NTN and the AIRMoN. Weekly wet deposition
samples are collected by the NTN, while daily or event-based  samples
were collected by the AIRMoN.  The NYSDEC (  HYPERLINK
"http://www.dec.state.ny.us"  http://www.dec.state.ny.us ) also collects
weekly wet deposition samples independently from the NADP.  The wet
deposition of SO42-, NO3-, and NH4+ from 43 NADP/NTN sites, 7
NADP/AIRMoN sites, and 19 NYSDEC sites are included in this model
evaluation database.  

Evaluation of CMAQ predictions

  are the average concentrations, respectively, and N is the sample
size.

Observed average, in ppb:

 

Predicted average, in ppb (only use Pi when Oi is valid):

 

Correlation coefficient, R2:

 

Normalized mean error (NME), in %:

 

Root mean square error (RMSE), in ppb:

 

Fractional error (FE), in %:

 

Mean absolute gross error (MAGE), in ppb:

 

Mean normalized gross error (MNGE), in %:

 

Mean bias (MB), in ppb:

 

Mean normalized bias (MNB), in %:

 

Mean fractionalized bias (MFB), in %:

 

Normalized mean bias (NMB), in %:

 

Daily maximum 8-hour O3 concentrations

Model evaluation statistics, based on daily maximum 8-hour average O3
levels on those days having (1) at least 18 valid observations, or (2)
fewer than 18 valid observations but the observed daily maximum O3
concentration was at least 85 ppb, are presented here for all sites
across the OTR and all of VA.  The data cover the period May 15 through
September 29, excluding July 6-9, when many sites across the eastern US
were affected by large forest fires in Quebec.  There are 208 SLAMS/NAMS
sites and 28 special sites.

These model evaluation statistics were computed using two different
threshold values for observed daily maximum 8-hour O3.  First, the
statistics were computed using only those days when the observed daily
maximum 8-hour O3 concentration exceeded 40 ppb.  Second, the statistics
were computed using only those days when the observed daily maximum
8-hour O3 exceeded 60 ppb.  This latter method focuses on the highest O3
days.

Figures 1-4 display time series of observed and predicted daily maximum
8-hour O3 concentrations averaged over all sites across the OTR, at
SLAMS/NAMS and special sites and for the daily maximum two thresholds. 
These averages were computed for each day considering all sites that met
the corresponding threshold criteria.  In general the observed and
predicted composite average O3 concentrations track each other rather
well, although there was fairly substantial underprediction during the
mid-August period.  Also, the model performance tends to be better when
the lower cutoff (40 ppb) was considered.

Figures 5-8 display spatial maps of fractional error and mean
fractionalized bias for the two threshold levels.  At each site the
statistics were computed over the entire modeling season.  Both the
SLAMS/NAMS and special monitors are displayed here.  In general, the
model performance was better in the vicinity of urban areas and along
the northeastern corridor, compared to the performance in rural areas
where the model tended to underpredict daily maximum concentrations. 
The other statistical metrics yielded similar results to FE and MFB.

Table 1 lists the median and range in fractional error, and the mean
fractionalized bias of daily maximum 8-hour O3 calculated at each site
over the season, for both observed thresholds (40 and 60 ppb), as well
as all sites versus just the SLAMS/NAMS sites.  Considering just
SLAMS/NAMS sites, FE was always less than 32% for the 40 ppb threshold,
and less than 40% for the 60 ppb threshold.  Similarly, the MFB at
SLAMS/NAMS sites ranged from -29 to +23% for the 40 ppb threshold, and
ranged from -40 to +22% for the 60 ppb threshold.  Adding the special
sites did not affect the statistics substantially.

Diurnal variations of gases

	

Figures 9-17 display the composite diurnal variations of the species
reported hourly – O3 (SLAMS/NAMS and other/special sites, displayed
separately), continuous PM2.5, CO, NO, NO2, SO2, ethene, and isoprene. 
The average diurnal variations are for the period of May 15-September 30
– again excluding July 6-9 – considering all sites in the OTR.  Note
that the O3 diurnal variations were computed from running 8-hour
averages, with hours denoting the start of the 8-hour block.  The number
of monitors used to compute each composite diurnal variation is shown in
each figure.

For O3, the composite diurnal pattern predicted by CMAQ is fairly
similar to that observed, especially at the more urban SLAMS/NAMS
monitors.  However, on average CMAQ predicts the daily maximum about an
hour earlier than observed.  For most of the other species presented
here, CMAQ tends to predict two daily peaks, one morning and one late
afternoon.  For some species, such as PM2.5 mass the observed
concentration on a composite basis has very little diurnal variation. 
On the other hand, primary pollutants like CO, NO, and ethane, CMAQ
exhibits qualitative agreement with the observations.

