A Stochastic Single-Particle Lagrangian

Model for the Concentration Fluctuations

in a Plume Dispersing Inside an Urban
Canopy

Alon Manor

Boundary-Layer Meteorology
An International Journal of Physical,
Chemical and Biological Processes in
the Atmospheric Boundary Layer

ISSN 0006-8314

Boundary-Layer Meteorol
DOI 10.1007/s10546-013-9871-5

1 23
Your article is protected by copyright and all
rights are held exclusively by Springer Science
+Business Media Dordrecht. This e-offprint
is for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy
Boundary-Layer Meteorol
DOI 10.1007/s10546-013-9871-5

RESEARCH NOTE

A Stochastic Single-Particle Lagrangian Model

for the Concentration Fluctuations in a Plume
Dispersing Inside an Urban Canopy

Alon Manor

Received: 6 January 2013 / Accepted: 15 October 2013

© Springer Science+Business Media Dordrecht 2013

Abstract A new approach for estimating concentration fluctuations intensity in dense built-
up environments using a Lagrangian stochastic (LS) particle model is described. Following
past success in modelling the dynamics of concentration variance as a diffusion-advection
process, the ensemble-averaged concentration variance is represented by particles that advect
and diffuse throughout the computational domain. The calculation of the concentration vari-
ance is addressed by assuming an appropriate distribution of effective variance sources for
a given mean concentration field. Dissipation is treated by allowing the variance carried by
every particle to decay exponentially with a locally-estimated decay time. The approach has
the benefit of easily handling complex boundary conditions. It can also be easily and naturally
implemented as an extension to an existing LS model, which is used for mean concentration
estimations. The method differs from existing two-particle methods that demand knowledge
of the structure function of the flow. It is also more computationally efficient than micro-
mixing approaches that involve maintaining high population levels of particles in every grid
volume. The model is compared with high frequency concentration measurements, taken as
part of the JU2003 (Joint Urban 2003) experiment that was carried out in Oklahoma City.
Good agreement is observed.

Keywords Built-up areas · Concentration fluctuations · Lagrangian diffusion

1 Introduction

Pollutant dispersion in dense built-up environments is commonly modelled using Lagrangian

stochastic (LS) particle models (Kaplan and Dinar 1996; Tinarelli et al. 2007; Balwinder et al.
2008 ). In these models, the passive scalar pollutant is represented by particles that transport
and diffuse throughout the computational domain. The velocity fluctuations, which cause
turbulent diffusion, are modelled by stochastic additive changes in the particle velocity, taken

