Building and Environment 223 (2022) 109506

Contents lists available at ScienceDirect

Building and Environment

journal homepage: www.elsevier.com/locate/buildenv

Partially averaged Navier-Stokes simulation of flow around an isolated

building model with a 1:1:2 shape
Hongyuan Jia *, Hideki Kikumoto
Institute of Industrial Science, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8505, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: This study analyzes the performance of a numerical model with partially averaged Navier–Stokes (PANS)
Partially averaged Navier-Stokes equations for simulating the flow field around an isolated building model with a 1 (length): 1 (width): 2 (height)
Isolated high-rise building shape. The PANS model was based on the k − ω SST unsteady Reynolds-averaged Navier-Stokes equations
Computational fluid dynamics
(URANS) model, and three different values were imposed on the coefficient fk , which represents the proportion of
Pedestrian-level wind environment
modeled turbulence kinetic energy (TKE) in PANS. The simulation results were validated by experimental
measurements and compared to those of large-eddy simulation (LES) and URANS. According to the results, it was
found that the resolution ability of PANS can be adjusted between URANS and LES by tuning fk . When compared
to URANS, PANS improved the performance in predicting the mean flow field around the target building and
separation layers in the wake region. The prediction error in the reattachment length in the wake region was
reduced by 88% compared to URANS. Because PANS can explicitly resolve more turbulence, the prediction
accuracy of TKE against the experiment (FAC2: 0.85) is higher than that of URANS (FAC2: 0.45). Through power
spectrum analysis, it was determined that PANS can reproduce the characteristic vortex shedding caused by the
building with specific frequencies. When compared to LES, it resolved approximately 99% of turbulence fluc­
tuations in the relatively low-frequency domain at the representative point in the wake region.

1. Introduction LES has become increasingly popular for urban wind simulations. Its
ability to resolve most turbulence structures explicitly makes its pre­
Computational fluid dynamics (CFD) has been used as an effective diction accuracy higher than that of RANS. With proper simulation
tool in urban wind environments for many years to analyze problems techniques, LES results can be considered as replacements for wind
such as pedestrian wind [1,2], pollutant dispersion [3–7] and turbulence tunnel experiment (WTE) measurements [12,13]. Many studies have
structures around buildings [6–9]. utilized LES to clarify the detailed flow patterns [14] and transport
Among the rich turbulence models in CFD, the most popular models mechanisms of pollutants around buildings [15,16]. However, the high
for building simulation are the Reynolds-averaged Navier–Stokes performance of LES cannot be guaranteed without a fine mesh and
(RANS) model and large-eddy simulation (LES). RANS has been used for sufficient calculation resources, which constrains its popularity in in­
decades to predict steady flow fields around buildings. Based on its dustrial applications.
modest requirements in terms of computational resources, a large Unsatisfied with the distinctive characteristics of RANS and LES,
number of industrial applications still rely on this model [10]. various researchers began to explore possible bridging methods between
However, it is well known that unsteady fluctuations cannot be re­ these techniques, hoping that both low computational complexity and
flected accurately by RANS. Additionally, based on the limitation of high accuracy could be realized simultaneously. Unsteady RANS
turbulence closure modeling in RANS, it is difficult to accurately predict (URANS) adds time-marching iteration to RANS to simulate variance of
time-averaged flow fields and turbulence properties near buildings. the velocity field and encouraging improvements compared to RANS
Furthermore, the separation at sharp corners of buildings and reat­ have been reported [11,17].
tachment behaviors on roofs disappear in most RANS models [11]. Regardless, the temporal or spatial scales of averaging filters become
Based on the continued development of computational technologies, unclear when RANS evolves into URANS. The turbulence modeling

* Corresponding author.
E-mail address: [email protected] (H. Jia).

https://doi.org/10.1016/j.buildenv.2022.109506
Received 17 May 2022; Received in revised form 4 July 2022; Accepted 14 August 2022
Available online 20 August 2022
0360-1323/© 2022 Elsevier Ltd. All rights reserved.
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

