ARTICLE IN PRESS

Atmospheric Environment 38 (2004) 4675–4687

FLACS CFD air quality model performance evaluation with

Kit Fox, MUST, Prairie Grass, and EMU observations
Steven R. Hannaa,*, Olav R. Hansenb, Seshu Dharmavaramc
a
Hanna Consultants, 7 Crescent Ave., Kennebunkport, ME, USA
b
GexCon, Bergen, Norway
c
DuPont Company, Wilmington, DE, USA
Received 4 January 2004; received in revised form 24 May 2004; accepted 24 May 2004

Abstract

The FLACS CFD model, which can be used to estimate the flow and dispersion around buildings and other large
roughness obstacles, has been evaluated with tracer data from several field experiments involving obstacle arrays. The
following experiments were used in the evaluation exercise: Kit Fox (52 trials with puff and plume releases of slightly
dense CO2 gas in arrays of billboard-shaped obstacles), MUST (37 trials with puff releases of neutrally buoyant tracer
gas in an array of 120 shipping containers), Prairie Grass (43 trials with continuous plume releases of neutrally buoyant
tracer gas over a flat agricultural field), and the EMU L-shaped building (a wind tunnel experiment involving a release
from an open door in the courtyard area of an L-shaped building). The primary focus is on the maximum concentration
on the monitoring arcs. The performance statistics are consistently fairly good, with a median of 86% of the predictions
within a factor of two of the observations, a median relative bias suggesting a 20% underprediction, a median relative
scatter of about 50%, and a median 20% underprediction of the overall experiment maximum. These results are all well
within the criteria of acceptance for dispersion models. Evaluations with the EMU L-shaped building data show that
72% of FLACS predictions are within a factor of two of observations, and that the model can predict the dimensions of
the recirculating cavity behind the building within a factor of two. It is suggested that these extensive data sets involving
tracer releases in obstacle arrays be used to evaluate other CFD models.
r 2004 Elsevier Ltd. All rights reserved.

Keywords: CFD model; Model evaluation; Air quality models; Urban dispersion models

1. Introduction The CFD model solves the basic time dependent

Navier–Stokes equations, but using a small grid size
Air quality models are used to predict the transport (of the order 1 m or even less). CFD models are
and the turbulent dispersion of gases or aerosols after especially useful when the plume is dispersing within
they are released into the atmosphere. Because of arrays of obstacles such as buildings in urban areas or
advances in computer speed and storage capabilities, it industrial areas, which also can have many pipe racks,
is now practical to apply Computational Fluid Dy- tanks, and other types of obstacles. Some CFD models
namics (CFD) models to some of these air quality are being run for specific urban building domains with
modeling scenarios involving short-range dispersion. links to urban neighborhood models and further links to
mesoscale meteorological models (e.g., Brown et al.,
*Corresponding author. Tel.: +207-967-4478; fax: +207- 2001). In the past, comprehensive evaluations of CFD
967-5696. models with field observations of tracer dispersion have
E-mail address: [email protected] (S.R. Hanna). often been hampered by the slow run times and by

1352-2310/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.atmosenv.2004.05.041
 ARTICLE IN PRESS
4676 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

problems with properly maintaining the turbulence (e.g., assessments (Hansen et al., 1999a, 2001), and is required
the experiences with the FEFLO CFD model reported by industry standards in certain countries (NORSOK,
by Hanna et al., 2000). Riddle et al. (2004) and 2001). For the land-based chemical process industries,
Sklavounos and Rigas (2004) describe recent CFD FLACS is used to assist in siting control rooms so they
model applications where comparisons are made with are less likely to be affected by any nearby explosions
some field data and there are sensitivity studies with (van Wingerden et al., 1999).
turbulence models used by the CFD models. Since small details of the obstacles on an oil platform
The current paper concerns the FLACS (FLame can have a significant impact on flame acceleration, the
ACceleration Simulator) CFD model, which is similar to proper representation of the obstacles was a key aspect
many other CFD models but which is unique in its use of the development of the FLACS explosion prediction
of a distributed porosity approach for parameterizing code. A so-called distributed porosity concept was
buildings and other obstacles. The porosity of a grid is developed, as a compromise between the need to
represented as a fractional coverage of each grid volume characterize the geometric details and the need to have
and each grid face with sub-grid obstacles. Turbulence the code run in a reasonable time. Obstacles such as
production terms are parameterized for sub-grid objects structures and pipes are represented as area porosities
(Arntzen, 1998). In the current paper, FLACS is (the opposite of blockages) on control volume (CV)
evaluated with extensive field observations involving faces, and are represented as volume porosities in the
tracer gas releases in three independent field experiments interior of the CV. Each CV surface or each CV volume
(Kit Fox, MUST, and Prairie Grass) and wind tunnel is either fully open, fully blocked or partly blocked. For
data from an L-shaped building. The numbers of the partly blocked surfaces or volumes, the porosity is
experimental trials used in the FLACS evaluations are defined as the fraction of the area/volume that is
much larger than the numbers used in any prior available for fluid flow. The resulting porosity model is
evaluations with atmospheric tracer data (e.g., the used to calculate the flow resistance terms, the turbu-
evaluations reported by Riddle et al., 2004). lence source terms from small objects, and the flame
speed enhancement due to flame folding in the sub-grid
wake. The flame folding parameter is very important for
2. Description of FLACS flow and dispersion model explosion calculations, but irrelevant for pure dispersion
calculations. There are several problems that had to be
A description of the field experiments and the results resolved when the porosity methodology was developed.
of the model evaluation exercise are given in Section 4. For example, it is desired that the porosity calculations
First, in the current section, the characteristics of the should be automatic within the code, and should not
FLACS CFD model and some justifications for its overly influence the results when the grids are translated
assumptions are given. FLACS was developed in the or their size and shape are changed. It is necessary that
early 1980s to simulate the initial dispersion of gas leaks closed surfaces or corners should remain closed with
and subsequent explosions in offshore oil and gas different assumed grids, and openings in walls should
production platforms. The conservation equations for not depend on the assumed grid. Sub-grid objects (with
mass, momentum, and enthalpy, in addition to con- sizes less than the grid size) must be handled differently
servation equations for concentration and for flammable than on-grid objects (with sizes greater than the grid
gas effects, are solved on a Cartesian grid using a finite size), and special care is needed if more than one sub-
volume method. The equations are closed using the k2e grid object is within the same CV. In FLACS, different
equations for turbulence, as described by Harlow and drag coefficients are used for cylindrical and rectangular
Nakayama (1967). The SIMPLE pressure correction sub-grid objects, and significant drag and turbulence
method (Patankar, 1980) is applied, and extended for are generated only behind an object, and not along an
compressible flows with source terms for the compres- object that partly blocks a CV. To handle all of these
sion work in the enthalpy equation. Hjertager (1985, conditions within the porosity algorithm in FLACS, ten
1986) describes the basic equations used in the FLACS coefficients are calculated for each control volume.
model, and Hjertager (1988a,b) present the results of A comprehensive description of this concept is given in
several laboratory and field experiments used in the Hjertager (1985, 1986) and Arntzen (1998).
development of FLACS. For the flow scenarios described in this article, an
Because accidental explosions around oil platforms important consideration is the drag formulation from
are normally preceded by the release and subsequent the partly porous objects and the modeled turbulence
dispersion of flammable materials in the atmosphere in production behind objects classified as sub-grid. For the
and around the platform, the FLACS CFD model smallest objects, the flow kinetic energy lost due to the
contains algorithms for calculating the initial dispersion drag is directly added as a production term for turbulent
of flammable gas. The use of CFD models to calculate energy. With increased size of the sub-grid object
dispersion is widespread in explosion quantitative risk relative to the grid size, the sub-grid turbulence
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4677