A. Manor (B)
Israel Institute for Biological Research, P.O.B. 19, Ness-Ziona, Israel
e-mail: [email protected]

123
 Author's personal copy
A. Manor

from a given distribution with a location-dependent standard deviation denoted by σ (r). The
turbulent integral time scale is introduced by a particle turbulent velocity decorrelation time,
which also depends on location and is denoted TL (r). Beside their stochastic movement, the
particles are advected according to a pre-calculated mean wind field. The topology of the
buildings is taken into account explicitly by representing each building by a set of planes
(roof and walls) off which the particles are reflected.
Dispersion calculations based on this approach have several advantages in comparison
to other methods that consist of solving the relevant prognostic equations on a grid. Firstly,
domain edge and surface boundary conditions are conveniently taken into account by absorb-
ing the particles through the edges of the domain and reflecting them off the planes that rep-
resent the complex surface underneath. Secondly, each particle trajectory does not depend on
other trajectories. This makes parallelization straightforward, allowing for rapid calculations
needed for real time emergency response applications.
The average wind field used for the advection of the particles can be calculated using a
common computational fluid dynamics method such as Reynolds-averaged Navier–Stokes
(Hsieh et al. 2007). However, for operational real-time risk assessment tasks, the wind field
can be generated using a fast diagnostic model for the flow field around an obstacle (Röckle
1990), which is then modified to satisfy the continuity equation (Sherman 1978).
Usually the above models are used to calculate the spatio-temporal distribution of the
average concentration, yet for some applications, the variance of concentration fluctuations
around its mean is required. For example, the fluctuating nature of the concentration in an
atmospheric plume is important for the estimation of odour nuisance (Mussio et al. 2001).
Toxicity assessments are also affected by the concentration fluctuations, since the toxic load
caused by the inhaling of contaminated air may differ substantially when the concentration
fluctuates (Ride 1995).
Since the pioneering work of Gifford (1959), the topic of concentration fluctuations has
been extensively studied both by theoretical modelling and through field and wind-tunnel
observations. Briefly, current models that exist in the literature can be divided into three
groups. The first consists of “meandering plume” models (see e.g.,Gifford 1959; Hanna 1967;
Sykes 1984). These models utilize estimations of the length scales of the time-averaged
and the instantaneous plumes generated by a continuous release of a passive tracer into
the atmosphere, to predict the fluctuations intensity and various other characteristics of the
concentration spectrum.
Another class of models are LS models that by turn fall into two groups. The first relates
to “particle pairs” models, where fluctuation intensity is calculated by releasing particle
pairs moving correlatively (Durbin 1980; Sawford 1983). Using this method relies on prior
knowledge of the structure function tensor for the specific considered flow, which for complex
flows is rarely known. In a second approach, frequently referred to as the “micromixing”
method (see e.g., Gonzalez 1997; Sawford 2004; Cassiani et al. 2005), every particle carries
a concentration that interacts with the whole local ensemble-average concentration, or, as
in some improved versions, only with the mean concentration determined by particles that
resemble its own state (in velocity phase space). The main shortcoming of this approach lies
in the demand for every grid element to contain a large number of particles throughout the
whole simulation.
Lastly, a third group of models is based on the Reynolds-averaged prognostic equation
for passive scalar variance
  2
∂c2 ∂c2 ∂  2 ∂C ∂c
= −Ui − u i c − 2c u i − 2ν (i = x, y, z) (1)
∂t ∂ xi ∂ xi ∂ xi ∂ xi

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

The terms that contain double and triple correlations (excluding c2 itself, of course) are
modelled by a closure scheme, and the resulting differential equation is solved simultane-
ously with the Reynolds-averaged equation for the mean concentration and the Navier–Stokes
equations to obtain a solution for c2 , C and Ui (see e.g., Csanady 1967; Sykes et al. 1984).
Recently Hsieh et al. (2007) and Milliez and Carissimo (2008) successfully utilized this
approach for concentration fluctuations in a built-up environment. Of course, all the diffi-
culties that characterize grid-based methods for solving differential equations, such as the
handling of boundary conditions and meshing, also apply here. This is especially true where
the surface boundary is complex, as is the case for an urban canopy.
The model presented in this work is, in a sense, a combination of LS methods with the
closure model approach. The model is based on approximating the solution of the Reynolds-
averaged equation for a passive scalar variance using a LS particle model instead of by
the conventional grid-based techniques. This approach has all the advantages of LS particle
models mentioned above, making it suitable for emergency response applications.
Another advantage of the approach suggested here stems from the fact that the procedure
for calculating the concentration variance uses the underlying framework of the LS particle
model, used to calculate the average concentration distribution. This makes it appealing for
use in applications where the average concentration is calculated by a LS model, as is the
case, for example, for the models described by Tinarelli et al. (2007) and by Balwinder et al.
(2008).
The model is implemented by emitting particles representing concentration variance from
multiple effective sources in the calculation domain. As will be explained shortly, these
effective sources of concentration variance are determined by the spatial distribution of the
mean concentration. The three-dimensional spatial distribution of concentration variance is
constructed from the statistics of the locations of these particles.

2 Formulation of the Model

In this work, the LS model developed by Kaplan and Dinar (1996) was utilized in two different
aspects. First, as further explained below, the model for the concentration variance demands
the three-dimensional average concentration distribution as an input. Secondly, the mean
wind field and turbulence parametrization used to move the particles that represent average
concentration are exactly the same for the variance-carrying particles. For this reason, before
the model for the concentration variance is described, a brief overview of the model for the
average concentration is in order. The reader will find more comprehensive information in
Kaplan and Dinar (1996).

2.1 The LS Model for the Average Concentration

The initial average wind field is generated by the approach first presented by Röckle (1990).
The flow field around each building is separated into three distinct zones: (a) the displacement
zone (on the upwind side), (b) the cavity zone (on the lee side), and (c) the wake zone
(further downwind). As described in Kaplan and Dinar (1996), the flow speed in each zone
is estimated by diagnostic formulae. These predict the horizontal wind as a function of the
spatial dimensions of the building and of the ambient wind field (that expected when the
building is absent).