Daily average concentrations of co-pollutant trace gases

Composite daily average predicted and observed concentrations of CO, NO,
NO2, SO2, C2H4, HCHO, and C5H8 across the OTR are displayed in Figures
18-24.  Daily average concentrations of the criteria gases, C2H4 and
C5H8 were computed from hourly averages, and only those days having at
least 12 hours of valid observed data were considered here.  The HCHO
data shown here are based on 24-hour average values every 6th day.  The
criteria gas data cover the period May 15 – September 30, whereas the
NMHC data only cover the June 1 – August 31 period, since these data
are predominantly PAMS data; however, excluded from this analysis is the
July 6-9 period when many sites across the eastern US were affected by
large forest fires in Quebec.

Table 2 lists the median and range in mean fractionalized bias
calculated at each site over the season used in this analysis.  The
values listed in Table 2 were computed at each site over the entire
season.  While the range in MFB is rather large for each species across
all sites, the median MFB was below 50% for all species except C2H4,
which is substantially overpredicted by CMAQ.  It should be noted that
these species can vary substantially from day to day, and days with very
low modeled or observed values can contribute to high MFB.

PM2.5 mass and speciation

Composite daily average predicted and observed concentrations of PM2.5
mass (both daily average FRM data and continuous data), as well as major
speciation –SO4, NO3, NH4, EC, OM (defined here operationally as
1.8×blank-corrected organic carbon), and crustal mass (sum of oxides of
Al, Ca, Fe, Si, and Ti) – across the OTR were compared in this
analysis. The data cover the period May 15 – September 30, and again
the July 6-9 period was excluded, when numerous sites in the eastern US
were affected by large forest fires in Quebec.  The continuous and FRM
PM2.5 data are shown every day, since there are ample daily FRM sites
across the OTR.  The speciation data included here are daily averages
every third day, and consist of the largely urban EPA STN and the
largely rural IMPROVE network.  The two speciation networks collect
PM2.5, SO4, NO3, EC, OM, and crustal mass, while only the STN reports
NH4 at a sufficient number of locations.

observed and predicted crustal concentrations were 4.59 μg m-3 and 1.74
μg m-3, respectively at the STN monitors, and 4.46 μg m-3 and 0.99 μg
m-3, respectively at the IMPROVE monitors.

	As with the gaseous co-pollutant data, there is a substantial spread in
MFB across the sites.  However, the median MFB for PM2.5 mass and SO4
was generally small (<12%) for both urban and rural sites.  CMAQ tends
to overpredict NO3, more so at the IMPROVE sites.  CMAQ also tends to
underpredict OM at both urban and rural sites, although some of this
discrepancy may be attributed to the fact that OM is operationally
defined and is highly dependent on the blank correction and multiplier
to account for other components of OM not directly measured.  CMAQ tends
to overpredict both EC and crustal mass, especially at urban sites;
similar to OM, the crustal mass overprediction is related to the fact
that this parameter is operationally defined.

Wet deposition of sulfate, nitrate, and ammonium

Observed and predicted wet deposition of SO4, NO3, and NH4 were compared
over the period May 14 – September 30.  For this analysis, weekly or
event-based wet deposition amounts from the NADP/NTN (43 sites),
NADP/AIRMoN (7 sites), and New York State DEC (19 sites) covering the
entire OTR plus all of VA and WV were integrated over the
four-and-a-half months.  Because the observed weekly wet deposition
samples did include July 6-9, the corresponding CMAQ predictions also
include this period.  Table 4 lists the model evaluation statistics for
integrated wet deposition of SO4, NO3, and NH4 at each site over the
season, while Figures 40-42 compare the observed and predicted weekly
values relative to the 1:1 line.

Overall CMAQ tended to overpredict wet deposition of these ions.  On a
percentage basis, the overprediction was least for SO4 and highest for
NO3.  The NME, MNGE, MNB, and NMB were less than 50% for the three ions.
 Given that precipitation is very difficult to predict, especially
during the summer months when rainfall can vary tremendously over a 12
km by 12 km area represented by this model grid, CMAQ did a rather good
job reproducing seasonal wet deposition over the OTR.

Upper-air O3, CO, and SO2 data

The University of Maryland operated an instrumented light aircraft
during the summer of 2002.  On 26 days from May-August meteorological,
trace gas, and particle scattering/absorption data were collected during
ascent or descent spirals over 41 regional airports.  In all, 144
spirals were performed from near the surface to about 3 km above ground
level.  For this analysis, composite average profiles of O3, CO, and SO2
were created over three time periods:  “morning” (08-11 EST),
“afternoon” (12-16 EST), and “evening” (17-19 EST).  The minute
average observed concentrations were aggregated into layer averages,
which correspond to the lowest 15 model layers.  Model layers are
increasingly thick away from the surface; the surface layer is about 20
m thick while the 15th layer is about 500 m thick (and centered about
2.8 km above the ground).

Figures 43-51 display the observed and predicted composite vertical
profiles of O3, CO, and SO2 for the three time periods.  In terms of
profile shape, CMAQ was in good qualitative agreement for all three
species above the surface during the afternoon hours.  For CO, the model
tends to greatly underpredict observed levels near the surface, whereas
the predicted O3 and SO2 concentrations are closer to the respective
observed values.