production is gradually decreased. Objects with a Previously, a logarithmic stationary wind profile with a
dimension of 1.5–2.0 CVs (where the exact limit depends specific turbulence intensity and length scale could be
on position on grid) in both cross-flow directions are specified. This would typically lead to an over-prediction
defined to be on-grid objects. For these objects, there is of the hazard distance (i.e., an over-prediction of
no sub-grid turbulence production, since the shear layers concentrations) from a gas release, because the simu-
handle the turbulent production. lated atmospheric boundary layer turbulence was
Modifications to the standard k2e model have been underestimated. Hanna et al. (2002) and Riddle et al.
implemented in FLACS, as briefly listed below: (2004) report similar results (i.e., underestimation of
turbulence) with other CFD models. Using a method
* For objects with width equal to about one control
suggested by Arya (2001) and Han et al. (2000), FLACS
volume (CV), typical numerical schemes will estimate
was modified to estimate the turbulent kinetic energy
zero turbulence production, as the du/dy terms
and dissipation rate based on input of Pasquill stability
behind the objects will be identical and opposite
class (from unstable A to stable F) or input of Monin-
and will cancel. This situation has been improved by
Obukhov length L:
a modification of the discretization for production of
In the case of the arrays of obstacles placed on in
turbulent energy.
relatively flat rural surroundings, such as the MUST and
* Wall functions with no slip condition are implemen-
Kit Fox and EMU experiments studied in the current
ted for solid surfaces. Analytical models estimate
paper, the surface roughness length, zo ; used for the
the total production of turbulent energy across the
wind inflow boundary condition is based on the assumed
boundary layer, and add production terms to the
roughness upstream of the experiment. For the unstable
relevant CVs across the boundary layer, even if
Pasquill classes (A–C) an average heat flux from the
the first grid cell is larger than the boundary layer.
ground (from Arya, 2001) is specified as an input, and is
* Source terms for the production of turbulent energy
used to calculate the upstream turbulence profiles.
due to Rayleigh–Taylor instabilities (when buoyant
It would be useful to account for the fact that the
gas is accelerating denser gas due to gravity) have
wind speed and direction in the atmospheric boundary
been included.
layer vary with time and space over a continuous
Except for these modifications, the k2e model with spectrum. Typical turbulence parameterizations in CFD
standard constants is applied in the turbulence model. models are primarily based on engineering approxima-
During the course of more than 20 years of develop- tions of sub-grid turbulence and may not be able to
ment and evaluation of the FLACS software, the represent atmospheric eddies (fluctuations) with periods
numerical methods have been steadily modified and of 10–100 s or more (the larger periods are associated
revised user guides have had to be developed. Of course with the so-called meandering flow field). If a CFD
this is the case for any code that is in wide use over many model does not satisfactorily account for the larger
years. Currently FLACS uses a second order central range of atmospheric eddies, it may therefore under-
difference scheme for the diffusive fluxes, and a second predict turbulence levels and consequently over-predict
order so-called kappa scheme for the convective fluxes. hazard distances. To better approximate the wide range
The kappa scheme is a hybrid scheme, which does a of atmospheric eddy sizes in the version of FLACS used
weighting between second order central and upwind here, preliminary approximations of larger eddies with
difference schemes. Second order schemes in time have periods from 10 to 100 s have been implemented. The
also been implemented but are not applied in the default periods of the fluctuations could have covered a broader
version of FLACS applied in this paper. The second range, but it is assumed that it is sufficient to use two
order terms in time are not required for the short time representative periods in the middle of the spectrum. It is
steps used in the current study, and tend to become anticipated that these approximations can be improved
unstable when the time steps become large. as more experience is gained. In the current method, two
A pressure correction method is applied in order to harmonic waves with periods 10–15 s (slightly different
obtain a solution that converges. Iterations are repeated for the three directions to avoid repetition) and 60–70 s
until a mass residual of less than 10
4 is obtained. (only for the two horizontal directions and also slightly
For the calculations of transport and dispersion of different to avoid repitition) have been chosen to
gases or small particles, the boundary conditions are approximate the meandering flow. Based on standard
usually either (1) no slip solid surfaces, (2) wind inflow assumptions in Arya (2001), the total magnitude of these
(or outflow), or (3) a passive outflow condition at velocity fluctuations is 2.4, 1.9 and 1.3 times the friction
ambient pressure. Prior to the evaluation exercise velocity, u, in the along wind, cross wind, and vertical
presented in this paper, the assumed atmospheric directions, respectively. Fluctuations are thus propor-
boundary layer turbulence inputs have been modified tional to the wind speed and depend on the surface
to account for the effects of stability and for the effects roughness. Future research will address the sensitivity of
of relatively low-frequency lateral meandering flow. the results to these choices and perhaps improve the
 ARTICLE IN PRESS
4678 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