limitations in RANS, like bad performance for shear flow or rotating 2. Methodology
flow, also occur in URANS, meaning a gap in simulation performance
relative to LES still exists. Because LES and RANS are complementary to The PANS model can be derived from the governing equations of
each other [10], hybrid RANS/LES modeling approaches have also been RANS [26]. PANS based on the standard k − ε RANS model has been
proposed, such as the detached eddy simulation (DES). DES behaves like firstly proposed and applied to the simulation of flow past a square
URANS near walls and like LES away from walls, which seems to balance cylinder [27], flow past a circular cylinder, driven cavity flow [26], and
calculation cost and accuracy reasonably [18]. Liu & Niu [19] applied jet flow [28]. It was found that PANS outperformed both URANS (k − ε)
DES to simulate the flow around an isolated high-rise building and and even LES when they all used a modest grid resolution. However, the
compared it to LES and RANS. They also testified the performance of flow field structures of buildings are different from that of the square
DES by simulating the pedestrian-level wind of an elevated building cylinder, which is mainly assumed to be infinitely long. The building
[20] and building array [21]. Their systematic research revealed that model is mounted on the bottom boundary with limited height. The end
DES can provide comparable accuracy to LES with a coarse mesh and effect causes the flow field to become strongly three-dimensional. Apart
fewer computing resources. This intriguing finding has also been from the typical structures like Karman vortex shedding in the hori­
confirmed by more applications of DES in the wind engineering like the zontal plane, the circulation and separation caused by the roof in the
high-rise structure with an elliptical shape [22] and steep hilly terrain vertical plane also deduce other arch-type vortex shedding and down­
[23]. Even though DES is potential to become a more affordable option wash effects [37,38]. In this case, it is vital to examine the
for LES, several points still need further improvements. In the results of three-dimensional features of the surrounding flow of the building.
[19], DES failed to produce the separation bubble on the roof, which is However, this k − ε model form may perform poorly when the target
an important characteristic of the flow field surrounding the building. In field has strong separation [39] or wall-bounded flows [32], which are
addition, DES is found to be sensitive to mesh quality in LES regions the main characteristics of flows around buildings considered in this
[23], where coarse grids may cause severe numerical instabilities. study.
Critical issues about the grid resolution and mechanism caused by the Additionally, this study aimed to compare the performances of
switching between LES and URANS in the bridging region need to be URANS, PANS, and LES, but k − ε URANS cannot reproduce the Karman
solved [24,25]. vortex shedding behind a building [11]. Therefore, it would be unfair to
In recent years, a novel bridging model between direct numerical compare URANS and PANS in the k − ε form. Therefore, we selected the
simulation (DNS) and RANS has been proposed. This model is called the k − ω SST model [40] for URANS, which has been confirmed as a suc­
partially averaged Navier-Stokes (PANS) model [26]. PANS applies a cessful model for predicting the flow separation around building corners
constant/dynamic averaging filter to the Navier-Stokes equation and [11]. Correspondingly, we applied the k − ω SST PANS model [41] as
uses two coefficients to control the turbulence resolution. When the shown below.
mesh is sufficiently fine, PANS behaves like DNS and directly resolves If we apply an arbitrary averaging filter 〈 • 〉 to the incompressible
fluctuations. When the mesh is coarse, PANS behaves like URANS to Navier-Stokes equation, according to the averaging invariance of tur­
model most turbulence [27]. Subsequent research [28] has connected bulence filtering [42], it becomes
these adjustments to grid dimensionality and the local turbulence scale. ( )
Therefore, PANS realizes a “smart” method for tuning performance and ∂ui ∂ui ∂τ vi , vj ∂ pu ∂2 ui
+ uj + =− +ν
balancing accuracy and mesh quality. Additionally, PANS uses an ∂t ∂xj ∂xj ∂xi ∂xj ∂xj
( ) (1)
averaging filter with intuitive physical meaning, making it more ∂ 2 pu ∂ui ∂uj ∂2 τ vi , vj
− = +
attractive than URANS. ∂xi ∂xi ∂xj ∂xi ∂xi ∂xj
PANS has been applied in many application scenarios, including flow
past square cylinders [27], flow past trucks [29] and high-speed trains where ui is the averaged velocity of i component, pu is the averaged
[24], flow in a centrifugal impeller [30], jet flow [28], lid-driven cavity pressure, and τ(vi , vj ) denotes the generalized central moment. The
flow [31], and flow around a hydrofoil [32]. In all of these cases, the second part of Eq. (1) is the averaged form of Poisson equation as a
performance of PANS was confirmed to be promising. It is more accurate substitution for the continuity equation.
than URANS, and more efficient than LES when coarse grid resolution In RANS, the filter 〈 • 〉 takes the ensemble average of turbulent
was adopted. It is reasonable to infer that PANS would be a powerful tool motions and τ(vi , vj ) evolves into Reynolds stress. The governing equa­
for simulating urban wind flow. However, to the best of our knowledge, tions for the k − ω SST RANS model are defined as follows [40]:
examples of applying PANS to a flow field around a three-dimensional ∂k ∂k
(
∂ νt ∂k
)
building in relatively high-turbulence flows are still sparse, even + u = P̃k − β∗ ωk +
∂t ∂x ∂x σk ∂x
though this is a typical problem in wind engineering. Unlike most of the ( )
(2)
∂ω ∂ω γ ∂ νt ∂ω 1 ∂k ∂ω
flow fields listed above, this type of field contains complicated
2
+ u = Pk − β ω + + (1 − F1 ) • 2σω2
∂t ∂x νt ∂x σ ω ∂x ω ∂x ∂x
three-dimensional vortex structures with different scales stemming from a1 k
strong separation and wall-bounded flows. The prediction of such tur­ νt =
max (a1 ω, SF2 )
bulent fields is useful for several important research problems, such as
pedestrian wind comfort and pollutant dispersion. Therefore, it is where u is the ensemble average of velocity, k is the modeled turbulence
meaningful to examine the applicability of PANS to urban wind kinetic energy. The relationship between ω and the dissipation term ε is
environments.
In this study, PANS was used to simulate the flow around an isolated ε = β∗ ωk (3)
building model with a 1 (length): 1 (width): 2 (height) shape, which is a
classic benchmark [2,11,33–36]. The results were compared to those of Pk is the production term for k. P̃k is defined as min (Pk , cl β∗ ωk). The
URANS, LES, and WTE measurements to evaluate prediction accuracy. values of the coefficients β∗ , σ k , β, σω , F1 , σ ω2 , a1 , and F2 can be found in
Section 2 introduces the governing equations of the PANS model. A the paper by Ref. [40].
detailed description of the case study and the settings for each CFD Unlike RANS, PANS assumes that a filter 〈 • 〉 only partially averages
model are presented in Section 3. The performance of PANS for the turbulent motions, where the unresolved energy and dissipation can be
simulation of the mean flow field, turbulent kinetic energy (TKE), and defined as
time series fluctuations is detailed in Section 4. Section 5 provides our ku = fk × k, 0 < fk ≤ 1
concluding remarks and discusses possible future research directions. (4)
εu = fε × ε, 0 < fε ≤ 1