Summary

	Various model evaluation statistics are presented here for a variety of
gaseous and aerosol species in addition to O3.  In general, the CMAQ
results were best for daily maximum O3 and daily average PM2.5 and SO4
mass.  Many other species vary tremendously over the course of a day, or
from day to day, and small model over- or underprediction at low
concentrations can lead to large biases on a composite basis.  It is
important to demonstrate that the model performs reasonably over the
diurnal cycle, not just in terms of daily maximum or average values. 
Also, it is important to demonstrate that the model can reproduce
concentrations above the ground level.Table 1.  Median and range in
fractional error (FE, %) and mean fractionalized bias (MFB, %) for daily
maximum 8-hour O3 using the 40 ppb and 60 ppb observed thresholds.  The
values using only SLAMS/NAMS sites are boldfaced, the values using all
sites are in regular font.

Metric, threshold	Range (%)	Median (%)

FE, 40 ppb	+10 to +34%

+10 to +32%	+15%

+15%

MFB, 40 ppb	-34 to +23%

-29 to +23%	-6%

-6%

FE, 60 ppb	+9 to +40%

+9 to +40%	+15%

+15%

MFB, 60 ppb	-40 to +22%

-40 to +22%	-12%

-11%



Table 2.  Median and range in mean fractionalized bias (%) for daily
average CO, NO, NO2, SO2, C2H4, HCHO, and C5H8.

Pollutant	Range in MFB (%)	Median MFB (%)

CO (97 sites)	-128 to +144%	-10%

NO (75 sites)	-182 to +116%	-46%

NO2 (97 sites)	-125 to +107%	+13%

SO2 (134 sites)	-139 to 140%	+3%

C2H4 (19 sites)	+28 to +168%	+86%

HCHO (18 sites)	-66 to +96%	-13%

C5H8 (19 sites)	-54 to +165%	+43%



Table 3.  Median and range in mean fractionalized bias (%) for daily
average PM2.5, SO4, NO3, NH4, EC, and OM.

Pollutant	Range in MFB (%)	Median MFB (%)

PM2.5 (FRM; 257 sites)	-59 to +119%	-4%

PM2.5 (continuous; 57 sites)	-39 to +85%	+5%

STN PM2.5 (49 sites)	-45 to +102%	-9%

IMPROVE PM2.5 (21 sites)	-36 to +19%	-10%

STN SO4 (49 sites)	-21 to +60%	+12%

IMPROVE SO4 (21 sites)	-26 to +16%	-7%

STN NO3 (49 sites)	-73 to +406%	+25%

IMPROVE NO3 (21 sites)	-57 to +358%	+64%

STN NH4 (49 sites)	-36 to +112%	+16%

STN EC (49 sites)	-42 to +269%	+34%

IMPROVE EC (21 sites)	-60 to +146%	-27%

STN OM (49 sites)	-82 to -25%	-58%

IMPROVE OM (21 sites)	-60 to +7%	-40%

STN crustal (49 sites)	+2 to +546%	+182%

IMPROVE crustal (21 sites)	-18 to +163%	+38%



Table 4.  Model evaluation statistics for integrated wet deposition of
SO4, NO3, and NH4

.

Parameter	SO4	NO3	NH4

Observed average, mg m-2	1063	704	185

Predicted average, mg m-2	946	367	117

Correlation coefficient, R2	0.17	0.22	0.12

NME, %	34	49	48

RMSE, mg m-2	490	417	109

FE, %	36	62	57

MAGE, mg m-2	365	344	89

MNGE, %	36	45	46

MB, mg m-2	-118	-337	-68

MNB, %	-3	-44	-28

MFB, %	-13	-61	-44

NMB, %	-11	-48	-37



Figure 1.

Figure 2.

Figure 3.

Figure 4.

Figure 5.

Figure 6.

Figure 7.

Figure 8.

Figure 9.

Figure 10.

Figure 11.

Figure 12.

Figure 13.

Figure 14.

Figure 15.

Figure 16.

Figure 17.

Figure 18.

Figure 19.

Figure 20.

Figure 21.

Figure 22.

Figure 23.

Figure 24.

Figure 25.

Figure 26.

Figure 27.

Figure 28.

Figure 29.

Figure 30.

Figure 31.

Figure 32.

Figure 33.

Figure 34.

Figure 35.

 

Figure 36.

Figure 37.

 

Figure 38.

Figure 39.

Figure 40.

Figure 41.

Figure 42.

Figure 43.

ë

ê

ë

̖j

„@

„Ð^„@

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

ˆ

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

혈F鐃᳿ꐋⰖ"蠆

蠆

蠆

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

”ÿô

@

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Ð

Figure 44.

	

Figure 45.

	

Figure 46.

	

Figure 47.

Figure 48.

	

Figure 49.

Figure 50.

	

Figure 51.

 PAGE   

 PAGE   1 