methodology. Eddies with periods much longer than 60 s like wide porous boxes near the ground) was added to
would be of less interest for the CFD scenarios in this the domain. It was found that the small sagebrush
paper because of the size of the domain (about 100 to obstacles did not significantly influence the results, since
1000 m) and the duration of the simulation period the MUST obstacles were much larger (about 2.5 m tall
(about 10–20 min). As stated earlier, in practice, slightly and 2.4 m wide and 12.2 m long) and there were 120 of
different frequencies are chosen for fluctuations along them. To conclude, if details are lacking on the
wind, across wind and vertically to avoid repeated flow geometry, and if there is known to be a systematic
patterns. These assumptions represent the default field, pattern in what is lacking, it is usually not a problem to
which was used for all the simulations reported. It is also represent these details approximately (e.g., trees in a
possible for the user to arbitrarily specify fluctuations park may be represented quickly by assuming a typical
with two sine waves with specified periods and shape, density and porosity).
magnitudes in the three directions. Questions also arise whether the k2e model is the
The gas explosion modules in FLACS have been most appropriate closure model. A set of (empirical)
evaluated with data from hundreds of gas explosion constants is needed for the k2e and for any other
experiments in the laboratory and in the field (Hansen turbulence model. The advantage of the k2e model is
et al., 1999b). Previous evaluation exercises involving the that there is a limited number of equations (two) and
dispersion algorithms in FLACS have focused on the constants. Improvements to the model to make it more
dispersion of flammable natural gas, with observed suitable for transient scenarios have been made, and
concentrations in the flammable range from 5 to 15%. It numerical problems have been corrected where, in the
was concluded from those studies that the diameter of Prairie Grass, Kit Fox, and MUST runs, little turbu-
the cloud after expansion to ambient pressure must be lence was generation behind an object with size equal to
resolved by at least 1 CV to obtain good results. For the about one grid size. As seen by other modelers (e.g.,
Kit Fox and EMU data evaluated later in this paper, Riddle et al., 2004) who have applied k2e closure to
these guidelines could be followed, since the initial atmospheric dispersion scenarios, too much dissipation
diameter of the release was 1.5 m for Kit Fox and 4 m of turbulence was observed for scenarios with no
for EMU. However, for the MUST and Prairie Grass buildings or other obstacles (e.g., the Prairie Grass
field data, also evaluated later, the small initial diameters experiment). For urban areas and building obstacles
(about 1 cm) and the relatively large domains (about represented by the MUST and Kit Fox scenarios, the
100–1000 m) made it more difficult to follow this regular patterns of obstacles ensured sufficient produc-
guidance, and much coarser grids than normally used tion of turbulent kinetic energy, as parameterized by the
were applied to ensure acceptable simulation times. sub-grid concept in FLACS. The weakness of the k2e
Consequently, we have to accept the fact that the gas in closure method for scenarios large numbers of obstacles
the CVs near the source will be too dilute. A concept for is less important. If another alternate turbulent closure
local grid refinement near the source has been developed model were used, such as a Reynolds stress model, more
(Gjesdal, 2000), but problems with stability limits the equations would be needed (at least six) and many more
use of these algorithms. (less well-established) constants are required.
The question can be asked, if the FLACS system relies As for all numerical models, the run times for FLACS
on high-resolution (e.g., 1 m resolution) geometric data are roughly proportional to the number of grid cells
to calculate the porosity values, will there be problems in times the number of seconds simulated times the number
urban areas because the data for large numbers of of time steps each second. The grids have been made
buildings are not sufficiently detailed? For a small relatively coarse in the current study, as the goal was to
number of buildings it is more likely that detailed CAD obtain relatively quick results. Within a few days, 40 or
data are available. Several organizations are currently 50 tracer release trials in a field experiment could be
sponsoring development of 1 m resolution geometry simulated, with about ten 1 GHz Pentium 3 PCs
data for larger cities. However, it is important to note available. The number of time steps each second
that FLACS can use whatever data are available to depends on the time step criteria (i.e. the CFL number),
approximate the porosity using empirical assumptions. the finest grid size, and the wind speed. If the grid is
For studies with FLACS of onshore and offshore refined locally, then shorter time steps may be needed to
chemical process facilities, the distribution of pipe satisfy the CFL criteria. Examples of numbers of CVs
lengths and diameters can be specified as much as and simulation times for the FLACS runs carried out in
possible based on the CAD model. However, if for the current paper are: For Kit Fox (assuming the coarse
instance, data on the 2’’ piping in the facility is lacking, grid used for this paper), where there are 33,000 CVs,
an expected amount of 2’’ piping is added to make the and 300–900 s of simulated time, the typical elapsed time
model realistic (this is called the artificial congestion on the PC for one run is two to three hours, but as large
concept). This method was used in the MUST runs as nine hours for one of the experiment trials. For the
described later, where sagebrush (0.40 m tall and shaped quality check runs for Kit Fox (assuming a finer grid)
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4679

where there are 308,000 CVs, and 350–600 s of simulated The use of maximum concentrations on arcs for the
time, the typical elapsed time on the PC for four runs is model evaluation exercise is fairly standard for evalua-
30, 50, 54 and 150 h. For MUST, where the grid used for tions of dispersion models and field experiments in open
calculations contained 55,000–75,000 CVs and where terrain. Even though the Kit Fox and MUST experi-
there was 500 s of simulated time, the typical elapsed ments involve obstacle arrays and tracer releases at
time on the PC for one run is 6–10 h (one case 15 h). For heights less than the obstacle heights, the monitoring
Prairie Grass, where the grid contained about 50,000 arcs were set up at distances beyond a few rows of
CVs, and where there were 1000–1200 s of simulated obstacles. Consequently, for those experiments, the flow
time, the typical elapsed time on the PC for one run is and dispersion around individual obstacles is not being
about 9–40 h, with the longer simulation times for investigated, and it is felt that the arc-maximum
stability classes A–C (i.e., unstable conditions). concentrations are appropriate for evaluation. However,
The initialization of the wind field consumes a the main use of a CFD model, and where it has
relatively small amount of time (about 10–50 s). The advantages over more standard models such as Gaus-
assumed initial wind field has a logarithmic mean wind sian plume models or Lagrangian puff models, is to
profile at the upwind boundary, and subgrid pus larger provide detailed three-dimensional, time dependent
turbulence components, as described earlier. Then, due information in the area near (within one obstacle height)
to the generation of flows and turbulence by the of the obstacle. The EMU L-shaped building (observed
presence of the obstacles, typically about 10–30 s in a wind tunnel) is the only data set in the current
are needed for the flow to adjust around the obstacles. exercise that provides sufficient detail in the area near
The presence of buildings will not significantly increase the building. Future evaluations of CFD models should
the requirement for simulation time for a given grid, focus more on the identification of important model
because of the efficiencies offered by the sub-grid outputs and development of methods to evaluate CFD
porosity assumption. models near obstacles.
Although the field experiments studied in the current The following equations define the statistical perfor-
paper all involve tracer gases, FLACS is also able to mance measures, which include the fractional bias (FB),
model aerosols based on the Eulerian transport equa- the geometric mean bias (MG), the normalized mean
tions. The aerosol models have been used to simulate the square error (NMSE), the geometric variance (VG), and
effect of water sprays on gas dispersion (Hansen, 2003). the fraction of predictions within a factor of two of
observations (FAC2):