2
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

fk and fε describe the resolution ability of PANS, which is dependent on bridging region between different resolutions, and a bridging region
the mesh size and can also be controlled by users. The coefficient fω can closure modification is necessary for PANS [43]. Therefore, for the
be similarly defined as simulation simplicity of this early attempt at a building model, this
research prescribed constant values 0.3, 0.5, and 0.7 to fk to investigate
ωu = fω × ω (5) its effects on the simulation.
According to Eq. (3), these three coefficients have the following
relationships: 3. Case description

εu β∗ ωu ku 3.1. Basic simulation settings

fε = = ∗ = fω × fk (6)
ε β ωk
In this study, the target building was set to a 1(x): 1(y): 2(z) block
When fk and fε change to one, PANS becomes RANS, meaning all TKE
model with a height H = 0.2 m, as shown in Fig. 1. This model is one of
and dissipation will be modeled. By introducing these partially averaged
the most basic benchmarks for urban wind environment research.
coefficients into Eq. (2), the transport equation for ku and ωu can be
Detailed WTE measurements have been shared by the Architectural
derived.
( ) Institute of Japan for this model (https://www.aij.or.jp/jpn/publish
∂ku ∂ku ̃ ∂ νu ∂ku /cfdguide/index_e.htm). Different CFD models have been applied to
+u = Pku − β∗ ωu ku +
∂t ∂x ∂x σku ∂x simulate the flow around this model [11,33,35] and the results have
( )
∂ωu ∂ωu γ 1 γ ∗ been used to adjust simulation strategies, which can function as guid­
+u = Pku + − 1 β ωu ku
∂t ∂x νu fω νu ance for the simulation of more complex building geometries. Therefore,
( )
ω 2
∂ νu ∂ωu fω 1 ∂ku ∂ωu we employed this model to verify the performance of PANS for the
− β u+
fω ∂x σωu ∂x
+ (1 − F1 ) • 2σω2
fk ωu ∂x ∂x (7) simulation of flow past a building.
Three different CFD models, namely URANS, PANS, and LES, were
fk
σ ku = σk implemented based on the finite volume method using the open-source
fω
software OpenFOAM v.2112 [44]. URANS used the k − ω SST model
fk proposed by Ref. [40]. PANS was implemented based on Eq. (7). LES was
σ ωu = σ ω
fω implemented as a standard Smagorinsky model with a Smagorinsky
constant [45] Cs = 0.12, which has been confirmed to be appropriate for
where P̃ ku , β ωu ku are the production and dissipation term for ku
∗
the flow around buildings [36].
respectively. The simulation settings followed experience from previous research
For the PANS simulation of Eq. (7), the most critical coefficients to be [35]. The calculation domain was 13H(x) × 6H(y) × 5H(z) and the inlet
determined are fk and fε (or fω ), which represent the ratios of unresolved was 3H away from the leeward surface of the block. A space of 10H was
energy and dissipation over the total amounts of energy and dissipation. reserved for the development of vortex shedding behind the block. The
In most simulations in urban wind environment areas, meshes are not Reynolds number based on the reference velocity uref and H is about
sufficiently fine to simulate dissipation explicitly, so the value of fε is 1.056 × 105 . uref was measured at (− 3H, 0, H). A structured grid with a
commonly set to one [26]. fk was imposed as a global constant when minimum size of H/40 expanded from the block surface to the bound­
PANS was originally proposed [26], and different values of fk may affect aries with an expansion ratio of 1.08, which is set as the basic mesh. To
the model accuracy. Girimaji & Abdol-Hamid [28] modified fk into a check the grid sensitivity, a fine mesh with a minimum size of H/80 and
variable related to the mesh size and length scale of local turbulence. a coarse mesh with a minimum size of H/20 were also generated.
However, it was clarified later that the spatial-temporal variants of fk in The selection of the discretization scheme is important for the ac­
non-stationary turbulence will lead to communication errors in the curacy of CFD and an inappropriate scheme may result in severe

Fig. 1. Schematic of a 1:1:2 high-rise building model and its calculation domain.

