10.

5 DEVELOPMENT OF A WRF-AERMOD TOOL FOR USE IN REGULATORY APPLICATIONS

Toree Myers-Cook*, Jon Mallard, and Qi Mao

Tennessee Valley Authority (TVA)

ABSTRACT available. For rural sources, the closest

representative NWS station may be greater than 50-
The Tennessee Valley Authority (TVA) has developed 100 kilometers away. Furthermore, NWS data
a WRF-AERMOD tool which transforms the Weather limitations - instrumentation limits, missing data, and
Research and Forecasting (WRF) model output into lack of surface parameters required by AERMET -
the meteorological input needed by AERMOD have motivated the EPA to explore the use of
(AERMAP/AERMET/AERMOD), the EPA- prognostic meteorological models, specifically MM5
recommended model for short-range dispersion (Brode 2008), which can provide a timely and spatially
modeling. The tool was used to model 2002 comprehensive meteorological dataset for input to
emissions from TVA’s Allen Fossil (ALF) Plant in AERMOD. The EPA has been testing an in-house
Memphis, Tennessee. Modeling using NWS data MM5-AERMOD tool for possible use in future
was also performed and results were compared. regulatory applications. However, since MM5 is no
Analyses of modeling results showed several longer supported by its developers and the WRF
differences between the two approaches, with much (Weather Research Forecasting) model is now
of the inconsistencies attributed to four main findings. considered the state-of-the-art meteorological model
First, the NWS-AERMOD approach tended to stay to replace it, TVA developed a WRF-AERMOD tool
more unstable in the summer and stable in the winter that takes WRF three-dimensional meteorological
with unrealistically high mechanical boundary layer output (surface and upper air) and transforms it into
(MBL) heights. Second, user-estimated (or the AERMET-ready surface and profile files for use in
AERSURFACE-derived) surface parameters - Bowen AERMOD. The tool was used to model 2002
ratio, albedo, and surface roughness - in the NWS emissions from TVA’s Allen Fossil Plant in Memphis,
approach were substantially different from WRF- Tennessee.
estimated values which led to significant differences
in the derived boundary layer parameters. Third, the 2. WRF-AERMOD SETUP
considerable percentage of calms in the NWS dataset
affected NWS-AERMOD results as boundary layer The WRF meteorological model is a mesoscale
parameters and estimated concentrations were not numerical weather prediction system that has been
calculated during calm conditions. Finally, when 24- designed to be a flexible, state-of-the art atmospheric
hour SO2 concentrations from both approaches were simulation system, suitable for use in a broad range of
compared to the onsite ALF SO2 monitor, WRF- applications across scales ranging from meters to
AERMOD estimated values much closer to observed. thousands of kilometers (Michalakes et al 2004). The
The WRF tool had several advantages over NWS as WRF model was set up using one course domain at
it was able to directly output more land use and 12 kilometer resolution and with 27 vertical sigma-
boundary layer parameters needed by AERMOD than pressure levels. The major physics options included
NWS, it had a more representative upper air profile, it the YSU PBL (Planetary Boundary Layer) scheme,
contained no missing data, and it was able to provide the NOAH LSM (Land Surface Model), the RRTM
comprehensive data at the precise location of the (Rapid Radiative Transfer Model) scheme for long-
source. wave radiation, the Dudhia scheme for shortwave
radiation, the KF (Kain and Fritsch) cumulus scheme,
1. INTRODUCTION and the WSM (WRF Single-Moment) 3-class simple
ice scheme (Skamarock et al 2008). The GCIP
For short-range regulatory dispersion modeling (less NCEP NAM / ETA 3D data and surface analysis data
than 50km), the U.S. Environmental Protection were used for input into the model. The WRF model
Agency (USEPA) recommends use of the AERMOD simulated meteorological conditions over North
1
modeling suite . The meteorological data America for the 2002 year in seven-day simulation
preprocessor AERMET incorporates air dispersion steps (with one overlap day which was used for ramp-
based on planetary boundary layer turbulence up) and produced the standard hourly meteorological
structure and scaling concepts and requires output. Most of the basic surface variables needed by
representative National Weather Service (NWS) AERMET / AERMOD were readily available from
surface and upper air data as well as onsite data, if WRF and were extracted using a standard WRF data
2
extraction program . Relative humidity and cloud

*CTR 1P-M, 1010 Reservation Rd, Muscle Shoals, AL

2
35662; email: tmcook@tva.gov The program, read_wrf_nc.f, can be obtained from the
1
The AERMOD modeling suite consists of the AERMET WRF download page:
meteorological processor, the AERMAP terrain data (USGS) http://www.mmm.ucar.edu/wrf/users/download/get_sources.
processor, and the AERMOD dispersion model. html