C% o
C% p
FB ¼
; ð1Þ
0:5 C% o þ C% p
3. Description of statistical model performance evaluation
methods

MG ¼ exp ln Co
ln Cp ; ð2Þ
The FLACS model is evaluated following the
2
Co
Cp
approaches for model performance measures for air NMSE ¼ ; ð3Þ
quality models suggested by Weil et al. (1992), Hanna C% o C% p
et al. (1993), and ASTM (2000). This section describes  

2
the specific approach used in the current paper, based on VG ¼ exp ln Co
ln Cp ; ð4Þ
a methodology suggested by Hanna et al. (1993) and
summarized by Chang and Hanna (2004). Cp
The Kit Fox (Hanna and Chang, 2001), MUST FAC2 ¼ fraction of data that satisfy 0:5p p2:0;
Co
(Biltoft, 2001), and Prairie Grass (Barad, 1958) field
ð5Þ
experiments used in the current paper all involve tracer
samplers or monitors installed on arcs at specific where
downwind distances (four or five distances are used in
these three field experiments). The evaluations in this Cp : model predictions of concentration,
paper focus on the maximum concentration observed Co : observations of concentration,
and predicted on a given arc during a given experimental %
overbar ðCÞ: average over the data set, and
trial. Note that the location of the monitor with the sC : standard deviation over the data set.
observed maximum is not necessarily the same as the
location of the monitor with the predicted maximum. A perfect model would have MG, VG, and
For the EMU L-shaped building (Hall, 1997), where FAC2=1.0; and FB and NMSE=0.0. Of course,
there are not well-defined monitoring arcs, the evalua- because of the influence of random atmospheric
tions focus on observed and predicted concentrations, processes, there is no such thing as a perfect model in
paired in space and time at several monitor locations. air quality modeling. In addition to the standard
 ARTICLE IN PRESS
4680 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

performance measures defined above, which use data Typical magnitudes of the above performance mea-
from a large number of experimental trials, the simple sures and estimates of model acceptance criteria have
ratio of the overall maximum observed concentration to been summarized by Chang and Hanna (2004) based on
the overall maximum predicted concentration on each extensive experience with evaluating many models with
arc is listed for each field experiment. Thus, for the many field data sets. It was concluded that, for research-
100 m arc of the Prairie Grass field experiment, and over grade field experiments such as those investigated in this
all 43 experimental trials, there is a single overall paper, ‘‘acceptable’’ performing models have the follow-
maximum observed concentration and a single overall ing typical performance measures. The fraction of
maximum predicted concentration. These two maxima predictions within a factor of two of observations is
may occur during different experiment trials. about 50% (i.e., FAC2>0.5). The mean bias is within
All five performance measures defined above should 7 30% of the mean (i.e., –0.3 oFB o0.3 or 0.7 oMG
be calculated and considered together, since each o1.3). The random scatter is about a factor of two of
measure has pros and cons. For example, the linear the mean (i.e., NMSE o4 or VG o1.6). However, these
measures FB and NMSE can be overly influenced by are not firm guidelines and it is necessary to consider all
infrequently occurring high observed and/or predicted performance measures in making a decision concerning
concentrations, whereas the logarithmic measures MG model acceptance. Since most of these criteria are based
and VG may provide a more balanced treatment of on research grade field experiments, model performance
extreme high values. would be expected to deteriorate as the quality of the
To help interpret the performance measures, the inputs decreases.
following examples of typical values may be useful:
For the fractional bias, FB=0.67 implies a factor of two
mean under-prediction, and FB=
0.67 implies a factor 4. Description of data sets and results of evaluations
of two mean over-prediction. The geometric mean, MG,
can be thought of as the ratio of the geometric mean of In order to demonstrate that a model is performing
Co to the geometric mean of Cp : Consequently, a factor satisfactorily, it is best to evaluate the model using
of two mean bias would imply that MG=0.5 or 2.0, and several independent field experiments. Because the
a factor of four mean bias would imply that MG=0.25 FLACS CFD model is intended for use at industrial
or 4.0. To interpret the normalized mean square error, sites and urban sites with numerous obstacles, such as
NMSE, assume that the mean of the observed concen- buildings, storage tanks, and pipe racks, the focus
trations equals the mean of the predicted concentrations. should be on field experiments involving obstacles. For
Then NMSE=1 if the typical error equals the mean and this reason, the Kit Fox (Hanna and Chang, 2001) and
NMSE=4 if the typical error equals two times the MUST (Biltoft, 2001) experiments were chosen. Also,
mean. As NMSE becomes much larger than 1.0, it can because of concerns about CFD models’ abilities to
be inferred that the distribution is not normal but is maintain the proper atmospheric turbulence levels over
closer to log-normal (e.g., many low values and a few an open field in the absence of buildings, the well-known
large values). The geometric variance, VG, expresses the Prairie Grass field experiment (Barad, 1958) is used as
scatter of a log-normal distribution, which can be part of our evaluations. Finally, to demonstrate the
expressed as, say, ‘‘plus or minus 20%’’, or ‘‘plus or FLACS model’s capabilities to simulate the flow and
minus a factor of 10’’. For example, a factor of 2 scatter dispersion patterns at close distances from a single
would imply VG=1.6, a factor of 4 scatter would imply building, the EMU L-shaped building wind tunnel data
VG=6.8, and a factor of 5 scatter would imply VG=12. set (Hall, 1997) is used. The characteristics of these field
For models where the relative bias and scatter is experiments are summarized below and the results of the
small, there are further approximations that can be statistical evaluations are given.
derived. For example, if C% o
C% p o0.1, then FB equals
ðC% o
C% p Þ=C% o ; which can be thought of as the relative 4.1. Must field experiment
under- or over-prediction. That is, a 2% over-prediction
leads to FB=
.02. Similarly, for small ðC% o
C% p Þ=C% o ; The mock urban setting test (MUST) field experiment
then MG ¼ 1þðC% o
C% p Þ=C% o ¼ 1 þ FB: The perfor- consisted of 37 releases of propylene tracer gas in an
mance measures for relative scatter, NMSE and VG, array of 120 obstacles at the Dugway Proving Ground
also have simple approximations for small ðC% o
desert site (Biltoft, 2001). The obstacles were shipping
C% p Þ=C% o : For example, since (NMSE)1/2 can be inter- containers, which are about the size of the trailer in a
preted as the relative scatter, then VG=1+NMSE. tractor-trailer rig (12.2 m long by 2.42 m wide by 2.54 m
These approximations are valid only for small ðC% o
high). Fig. 1 shows the arrangement of the array of
C% p Þ=C% o : However, since ðC% o
C% p Þ=C% o is small for many obstacles as simulated by FLACS. The wind speed
of the FLACS evaluations, it will be seen that these assumed for each release trial was determined by
approximations are roughly valid. averaging four wind observations at 6 m elevation near
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4681