3
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

numerical fluctuations or viscosity. Ikegaya et al. and Ono et al. [33,46] verted to ku by fk , three dominant components of the Reynolds stress
discussed the effects of advection schemes on the LES results for isolated were multiplied by 1 − fk and then input into DFM to generate ‘filtered’
buildings and determined that blending schemes with 90% second-order inflow fluctuations. Meanwhile, ku and ωu profiles based on Eqs. (4) and
central and 10% first-order upwind schemes can provide better pre­ (5) were imposed on to inlet as shown in Fig. 2(b) to represent the
dictions than pure central or upwind schemes. Additionally, there has turbulence that was unresolved.
been limited research exploring the relationship between advection In order to check the sustainability of the inflow boundary during the
schemes and PANS performance. To ensure consistency in the compar­ development, Fig. 2(c ~ h) shows the vertical profiles of mean velocity
ison, this blending advection scheme called ‘filteredLinear2’ in Open­ and TKE at different locations in the LES, URANS, and PANS (fk = 0.7)
FOAM was employed in all three models. Similarly, the second-order with the basic mesh. It can be observed that the mean velocity profiles
implicit time marching scheme ‘Backward’ with one step of 0.0004 s are well sustained in all three cases. The difference between the three
(basic mesh) was imposed on all cases to strictly limit the Courant locations and the WTE is acceptable. As for the TKE, the profiles in the
Number below 1. For all three simulations, the first 30 s were used for LES become closer to the WTE measurements in the lower part when the
initialization to ensure the convergence of the simulations, and the inflow is approaching the building model as shown in Fig. 2(f). How­
following 60 s were used for data sampling to ensure that the results ever, k and ku in URANS and PANS tend to slightly decrease near the
were stochastically steady. bottom boundary, which is considered to be related to the friction
Regarding the boundary conditions, the bottom and block surfaces dissipation and wall function effects. Despite that, the changings in
were defined as nonslip walls with the Spalding wall function. The two profiles are limited, so the generated inflows properly developed and are
sides of the calculation domain were also defined as nonslip walls to reliable for the simulations.
correspond to the walls in a WTE. The top wall was defined as a free-slip
boundary for velocity and a zero gradient for pressure. The outlet was
set with the advective boundary condition with zero pressure and a zero 3.3. Grid independence
gradient for velocity. The above boundary conditions were kept the
same for URANS, PANS, and LES, while different settings for the inlet are The grid resolution testing was conducted based on the LES with
introduced as below. three mesh sets: LES_fine, LES_basic, and LES_coarse. Fig. 3 compares the
profiles of the streamwise component of mean velocity and the mean
normal stress in the vertical plane (y = 0) and the horizontal plane (z =
3.2. Inflow generation
H /16). The results of LES_coarse show evident disparity from the other
two cases. The second-order turbulence statistics were underestimated
The inflow profiles designed according to WTE measurements were
on the roof and overestimated in the top part of the wake region by the
used for the inlet in URANS. The vertical profiles of velocity and TKE at
LES_coarse. Meanwhile, the LES_basic almost yielded the same result as
the inlet are compared to the WTE measurements in Fig. 2. One can see
the LES_fine except that the normal stress of the separation layer at the
that the properties of the approaching flow in the simulation are close to
side of the building was slightly smaller. Considering the computational
those in the WTE.
cost, the basic grid set is considered to be the economical choice in this
In contrast, this profile cannot be applied to LES and PANS. On the
study, which is the same result as the previous research [34]. This mesh
one hand, LES requires the turbulent inflow boundary condition, which
setting was applied to all three simulations in the following contents to
could significantly influence the turbulence field and energy production
discuss the simulation results.
around an isolated building model [6]. One common inlet setting for LES
is loading a turbulent inflow database, which can be obtained through
preliminary LES simulations to reproduce the inflow generated in a wind 4. Results & discussion
tunnel with spires and roughness [34].
On the other hand, PANS has rarely been applied in urban wind 4.1. PANS with different f k values
environment areas before. The method to set the inlet boundary con­
ditions for the velocity, unresolved ku , and unresolved ωu to realize Firstly, we compared the simulation results of PANS with different fk
turbulent inflow is still an unsettled issue. To the best of our knowledge, values to investigate the effects of resolution coefficients on the accu­
existing research on PANS simulation has mainly applied laminar inflow racy. Fig. 4 shows the time-averaged profiles of vertical velocity com­
without turbulence [47] or fixed time-averaged profile for ku and ωu to ponents on the central vertical plane in all PANS cases. The caparisons of
the inlet boundary [24]. However, PANS lies between LES and URANS velocity profiles in other directions as well as the results on the hori­
that the TKE of the inflow is divided into two parts: temporal velocity zontal plane can be found in the Appendix. A. It can be observed that
fluctuations which are explicitly simulated and unresolved ku . Thereby, the averaged velocity distributions are similar in all cases. Although
both the time-series turbulent database for velocity and profile for ku are small deviations exist along with the separation layer in the top part of
necessary. the wake region, the simulated velocities are close to the WTE mea­
To generate different turbulent inflow corresponding to different fk , surements. Among all cases, fk = 0.3 generally resulted in the best
this research applied a digital filter method (DFM) [48] to produce agreement with the WTE.
inflow data for LES and PANS. This method was proposed to generate However, the difference turned out to be conspicuous in the TKE
artificial velocity data for LES or DNS by specifying first and results as shown in Fig. 5. Because PANS is based on an average filtering
second-order statistics at each point as well as an autocorrelation approach, the total TKE kt is composed of two components: the explicitly
function. It was verified to be a useful tool to provide the inflow resolved TKE ke caused by large-scale flow and the implicit sub-filter
boundary condition for LES [49,50]. Users need to provide the Reynolds TKE ku that is modeled.
stress tensor, length scale tensor and mean velocity profile to drive the
generation process. In this research, the WTE measurement values were ( kt = ke + ku )/ (8)
input into this method to produce the turbulent inflow data for LES. To ke = u’x 2 + u’y 2 + u’z 2 2
examine the applicability of the generated inflow, we extracted the
mean velocity and TKE profiles of it downstream of the inlet (x = − Here, ui denotes the velocity fluctuations explicitly resolved by the
′