1
cover were not readily available from WRF. Instead, averaged SO2 monitor observations. A cumulative
they were derived from temperature, mixing ratio and frequency distribution (CFD) was performed to
pressure based on a UCAR algorithm 3 . All of the present a running total of SO2 concentrations in order
AERMET-required upper-air variables were readily to see how well each approach compared to
available from WRF and were extracted for all 27 observations (Figure 2). The CFD showed that the
levels. WRF-AERMOD tool estimated the observed SO2
concentrations much better than the NWS-AERMOD
The AERMET processor requires NWS surface and approach. Additional statistical analyses of the mean
upper air observations as well as user estimation of bias, mean error, and root mean square error (RMSE)
land surface characteristics (albedo (r), Bowen ratio showed that independent of receptor location, WRF-
(B0), and surface roughness length (z0)) to derive estimated 24-hour SO2 concentrations with a lower
several boundary layer parameters: sensible heat bias, error, and RMSE than NWS (Figure 3).
flux (H), surface friction velocity (u*), convective
velocity (w*), the vertical potential temperature Chi over Q (χ/Q) values were calculated on the
gradient above the PBL (VPTG), the height of the AERMOD results to see if any patterns or similarities
convectively-generated boundary layer (CBL), the existed in the concentration averages. They were
height of the mechanically-generated boundary layer calculated for 1-hour, 3-hour, 8-hour, 24-hour, and
(MBL), and Monin-Obukhov length (L). These annual averages. The χ/Q values, or dispersion
boundary layer variables were not standard output factors, were obtained by dividing the maximum
from WRF. Therefore, several AERMET conversion impact concentrations by the emission estimates in
routines were modified to take the necessary WRF units of seconds / cubic meter. The χ/Q values
surface variables and calculate them. Once all provide a relatively simple means to estimate the
surface and upper-air variables were extracted or magnitude of pollutant impacts from a source; they
derived, they were reformatted into AERMET Stage 3 are better representative of maximum concentration
format for ready use into AERMOD. impacts from dispersion modeling when no chemistry
is included. The χ/Q results showed that for short-
The WRF-AERMOD tool was used to model 2002 term averages (< 24-hour), the WRF-AERMOD tool
emissions of criteria pollutants with their standard predicted higher maximum concentrations than NWS-
averaging times from the TVA Allen Fossil (ALF) Plant AERMOD (Figure 4). For the longer-term averages
in Memphis, Tennessee. The NWS 2002 Memphis (24-hour and annual), NWS-AERMOD predicted
surface and Little Rock NWS upper air data were also higher maximum concentrations. Furthermore, both
processed through AERMET and used in AERMOD to the NWS approach and WRF approach predicted the
model emissions from ALF. Both approaches were highest 3-hour and 24-hour concentrations in
modeled using a four-nested AERMOD domain with December, with the 24-hour concentrations predicted
the finest receptor grid closest to the source and on the same day. An analysis of the highest 98th
extending out 20 kilometers (Figure 1). percentile of concentrations for all averaging periods
revealed that the majority (75%) of highest
3. AERMOD ANALYSES concentrations estimated with the WRF-AERMOD
tool fell in December. The NWS approach tended to
The ALF site was chosen as the modeling source for spread the highest 98th percentile of concentrations
several reasons: 1) its location in an urban setting over several months, with December ranking highest
(Memphis) close to a NWS station, which allowed for (26%) followed by March (21%) (Figure 5).
a good comparison between WRF and NWS
meteorology, and 2) a SO2 monitor located at the ALF Finally, spatial plots were constructed to view the
site, which allowed for comparison between estimated distribution of concentrations around the ALF site.
concentrations and observations. In Memphis, there For all averaging periods <= 24 hours, the WRF-
are three large sources of SO2 emissions, of which AERMOD tool tended to spread concentrations much
ALF is the largest. The closest source of SO2 farther downwind from the source than NWS-
emissions to the ALF site is from the Cargill Corn AERMOD. For the 24-hour average, even though
Milling plant located to the northeast. It emits less NWS-AERMOD produced a higher maximum, the
than 1/6th of the SO2 emissions at ALF. With spatial plots revealed that WRF-AERMOD tended to
prevailing winds from the southwest and west, predict a larger distribution of high concentrations that
impacts from this plant to the ALF monitor are most extended much farther from the source than the NWS
likely minimal. Therefore, AERMOD-estimated approach (Figure 6). For the annual average, NWS-
maximum 24-hour SO2 concentrations from both the AERMOD predicted both a higher maximum annual
WRF tool and the NWS approach (independent of concentration and a larger spread of high
receptor location) were compared to 24-hour concentrations away from the source than the WRF-
AERMOD tool.
3
The UCAR algorithm, cloudf.F, is found in the GRAPH
program (/mesouser/MM5V3/GRAPH.TAR.gz) which can be
obtained from the UCAR ftp site: anonymous@ftp.ucar.edu.