Fig. 1. Locations of 120 obstacles in MUST experiment (described by Biltoft, 2001). This plot was generated by FLACS. The obstacles
are 2.54 m high. The numerous groups of three smaller obstacles are intended to represent sagebrush and other bushes on the desert
floor. Some towers are shown as vertical lines. The tracer gas was released from various locations between the first and third rows on
the upwind (foreground) edge of the array. The four monitoring ‘‘arcs’’ were located between rows 3 and 4, 5 and 6, 7 and 8, and 9 and
10. Downwind distances of the arcs varied with release location but averaged about 25, 60, 95, and 120 m.

Table 1
Performance measures for the 37 trials in the MUST obstacle array, where maximum observed and predicted concentrations on four
monitoring arcs are compared

Median over 4 arcs Arc 1 (25 m) Arc 2 (60 m) Arc 3 (95 m) Arc 4 (120 m)

Max Co 99.5 35.7 17.2 10.1

Max Cp 47.2 13.9 7.73 10.0
Max Co /Max Cp 2.15 2.08 2.56 2.22 1.01
Mean Co 25.5 8.9 5.5 4.2
Mean Cp 14.6 5.3 3.0 3.7
FB 0.53 0.55 0.51 0.60 0.45
NMSE 1.64 2.03 1.85 1.44 1.24
MG 1.57 1.44 1.43 1.72 1.70
VG 1.69 1.71 1.44 1.67 2.65
FAC2 0.64 0.68 0.60 0.78 0.59

the corners of the obstacle array. Average wind speed numbers are within the range of acceptable model
was 3 m/s and the wind direction generally blew from performance. There was little trend in model perfor-
the foreground to the background of the array in Fig. 1. mance with downwind distance.
The release locations were altered slightly from trial to
trial, but were always near the first three rows of 4.2. Kit fox field experiment
obstacles (near the foreground of Fig. 1). There were
four sets of downwind monitoring arrays (at downwind The Kit Fox field experiment took place at the
distances of about 25, 60, 95, and 120 m), and the Nevada Test Site, where two types of flat ‘‘billboard-
maximum observed and predicted concentrations on shaped’’ obstacle arrays were used—the larger ERP
each array were compared. array (with height 2.4 m) and the smaller URA array
Table 1 summarizes the FLACS model performance (with height 0.2 m). Fig. 2 shows the arrangement of the
for its predictions of maximum concentrations on each ERP and URA obstacles as simulated by FLACS. The
arc for the MUST obstacle experiments. The overview in experiments and some analyses of the data, including
Table 1 suggests that there is an approximate factor of comparisons with the HEGADAS dense gas model, are
two under-prediction for Max C and about a 36% described by Hanna and Chang (2001). All experiments
under-prediction on average. The relative scatter is took place in the evening near sundown, and the average
about 1.5 times the mean. 64% of the predictions are wind speed was about 2.5 m/s. There was a total of 52
within a factor of two of the observations. These experiments, but these have been split into four sets for
 ARTICLE IN PRESS
4682 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

Fig. 2. Locations of obstacles in Kit Fox experiment (described by Hanna and Chang, 2001). This plot was generated by FLACS. The
inner array of larger (2.4 m high) obstacles is called the ERP array. The large rectangular area is covered by 0.2 m high obstacles called
the URA array. A line of Counihan turbulence generators is seen on the upwind (left) edge. Meteorological towers and vertical
sampling towers are seen as vertical lines. The release was from a 1.5 m  1.5 m area source near the middle of the ERP array.

Table 2
Performance measures for Kit Fox field experiment, where maximum observed and predicted concentrations on monitoring arcs (at 25,
50, 100, and 225 m) are compared, and statistics represent medians over the four arcs

Data subset ERP Plumes ERP Puffs URA Plumes URA Puffs Overall

N 6 13 12 21 52
Mean wind speed 1.7 m/s 2.1 m/s 2.9 m/s 3.0 m/s 2.5 m/s
Max Co /Max Cp 0.64 0.88 1.81 1.51 1.22
FB 0.03 0.07 0.37 0.09 0.08
NMSE 0.30 0.15 0.22 0.12 0.18
MG 1.05 1.19 1.41 1.06 1.12
VG 1.17 1.36 1.22 1.14 1.20
FAC2 1.00 0.90 0.96 0.92 0.94

Results are given for the overall comparisons for the 52 trials and for four subsets of the data for plume versus puff releases and for
ERP (2.4 m high) obstacles and for URA (0.2 m high) obstacles.

our statistical analysis: 6 ERP trials with ‘‘plume’’ concentration for the sets of trials, with a 30 or 40%
releases (duration of 120 s or greater), 13 ERP trials with over-prediction for the ERP (2.4 m high) array and a 30
‘‘puff’’ releases (duration of 20 or 25 s), 12 URA trials or 40% under-prediction for the URA (0.2 m high)
with ‘‘plume’’ releases, and 21 URA trials with ‘‘puff’’ array. There is little trend with wind speed, stability, or
releases. CO2 gas was released at ground level from a downwind distance.
1.5 m  1.5 m square opening near the middle of the
obstacle array. In all experiments, the maximum 4.3. Prairie grass field experiment
observed and predicted concentrations on each of the
four monitoring arcs (at 25, 50, 100, and 225 m) were The Prairie Grass field experiment has become the
evaluated. standard database used for evaluation of models for
Table 2 presents the Kit Fox performance measures continuous plume releases near the ground over flat
separately for the four sets of release types and obstacle terrain (ASTM, 2000). The site consisted of an
types. The Kit Fox field experiment can be considered to agricultural field where the grass had been cut and was
represent four independent sets of trials. The FLACS short dry stubble at the time of the experiments. The
model performance is seen to be fairly good for each of fundamental data report is by Barad (1958), although
the four sets of trials, with a relative mean bias less than there are hundreds of published papers describing
720% and a relative scatter of 50% or less. Over 90% various methods of analysis of the data. A continuous
of the model predictions are within a factor of two of trace amount of neutrally buoyant gas was released from
observations. A slight trend is seen for the maximum a small tube at a height of 0.46 m. There were 43 trials in
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4683

Fig. 3. Locations of five arcs (x=50, 100, 200, 400, and 800 m) in the Prairie Grass experiment (described by Barad, 1958) and an
example of a FLACS prediction (PG test 21, Stability D, 6 m/s). Predicted concentration is shown in a logarithmic scale where
concentrations range from ppm (
6) to 100 ppm (
4).