123
 Author's personal copy
A. Manor

The wind field is further refined by a multigrid mass consistent algorithm (Dinar 1984),
that alters the wind field to obey ∇ ·u = 0, while keeping it as close as possible to the original
wind field. In this work, a neutral wind profile is used as an ambient wind field. The wind
velocity at a certain reference height over rooftops is given as input.
Turbulent diffusion is modelled by the stochastic motion of the particles, governed by the
equations
dr
= u(r, t) + U(r, t), (2a)
dt
du = a(r, t)dt + b(r, t). (2b)
The first equation describes the particle’s change of location according to the average (U(r, t))
and fluctuating (u(r, t)) wind components. In the second equation, b(r, t) is a stochastic part
given by

dt
bi (r, t) = σi (r) ξi , i = x, y, z, (3)
TL i (r)
with ξ representing white noise, where σ (r) is the standard deviation of wind fluctuations
and TL (r) is the Lagrangian time scale. As further explained in Sect. 2.3, the mean wind field
U is utilized to estimate these two parameters of turbulence. Also, a(r, t) is the deterministic
term given by
 
u i (r, t) 1 ∂(σi (r))2 u i (r, t)
ai (r, t) = − + 1+ , i = x, y, z, (4)
TL (r) 2 ∂ri σi (r)
 i    
1 2

where term 1 is the velocity correlation decay term, and term 2 is a correction term, needed
in order to satisfy the well-mixed principle (Thomson 1987).

2.2 Estimating the Turbulence Parameters

The scheme used here to estimate the turbulent integral time scale by the mean flow U(r, t) is
slightly different than that described in Kaplan and Dinar (1996). For that reason, it is hereby
explained in some detail.
Recent measurements inside a street canyon, reported by Nelson et al. (2011), show that
in the turbulent kinetic energy budget, the shear production term and the energy dissipation
term are of the same order of magnitude. Nelson et al. found that this is true especially in
daytime. Here, the measured shear production term was the most dominant positive term in
the turbulent kinetic energy budget equation, balancing the negative dissipation term. Slightly
above the urban canopy and throughout the rest of the boundary layer, the similarity of the
mechanical production and the dissipation terms is further enhanced (see e.g., Castro 2006).
We therefore write
∂Ui
u i u j ≈ . (5)
∂x j
With the eddy viscosity approximation and writing  in terms of the turbulent time scale TL
and of the turbulent kinetic energy k, one has
2
νt ∂Ui ∂U j k
+ ≈C . (6)
2 ∂x j ∂ xi TL

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

Recalling that the eddy diffusivity νt is proportional to k 2 /, we finally obtain

−1
∂Ui ∂U j
TL = C  + , (7)
∂x j ∂ xi

with C  a constant set to unity to make (7) converge to the known shear dependence of TL
 −1
for altitudes that are far above roof tops in a neutral atmosphere, i.e., TL = dU /dz .
As in Kaplan and Dinar (1996), the standard deviation of wind fluctuations is estimated
by
L
σ (r) = 0.6 , (8)
TL (r)
where L is a turbulent length scale, given by

L = min(dmin , Z ), (9)

and where dmin is the distance between the location of the particle and the closest obstacle,
and Z is the height of the particle relative to the ground.

2.3 The Model for the Concentration Variance

The starting point of this model is the steady-state version of Eq. 1, the Reynolds-averaged
equation for the concentration variance,
  2
∂c2 ∂  2 ∂C ∂c
−Ui − u i c −2c u i −2ν = 0, i = x, y, z, (10)
∂ xi ∂ xi ∂ xi ∂ xi
           