2
4. METEOROLOGICAL ANALYSES between predictions. However, FB results must be
interpreted carefully because the sign of FB can be
The results of the CFD and the χ/Q analysis led to opposite of the sign for the computed bias. In other
further investigation of the AERMET-generated words, one can have an over-prediction of a variable
meteorological conditions that were present during (a positive bias) but have a negative FB. Because of
the highest estimated concentrations. Surface the large number of meteorological surface variables
conditions were primarily investigated since they were that are output from AERMET (approximately 20), the
most likely to influence results. A review of the FB was used as a screening tool to try and isolate
meteorological variables present during the maximum those variables that appeared to have the largest
averages showed no obvious concurrent, key disagreement between NWS and WRF approaches
meteorological drivers present for either approach. (Figure 7). The FB was calculated for both calm and
For the highest 24-hour average concentrations that no-calm conditions. Once the key variables were
occurred on the same day, both approaches had isolated, mean bias, mean error, the root mean
similar values for wind speeds and direction, MBL square error (RMSE), and monthly averages were
heights, and CBL heights. Again, there were no calculated to further evaluate performance (Table 2).
obvious variables that stood out as key drivers. For the key variables, most of the results of the
statistics and averages reflected calm conditions
In an attempt to isolate the key meteorological drivers since results from no-calm conditions were very
that were producing the maximum concentrations, a similar.
linear regression analysis was performed on both
meteorological datasets. At first, the analysis Based on the results of the FB analysis, the following
revealed inconclusive results as each approach surface parameters had a disagreement by at least a
yielded different key predictors. The analysis of the factor of 2 (+/- 0.67): sensible heat flux (H), Bowen
NWS approach estimated w* as the key predictor ratio (B0), surface roughness length (z0), Monin-
whereas the WRF approach estimated VPTG as the Obukhov length (L), and convective velocity scale
most significant predictor. A closer look at the (w*). The largest disagreement was with L (1/L) as
AERMET output from both approaches revealed that FB results showed almost no agreement. However,
the NWS data contained approximately 20% calm L, H, and B0 ratio can become unbounded with large
wind conditions. Calm conditions are essentially extremes and vary widely throughout a day.
treated as missing in AERMOD and the model will not Therefore, further evaluation of these parameters
calculate any derived boundary layer parameters or using other statistical measures and monthly patterns
concentrations during calms. Therefore, all hours was crucial.
containing calms were removed from the NWS
dataset with corresponding hours removed from the The Monin-Obukhov length (L) is a derived convective
WRF dataset and another regression analyses was boundary layer (CBL) parameter that is estimated in
performed on the matched data set. The results of AERMET through an iterative procedure with surface
this analyses showed that the MBL height was the friction velocity, u*. It is a stability parameter that
most significant predictor for both approaches (Table represents the height above which convectively driven
1). Clearly, calms strongly influenced the statistical turbulence dominates over mechanically driven
analysis and model performance. However, the turbulence (USEPA 2000). In AERMET, H and u* are
disparity between the level of importance attributed to first calculated, and then used to calculate L using
2
MBL in the NWS versus WRF AERMOD (partial r of similarity theory. The inverse length is more
0.49 and 0.15, respectively) results was striking. commonly used in evaluations as L can become
unbounded and large. The inverse length is more
To further evaluate the performance of both often expressed in relation to surface roughness
approaches and potentially explain the patterns seen length, z0. Large negative values of z0/L correspond
in the concentration averages, several statistical to large instability due to buoyancy; positive values
indicators were calculated for the AERMET output correspond to stable conditions (Golder 1972). For all
variables. For screening purposes, a measure of comparisons involving L, either the inverse or z0/L
performance recommended by the USEPA (USEPA was used. Statistical evaluation of a parameter that
1992) is the fractional bias (FB): has widely varying, large extremes can often produce
misleading conclusions. To complicate matters,
2 NWS-AERMET only computes L during non-calm
. conditions whereas the WRF-AERMOD tool
calculates it for all wind speeds. Therefore, a Pasquill
stability analysis which compared 1/L with z0 only
It is often considered a useful statistic since it is during non-calm conditions was performed (Figure 8).
symmetrical and bounded with limits from -2.0 It revealed that the NWS approach kept a slightly
(extreme under-prediction) to +2.0 (extreme over- more stable regime than WRF with as much as 5%
prediction). Bias values equal to +/-0.67 reveal over / more E-F stability classes present throughout the
under-prediction by a factor of 2, whereas bias values 2002 year. A monthly breakdown of the stability
close to 0 (zero) indicate a high level of agreement classes for each approach using the criteria of 1/L > 0