Table 3
Performance measures for Prairie Grass field experiment, where maximum observed and predicted concentrations on five monitoring
arcs (50, 100, 200, 400, and 800 m) are compared. There are 43 trials

Median over five arcs Arc 1 (50 m) Arc 2 (100 m) Arc 3 (200 m) Arc 4 (400 m) Arc 5 (800 m)

Max Co 339 195 111 49.9 37.0

Max Cp 259 174 93.7 65.7 30.2
Max Co /Max Cp 1.18 1.31 1.12 1.18 0.76 1.22
Mean Co 122.1 47.8 20.4 8.10 3.69
Mean Cp 65.7 28.7 17.1 8.19 4.11
FB 0.18 0.60 0.44 0.18
0.01
0.11
NMSE 0.43 0.543 0.426 0.236 0.437 0.348
MG 1.53 2.63 2.20 1.53 0.84 0.35
VG 2.75 4.31 2.75 1.61 1.94 16.9
FAC2 0.49 0.49 0.48 0.62 0.67 0.47

a variety of stability conditions and with an average 4.4. EMU L-shaped building wind tunnel experiment
wind speed of 5 m/s. In all experiments, the maximum
observed and predicted concentrations were evaluated at The evaluation of model uncertainty (EMU) study
a height of 1.5 m on each of the five monitoring arcs (at involved a comprehensive evaluation of models in a
50, 100, 200, 400, and 800 m). Fig. 3 shows the locations European Commission—sponsored study described by
of the samplers on the monitoring arcs, and gives an Hall (1997). Several research laboratories participated in
example of a FLACS simulation. the wind tunnel experiments, the model runs, and the
Table 3 contains a summary of the performance evaluations. The current evaluation used only one of the
measures for the Prairie Grass experiment. Emphasis is EMU scenarios—a single L-shaped building (see Fig. 4)
on maximum concentrations at given downwind dis- located on a flat surface. Neutral ambient conditions
tance arcs. FLACS shows a slight (20%) average relative were assumed and a continuous release of neutrally
under-prediction. The relative scatter averages about buoyant gas took place from a ‘‘courtyard’’ door.
70% to a factor of two. Although it cannot be seen in Detailed wind tunnel observations were made of flow
Table 1, a break-down by stability class would reveal an and dispersion. In the current paper, the concentration
over-prediction tendency at the more distant arcs (400 predictions at a few locations are compared with the
and 800 m) during unstable (daytime) conditions, as wind tunnel observations. Also, the approximate dimen-
found for most other models. It is thought (see Weil sions of the predicted recirculation zones are compared
et al., 1992) that convective eddies were ‘‘lifting’’ the with observations.
observed plume off the ground during the most unstable Predicted concentrations at a few cross-wind locations
trials. However, the performance statistics for stable on the cross-section at a distance H downwind of the lee
conditions were satisfactory at the 400 and 800 m arcs. edge of the L-shaped building were compared with the
 ARTICLE IN PRESS
4684 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

Fig. 4. View from above of EMU L-shaped building, showing coordinate system and example of predicted FLACS concentration
distribution. Note that the building height, H; equals 10 m. The tracer gas release was from a door shown as a thick line at y=0.0 m
and between about x=
18 m and x=
14 m.

Table 4
Comparisons of FLACS predicted concentrations versus observed concentrations (from wind tunnel) for EMU L-shaped building

y/H
2
2
2
1.5
1.5
1.5
1.0
1.0
1.0

Co Cp Co /Cp Co Cp Co /Cp Co Cp Co /Cp

z/H=.16 0.84 0.60 1.40 1.1 0.36 3.06 0.88 0.36 2.44
z/H=.37 0.62 0.34 1.82 0.95 0.30 3.17 0.83 0.32 2.59
z/H=.67 0.27 0.13 2.08 0.50 0.22 2.27 0.63 0.29 2.17
z/H=1.02 0.05 0.034 1.47 0.25 0.079 3.16 0.44 0.23 1.91
z/H=1.47 0.01 0.010 1.00 0.03 0.021 1.43 0.13 0.06 2.17
z/H=1.99 0.005 0.004 1.25 0.01 0.009 1.11 0.017 0.026 0.65
Median 1.43 2.72 2.17
FAC2 0.83 0.33 0.33

y/H
0.5
0.5
0.5 0.0 0.0 0.0 0.5 0.5 0.5
Co Cp Co /Cp Co Cp Co /Cp Co Cp Co /Cp
z/H=.16 0.63 0.38 1.66 0.3 0.45 0.67 0.18 0.36 0.50
z/H=.37 0.63 0.38 1.66 0.5 0.53 0.94 0.32 0.53 0.60
z/H=.67 0.63 0.36 1.75 0.6 0.60 1.00 0.48 0.79 0.61
z/H=1.02 0.58 0.34 1.71 0.7 0.75 0.93 0.62 1.05 0.59
z/H=1.47 0.45 0.24 1.88 0.6 0.71 0.85 0.36 0.63 0.57
z/H=1.99 0.05 0.08 0.63 0.1 0.12 0.83 0.067 0.11 0.61
Median 1.68 0.89 0.60
FAC2 1.00 1.00 1.00

At x/H=1.0 (i.e., one building height downwind of the lee of the building) for six different heights (z/H=0.16, 0.37, 0.67, 1.02, 1.47,
and 1.96), and for six different lateral positions (y/H=
2.
1.5,
1.0,
0.5, 0.0, and 0.5). y=0.0 along the center of the lee building
wall and y is positive towards the longer side of the building. These data were used to generate the scatter plot in Fig. 6.

