Journal of the Air & Waste Management Association

ISSN: 1096-2247 (Print) 2162-2906 (Online) Journal homepage: http://www.tandfonline.com/loi/uawm20

Statistical evaluation of a new air dispersion

model against AERMOD using the Prairie Grass
data set

Fernando Augusto Silveira Armani , Ricardo Carvalho de Almeida & Nelson

Luís da Costa Dias

To cite this article: Fernando Augusto Silveira Armani , Ricardo Carvalho de Almeida & Nelson
Luís da Costa Dias (2014) Statistical evaluation of a new air dispersion model against AERMOD
using the Prairie Grass data set, Journal of the Air & Waste Management Association, 64:2,
219-226, DOI: 10.1080/10962247.2013.852996

To link to this article: http://dx.doi.org/10.1080/10962247.2013.852996

Published online: 21 Jan 2014.

Submit your article to this journal

Article views: 302

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=uawm20

Download by: [78.97.116.67] Date: 12 July 2017, At: 00:28

TECHNICAL PAPER

Statistical evaluation of a new air dispersion model against AERMOD

using the Prairie Grass data set
Fernando Augusto Silveira Armani, Ricardo Carvalho de Almeida,*
and Nelson Luís da Costa Dias
Universidade Federal do Paraná, Programa de Pós-Graduação em Engenharia Ambiental, Centro Politécnico, Curitiba, PR, Brazil
⁄Please address correspondence to: Ricardo Carvalho de Almeida, Universidade Federal do Paraná, Programa de Pós-Graduação em Engenharia
Ambiental, Centro Politécnico, Caixa Postal 19022, Curitiba, PR, CEP: 81531-980, Brazil; e-mail: rcalmeida@ufpr.br

In this work, the authors present a statistical assessment of two atmospheric dispersion models. One of them, AERMOD
(American Meteorological Society/Environmental Protection Agency Regulatory Model), adopted by the U.S. Environmental
Protection Agency, is widely used in many countries and here is taken as a baseline to assess the performance of a newly
proposed model, MODELAR (Modelo Regulatório de Qualidade do Ar). In terms of parameterizations and modeling options,
MODELAR is a somewhat simple model. It is currently being considered for adoption as the regulatory model in Paraná State,
Brazil. The well-known Prairie Grass data set, already used in earlier evaluations of the same version of AERMOD analyzed here,
was used to perform model assessment. The evaluations employed well-established statistical performance descriptors and
techniques. The results indicate that MODELAR is a slightly better predictor, for the Prairie Grass data set, of concentrations
under unstable conditions, whereas AERMOD has a better performance under near-neutral and stable conditions. Moreover, cases
of severe overestimation and underestimation, as detected by the Factor of Two index, are clearly associated with extreme stability
conditions (both unstable and stable), stressing the need for better parameterizations under these conditions.

Implications: Statistical evaluation of a proposed numerical dispersion model, called MODELAR, and of the well-established
dispersion model AERMOD was performed, using the Prairie Grass data set as reference. The results showed that MODELAR
performs well under unstable conditions, making it competitive with other dispersion models. However, the results also showed that
both models require improvement of the dispersion parameterizations for extreme stability conditions, both unstable and stable.

Introduction
In this work, we present a new dispersion model that is under
development as a candidate for the regulatory air dispersion
model of Paraná State, Brazil. Given the widespread use of
U.S. Environmental Protection Agency (EPA) models in Brazil
for licensing purposes, the new model is compared against EPA’s
AERMOD (American Meteorological Society/Environmental
Protection Agency Regulatory Model; Cimorelli et al., 2004a)
using a commonly used data set for dispersion model evaluation
(Prairie Grass; Barad, 1958). Both the Prairie Grass data set and
the AERMOD simulation results used in this work were obtained
by the work of Olesen et al. (2007).
Most regulatory atmospheric dispersion models in use are
based on gaussian approximations. A basic assumption of those
models is that wind speed and turbulence diffusion coefficients
are constant to produce a gaussian concentration distribution
about the plume’s centerline with lateral and vertical standard
deviations sy and sz, respectively, which are usually estimated
from meteorological data (Hanna et al., 1982; Tirabassi, 1989).
Such gaussian models have the advantage of being analytical,
easy to implement and apply, and have performed reasonably
well in several field experiments where they were put to test
(Weil and Jepsen, 1977; Harrison and Mac Cartney, 1980).
However, in their original form they failed to consider the
mean wind variation with height and the nonhomogeneity of
turbulence. It has been recognized since the 1960s that these
simplifications are inadequate in applications where the wind
speed varies considerably with height, or in conditions close to
local free convection (Gifford, 1968; Pasquill, 1974).
In an effort to address these shortcomings, several studies
were developed to include the effects of mean wind and turbulent
diffusion dependence with height (Demuth, 1978; Huang, 1979;
Lin and Hildemann, 1996). The resulting family of analytical
models includes the current regulatory models in use by the EPA,
the AERMOD modeling system (Cimorelli et al., 2004a) and
CALPUFF modeling system (Scire et al., 2000).
However, although there are many turbulence and wind para-
meterizations that show reasonable agreement with observa-
tional data (Ulke, 2000; Demuth, 1978; Huang, 1979; Lin and
Hildemann, 1996), analytical models use simpler approxima-
tions of these profiles to obtain exact solutions, such as
AERMOD, whose concentration formulations consider these
profiles by an equivalent (effective) value constructed by aver-
aging over the layer through which plume material travels
directly from the source to receptor (Cimorelli et al., 2004b).