3
for stable conditions showed that NWS tended to stay direct user-calculation via reference tables (USEPA
more stable than WRF, especially during the winter 2000). Either procedure estimates hourly B0 values
months. that are dependent on time (i.e., annually, seasonally,
or monthly) and by land use characteristics that are
Sensible heat (H) is energy transferred between the defined within user-specified wind sectors. The mean
surface and air when there is a difference in bias results showed that NWS-AERMET estimated
temperature between them (Wallace and Hobbs higher values of B0 than the WRF-AERMOD tool.
2006). It can be negative or positive, with negative Average monthly B0 values confirmed that the NWS
values usually indicative of warmer air temperatures approach estimated higher values for all months
than the surface (inversions). It is important in except March (Figure 10). From January-February
AERMET as it is used to calculate several and in December, the NWS-estimated B0 values were
parameters: u*, w*, CBL, and L. In WRF, it is directly approximately four (4) times higher than WRF-
calculated and output. When NWS data is used, estimated values. The WRF-estimated values
AERMET estimates it using the user-defined Bowen typically ranged from 0.2 during the summer to 0.6 in
ratio and observed net radiation. If net radiation is not the winter, with NWS values ranging from 0.3 in the
available, solar radiation, temperature, and cloud summer to 1.6 in the winter. Given typical
cover are used. The mean bias, mean error, and documented winter B0 values of 1.0 - 1.6 for urban /
RMSE evaluations were difficult to interpret given grassland categories, it appears as if the WRF-
extreme limits on H; they tended to show that on AERMOD tool underestimated values, resulting in
average, the WRF-AERMOD tool had larger positive extra evaporation from surface wetness in addition to
values of sensible heat than NWS (Table 2). A the normal transpiration rates which would keep the
monthly analysis of day/night heat fluxes between boundary layer more moist throughout the year (Paine
WRF and NWS revealed that WRF kept positive 1987; AERMET 2000). It was discovered that the
values of sensible heat flux both day and night Bowen ratio calculated from WRF’s sensible and
throughout the year (Figure 9). For the NWS latent heat fluxes can sometimes become extremely
approach, large negative night fluxes dominated the large, unlike the NWS-AERMET approach which
monthly averages from January-March and defines limited B0 values. Since analyses of H tended
November-December. Since negative heat fluxes to show that WRF had higher values than NWS,
correspond to a more stable boundary layer, NWS- larger latent heat flux values must have dominated
AERMET was estimating strong inversions at night the energy balance equation to keep B0 smaller.
during the winter and thus keeping a more stable However, WRF uses H to derive B0 - unlike the NWS
regime than WRF. On the flip side, NWS-AERMET approach which uses B0 to derive H. Therefore, its
estimated larger average fluxes than WRF in the effects on derived boundary layer parameters were
summer, which reflected a more unstable boundary most likely less significant. Furthermore, studies
layer. The discovery that the WRF-AERMOD show that AERMOD concentrations are not highly
estimated no negative heat flux values was alarming. sensitive to changes in r and B0 (Karvounis et al.
Sensible heat flux remains positive at night only under 2007).
certain conditions (i.e., warm air advection). Further
investigation revealed that WRF does directly output The surface roughness length, z0, was another
both positive and negative values. When heat flux surface parameter that appeared to show much
values are read into the AERMET subroutine disagreement between WRF and NWS. The z0 is the
MPPBL.F, the program sets a positive value (limit of height at which the mean horizontal wind speed is
2
0.1 W/m ) under convective conditions. However, the zero. It quantifies the obstacles to wind flow from
AERMET subroutine changed WRF’s negative heat various surfaces. Typical values range from 0.001
flux values to the positive limit even when convective meters over water to 0.2-3.0 meters for an urban area
conditions were not present. This important finding (Stull 1988). In WRF, z0 is determined from the
reveals that there may be some inconsistency in the model’s land use reference tables and remains
AERMET criteria used to derive and limit heat flux constant for all hours at a particular location. For the
values. Once this discovery was known, a second NWS approach, z0 is determined by the
WRF-AERMOD run was made setting the limited H AERSURFACE tool or by direct user estimation using
values to negative. Results showed no change in the AERMET user’s guide land use reference tables.
pollutant concentrations. As with B0, either estimation produces hourly z0
values which may vary substantially depending on the
The Bowen ratio (B0) is the ratio of energy fluxes by seasonal frequency and land surface characteristics.
sensible and latent heating and is sensitive to The AERSURFACE-estimated values vary by season;
boundary layer surface moisture. The AERMET they do not remain constant for all hours. For the ALF
processor uses B0 to calculate H, which is further site, WRF estimated a constant z0 of 0.2 meters;
used to calculate additional boundary layer NWS estimated z0 that varied from 0.3 meters in the
parameters (as previously mentioned). Since WRF summer to 1.6 meters in the winter (Figure 11). The
directly calculates sensible and latent heat fluxes, B0 NWS-AERMET’s large variation in z0 seemed
was easily derived. For the NWS approach, B0 was counterintuitive. For a downtown urban site, z0 would
estimated by use of the AERSURFACE tool or by be higher from the close proximity of structures