0.25H) and compared them to those of WTE in Fig. 2(a&b). Although the simulations. kt values are presented in Fig. 5.
TKE in the bottom part is underestimated, most of the values agree well On both planes, it was confirmed that more TKE was resolved with
with the WTE measurements. the increase of fk value. In this case, the result of fk = 0.3 is closer to the
Then, as for PANS, because part of velocity fluctuations was con­ measurements than those of the other two cases, especially in the sep

4
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 2. Comparisons of vertical inflow profiles between simulations and WTE. (a): mean velocity profile at the inlet; (b): total TKE in WTE and LES, unresolved TKE in
PANS at the inlet; (c)~(e): developed mean velocity profiles at different locations; (f)~(h): developed TKE profile of LES and URANS, unresolved TKE of PANS at
different locations.

5
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 3. Comparisons of mean streamwise velocity and variance u′x 2 predicted by LES with coarse, basic and fine mesh. Top: vertical plane(y = 0), bottom: horizontal
plane (z = H /16), left: streamwise velocity, right: variance of streamwise velocity u′x 2 .

aration layers on the roof and side of the block. Cases with larger fk tend
to underestimate TKE in these separation areas. In addition, even fk =
0.3 cannot ensure that the simulation is consistent with WTE on the
horizontal plane. TKE was obviously underestimated in the wake region.
Considering that the horizontal plane is close to the bottom wall, small-
scaled turbulence is difficult to predict. Using smaller meshes and fk
value may improve the current result. At x = − 0.75H, a large deviation
between simulation and WTE has appeared. It has been pointed out that
this problem may lie in the experimental measurements because the
presence of a probe may interfere with the block and bottom wall [33].
Fig. 6 summarized the profiles of ku in each PANS case in the vertical
and horizontal planes. Contrary to the results of kt , the values of ku in­
creases when fk is larger since more TKE is modeled, especially in the
vertical plane. In the horizontal plane, the difference of ku mainly con­
centrates on the front region and side of the block, where fk = 0.7 still
has the largest value. However, the three cases all have limited ku in the
wake region, which indicates that the modeling of turbulence is insuf­
ficient, so the prediction of TKE in the wake region can only rely on the
explicit simulation.
Based on the velocity and TKE results, it is concluded that smaller fk
can bring better simulation accuracy because more turbulence can be
explicitly simulated with the current mesh, which is fine enough for LES.
Fig. 4. Comparisons of mean vertical velocity fields on a vertical plane (y = 0) This finding agrees with other PANS simulations in the previous
between PANS cases with different fk values and WTE. research [27,41]. Consequently, we selected fk = 0.3 to represent PANS

6
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 5. Comparisons of total TKE kt predicted by PANS cases with different fk

values and the WTE. Top: vertical plane(y = 0), bottom: horizontal plane
Fig. 6. Comparisons of unresolved TKE ku predicted by PANS cases with
(z = H /16).
different fk values. Top: vertical plane(y = 0), bottom: horizontal plane
(z = H /16).
in the following comparison between LES and URANS.
Fig. 8 compares the mean velocity fields in the horizontal plane with
z = H/16 between the three simulations and WTE. This plane is close to
4.2. Comparisons of LES, PANS, and URANS
the bottom wall boundary. One can see that PANS reaches the same level
of accuracy as LES. However, both PANS and LES slightly underestimate
In this part, the performance of LES, PANS (fk = 0.3), and URANS is
the streamwise velocity of the separation layer on the side of the block
compared to each other in the aspects of the time-averaged flow field
when y > 0.5H while such a phenomenon did not appear in the URANS.
and turbulence properties.
This may result from the different inflow boundary conditions that
URANS used a velocity profile, while LES and PANS used generated
4.2.1. Mean velocity field
turbulent database. In the result of spanwise velocity, LES and PANS
Fig. 7 compares the mean velocity fields in the vertical plane with
kept a high accuracy. URANS exhibits large deviations in the wake re­
y = 0 between the simulations of three models and WTE. In general, LES
gion where y < 0.5H because it cannot correctly reproduce the separa­
yields the best agreement with the WTE at most positions. However, it is
tions and wake vortex shedding.
noteworthy that the predictions of PANS have the same level of accuracy
As a result, with the same mesh and advection schemes, the pre­
as those of LES except that it slightly deviates from LES at x = 1.25H for
diction accuracies of the mean velocity field for PANS and LES are
the velocity in the vertical direction. In the wake region, both PANS and
similar, whereas URANS exhibits relatively large errors due to the model
LES exhibit good agreements with the WTE, whereas URANS has rela­
limitations.
tively large prediction errors.