219
Journal of the Air & Waste Management Association, 64(2):219–226, 2014. Copyright © 2014 A&WMA. ISSN: 1096-2247 print
DOI: 10.1080/10962247.2013.852996 Submitted May 3, 2013; final version submitted September 23, 2013; accepted October 1, 2013.
220 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226

In this work, in an effort to overcome those simplifications The numerical solution employs an implicit trapezoidal
inherent to some analytical solutions, the dependence with method, which allows relative large grid spacing in x. The
height of the mean wind and turbulent diffusivity is represented horizontal grid spacing is constant in x. The grid points in the
explicitly in a numerical solution of the advection-diffusion vertical are unevenly distributed, with smaller spacing close to
equation. The model has reached a stage where a first operational the ground where velocity and concentration gradients are larger.
version is available for initial tests. The operational version The solution proceeds column-wise (in the z-direction) from
includes the ability to incorporate an arbitrary number of sources the first column for x > 0 , i.e., for positive x values. Each column
at different heights, and to spread a two-dimensional (2D) solu- is solved with the successive overrelaxation iterative method.
tion laterally with a gaussian plume, using the same strategy of The iterations stop when the relative error in concentration is
the models AERMOD (Cimorelli et al., 2004b) and OML smaller than 1010.
(Operationelle Meteorologiske Luftkvalitetsmodeller) (Olesen Zero-flux boundary conditions are prescribed at the bottom
et al., 2007), but (unlike them) without inclusion of the meander (z ¼ z0 ) and top (z ¼ h) of the domain, viz.,
effect. In this paper, only the 2D version without lateral spread-
ing, and with a source located close to the ground, is used, since @C
this is enough for comparison with the Prairie Grass data. Kz ¼0 ð2Þ
@z
The evaluation of the performance of the resulting Eulerian
numerical model, called MODELAR (Modelo Regulatório de at both heights.
Qualidade do Ar), as well as its comparison with AERMOD, are For unstable conditions, the vertical turbulent diffusion is the
the central objectives of this paper. As a first step towards same as used by Velho et al. (1997):
establishing MODELAR’s capacity to predict pollutant concen-

z 13
  
trations with a skill comparable to models widely used in Brazil Kz ðZÞ z 13 4z
for regulatory purposes, we compare both MODELAR and ¼ 0:22 1 1 ¼ exp
w:h h h h
AERMOD against the Prairie Grass data set.
Prairie Grass is considered a standard database for atmo-
spheric dispersion model validation (American Society for  
8z
Testing and Materials [ASTM], 2000). With the Prairie Grass 0:0003 exp ; ð3Þ
data set, it is possible to evaluate the model performance in h
predicting the field concentration emitted continually over flat
terrain, near the ground and near the source. where
The criteria used here to assess the model’s skill follow the  13
recommendations in “Standard Guide for Statistical Evaluation w: ¼
g
wuv jo h ð4Þ
of Atmospheric Dispersion Model Performance” (ASTM, 2010)
v
and Chang and Hanna (2004). Model evaluation is performed
independently of wind direction, and only vertical dispersion is is the convective velocity scale, h is the height of the atmospheric
taken into account, with all measured concentrations averaged boundary layer,
 v is the mean virtual potential temperature, and
laterally. wuv jo is the surface kinematic virtual sensible heat flux.
In Description of Model MODELAR and Description of Reynold’s separation is written A ¼ A  þ a; with the overbar
Model AERMOD, we briefly describe the relevant aspects of indicating the mean and the lowercase letter the turbulent
AERMOD and MODELAR for this work. In sections The fluctuation.
Prairie Grass Experiment and Comparisons Criteria, we describe For stable conditions, the turbulence is characterized by local
in more detail both the data set and statistical measures of model values, such as the local Monin-Obukhov length ^ ðzÞ (Degrazia
performance, respectively. In Results, we give the main results of and Moraes, 1992):
the comparisons between models and data. Then, we summarize