4
(buildings), but it would vary little by season. Having deep of a MBL rivaled the CBL, and it caused much
a seasonal variation would indicate vegetation more plume dilution for stable plumes which most
changes, but z0 would be higher in the summer than likely lowered pollutant concentrations and kept
winter. Average bias results confirmed that WRF plumes closer in to the source.
estimated lower z0 values than NWS (Table 2).
Typically, higher z0 values correspond to more wind For the CBL, first glance of the mean bias, error, and
flow obstacles which, in turn, enhance turbulence and RMSE results showed that WRF tended to keep
ultimately affect mixing height (Wallace and Hobbs slightly higher CBL heights than the NWS approach
2006). The increased turbulence associated with (Table 2). However, monthly averages of CBL
higher roughness lengths increases the dispersion of heights revealed that NWS-AERMET estimated
the plume away from the centerline height. The higher CBL heights in the summer (Jun-Sep) and
AERMOD model is extremely sensitive to changes in lower CBL heights in the winter when compared to
surface roughness and wind speeds; small changes WRF (Figure 13). This is further confirmed by a
in these variables have been shown to affect the frequency distribution of CBL heights which showed
distance within which concentration limits are that the NWS approach estimated a larger percentage
exceeded by several hundred meters (Grosch and of very low and extremely high CBL heights (Table 3).
Lee 1999; Faulkner et al 2008). Studies have shown The NWS-AERMET maximum estimated CBL heights
that lower z0 values in AERMOD tend to increase were roughly 300m higher than WRF. These
short-term concentrations (1-hour, 3-hour) and lower extremes were coupled with the discovery that NWS-
the long-term maximum averages (24-hour, annual) AERMET estimated fewer convective hours during
(Grosch and Lee 1999). the day than WRF. Throughout the year, the WRF-
AERMOD tool estimated approximately 2-4 more
In order to estimate turbulence in the CBL, AERMET convective daytime hours than NWS-AERMET. So,
calculates a convective velocity scale, w*, which is even though the WRF-AERMOD tool calculated more
directly proportional to H and the CBL and inversely hours of CBL heights, the NWS approach had more
proportional to temperature (T) and density (ρ). It extremes. The CBL results coincide with NWS-
indicates the amount of turbulent kinetic energy in AERMET’s larger w* and higher H during the
thermal updrafts in the CBL, with typical values for summer.
deep layers on the order of 1-2 m/s (Stull 1988). It is
only calculated during daytime convection. Both the Finally, based on AERMOD’s high sensitivity to winds
mean bias results and the monthly averages of w* (Steib 2005), the distribution of surface wind fields
indicated that the WRF-AERMOD tool tended to from both approaches was evaluated. Wind roses
produce lower values of w* than NWS-AERMET were constructed to determine the frequency of wind
(Figure 12). The WRF-AERMOD tool’s lower w* speed and direction throughout the year (Figure 15).
values were most likely due to the consistently lower Calm conditions in the NWS dataset (< 1.3 m/s) were
estimated CBL heights and a more moist boundary excluded since calms did not influence
layer. Lower values of w* typically indicate a more concentrations. The wind analysis showed that NWS-
neutral regime during daytime convection (Wallace AERMET winds tended to blow either from the
and Hobbs 2006; Stull 1988). SSW/S or the NNE/NE most of the time; these
directions comprised almost 50% of the total hourly
Even though the FB results did not indicate large MBL winds. The WRF-AERMOD tool’s wind distribution
and CBL disagreement, monthly average snapshots was more evenly spread; the largest percentage of
of the data combined with additional statistics were the distribution (24%) came from the SW/SSW/S
computed for the two variables. The MBL was directions. A frequency analysis confirmed that WRF
especially examined since the regression analysis tended to have slightly higher wind speeds for all wind
seemed to indicate that it was one of the most speed categories. The higher wind speeds could be a
important predictors of concentrations. According to possible explanation for the greater dispersion of
the mean bias results, the WRF-AERMOD tool pollutants from the WRF-AERMOD tool. With more
generally tended to produce lower MBL heights than persistent wind directions, maximum pollutant
NWS-AERMET. A frequency analysis of plume rise in concentrations are more likely to impact the same grid
the MBL between both approaches revealed several cells, in turn producing higher long-term averages.
important findings (Figures 14). First, NWS-AERMET The NWS-AERMET’s more restrictive wind
estimated low plume heights - well below the MBL - distribution was another likely explanation for the
throughout the year, whereas the WRF-AERMOD tool higher long-term maximum average concentrations.
estimated higher plume heights with some above the
MBL (July). Furthermore, the average MBL heights The wind analyses unexpectedly revealed that there
from NWS-AERMET were significantly higher than the was some discrepancy between the NWS raw wind
WRF-AERMOD tool, especially in the winter. With data and the AERMET-produced wind data. Hourly
typical MBL heights ranging a few hundred meters raw NWS input variables did not always correspond to
deep (Yu 1977), NWS-AERMET estimated values the AERMET output variables. This discovery led to
unrealistically high, with averages reaching 2000 an investigation of the NWS raw input data and the
meters in December, January, and February. That AERMET-produced data to determine the extent of