7
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 7. Comparisons of mean velocity fields on a vertical plane (y = 0) be­

tween LES, PANS (fk = 0.3), URANS and WTE. Top: streamwise velocity, bot­ Fig. 8. Comparisons of mean velocity fields on a horizontal plane (z = H/16)
tom: vertical velocity. between LES, PANS (fk = 0.3), URANS and WTE. Top: streamwise velocity,
bottom: spanwise velocity.
4.2.2. Streamline and reattachment
One significant requirement for CFD models in the simulation of flow measurement results. Additionally, the separation bubble on the roof
past buildings is the reproduction of turbulent structures induced by the has become larger. The reattachment is too close to the leeward side,
sharp edges of buildings. The representative characteristics are the where part of the structure is mixed with the wake vortex. However, the
separation layer on the roof, geometry of the wake vortex, and reat­ URANS results are still much better than those of steady RANS, where
tachment positions on the roof and in wake regions. Steady RANS the reattachment position in the wake is greater than 2H and no reat­
models have been criticized because of their failure to reproduce reverse tachment occurs on the roof [11]. It is worth noting that the results here
flows on roofs and their overestimation of reattachment positions are also better than the performance of k − ω SST URANS reported by
behind buildings [11]. Therefore, it is necessary to verify the perfor­ Ref. [11], where reattachment on the roof could not be observed. The
mance of PANS in this aspect. URANS streamline in Fig. 9 is similar to that of URANS with the RNG k −
Fig. 9 presents the time-averaged streamlines in the central vertical ε model and modified ε equation presented by Ref. [11]. This improved
plane produced by the three models and the WTE measurement. The performance may stem from different advection schemes because an
capabilities of LES have been confirmed in several previous studies [6, upwind scheme was used in the previous simulations, which could result
33,36] and the main flow field characteristics can be clearly observed in in redundant numerical viscosity.
the simulated flow field. This finding was successfully reproduced in this The streamline of PANS is much closer to the WTE measurements
study and one can see that the reattachment points are at approximately than those of URANS. The reattachment point in the wake region is
0.3H and 0.7H on the roof and wake region, respectively. These values approximately x = 0.75H, which is the same as that of LES. The devia­
agree with previous experimental results [51]. tion from the experiment is reduced by approximately 88% compared to
In contrast, for URANS, the reattachment length in the wake region URANS. Besides, the separation bubble on the roof is complete in the
reaches above 1.1H, which is almost double the value in the WTE PANS result even though it is slightly longer than that in the LES.

8
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 9. Comparisons of mean streamline and reattachment positions in the vertical plane (y = 0) predicted by three turbulence models and the WTE.