the results and conclude the work in Conclusion. z 1:5a1 a2
^ ðzÞ ¼ L 1  ; ð5Þ
h
Description of Model MODELAR where L is the Monin-Obukhov length calculated with surface
flux values. For a shear-dominated stable boundary layer, a1 ¼
MODELAR solves the advection-diffusion equation numeri- 1.5 and a2 = 1.0 (Nieuwstadt, 1984). Furthermore, there is a
cally, neglecting longitudinal diffusion, aligning the x-axis with local velocity scale given by
the mean wind direction, and the z-axis with the vertical direc-
tion. The resulting differential equation is the same as solved by
z  a3
Wortmann et al. (2005): u2 ðzÞ ¼ u20 1  ; ð6Þ
h

U ð@CÞ @ 2 C ð@Kz Þð@CÞ in which a3 ¼ 1.5 and uo is the friction velocity at surface
¼ Kz 2 þ @z ; ð1Þ (Wyngaard et al., 1974).
@x @z @z
The vertical turbulent diffusion for stable or neutral condi-
where Kz is the vertical turbulent diffusion coefficient. tions is the same as used by Degrazia et al. (2000):
 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226

13
0:4 1 þ 3:7^z uz
Kz ðzÞ ¼
43
1 þ 15 ufz0 þ 3:7^z
where f is the Coriolis parameter.
Description of Model AERMOD
AERMOD model 04300 was used in this work; this is the
same version used by Olesen et al. (2007), whose description is
given by Cimorelli et al. (2004a).
Essentially, AERMOD solves eq 1, but with a constant wind
speed and turbulent diffusion coefficient. These values, however,
are calculated from mean values of the corresponding profiles
predicted by similarity theory, for the centerline of the plume
down to the receptor. They are called “effective” parameters, and
~ and s
denoted here by a tilde, e.g., U ~.
The general expression for the calculation of AERMOD’s
concentrations is
Q height (h), Monin-Obukhov length (L), and convective velocity
Cðx; y; zÞ ¼ P ðx; yÞPz ðx; zÞ;
~ y
U scale (w*). A constant momentum roughness (z0 ¼ 0:006 m) is
where Q is the source emission rate, U ~ is the effective wind
speed, and Py and Pz are probability density functions that
describe the lateral and vertical concentration distributions,
respectively.
Under stable conditions, the plume is represented by gaussian
distributions both in the horizontal and vertical directions. Under
unstable conditions, the horizontal distribution remains gaus-
sian, whereas the vertical distribution is bigaussian, in order to
better represent the asymmetrical effect of positive and negative
vertical velocity fluctuations.
Furthermore, in the unstable boundary layer, AERMOD uses
two additional sources to describe the pollutant’s distribution
from the actual source: an “indirect” source to address the lofting
behavior and dispersion of “nonpenetrating” plumes, and a
“penetrated” source to account for plume material that initially
penetrates the inversion but subsequently fumigates into the
CBL (Convective Boundary Layer). In addition, image sources
are included to satisfy the zero flux conditions at z ¼ 0; h.
The vertical concentration distributions of the indirect and
actual sources are bigaussian, and the vertical concentration
distribution of the penetrated source is gaussian. Regarding the
distribution of the actual source, the vertical dispersion in the
z
ð u  Þ x2
unstable surface layer is parameterized according to sz a U~jLj
so as to provide good agreement between the modeled and
observed concentrations from the Prairie Grass experiment
(Cimorelli et al., 2004b).
The Prairie Grass Experiment
The data used in this work come from the Prairie Grass
experiment, which took place near O’Neil, Nebraska, during
the summer of 1956. The site altitude is 596.37 m and its location
is 42 29.60 N, 98 34.30 W (Barad, 1958).
221
Sixty-eight releases “(runs)” were made under different sta-
bility conditions. For each run, SO2 tracer was continuously
ð7Þ
released without buoyancy from a height of 0.46 m above
ground, except four runs 65, 66, 67, and 68, whose release height
was 1.5 m. Concentration measurements were made by receptors
placed at 1.5 m above ground and in arcs at 50, 100, 200, 400,
and 800 m from the source, and the averaging time of the
observed concentration values was 10 min. Simultaneous mea-
surements of wind and temperature profiles in the surface layer,
and radiosoundings, are also available.
From the measured velocity and temperature profiles in the
surface layer, Olesen et al. (2007) estimated a roughness length
of 6 mm. For runs 3, 4, 13, and 14, during which the wind speed
was below 1 m sec1, a much larger roughness of 10 cm was
obtained. We will follow the common practice of excluding these
runs from model evaluations using the Prairie Grass data set.
There is not an official digital source of the Prairie Grass data
set, but it can be obtained from http://www.dmu.dk/
International/Air/Models/Background/oml.htm. This is a digital
annex to the work by Olesen et al. (2007). The data set includes
for each 10-min run: friction velocity (u*0), boundary-layer
ð8Þ
adopted. Also provided at that Web site are computed values of
the crosswind integrated concentration values both from the
observed data and from modeling results for AERMOD version
04300. These are the same data as used by Marie et al. (2012)
and Sharan and Kumar (2009).
Comparison Criteria
We compare model simulations with observed concentrations
using a set of statistical scores. First, several indices defined in
Olesen (2001) are used:

2
Coy  Cpy
NMSE ¼ ð9Þ
Coy Cpy


GM ¼ exp ln Cpy  ln Coy ; ð10Þ
h i
2
GV ¼ exp ðlnCpy  lnCoy Þ ; ð11Þ


2 Cpy  Coy
FB ¼ ð12Þ
Cpy þ Coy
Fa2 ¼ Fraction of data that satisfy
Cpy
0:5< <2; ð13Þ
Coy
where NMSE is the normalized mean square error, GM is the
geometric mean bias, GV is the geometric variance, Fa2 is the
222 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226
Factor of Two, and FB is the fractional bias. The overbar indi-
y
cates means over the runs. In eqs 9–13, Co are the observed
y
crosswind integrated concentrations, and Cp are their counter-
parts predicted by the model. The best results for a model are
GM, GV, Fa2 ¼ 1.0, and FB, NMSE ¼ 0.0.
The indices defined above should be considered as a whole,
as they obviously address different kinds of errors, such as bias
(in the case of FB, GM) and dispersion (in the case of NMSE,
GV). Moreover, it is well known that all averages can be heavily
influenced by the presence of outliers (Hanna et al., 2004).
Chang and Hanna (2004) established criteria for a good dis-
persion model, based on eqs 9–13, namely Fa2 > 0.5, 0.3< FB
< 0.3, 0.7 < GM < 1.3, NMSE < 0.5, and GV < 1.6.
Due to the short averaging time of the Prairie Grass observa-
tions (10 min), we are providing comparisons of observed and
simulated crosswind integrated concentration values. This
allows evaluation of model performance for vertical dispersion.
This also avoids having to deal with problems associated with the
horizontal dispersion, which is a strong function of averaging
time, especially for averaging times of 1 hr or less. We use a ratio:
Cpy
ð14Þ
Coy
to assess the relative difference in the crosswind integrated
values as a function of downwind distance and stability.
In order to assess whether the performance statistics for the
two models are significantly different, we are using bootstrap
resampling to compare those statistics based on Student’s t test
with a 90% level of confidence, as recommended by ASTM
(2010).
In order to do that, we generated bootstrap samples composed
of the observed values and the associated simulation results for
both models, respecting the “concurrent sampling” recom-
mended by ASTM (2010). In our case, around 500 resamplings
with replacement were utilized, and the statistics given by eqs 9–
12 were calculated separately for stable and unstable runs. Each
Cy
Figure 1. Scatterplots of Q simulated by MODELAR (a) and AERMOD (b) against
bootstrapped sample size was 33 runs  5 arcs for unstable, and
30 runs  5 arcs for stable conditions, the same values of the
experiment.
Each bootstrapped sample generates a difference 
between the performance statistics for the two models. We
compute the average ðÞ  and standard deviation (s) to