5
the differences between datasets. The AERMOD analysis seemed to indicate that MBL heights are a
modeling system does use smoothing in derived key player in predicting concentrations.
boundary layer parameters. However, substantial
differences between input and output meteorological Bowen ratio (B0) values between both approaches
variables would not be expected. Wind direction were significantly different. Based on typical
(WDIR), wind speed (WS), temperature (T), cloud documented ranges for B0 in urban areas, the WRF-
cover (CC), and surface pressure (PRES) were AERMOD tool appeared to underestimate values,
statistically compared between raw input and keeping a more moist boundary layer than NWS. It
AERMET output (Table 4). Substantial differences was discovered that the Bowen ratio calculated from
were seen in winds and cloud cover with smaller WRF’s sensible and latent heat fluxes sometimes
differences seen with pressure. became extremely large, unlike the NWS-AERMET
approach which set limited B0 values. It was
5. CONCLUSIONS inconclusive as to the significance WRF’s
underestimated B0 values had on estimated
Comparisons between the WRF-AERMOD tool and concentrations. Unlike the NWS approach which
the current EPA-recommended NWS approach used B0 to derive H, the WRF-AERMOD tool had H
revealed several significant findings about the and derived B0. With the AERMET processor relying
performance of each option. For the 2002 episode, on H to derive additional boundary layer parameters,
the WRF-AERMOD tool estimated higher short-term obtaining representative H values would seem to be
maximum concentrations (< 24 hour) whereas NWS- of greater importance. However, several studies have
AERMOD estimated higher long-term maximum shown that AERMOD concentrations are not as highly
concentrations (>= 24 hour). The spatial plots also sensitive to changes in H, r, and B0 as they are to
revealed that even though NWS produced a higher surface roughness length, z0 (Karvounis et al. 2007).
maximum 24-hour average, WRF-AERMOD tool The AERMOD users manual states, “The sensible
estimated a larger distribution of higher heat flux, Bowen ratio and albedo are not used by
concentrations. These patterns were most likely due AERMOD, but are passed through by AERMET for
to the differences seen in estimated z0, wind information purposes only” (USEPA 2004).
distribution, and MBL heights. Studies have shown
that pollutant concentrations tend to be highly The significant number of calms (20%) in the NWS
sensitive to variations in z0 and wind speeds in dataset made direct comparisons with WRF very
particular (Faulkner et al. 2008). Furthermore, difficult and compromised some of the accuracy of the
AERMOD studies have revealed that obtaining analyses since the NWS approach did not calculate
representative z0 values is crucial to model accuracy boundary layer parameters during calm conditions -
(Karvounis et al. 2007; Grosch and Lee 1999). unlike WRF which calculated parameters for every
Changes in z0 generate changes in mixing heights hour (with the exception of those calculated only
and alter the profiles of various meteorological during daytime convection). More importantly,
parameters; therefore, representative z0 values are estimated concentrations were not calculated during
crucial. Both approaches relied on reference tables calm conditions. Removing 20% of the hourly
for z0 estimation. With typical z0 values for urban estimations from the dataset undoubtedly affected the
environments documented around 0.2-1.0 meters and resulting concentration averages. The WRF-
having little seasonal variation (USEPA 2004), it AERMOD tool’s ability to output all hourly variables
appears as if the NWS approach did not represent the with the exception of those that are calculated during
ALF site most accurately. daytime convection is a major advantage over the
NWS-AERMOD approach. Even though the upper air
Both the WRF-AERMOD tool and the NWS approach variables were not analyzed, the WRF-AERMOD
estimated the majority of highest concentrations in tool’s 27-layer atmospheric profile was another
December. For both approaches, December had the significant improvement over the NWS-AERMOD 1-
lowest average CBL heights, w*, and H; and had the layer profile. A detailed vertical stratification
highest RH and MBL heights - all indicators of a unquestionably produced a more representative
stable regime. The largest disagreement between atmospheric profile than the “standard” approach.
approaches appeared to be with the derived boundary
layer parameters. Furthermore, all of the derived The most important finding in the analyses was the
boundary layer parameters were interdependent, so a result of the CFD which revealed that the WRF-
variation in one variable greatly influenced the AERMOD tool estimated 24-hour SO2 concentrations
variation of another variable. Analyses of the stability more closely to observed than the NWS approach.
parameters showed that NWS-AERMET remained The significant number of calms and missing data, the
more unstable in the summer and more stable in the unrealistic MBL heights, the highly variable z0, and a
winter than WRF, with unrealistically deep MBL 1-layer upper air profile most likely impaired the ability
heights. The large MBL heights allowed for of the NWS-AERMOD approach to fully capture
substantial plume dilution for the stable plumes which boundary layer and pollutant behavior. Furthermore,
most likely lowered pollutant concentrations. This is a estimation of surface characteristics via land use
significant discovery, especially since the regression tables or user estimation as required by both

6
approaches may not be the best method, especially 100%

for z0 as the parameter has been documented as 90%

highly influencing concentration averages (Karvounis 80%

SO2 Concentrations  (ug/m3)
et al. 2007). Finally, it is uncertain whether the NWS 70%
approach is truly representative of a modeled site 60%
when the source lies in a location far removed from 50% Observed
an NWS station or other suitable meteorological 40%
NWS

monitoring site. Locally-influenced meteorological 30%

WRF

conditions such as radiation, cloud cover, 20%

precipitation, and winds may differ widely between an 10%
existing or potential source and meteorological data 0%
used for dispersion analysis. Some refinements to

2.7
3.1
5.3
8.1
8.6
9.0
9.6

More
10.1
10.8
11.6
12.3
13.0
13.9
15.2
15.8
17.3
20.0
23.9
26.0
31.3
33.7
37.7
63.6
the WRF-AERMOD tool’s approach in estimating B0
values closer to those documented in literature as Figure 2: Cumulative Frequency Distributions of 24-
well as modification to the AERMET subroutine to hour AERMOD SO2 Concentrations from NWS-
keep H values negative and limit B0 are necessary. AERMOD and WRF-AERMOD at ALF Monitor
Overall, the tool seemed to provide a better (Independent of Receptor Location).
representation of boundary layer physics and
pollutant behavior at the ALF site than the current
EPA-recommended approach. With some studies 18
backing WRF over MM5 in capturing boundary layer NWS WRF 16.29
16
processes (Kwun et al. 2009) and with MM5 no longer
supported by its developers, the WRF-AERMOD tool 14

may prove to be a valuable meteorological processing 12

11.93 11.91

tool for AERMOD applications. Final assessment of

10
tool performance would definitely require further 8.14
analyses using multiple sites and episodes. 8 7.22