1 2 3 4

with term 3 a production term, which, by the K-gradient closure c u i = −K i ∂∂C

xi , can be
approximated as
2
∂C ∂C
− 2c u i ≈ 2σi2 TL i . (11)
∂ xi ∂ xi

In Eq. 11, the relation K i = σi2 TL i has been used. It is worth pointing out that if counter-
gradient fluxes are neglected, this term is always positive, i.e., it is always a source term for
concentration variance.
The common way to model the dissipative term (term 4 in Eq. 10) is through replacing it
with a first-order decay term (see e.g., Csanady 1967; Sykes et al. 1984; Hsieh et al. 2007)
 2
∂c c2
− 2ν ≈− , (12)
∂ xi τd
where τd is a decay time, that is further discussed in Sect. 2.4.
Substituting (11) and (12) into (10) we have
2
∂c2 ∂  2 ∂C c2
− Ui − u c + 2σi2 TL i − = 0. (13)
∂ xi ∂ xi i ∂ xi τd

123
 Author's personal copy
A. Manor

Term number 2 in Eq. 10 is a turbulent transport term. In many studies (Sykes et al. 1984;
Hsieh et al. 2007), this term is modelled by assuming proportionality between the turbulent
flux of c2 and the gradient of c2 , i.e., by writing
 
∂  2 ∂ 2
(v) ∂c
uc = Ki , (14)
∂ xi i ∂ xi ∂ xi
(v)
where K i is a diffusion coefficient for the concentration variance.
Naturally, Eq. 13 is usually solved simultaneously with the equation for the mean con-
centration with a source term Q(r)
∂C ∂  
− Ui − u c + Q(r) = 0 . (15)
∂ xi ∂ xi i
In Eq. 15, the turbulent transport term is also commonly modelled through the relation
 
∂   ∂ ∂C
uc = Ki . (16)
∂ xi i ∂ xi ∂ xi

The ability of modelling the concentration variance using a diffusion equation supports
the idea, implemented here, that in analogy to the utilization of LS models to simulate the
diffusion equation for the mean concentration, these models can also be used to simulate the
diffusion equation for the concentration variance.
Another assumption made by many authors, and used here, has to do with the turbulent
diffusion rate of the concentration variance. In all works known to us, in which equations
(v)
similar to Eqs. 13–16 are used, it is assumed that K i ≈ K i , so the same numerical value
is used for both of them. In other words, although the total concentration C (C = C + c )
is a passive scalar, carried by the flow, while c (its deviation from the local mean value) is
not, it is still assumed that the turbulent diffusion processes that expand the plume of mean
concentration operate in a very similar way to diffuse the concentration variance.
Adopting this assumption in the current work allows us to introduce the turbulent diffusion
of the concentration variance in the same way as for the mean concentration, as will be
explained shortly.
In Eq. 13, the variance source term (marked Q v (r) to distinguish it from the “real” source
Q(r)) is
2
∂C
Q v (r) = 2σi2 TL i . (17)
∂ xi

Since the solution of Eq. 15 can be assessed using a LS particle model for any distribution
of sources, Eq. 13 can also be solved by the same LS model, but with the source distribution
Q v (r) and with particles that carry variance, and which decay at a rate dictated by τd (r)
(where r is the location of the particle). In other words, another equation is added to the set
of equations of motion (Eq. 2a),
dη
= −η/τd (r), (18)
dt
where η describes the time-decaying concentration variance carried by a specific particle.
Before the calculation details are described, it is important to note that the use of the LS
model with particles that carry c2 rather than C deviates from the normal usage of LS models

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

for modelling turbulent diffusion. Formally, a trajectory of one particle describes a single-
realization approximation for a trajectory of an air parcel in the turbulent flow. The parcel can
carry various physical properties such as concentration, temperature, energy or momentum.
Thus, the ensemble-averaged spatial distribution of the physical quantity considered is a
consequence of the combination of the different probabilities for each trajectory. Here, the
interpretation of the trajectories of the particles is quite different. As mentioned above, since
c is defined through C, its momentary value is not a conserved quantity along a Lagrangian
trajectory in the flow. Thus, the usage of a LS model for describing the diffusion of the
variance in the context of this work is an approximation, pertaining for representing the
diffusion of the ensemble averaged variance, the quantity Eq. 1 addresses. In other words,
when the LS model is used for variance-carrying particles, the physical meaning lies in the
addition of all trajectories rather than in any particular trajectory.
As mentioned above, the common assumption that the turbulent diffusion of the concen-
tration variance resembles that of the average concentration, makes it reasonable to use the
same mechanism to diffuse both the concentration carrying and variance carrying particles.
Summarizing, the method for calculating the concentration variance that has been applied
consists of the following stages:

1. The mean wind field was calculated, using a diagnostic mass-consistent model.
2. The turbulence parameters, TL (r) and σ (r), were estimated for the given mean wind field.
3. The mean concentration of a passive scalar, given a constant point source Q(r), was
calculated. The LS model of Kaplan and Dinar (1996) was used for this task.
4. The resulting distribution of the mean passive scalar concentration, C(r), and the fields
describing the turbulence (TL (r) and σ (r)) were utilized to evaluate the spatial distribution
of concentration variance sources, Q v (r), using Eq. 17.
5. The spatial distribution of the concentration variance was obtained using the same LS
model that was used to find the average concentration, and by using the same fields for
mean flow and turbulent diffusion. The effective sources distribution was determined by
Eq. 17, and the variance carried by the particles decayed exponentially with the time scale
τd , as expressed in Eq. 18.

2.4 The Decay Time τd

Generally, as noted by many authors (e.g. Sykes et al. 1984), the decay time τd is associated
with the time scale of eddies that have a similar length scale to that of the instantaneous
plume. This length scale is sometimes referred to as an inner scale, which differs from the
outer plume scale that takes into account the whole range of the energy spectrum, including
large meandering motions. In other words, τd ∼ σ d / d , where d is the inner length scale
of the plume, and σ d is the turbulent velocity scale associated with d . Sykes et al. (1984)
presented a model for the inner length and velocity scales with the assumption of an inertial
range energy spectrum.
In this work, we limit our attention to concentration fluctuations deep inside the urban
canopy. Far enough from the source, it is reasonable to assume that the inner length scale
of the plume converges to the length scale dictated by the urban canyon bounding it. This
means that the dissipation cascade of concentration fluctuations is governed by the eddies
that have the length scale of the street itself.
Following this assumption, and taking into account streets with more or less fixed widths,
we model τd as

123
 Author's personal copy
A. Manor

τd (r) = ATL (r), (19)

with A being a model constant, and TL (r) is the turbulent integral time scale at the parti-
cle location. The appropriate value for A will be determined in the following sections by
comparing the results of the model to measurements.

3 Computational Details

The domain used in this work is a 800 m × 800 m × 200 m domain, where the size of every
grid cell used for particle counting is x = y = z = 3.125 m. The same grid is used
both for the meteorological fields and for particle counting.
The spatial derivatives ∂C/∂ xi needed for evaluating Eq. 17, are evaluated using centre
differencing, and the variance source strength in each grid element is calculated. This results
in an enormous number of concentration variance sources. To reduce the computational load,
without ignoring important sources, a normalized concentration-variance source strength is
defined as
√
Q v (r)
Q (n)
v (r) ≡ . (20)
C(r)

Since the relevant resulting quantity is the concentration fluctuation intensity i ≡ c2 /C,
and not the variance itself, the importance of a source should not be estimated in terms  of
the amount of variance it emits, but rather by the ratio between its contribution to c2
and the mean concentration in its vicinity. Since the variance carried by the particles quickly
dissipates, estimating the mean concentration along their trajectory by the mean concentration
at the location of their source is quite reasonable.
(n)
After the normalized source strength Q v (r) is evaluated for each grid point, it is used to
sort the variance sources and the most important N sources are selected. Each of the selected
sources emits M concentration-variance particles. In order to ensure that choosing N sources
is sufficient, a fraction f is defined by
 (n)
Q v (i)
i⊂A
f ≡  (n)
, (21)
Q v (i)
i

where A is the group of chosen sources. High values of f (close to unity) imply that the
procedure worked well, and that the remaining unpicked sources are relatively negligible. In
the results to follow, the values of N and M are 100,000 and 100, respectively, and f > 0.97
for all cases. It is worth noting that the number of sources chosen is less than 3% of the total
collection of sources that fill the domain.
It is also important to point out that, though the overall number of variance particles needed
is N × M, which is a large number, the numerical load is relatively modest, since the particles
dissipate and can be terminated after the elapse of two or three dissipation time scales. Also,
the total number of particles needed is still much less than needed in micromixing models
(mentioned above in the introduction), in which every grid cell has to be populated by a
sufficient number of particles throughout the simulation.