Therefore, the time-averaged flow structures were well captured by cient when compared to WTE, especially in the wake region. This
PANS. insufficiency is expectable since fk = 0.3 is not small enough to promise
that the resolution ability of PANS is the same as LES, especially in the
4.2.3. Turbulence kinetic energy vicinity of the bottom boundary, where small-scaled turbulence
In addition to the time-averaged flow field, the turbulence model was frequently appears.
also expected to predict the turbulent energy and fluctuations correctly.
Fig. 10 summarizes the TKEs calculated by the three models and com­ 4.2.4. Quantitative comparison
pares them to the WTE measurements. Since a significant amount of TKE The quantitative analysis of the flow field simulated by three models
is modeled as k or ku in URANS and PANS, their kt following Eq. (8) is is conducted in this part by means of FAC2 [52] and RMSE (Root Mean
presented in Fig. 10. The contribution of unresolved TKE to the total kt is Square Error) values:
limited in LES by the current mesh, ke is regarded as kt for LES in Fig. 10. ⎧
⎪ Pi
In the vertical plane, PANS and LES accurately predict the TKE at ⎪
⎪
⎪ 1, for 0.5 < < 2,
⎨ Oi
most locations, including the peak on the roof and the wake region. It is 1 ∑N
FAC2 = ni with ni = (9)
notable that LES overestimated the peak value of TKE on the top of the N i=1 ⎪
⎪
⎪
1, for |Oi | ≤ W and |Pi | ≤ W,
⎪
block. This ripple was also found in Ref. [33] and was believed to result ⎩
0, otherwise.
from numerical oscillation brought by a large portion of the
second-order central scheme in the advection term. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑N
1 2
i=1 (Oi − Pi )
URANS provides the worst estimate, where the TKE is significantly RMSE =
N
(10)
underestimated for two reasons. First, because URANS theoretically V
applies an averaging filter with a large time scale, only large-scale pe­
Here, N is the total number of measurement points in the WTE, P is the
riodic fluctuations are directly reflected. Most stochastic turbulence is
prediction values of each model, O is the observation data. W represents
modeled in the same manner as steady RANS, which is apparently
the allowed ranges and takes the value of 0.005 for kt and 0.05 for other
insufficient for strong circulation flow. Second, the k− ω SST model
physical quantities according to the measurement uncertainty. V is a
includes a maximum limit on the production term of kt in the vicinity of
normalized factor, which takes the value of uref for velocity and u2ref for
the building, such as the rooftop, which may lead to the excessive
reduction of TKE [11]. kt . We calculated these two indices of ux , uz and kt for 207 measurement
For the profiles of x = − 0.75H in the horizontal plane at z = H/ 16, points in the vertical plane (y = 0), ux , uy and kt for 326 measurement
none of the models reproduces as much energy as the experimental re­ points in the horizontal plane (z = H /16). Detailed distribution of the
sults which may result from the measurement noise discussed above measurement points can be found in Ref. [13]. Table 1 summarized all
[33]. Except for this location, the scenario is identical to that of the the results of three simulations.
vertical plane. LES and PANS can predict much more TKE than URANS. According to FAC2 results, it can be noticed that values of LES are
By comparing Figs. 5 and 10, it can be observed that TKE results of all close to 1, which is superior to PANS, but their differences in most
PANS cases lie between LES and URANS. Larger fk makes PANS physical quantities are within 0.05 range. The streamwise velocity
approach to URANS and underestimate the TKE in the separation layer prediction of PANS in the vertical plane is even better than LES. What’s
and wake region, while smaller fk makes PANS perform like LES. more, the TKE in the horizontal plane was not perfectly predicted by
Despite that, it is worth noting that results of PANS are still insuffi­ PANS due to that the resolution ability is inadequate with fk = 0.3 as

9
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

two times of that of PANS, which indicates poor performance in the

velocity variance. In general, this quantitative analysis confirmed that
prediction accuracy is notably improved when we changed from URANS
to PANS.

4.2.5. Power spectrum density (PSD) function of velocity

Finally, we focus on the PSD function of velocity at four represen­
tative points to reveal how the three models behave in turbulent fluc­
tuation simulations. Fig. 11 presents the positions of the four points and
PSD curves. The horizontal axis uses a logarithmic scale and was non­
dimensionalized based on uref and the side width b = H/2 of the block.
All PSDs were processed by a 1.2 × 104 points Hamming window using
50% overlapping segments. Considering length restrictions, we do not
present the PSD results in all three velocity directions, but only the
dominant component at each point.
Point A is located in the upper region above the building, meaning it
does not receive large effects from the turbulence generated by the
block. Therefore, the LES curve exhibits a typical shape for turbulent
inflow that is relatively flat without dominant peaks. The energy con­
tained in LES and PANS is much larger than that in URANS because
URANS only apply inflow profiles with a constant velocity at the inlet.
Meanwhile, it should be noted that URANS isn’t expected to have sig­
nificant fluctuations since most of the turbulence has been modeled. It is
a correct behavior that all PSD curves of URANS exhibit weaker fluc­
tuation than LES and PANS in the following contents.
Because the prescribed Reynolds stress of inflow generation was
reduced to correspond with fk = 0.3, part of turbulence fluctuations in
the inflow was transferred into ku in PANS. This adjustment was re­
flected in the PSD curve that it starts to descend from the LES curve at
fb
uref = 0.1, which means the high-frequency fluctuations were cut off in
the PANS. However, when the inflow was generated by DFM, fluctua­
tions with all scales were adjusted by multiplying Reynolds stress with
fk . It is interesting to note that PANS only lost the high-frequency fluc­
tuations while it contains most of the low-frequency flow structures
since the PSD curve is almost equal to that of LES. The reason could be
that during the development of the inflow profile from the inlet to the
block, PANS changes the proportion of modeled turbulence to the total
in different frequencies. Small-scaled fluctuations were transferred to ku
while part of ku was recovered into explicit, large-scaled fluctuations.
Point B is located in the wake region and is strongly dominated by
vortex shedding from the side of the building. One can see that all three
fb
curves exhibit a noticeable peak at approximately uref = 0.09, which is
Fig. 10. Comparisons of total TKE kt predicted by three simulations and the
close to the Strouhal number obtained in previous WTEs [53] and sim­
WTE. Top: vertical plane(y = 0), bottom: horizontal plane (z = H /16).
ulations [11,33]. The peak in the LES curve is not as obvious as that in
the PANS curve because more scales of turbulence are resolved and
discussed above. FAC2 values of URANS were about 0.1 lower than
mixed. Additionally, the large difference in the high-frequency area
those of PANS for most terms. Especially, it poorly predicted the TKE in
between PANS and LES at point A decreases significantly due to two
the horizontal plane, where the FAC2 is only 0.45.
reasons. The first one is that the influence of turbulent inflow is
RMSE can reflect the normalized error of predictions against obser­
destroyed by vortex shedding from the block. The modeled part ku in the
vations. The situation is like that of FAC2 that the difference between
inflow has gradually transferred to velocity fluctuations and was
PANS and LES is limited. It is indicated that their errors are smaller than
explicitly resolved by PANS in the fine mesh area. Besides, the fluctua­
those of URANS in the vertical plane. kt errors of all three models
tions are dominated by the vortex shedding in the low-frequency
became too small because the denominator in Eq. (10) is rather large. It
domain, and the small-scaled vortex with high-frequency became
is interesting that the velocity prediction of URANS is better than PANS
fewer. Despite that, in the high-frequency domain, PANS resolves less
in the horizontal plane. However, its RMSE of kt reaches up to almost