obtain the t value, t ¼ s . When t is larger than 1.645, it
is possible to conclude with 90% of confidence that there is
a significant difference between the two models for the
particular statistic being tested.
Results
Following ASTM (2010), the data are assessed separately for
unstable (L < 0) and stable (L > 0) conditions. In Figure 1a and b
we show the results of observed versus predicted crosswind
integrated concentrations values for MODELAR and
AERMOD, respectively, for unstable conditions. Then, in
Figure 2a and b results are shown for stable conditions.
Note in Figures 1 and 2 three dotted lines, which are
y
y y y y p
Cp ¼ 2 Co , Cp ¼ Co , and Cy ¼ C2o , going from bottom to top.
Furthermore, we plotted the lines of least-squares fits forced to
intercept the origin of the graph (on the linear scale) to see
quickly whether the model is over or underestimating the
observations.
As expected, for unstable conditions the highest observed and
predicted concentrations occur at the arc closest to the source (50
m), whereas for stable conditions the largest concentrations are
more spread out along the 50-, 100-, and 200-m arcs.
A key feature of Figures 1–2 is that MODELAR has a
tendency for underestimating the vertical dispersion under
unstable conditions, the opposite occurring with
AERMOD. For stable conditions, MODELAR has a tendency
for overestimating and AERMOD for underestimating the ver-
tical dispersion. Furthermore, with regards to the direct compar-
ison of the observed and predicted concentrations, those
Cy
Q observed in the Prairie Grass experiment, in unstable conditions.
 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226 223

Cy Cy
Figure 2. Scatterplots of Q simulated by MODELAR (a) and AERMOD (b) against Q observed in the Prairie Grass experiment, in stable conditions.

estimates exceeding a factor of 2 are all overestimates for
AERMOD, whereas for MODELAR there are both over- and
underestimates of the observed concentrations.
A wider comparison between predicted and observed concen-
trations consists of the scores given by eqs 9–13 and by the so-
called residual plots (Hanna et al., 2004). For unstable condi-
tions, they are shown in Table 1 and in Figure 3, and for stable
conditions in Table 2 and in Figure 4.
For unstable conditions, we find that the mean value of all
scores are slightly better for MODELAR than for AERMOD and
that they satisfy the quality criteria suggested by Chang and
Hanna (2004), reviewed in section Comparison Criteria.
In these conditions, MODELAR overestimates Cy for the 50-
and 100-m arcs, and underestimates Cy for 200 through 800 m.
Although on average MODELAR looks unbiased, there is a clear
variation in the bias as a function of distance. This implies that
MODELAR is underestimating the vertical dispersion for the
50- and 100-m arcs, and is overestimating the vertical dispersion
for 200–800 m downwind.
On the other hand, AERMOD underestimates Cy under
unstable conditions initially, with a definite trend in the bias
with distance, so that by 800 m it is overestimating Cy. This
implies that AERMOD is overestimating the vertical dispersion
until the 800-m arc. In that arc, AERMOD underestimates the
vertical dispersion.
Although the unstable surface layer parameterization used by
AERMOD was designed to provide good agreement between the
modeled and observed concentrations from the Prairie Grass
experiment, it is noticeable in Figure 3 and Table 1 that the
AERMOD predictions worsens with distance. Nieuwstadt (1980)
and Weil et al. (2012) suggest a weaker dependence of vertical
dispersion, sz, with the distance in unstable conditions and within
the convective matching region. Then, it is possible that changing
z
ðu Þ x2
the x2 dependence in sz a U~jLj to a power law might improve

Table 1. Results of model performance statistics under unstable conditions

Arc Model NMSE FA2 FB GM VG

50 m MODELAR 0.16 0.94 0.20 1.24 1.16
AERMOD 0.11 1.0 0.33 0.71 1.14
100 m MODELAR 0.10 0.97 0.04 1.02 1.07
AERMOD 0.15 1.0 0.34 0.70 1.19
200 m MODELAR 0.09 0.97 0.15 0.85 1.09
AERMOD 0.27 0.91 0.31 0.74 1.28
400 m MODELAR 0.15 0.94 0.21 0.86 1.20
AERMOD 0.30 0.76 0.16 0.92 1.38
800 m MODELAR 0.31 0.76 0.21 0.98 1.47
AERMOD 0.43 0.67 0.01 1.25 1.89
Mean MODELAR 0.16 0.91 0.07 0.99 1.20
AERMOD 0.25 0.87 0.23 0.86 1.38
All data MODELAR 0.15 0.92 0.10 0.98 1.19
AERMOD 0.13 0.87 031 0.84 1.35
224 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226

Cy
Figure 3. Residual plot of Cpy against run number for an unstable atmosphere, for Cyp
o Figure 4. Residual plots of Cyo
against run number for a stable atmosphere, for the
the 50-, 200-, and 400-m arcs.
50-, 400-, and 800-m arcs.