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

Fig. 1 An illustration of the Oklahoma City urban domain (looking north) as introduced into the model

4 Comparison to the Joint Urban (JU2003) Field Experiment Results

In July 2003, a comprehensive field campaign was carried out in Oklahoma City, U.S.A. The
program was designed to study transport and diffusion of pollutants in the urban boundary
layer. A large number of tracer releases and sampling, using various sampling methods were
conducted, along with extensive meteorological measurements. In Fig. 1 the urban canopy in
the our computational domain is presented. The whole field program consisted of 10 intensive
observational periods (IOP), where the first six were devoted to daytime observations, and
the other four took place at night. Each IOP was comprised of several tracer releases, where
some were continuous 30-min releases, and the rest were instantaneous (puff) releases.
Here, two datasets of fast response S F6 tracer gas analyzers (TGA), based upon ECD
technology, are used. The first one was obtained by the Air Resources Laboratory Field
Research Division and consists of the data obtanied from 10 truck-mounted analyzers. The
second dataset is based on measurements conducted by ITT Industries, who operated five
more TGA samplers. The used TGA samplers were located at distances of more than 100 m
from the source. The time constant of these instruments is about 1 s and the sampling rate is
about 2 Hz. The results of these measurements were recently reported by Finn et al. (2010).
These authors thoroughly analyzed the observations, giving attention to several important
issues including the second, third and fourth moments of the concentration signals, the level
of intermittency and the frequency distributions.
For comparison with the model presented in this work, release number 2 in IOP 6 (day-
time), and release number 1 in IOP 8 (nighttime) were chosen. The reference wind speed
over rooftops was taken from the averaged wind speed measured at the top of the OKT
building (height 127 m). The constant in Eq. 19 was set to A = 22 for which best fit was
observed for each of the two datasets separately. The two chosen datasets reflect two very
different conditions (day versus night meteorology, different release locations, and different
wind directions). Consequently, the fact that the same value of A can be deduced from each
of the datasets separately implies that within the framework used here, its value is universal,
at least in respect to the considered environmental variables.
The fluctuation intensity i was extracted for each detector by dividing the standard devia-
tion of the concentration signal measured by the detector by its average value. The start and
end times for this calculation were determined by cloud arrival and departure times that are,
of course, different for every detector.

123
 Author's personal copy
A. Manor

Fig. 2 Comparison of concentration fluctuation intensity as calculated by sampler measurements (circles), to

the intensity estimated by the model (contours). The upper panel shows the results for the night release (IOP8
release 1), while the lower panel refers to the day release (IOP6 release 2). The release locations are shown
with an arrow. Wind speed at the reference height was set to 5.0 and 8.5 m s−1 and wind direction to 175◦
and 160◦ , for the daytime and nighttime releases, respectively

In Fig. 2, the concentration fluctuation intensity as assessed from sampler data is shown
(in circles) along with the prediction of the model (contours). The measure of agreement
between the predictions and the experimental data is presented in Fig. 3. Good agreement
for most of the measurement locations is evident.
It is important to note that the ratio between concentration fluctuation dissipation time and
the turbulent Lagrangian time scale, which has been used here to yield optimal agreement with
observations, is significantly higher than the value used in other studies. For example, both

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

Fig. 3 Comparison between

fluctuation intensity predicted by
the model (vertical axis) and
extracted from sampler
concentration measurements
(horizontal axis), for the day
release (◦), and for the night
release (+)

Fig. 4 Typical concentration signals as a function of time, for the nighttime release (IOP 8, release 1, left
panel) and for the daytime release (IOP 6, release 2, right panel)

Hsieh et al. (2007) and Milliez and Carissimo (2008) estimate the concentration fluctuation
dissipation time τd to scale with the Lagrangian turbulent integral time scale TL with a
multiplying factor close to unity. This discrepancy, together with a detailed sensitivity analysis
of the different model parameters, will be discussed in future work.