wind tunnel observations, and the results are given in and the cross-wind and vertical profiles were well-
Table 4. The ‘‘steady-state’’ solution is given, without simulated. The median value of Co /Cp is 1.55, implying
accounting for meandering fluctuations, since a wind about a 35% under-prediction. Fig. 5 shows the scatter
tunnel does not tend to have as many meandering plot of the data in Table 4, again showing the good
fluctuations as the real atmosphere. Concentrations at agreement but the slight under-prediction tendency. In
36 locations are compared in the table (at y/H=
2.0, addition, based on the systematic wind vector variations

1.5,
1.0,
0.5, 0.0, and 0.5; and at z/H=0.16, 0.37, in Fig. 6, the dimensions of the predicted re-circulating
0.67, 1.02, 1.47, and 1.99). In general 72% of the cavities were in agreement (750%) with known
predictions are within a factor of two of the observations similarity relations based on wind-tunnel observations
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4685

Fig. 5. Scatter plot of FLACS predicted concentrations versus observed concentrations for EMU L-shaped building, at x/H=1.0 (i.e.,
one building height, H; downwind of the lee of the building) for six different heights (z/H=0.16, 0.37, 0.67, 1.02, 1.47, and 1.96, and
for six different lateral positions (y/H=
2.
1.5,
1.0,
0.5, 0.0, and 0.5). y=0.0 along the center of the lee building wall and y is
positive towards the longer side of the building.

Fig. 6. Cross-section in x and z of wind vectors predicted by FLACS for EMU L-shaped building.

(Hanna et al., 1982). The wind tunnel observations and arcs for Prairie Grass). The medians for the max Co /max
empirical formulas in Hanna et al. (1982) suggest that Cp ; FB, and MG suggest a general under-prediction
the length of the re-circulating wake or cavity in the lee tendency of about 20%. The medians for NMSE and
of the L-shaped building is about 1.0 to 1.5 times H, VG suggest a relative scatter of about 50%. About 86%
whereas the FLACS prediction in Fig. 6 for this length is of the predictions are within a factor of two of the
about 1.5–2.0. observations. These results are all well within the range
of acceptable model performance. The table also
4.5. Summary of evaluation results contains two estimates of the range of the 25 ranked
numbers. First the total range is given, and then the
Table 5 contains a summary of the performance range for 50% of the data (the range from the 7th to the
measures for the FLACS model applied to the MUST, 18th ranked values) is given. The 50% range is fairly
Kit Fox, and Prairie Grass field experiment data. The tight for field data, suggesting only about a720% range
medians and ranges of the statistics were determined in the relative mean bias and the relative scatter.
from the 25 sets of ranked numbers (4 arcs for MUST, 4 In addition, FLACS was evaluated with the wind
subsets of trials times 4 arcs=16 arcs for Kit Fox, and 5 tunnel observations of the EMU L-shaped building,
 ARTICLE IN PRESS
4686 S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687

Table 5 predictions are within a factor of two of the observa-

Median performance measures and ‘‘total range’’ and ‘‘inner tions, the relative mean bias is about 20% from the
50% range’’ or range for 50% of data closest to median over perspective of FB and about 30% from the perspective
Kit Fox, MUST, and Prairie Grass field experiments for of MG, and the relative scatter is about 50 or 60%. The
FLACS CFD model
FLACS performance measures are consistent across the
Median Total range Inner 50% range three field experiments and the one wind tunnel
experiment, and there are minimal trends with distance.
Max Co /Max Cp 1.22 0.56 to 2.56 0.94 to 1.86
The large numbers of independent atmospheric
FB 0.18
0.32 to 0.60 0.03 to 0.44
NMSE 0.29 0.07 to 2.03 0.17 to 0.44
dispersion experimental trials that have been evaluated
MG 1.32 0.35 to 2.63 1.06 to 1.47 with the FLACS CFD model are much more extensive
VG 1.28 1.07 to 17.9 1.15 to 1.94 than previously reported in the literature for any CFD
FAC2 0.86 0.47 to 1.00 0.67 to 1.00 model. The authors are already in contact with other
CFD modelers who have requested the data sets and
have their own evaluation exercises underway. It is
hoped that the major contribution of this paper can be
showing that about 72% of the predictions are within a to establish an evaluation methodology that will allow
factor of two of the observations, and that there is a comparative CFD model evaluations and spur colla-
tendency towards under-prediction by about 35%. The borative efforts to improve model formulations.
predicted dimensions of the re-circulating cavity behind
the building are within 50% of the known size of the
cavity.
Acknowledgements
For the MUST and Prairie Grass field experiments,
where the source aperture was relatively small (on the
This research was supported by the DuPont Com-
order of 1 cm), the FLACS simulation guidelines would
pany. Steven Hanna’s research on model evaluation
require a much better grid resolution near the source
methods was also supported by the Defense Threat
than has been applied in the calculations reported in this
Reduction Agency (DTRA), the Department of Energy,
paper. However, this increased resolution would have
and the Department of Homeland Security. The authors
required very much longer simulation times on the PCs.
appreciate the cooperation of Richard Fry of DTRA in
The lack of local grid refinement may have been partly
providing access to the MUST data.
responsible for the slight under prediction tendency
reported above. To improve the results, methods with
local grid refinement near the source or analytical
models for the initial plume development could be References
considered.
Arntzen, B.A., 1998. Modelling of turbulence and combustion
for simulation of gas explosions in complex geometries,
5. Conclusions PhD Thesis, NTNU, Trondheim, Norway, ISBN 82-471-
0358-3.
This paper describes the evaluation of the FLACS air Arya, S.P., 2001. Introduction to Micrometeorology, Second
dispersion model. The model was evaluated using four Edition. ISBN 0-12-059354-8, Academic Press, London, 420
pp.
data sets and using quantitative performance measures,
ASTM, 2000. Standard guide for statistical evaluation of
including the fractional bias (FB), the geometric mean atmospheric dispersion model performance. American
(MG), the normalized mean square error (NMSE), the Society for Testing and Materials, Designation D 6589-00.
geometric variance (VG), and the fraction of data where ASTM, 100 Barr Harbor Drive, West Conshohocken, PA
predictions are within a factor of two of observations 19428–2959.
(FAC2). The performance measures reported in many Barad, M.L.(Ed.), 1958. Project Prairie Grass. A field program
model evaluation exercises with many sets of field data in diffusion. Geophys. Res. Paper No. 59, Vols. I and II,
were reviewed in order to determine the characteristics AFCRF-TR-58-235, Air Force Cambridge research Center,
of a ‘‘good’’ or ‘‘acceptable model. It was concluded Bedford, MA.
that, for research-grade field experiments, a ‘‘good’’ Biltoft, C.A., 2001. Customer Report for Mock Urban Setting
Test (MUST). DPG Doc. No. WDTC-FR-01-121, West
model would be expected to have at least 50% of the
Desert Test Center, U.S. Army Dugway Proving Ground,
predictions within a factor of two of the observations, a Dugway, UT 84022-5000.
relative mean bias within 730%, and a relative scatter Brown, M., Leach, M., Calhoun, R., Smith, S., Stevens, D.,
of about a factor of two or three. Reisner, R., Lee, R., Chan, S., deCroix, D., 2001. Multiscale
The performance of the FLACS model is well within modeling of air flow in Salt Lake City and the surrounding
these performance criteria, since about 86% of the region. Proceedings, ASCE Structures Congress 2001,
 ARTICLE IN PRESS
S.R. Hanna et al. / Atmospheric Environment 38 (2004) 4675–4687 4687