Table 2. Results of model performance statistics under stable conditions

difference  and the t statistic. Interestingly, for all statistics
Arc Model NMSE FA2 FB GM VG except NMSE under unstable conditions, the tests show that the
model’s performances are significantly different. Thus, for
50 m MODELAR 0.41 0.87 0.30 1.14 1.14 unstable conditions, MODELAR is better in terms of FB, MG,
AERMOD 0.55 0.87 0.39 1.20 1.20 and VG. For stable conditions, on the other hand, MODELAR is
100 m MODELAR 0.11 1.0 0.07 0.82 1.13 better for NMSE and FB, and AERMOD is better for MG and
AERMOD 0.23 0.93 0.20 1.02 1.12 VG.
200 m MODELAR 0.13 0.97 0.25 0.68 1.27 It is also important to look closely at the cases where the
AERMOD 0.22 0.93 0.20 1.02 1.12 models errors are largest: both in Figures 3 and 4 we can see that
400 m MODELAR 0.21 0.70 0.34 0.62 1.45 some runs overestimate by a factor of 2 or larger. We now turn
AERMOD 0.28 0.93 0.26 1.10 1.20 our attention to these cases and try to detect possible connections
y
800 m MODELAR 0.46 0.47 0.38 0.60 1.72 C
with atmospheric stability. Thus, Cpy is plotted against the 1/L in
AERMOD 0.54 0.87 0.33 1.22 1.39 o
Figure 5. Clearly, the largest overestimations occur for very
Mean MODELAR 0.26 0.80 0.15 0.77 1.34
unstable and for very stable conditions, i.e., for the smallest
AERMOD 0.36 0.91 0.28 1.10 1.20
values of Monin-Obukhov length.
All data
MODELAR 0.30 0.80 0.03 0.75 1.33
AERMOD 0.41 0.91 0.29 1.11 1.20 Conclusion
We performed a comparison of two models, AERMOD and
MODELAR, using the Prairie Grass Project data set. As usual,
AERMOD’s predictions. However, this was not tested, since mod- the data were analyzed separately for unstable and stable condi-
ifying AERMOD was not the focus of this work. tions. Further specific analyses to highlight different model
For stable conditions, AERMOD overestimates Cy for the 50-, capabilities were then conducted.
400-, and 800-m arcs, and is nearly perfect for the 100- and 200-m From our comparisons of observed and simulated crosswind
arcs, whereas MODELAR underestimates Cy, starting at 100-m integrated concentration values, we see that, under unstable
arc onwards. Thus, MODELAR increasingly overestimates the conditions, MODELAR is underestimating the vertical disper-
vertical dispersion as the distance downwind increases. sion for the 50-m arc; it is overestimating the vertical dispersion
Table 3 presents the statistics computed over bootstrap resam- for 200–400 m downwind, and it is nearly perfect on the other
pling. The values of the MODELAR and AERMOD statistics for arcs. Under unstable conditions, AERMOD is overestimating
both unstable and stable conditions are listed, together with the the vertical dispersion initially, but when it reaches the 800-m
 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226 225

Table 3. Model statistics, mean difference, and t ¼ s by bootstrap resampling

Unstable Stable
Parameter MODELAR AERMOD 
 t MODELAR AERMOD 
 t
NMSE 0.15 0.13 0.02 0.56 0.29 0.40 0.11 4.42
FB 0.10 0.31 0.41 9.00 0.02 0.28 0.30 10.82
MG 0.98 0.84 0.14 4.83 0.75 1.11 0.36 18.15
VG 1.19 1.35 0.16 4.48 1.33 1.20 0.13 2.65
Note: The critical value is 1.645.

overestimates/underestimates are systematically related either

to very stable or very unstable conditions, where the Monin-
Obukhov Theory does not perform well. These results indicate
the need for better parameterization of the dispersion parameters
under those conditions.

Acknowledgment
The authors thank the suggestions by three reviewers, which
helped to improve this work considerably.