5 Discussion

Finn et al. (2010) report several differences between daytime and nighttime observations.
Concentration fluctuation intensities observed during the day were substantially higher in
comparison to those measured at night. Furthermore, differences in the frequency distribu-
tions were noted. At night, the frequency distributions tended toward Gaussian-like distrib-
utions whereas at daytime the distributions more closely matched a log-normal behaviour.
Finn et al. (2010) also report that the concentration time series measured by detectors
close to the plume centreline look qualitatively different during daytime than at nighttime.
Concentration signals during daytime (see, for example, Fig. 4, right panel) consisted of
intermittent pulses separated by minutes of clear air entrained into the plume. However, at
night (Fig. 4, left panel) the concentration fluctuates with low intermittency around its mean
value.

123
 Author's personal copy
A. Manor

Finn et al. stress that these differences between day and night exist in spite of similar
stability and turbulence conditions in the urban core for both day and night. They explain the
differences between day and night dispersions as a consequence of a combination of factors
including wind direction and meander, stability of the approaching flow and the different
release locations.
In this work, the simplified model that has been used does not consider any effects that
have to do with the static stability of the flow. The use of a stationary averaged wind
field also does not allow the introduction of the effects of large-scale meander. Thus,
the only difference between the two cases that has been studied involves the combina-
tion of wind speed and direction, relative to the built-up coverage, and to the release
location.

Fig. 5 Variance-source intensity (s−1/2 ) according to Eq. 20, at a height of 5 m above street level, for the
night release (above) and for the day release (below). In the latter case, there can be found relatively strong
source intensity, located close to the centreline of the plume, whereas in the first case, strong variance sources
are located further away from the centreline of the plume

123
 Author's personal copy
A Stochastic Single-Particle Lagrangian Model

(n)
The spatial distribution of Q v (r) (Eq. 20) 5 m above street level (Fig. 5) reveals what
seems to be the explanation of the difference between the fluctuation intensity at daytime
and at nighttime, in terms of the underlying framework of the model. From Eqs. 20 and
17, we should expect strong fluctuations in the vicinity of points for which the gradient in
average concentration is considerable. In the day release, the flow is channelled along the
street from south to north. Strong gradients in the wind speed are formed near street corners,
these gradients leading to the formation of strong gradients in the average concentration field,
very close to the centreline of the plume. These, in turn, create (through relation 17) dominant
sources for concentration variance.
In the night release, this phenomenon is much less pronounced. The wind direction over
rooftops does not follow any specific street direction, and thus channelling is not as strong
as in daytime. Consequently, gradients in the wind speed and in the average concentration
fields near the plume centreline are weaker, which implies the same for the concentration
variance sources.

6 Conclusions

In this study, a novel approach for estimating concentration fluctuation intensity in dense
built-up environments was described. The method is based upon treating the creation, dis-
sipation, transport, and diffusion of the ensemble-averaged concentration variance through
a LS particle scheme. Based on the time-averaged concentration field and on the turbulent
diffusion rate, effective sources of concentration variance are calculated throughout the com-
putational domain. This is followed by the emission of variance-carrying particles from these
sources, and the deduction of the spatial structure of the concentration variance using the
statistics of the particles.
As stressed, treating the concentration variance as a passive scalar such as the concentration
itself is definitely an approximation. However, other studies (e.g. Hsieh et al. 2007) show that
the transport and diffusion terms in the Reynolds-averaged concentration variance equation
can successfully be treated in exactly the same way as their counterparts in the equation for
the concentration.
The model possesses the advantages of LS particle models, useful especially in complex
urban environments and for practical risk assessment applications. Another advantage of the
new approach lies in its ability to be easily added as an extension for existing LS models in
common usage for mean concentration calculations.
The model is tested against observations from the JU20003 field experiment. One day-
time and one nighttime release were chosen for the comparison. In both cases, the model
satisfactorily reproduces the fluctuation intensity observed at sampler locations. Specifi-
cally, the model captures the observed difference between nighttime and daytime concen-
tration fluctuations in terms of the fluctuation intensity in the vicinity of the centreline of
the plume. This supports the hypothesis that the differences between nighttime and day-
time behaviour lies mostly in the combination of the buildings-wind orientation and the
release location, since these have significant influence on the magnitude of time-averaged
concentration gradients in the plume, which, in turn, serve as sources for concentration
variance.