4
6. FIGURES AND TABLES 2.46
2

6.1 Figures 0
BIAS ERROR RMSE

Figure 3: Mean Bias, Mean Error, and Root Mean

Square Error (RMSE) Statistics of NWS-AERMOD
and WRF-AERMOD estimated 24-hour SO2
Concentrations at ALF Monitor (Independent of
Receptor Location).

Figure 1: AERMOD domain with ALF location.

AUG 5 DEC 26 DEC 11
MAR 17 DEC 30 AUG 28 DEC 18

Figure 4: χ/Q estimates of pollutant concentrations

for different averaging periods from NWS-AERMOD
(blue) and WRF-AERMOD (red). Dates (yellow
boxes) indicate days reporting highest concentrations.

7
 80%
NWS 75%
WRF
70%

60%
Percent Distribution

50% % STABILITY A‐B C‐D E‐F

NWS 0.2% 93.1% 6.6%
WRF 0.1% 98.5% 1.4%
40%

30%
26%
21%
20%
16% 14%
11% 11%
10% 7% 5% 1/L > 0 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
3% 4%
2% 2% NWS 5.8% 4.8% 5.4% 4.6% 3.7% 2.9% 2.6% 3.6% 3.4% 4.6% 5.0% 6.6%
0% 1% 0%0% 0% 1% 0% 0% 1%0% 0%
0% WRF 4.7% 4.1% 4.9% 4.5% 3.2% 2.7% 2.4% 3.5% 3.2% 4.3% 4.5% 5.2%

JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
Figure 8: Pasquill Stability Analysis comparing z0 with
th
Figure 5: Monthly Distribution (%) of the Highest 98 1/L from NWS-AERMET and the WRF-AERMOD tool.
Percentile of Concentrations from NWS-AERMOD
(blue) and WRF-AERMOD (red).
60

NWS
50 WRF

40

30
Sensible Heat Flux (W/m2)

20

10

NWS
0

Figure 6: Spatial Plot of 24-hour Average JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Concentrations from WRF-AERMOD and NWS- ‐10

AERMOD ‐20

‐30

Figure 9: Average monthly sensible heat flux (H)

values from NWS-AERMET and the WRF-AERMOD
tool.

1.8

NWS WRF
1.6

1.4

1.2
Bowen Ratio (B0)

0.8

0.6

0.4
Figure 7: FB Results for Calm Conditions and No
Calm Conditions. 0.2

0
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Figure 10: Average Monthly Bowen Ratio (B0) values

from NWS-AERMET and the WRF-AERMOD tool.

8
 1.6
NWS WRF NWS
WRF
1.4

1.2
Surface Roughness (m)

0.8

0.6

0.4

0.2

0
Figure 14: Average Monthly Plume Rise and MBL
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC Heights from the WRF-AERMOD tool and NWS-
Figure 11: Average Monthly Surface Roughness (z0) AERMET.
values from NWS-AERMET and the WRF-AERMOD
tool.

1.4

NWS
WRF
1.2

1
NWS WRF
Convective Velocity (m/s)

0.8

0.6
Figure 15: Wind Roses and Wind Speed Distribution
0.4
from NWS-AERMET and the WRF-AERMOD tool.

0.2

6.2 Tables
0
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
CALMS INCLUDED
Figure 12: Average Monthly Convective Velocity (w*) STEP
REGRESSION ANALYSIS ‐‐ NWS DATA
VARIABLE PARTIAL R‐SQ MODEL R‐SQ STEP
REGRESSION ANALYSIS ‐‐ WRF DATA
VARIABLE PARTIAL R‐SQ MODEL R‐SQ
values from NWS-AERMET and the WRF-AERMOD 1
2
w*
u*
0.4415
0.2672
0.4415
0.7087
1
2
VPTG
MBL
0.0409
0.0536
0.0409
0.0945
tool. 3
4
H
T
0.0024
0.0029
0.711
0.7139
3
4
WS
1/L
0.0654
0.007
0.1599
0.1669
5 MBL 0.0026 0.7166 5 CBL 0.0036 0.1705
6 z0 0.0014 0.718 6 T 0.0086 0.1791

CALMS REMOVED
REGRESSION ANALYSIS ‐‐ NWS DATA REGRESSION ANALYSIS ‐‐ WRF DATA
1200 STEP VARIABLE PARTIAL R‐SQ MODEL R‐SQ STEP VARIABLE PARTIAL R‐SQ MODEL R‐SQ
1 MBL 0.4853 0.4853 1 MBL 0.1520 0.152
NWS CBL 2 CBL 0.3494 0.8348 2 VPTG 0.0159 0.1679
WRF CBL 3 1/L 0.0084 0.8431 3 CBL 0.0138 0.1817
4 w* 0.0052 0.8484 4 WS 0.0077 0.1894
1000
5 VPTG 0.0429 0.8913 5 u* 0.0217 0.211
6 u* 0.0061 0.8974 6 T 0.0035 0.2145

Table 1: Results of the Regression Analysis with

800
Calms Included and Calms Removed.
CBL Height (meters)

600

400

200

0
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Figure 13: Average Monthly CBL heights (CBL) from

NWS-AERMET and the WRF-AERMOD tool.