Funding
This work was partially supported by IAP (Instituto
Ambiental do Paraná) through research project FUNPAR 2629,
and by the Federal University of Paraná through the Graduate
School and the Graduate Program of Environmental Engineering
(UFPR/PRPPG/PPGEA).

Cyp
Figure 5. Residual plot of Cyo
against 1/L.
arc it underestimates vertical dispersion. Under stable condi-
tions, AERMOD behaves better than MODELAR because
MODELAR tends to overestimate the vertical dispersion starting
at the100-m arc onwards.
In order to assess the quality of the predictions as a function of
distance from the source, we used residual plots for all runs in
each measuring arc, as well as standard statistics for model
performance: NMSE, GM, GV, Fa2, and FB (for definitions,
see section Comparison Criteria). Furthermore, using the
Student’s t test based on bootstrap resampling, we compared
the performance statistics between the models: they are signifi-
cantly different (in a statistical sense) most of the time, with
MODELAR being better under unstable conditions. For stable
conditions, MODELAR “wins’’ in NMSE and FB, and
AERMOD in MG and VG.
The Fa2 (Factor of Two) results show several predictions in
excess of 200%, or less than 50%, of the observed values.
Residual analysis of the data shows that these extreme
References
American Society for Testing and Materials. 2010. Standard Guide for Statistical
Evaluation of Atmospheric Dispersion Model Performance. West
Conshohocken, PA: American Society for Testing and Materials.
Barad, M.L. 1958. Project Prairie Grass, A Field Program in Diffusion.
Geophysical Research, No. 59, Vol. I.Report AFCRC-TR-58-235(I).
Bedford, MA: Air Force Cambridge Research Center.
Chang, J.C., and S.R. Hanna. 2004. Air quality model performance evaluation.
Meteorol. Atmos. Phys. 87: 167–196. doi:10.1007/s00703-003-0070-7
Cimorelli, A.J., S.G. Perry, A. Venkatram, J.C. Weil, R.J. Paine, R.B. Wilson, R.F.
Lee, W.D. Peters, and R.W. Brode. 2004a. AERMOD: A dispersion model for
industrial source applications. Part I: General model formulation and bound-
ary layer characterization. J. Appl. Meterol. 44: 682–693.
Cimorelli, A.J., S.G. Perry, A. Venkatram, J.C. Weil, R.J. Paine, R.B. Wilson, R.F.
Lee, W.D. Peters, R.W. Brode, and J.O. Paumier 2004b. AERMOD:
Description of Model Formulation. Research Triangle Park, NC: U.S.
Environmental Protection Agency.
Cox, W.M., and J.A. Tikvart 1990. A statistical procedure for determining the
best performing air quality simulation model. Atmos. Environ. 24A:2387–
2395.
Degrazia, G.A., D. Anfossi, J., C., M. Carvalho, T. Tirabassi, and H. Campos
Velho 2000. Turbulence parameterization for PBL dispersion models in all
stability conditions. Atmos. Environ. 34: 3575–3583. doi:10.1016/S1352-
2310(00)00116-3
Degrazia, G.A., and O. Moraes 1992. A model for eddy diffusivity in a stable
boundary layer. Boundary Layer Meteorol. 58: 205–214. doi:10.1007/
BF02033824
Demuth 1978. A Contribution the analytical to steady solution of the diffusion
equation for line sources. Atmos. Environ. 12: 1255–1258.
226 Silveira Armani et al. / Journal of the Air & Waste Management Association 64 (2014) 219–226
Dresser, A.L., and R.D. Huizer 2011. CALPUFF and AERMOD model valida-
tion study in the near field: Martins Creek revisited. J. Air Waste Manage.
Assoc. 61: 647–659. doi:10.3155/1047-3289.61.6.647
Gifford, F.A. 1968. An outline of theories of diffusion in the lower layers of the
atmosphere. In Meteorology and Atomic Energy, ed. D.H. Slade, 65–116.
Washington, DC: U.S. Department of Energy.
Hanna, S.R., G.A. Briggs, and R.P. Hosker Jr. 1982. Handbook on Atmospheric
Diffusion. Washington, DC: Technical Information Center, U.S. Department
of Energy.
Hanna, S.R., O.R. Hansen, and S. Dharmavaram 2004. FLACS CFD air quality