Washington, DC. Report LA-UR-01-509, Los Alamos Fall Symposium Proceedings, Mary Kay O’ Connor Process
National Lab, Los Alamos, NM 87545 Safety Centre, Dept. of Chemical Engineering, Texas A&M
Chang, J.C., Hanna, S.R., 2004. Air quality model performance University, 3122 TAMU, College Station, TX 77843-3122,
evaluation. Meteorol. Atmos. Phys. to appear in. pp. 395–410.
Gjesdal, T., 2000. Local grid refinement for improved descrip- Hansen, O.R., 2003. Water deluge and influence on dispersion.
tion of leaks in industrial gas safety analysis. Compute FABIG Newsletter Article R485, FABIG, The Steel
Visual Sci. 3, 25–32. Construction Institute, Silwood Park, Ascot, Berkshire,
Hall, R.C.(Ed.), 1997. Evaluation of Model Uncertainty SL5 7QN, UK.
(EMU)—CFD modelling of near-field atmospheric disper- Harlow, F.H., Nakayama, P.I., 1967. Turbulence Transport
sion. Project EMU Final Report to the European Commis- Equations. Phys. Fluids 10, 2323–2332.
sion (Data CD also available with report)7. WS Atkins Doc Hjertager, B.H., 1985. Computer simulation of turbulent
No. WSA/AM5017/R7, WS Atkins, Woodcote Grove, reactive gas dynamics. J. Model. Identification Control 5,
Ashley Road, Epsom, Surrey KT18 5BW, UK. 211–236.
Han, J., Arya, S.P., Shen, S., Lin, Y.-L., 2000. An estimation of Hjertager, B.H., 1986. Three-dimensional modeling of flow,
turbulent kinetic energy and energy dissipation rate based heat transfer, and combustion. Handbook of Heat and
on atmospheric boundary layer similarity theory. NASA/ Mass Transfer. Gulf Publishing Company, P.O. Box 2608,
CR-2000-210298, North Carolina State University, Raleigh, Houston, Texas 770011, pp. 304–350 Chapter 41
North Carolina. Hjertager, B.H., Bjørkhaug, M., Fuhre, K., 1988a. Gas
Hanna, S.R., Chang, J.C., 2001. Use of the Kit Fox field data to explosion experiments in 1:33 scale and 1:5 scale; offshore
analyze dense gas modeling issues. Atmos. Environ. 35, separator and compressor modules using stoichiometric
2231–2242. homogeneous fuel–air clouds. J. Loss. Prev. Process Ind. 1,
Hanna, S.R., Chang, J.C., Strimaitis, D.G., 1993. Hazardous 197–205.
gas model evaluation with field observations. Atmos. Hjertager, B.H., Bjørkhaug, M., Fuhre, K., 1988b. Explosion
Environ. 27A, 2265–2285. propagation of non-homogeneous methane-air clouds inside
Hanna, S.R., Tehranian, S., Carissimo, B., Macdonald, R.W., an obstructed 50 m3 vented vessel. J. Haz. Mater. 19,
Lohner, R., 2002. Comparisons of model simulations with 139–153.
observations of mean flow and turbulence within simple NORSOK Z-013, 2001. Risk and emergency preparedness
obstacle arrays. Atmos. Environ. 36, 5067–5079. analysis, Norsok standard. Available from Standard Norge,
Hansen, O.R., Talberg, O.,Bakke, J.R., 1999a. CFD-based Postboks 242, N-1326 Lysaker, Norway.
methodology for quantitative gas explosion risk assessment Patankar, S.V., 1980. Numerical Heat Transfer and Fluid Flow,
in congested process areas: examples and validation status. Hemisphere Pub, ISBN: 0070487405
Proceedings, AIChE/CCPS International Conference and Riddle, A., Carruthers, D., Sharpe, A., McHugh, C., Stocker,
Workshop on Modeling the Consequences of Accidental J., 2004. Comparisons between FLUENT and ADMS for
Releases of Hazardous Materials, September 28-October 1, atmospheric dispersion modeling. Atnos. Environ. 38,
1999, San Francisco, USA, ISBN 0-8169-0781-1, 1029–1038.
pp. 457-477. Sklavounos, S., Rigas, F., 2004. Validation of turbulence
Hansen, O.R., Storvik, I.,van Wingerden, K., 1999b. Validation models in heavy gas dispersion over obstacles. J. Haz.
of CFD-models for gas explosions, where FLACS is used as Mater. A. 108, 9–20.
example; model description and experiences and recom- Van Wingerden, K., Hansen, O.R.Foisselon, P., 1999. Predict-
mendations for model evaluation. Proceedings European ing blast over pressures caused by vapor cloud explosions in
Meeting on Chemical Industry and Environment III, the vicinity of control rooms. Process Safety Progress 18,
Krakow, Poland, available at the Department of Environ- 14–17, Available from American Institute of Chemical
mental Systems Engineering, Faculty of Process and Engineers, Center for Chemical Process Safety, 3 Park Ave,
Environmental Engineering, Technical University of Lodz, New York, N.Y., 10016–5991.
ul. Wolczanska 175, 90-924 Lodz, Poland, pp. 365–382. Weil, J.C., Sykes, R.I., Venkatram, A., 1992. Evaluating air-
Hansen, O.R., Renoult, J., Bakke, J.R., 2001. Explosion risk quality models: review and outlook. J. Appl. Meteor. 31,
assessment: how the results vary with the approach chosen. 1121–1145.