Acknowledgments The author wants to express his gratitude to the Defence Threat Reduction Agency
(DTRA) for the JU2003 experiment data, to Prof. Dennis Finn for his useful directions, and to Dr. Haddassah
Kaplan for her important comments.

123
 Author's personal copy
A. Manor

References

Balwinder S, Hansen B, Brown M, Pardyjak E (2008) Evaluation of the QUIC-URB fast response urban wind
model for a cubical building array and wide building street canyon. Environ Fluid Mech 8:281–312
Cassiani M, Franzese P, Giostra U (2005) A PDF micromixing model of dispersion for atmospheric flow. Part
I: development of the model, application to homogeneous turbulence and to neutral boundary layer. Atmos
Environ 39(8):1457–1469
Csanady GT (1967) Concentration fluctuations in turbulent diffusion. J Atmos Sci 24:21–28
Dinar N (1984) Mass consistent models for wind distribution in complex terrain—fast algorithms for three
dimentional problems. Boundary-Layer Structure. Springer, Dordrecht, pp 177–199
Durbin PA (1980) A stochastic model of two-particle dispersion and concentration fluctuations in homogeneous
turbulence. J Fluid Mech 100(02):279–302
Finn D, Clawson KL, Carter RG, Rich JD, Biltoft C, Leach M (2010) Analysis of urban atmosphere plume
concentration fluctuations. Boundary-Layer Meteorol 136(3):431–456
Gifford F (1959) Statistical properties of a fluctuating plume dispersion model. Adv Geophys 6:117
Gonzalez M (1997) Analysis of the effect of microscale turbulence on atmospheric chemical reactions by
means of the p.d.f. approach. Atmos Environ 31(4):575–586
Hanna SR (1967) Spectra of concentration fluctuations: the two time scales of a meandering plume. Atmos
Environ 20(6):1131–1137
Hsieh K, Lien F, Yee E (2007) Numerical modeling of passive scalar dispersion in an urban canopy layer.
J Wind Eng Ind Aerodyn 95(12):1611–1636
Kaplan H, Dinar N (1996) A lagrangian dispersion model for calculating concentration distribution within a
built-up domain. Atmos Environ 30(24):4197–4207
Milliez M, Carissimo B (2008) Computational fluid dynamical modelling of concentration fluctuations in an
idealized urban area. Boundary-Layer Meteorol 127(2):241–259
Mussio P, Gnyp AW, Henshaw PF (2001) A fluctuating plume dispersion model for the prediction of odour-
impact frequencies from continuous stationary sources. Atmos Environ 35:2955–2962
Nelson MA, Pardyjak ER, Klein P (2011) Momentum and turbulent kinetic energy budgets within the Park
Avenue Street Canyon during the joint urban 2003 field campaign. Boundary-Layer Meteorol 140:143–162
Ride D (1995) A practical method of estimating toxic loads in the presence of concentration fluctuations.
Environmetrics 6:643–650
Röckle R (1990) Bestimmung der Strömungsverhältnisse im Bereich komplexer Bebauungsstrukturen. Ph.D.
dissertation.
Sawford BL (1983) The effect of Gaussian particle-pair distribution functions in the statistical theory of
concentration fluctuations in homogeneous turbulence. Q.J.R. Meteorol Soc 109(460):339–354
Sawford BL (2004) Micro-mixing modelling of scalar fluctuations for plumes in homogeneous turbulence.
Flow Turbul. Combust. 72(2):133–160
Sherman CA (1978) A mass consistent model for wind fields over complex terrain. J Appl Meteorol 17:312–
319
Sykes RI, Lewellen WS, Parker SF (1984) A turbulent-transport model for concentration fluctuations and
fluxes. J. Fluid Mech. 139:193–218
Sykes RI (1984) The variance in time-averaged samples from an intermittent plume. Atmos Environ 18(1):121–
123.
Thomson DJ (1987) Criteria for the selection of stochastic models of particle trajectories in turbulent flows.
J Fluid Mech 180:529–556
Tinarelli G, Brusasca G, Oldrini O, Anfossi D, Trini Castelli S, Moussafir J (2007) Micro-swift-spray (MSS):
a new modelling system for the simulation of dispersion at microscale. General description and validation.
In:Air pollution modeling and its applications XVII, Springer, New York, pp 449–458

123