9
AERMET VARIABLE MEAN BIAS MEAN ERROR FRACT BIAS FRACT ERROR RMSE Karvounis G., D. Deligiorgi and K. Philippopoulos,
H 25.29 42.03 ‐1.11 3.47 53.75 2007: On the Sensitivity of AERMOD to Surface
U* ‐0.18 0.26 ‐0.28 1.00 0.36 Parameters under Varios Anemological Conditions,
W* ‐0.73 0.73 ‐1.23 2.49 0.87 Proceedings of the 11th International Conference on
VPTG 0.01 0.01 0.06 0.83 0.03
Harmonisation within Atmopsheric Dispersion
CBL 150.91 381.13 0.34 1.23 500.69
MBL ‐522.18 680.41 ‐0.42 1.36 973.74
Modelilng for Regulatory Purposes, 5 pp.
1/L 0.002 0.013 ‐1.657 3.523 0.053
Z0 ‐0.54 0.54 ‐0.91 1.85 0.73 Kwun, J.H., Y.K. Kim, J.W. Seo, J.H. Jeong and S.H.
B0 ‐0.52 0.72 ‐1.24 2.91 1.77 You, 2009: Sensitivity of MM5 and WRF mesoscale
R ‐0.02 0.04 ‐0.03 0.24 0.09 model predictions of surface winds in a typhoon to
WS 0.13 1.32 0.01 0.71 1.69
planetary boundary layer parameterizations, Journal
T(K) ‐1.82 2.48 ‐0.01 0.02 3.04
PAMT ‐0.11 0.29 ‐0.03 0.49 1.63
Natural Hazards. ISSN 0921-030X (Print), 1573-0840
RH 2.11 10.10 0.04 0.32 13.09 (Online). vol 51, no. 1, pp. 63-77.
PRES ‐5.41 5.44 ‐0.01 0.01 5.73
CCVR ‐0.39 2.55 0.01 1.28 3.62 Paine, R. J., 1987: User's Guide to the CTDM
Table 2: Mean Bias, Mean Error, Fractional Bias, Meteorological Preprocessor (METPRO)
Fractional Error, and RMSE (WRF-NWS) for Program. EPA-600/8-88-004, U.S. Environmental
AERMET variables. Protection Agency, Research Triangle Park, NC.
(NTIS No. BP 88-162102).

CBL HEIGHT DISTRIBUTION Schroeder, Tony and G. Schewe, 2009: Sensitivity of

AERMOD to Meteorological Data Sets Based on
MET CBL<100m CBL>500m CBL>1000m CBL>1500m CBL>2000m
Varying Surface Roughness. Presented at 2009
NWS 4.6% 23.0% 11.1% 4.7% 1.1% Annual Air and Waste Management Association
WRF 2.1% 31.9% 15.2% 4.9% 0.8% Conference, June 18, 2009. Paper No. 2009-A-168-
Table 3: Distribution of CBL heights from NWS- AWMA, 15 pp.
AERMET and the WRF-AERMOD tool.
Steib R., 2005: Regulatory Modelling Activity in
Hungary, Advances in Air Pollution Modelling for
Environmental Security, pp 337-347.

Stull, R.B., 1988: An Introduction to Boundary Layer

Meteorology, 666 pp.

Wallace J.M. and P.V. Hobbs, 2006: Atmospheric

Science, Volume 92, Second Edition: An Introductory
Survey (International Geophysics), 505 pp.

USEPA, 2000: Revised Draft User’s Guide for the

AERMOD Meteorological Preprocessor (AERMET),
U.S. Environmental Protection Agency.

Table 4: Comparison of Basic Meteorological USEPA, 1992: Protocol for Determining the Best
Variables from NWS Raw Data and AERMET Output Performing Model, U.S. Environmental Protection
Data. Agency.

Yu, T.W., 1977: Determining Height of the Noctournal

7. REFERENCES Boundary lLayer. Journal of Applied Meteorology.
vol. 17, pp. 28-33.
Faulkner W.B., B.W. Shaw, and G.T. Grosch, 2008:
Sensitivity of Two Dispersion Models (AERMOD and
ISCST3) to Input Parameters for a Rural Ground-
Level Area Source, Journal of the Air & Waste
Management Association. ISSN 1096-2247. vol. 58,
no. 10, pp. 1288-1296.

Grosch, G.T. and R.F. Lee, 1999: Sensitivity Of The

AERMOD Air Quality Model To The Selection Of Land
Use Parameters, WIT Transactions on Ecology and
the Environm., 37.

10