model performance evaluation with Kit Fox, MUST, Prairie Grass, and EMU
observations. Atmos. Environ. 38: 4675–4687. doi:10.1016/j.
atmosenv.2004.05.041
Harrison, R.M., and H.A. MacCartney 1980. A comparison of the predictions of
a simple gaussian plume dispersion model with measurements of pollutant
concentration at ground-level and aloft. Atmos. Environ. 14: 589–596.
Huang, C.H. 1979. A theory of dispersion in turbulent shear flow. Atmos.
Environ. 13: 453–463.
Lin, J.-S., and L.M. Hildemann 1996. Analytical solutions of the atmospheric
diffusion equation with multiple sources and height-dependent wind speed
and eddy diffusivities. Atmos. Environ. 30: 239–254. doi:10.1016/1352-2310
(95)00287-9
Marie, E.E.J., B.-B. G. Hubert, and O.A. Pierre 2012. A First-order WKB
approximation for air pollutants dispersion equation. Atmos. Res. 120– 121:
162–169.
Nieuwstadt, F. 1980. Application of mixed-layer similarity to the observed
dispersion from a ground-level source. J. Appl. Meteorol. 19: 157–162.
Nieuwstadt, F. 1984. The turbulent structure of the stable nocturnal boundary
layer. J. Atmos. Sci. 41: 2202–2216. doi:10.1175/1520-0469(1984)
041<2202: TTSOTS>2.0.CO;2
Olesen, H. 2001. Ten years of harmonization activities: Past, present, and future.
Presented at the 7th International Conference on Harmonization within
Atmospheric Dispersion Modelling for Regulatory Purposes, Belgirate,
Italy, May 28–31.
Olesen, H.R., R. Berkowicz, and P. Lofstrom 2007. Evaluation of OML and
AERMOD. Aarhus, Denmark: National Environmental Research Institute.
Pasquill, F. 1974. Atmospheric Diffusion. New York, NY: Ellis Horwood Ltd.
Scire, J.S., Robe, F.R., Fernau, M.E., and Yamartino, R.J. 2000. A User’s Guide
for the CALMET Meteorological Model (Version 5). Concord, MA: Earth
Tech, Inc. http://www.src.com/calpuff/download/CALPUFF_UsersGuide.
pdf (accessed August 2013).
Sharan, M., and P. Kumar 2009. An analytical model for crosswind integrated
concentrations released from a continuous source in a finite atmospheric
boundary layer. Atmos. Environ. 43: 2268–2277. doi:10.1016/j.
atmosenv.2009.01.035
Tirabassi, T. 1989. Analytical air pollution advection and diffusion models. Water
Air Soil Pollut. 47: 19–24. doi:10.1007/BF00468993
Ulke, A.G. 2000. New turbulent parameterization for a dispersion model in the
atmospheric boundary layer. Atmos. Environ. 34: 1029–1042. doi:10.1016/
S1352-2310(99)00378-7
U.S. Environmental Protection Agency. 2003. AERMOD: Latest Features and
Evaluation Results. EPA-454/R-03-003. Research Triangle Park, NC: U.S.
Environmental Protection Agency.
Velho, H.F.C., J.C. Carvalho, and G.A. Degrazia 1997. Nonlocal exchange
coefficients for the convective boundary layer derived from spectral proper-
ties. Atmos. Phys. 70: 57–64.
Weil, J.C., and A.F. Jepsen 1977. Evaluation of the gaussian plume model at the
Dickerson Power Plant. Atmos. Environ. 11: 901–910.
Weil, J.C., P.P. Sullivan, E.G. Patton, and C. Moeng 2012. Statistical variability
of dispersion in the convective boundary layer: Ensembles of simulations and
observations. Boundary Layer Meteorol. 145: 185–210. doi:10.1007/s10546-
012-9704-y
Wortmann, S., M. Vilhena, D.M. Moreira, and D. Buske, 2005. A new analytical
approach to simulate the pollutant dispersion in the PBL. Atmos. Environ. 39:
2171–2178. doi:10.1016/j.atmosenv.2005.01.003
Wyngaard, J.C., O. Cote, and K.S. Rao 1974. Modeling of the Atmospheric
Boundary Layer. Advances in Geophysics, 18A. San Diego: Academic Press.
About the Authors
Fernando Augusto Silveira Armani is a graduate student at the Graduate
Program in Environmental Engineering (PPGEA), Federal University of Paraná
(UFPR), Curitiba, Brazil.
Ricardo Carvalho de Almeida and Nelson Luís da Costa Dias are associate
professors at the Department of Environmental Engineering of the Federal
University of Paraná.