Table 1
Quantitative analysis results of three simulations.
Model Index

FAC2_vertical FAC2_horizontal RMSE_veritcal RMSE_horizontal

ux uz kt ux uy kt ux uz kt ux uy kt

LES 0.90 0.97 1 0.94 0.90 0.96 1.3% 0.7% 4 × 10− 4 1.6% 1.1% 5 × 10− 4
PANS 0.92 0.95 0.99 0.94 0.88 0.85 1.3% 0.7% 4 × 10− 4 1.8% 1.2% 7 × 10− 4
URANS 0.83 0.85 0.88 0.86 0.77 0.45 1.8% 0.9% 9 × 10− 4 1.6% 1.2% 13 × 10− 4

10
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. 11. PSD functions of dominant velocity components at four representative points predicted by three different models.

energy than LES because small turbulence is filtered with fk = 0.3. The is also identical to the above descriptions for points B and C. It can be
situation for URANS does not change significantly after point A and the confirmed that the development of main vortex shedding begins at this
energy of critical flow structures is reflected in the PSD curve, which is point because URANS already has a peak at approximately fb
= 0.09.
uref
apparently inadequate for some application scenarios. To evaluate the
However, the shedding in LES and PANS was blurred by other structures,
energy in the low-frequency area of each PSD curve, each curve was
so the peaks were unclear.
integrated with ufbref from zero to one. Approximately 99% of the energy of
LES was resolved by PANS. This proportion is much larger than the 15%
of URANS. 4.3. Discussion about the computing resources
Point C is inside the separation bubble on the roof, where strong
fluctuations appear in the streamwise and vertical directions. The raised In the last part, we discussed PANS’s potential of saving computa­
fb tional costs. Based on the basic mesh above, we simply changed the
curvature in LES and peak in PANS appear when = 0.5, which agrees
uref minimal grid size in the x direction to H/30 and kept the resolutions of
well with the value confirmed in previous studies [11,33]. Therefore, the other two directions unchanged to produce a coarser mesh set. The
PANS successfully reproduced the turbulence resulting from the sepa­ mesh number of this configuration is 4.86 × 105 , which is about 11%
ration of building corners. The difference between PANS and LES in the smaller than that of the basic mesh. This coarse mesh was used in a PANS
high-frequency became larger than point B because much more simulation with fk = 0.3 (PANS_C30), and an LES (LES_C30). The
small-scaled turbulence was produced in the roof region, most of which calculation time of PANS_C30 in the platform with AMD EPYC 7452 CPU
were implicitly modeled in PANS. Again, URANS cannot predict the (16 cores used) and 128 GB RAM is about 23 h. As a comparison, the LES
fluctuations in this area and most of the TKE is directly modeled. with basic mesh in the same platform is about 32 h, which means that
The main characteristic at point D is the turbulence shedding from PANS_C30 saved about 28% of the computing time.
the side edge of the building. The relationship between the three curves Fig. 12 compared the simulation results between LES_C30,

11
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

PANS_C30, and WTE in the vertical plane (y = 0) and horizontal plane physically adjusts the proportion of explicitly resolved velocity fluctu­
(z = H /16). It can be confirmed that there is barely any difference be­ ations at each place, its value should be related to the local length-scale
tween LES_C30 and PANS_C30 in the results of mean streamwise ve­ of the vortex and the grid size in order to reach higher efficiency of
locity profiles. Both of them could still accurately reproduce the PANS. In this case, one of the main advantages of PANS can be realized
distribution as the WTE measurements. As for the TKE profiles, there are that the simulation can adjust the resolution levels according to the
noticeable distinctions between LES_C30 and PANS_C30. In the vertical existing mesh. It should be noted that the previous research where PANS
plane, the peak value on the roof was still slightly underestimated by evidently outperformed LES mainly utilized spatially varying fk [24,27].
PANS_C30. However, in the wake behind the building, PANS_C30 has Moreover, this benchmark only includes one single block, which is
better consistency with WTE than LES_C30 in the lower and downstream relatively simple compared to the real building simulation. It should be
regions. The possible reason is that the grid size of these two regions is reassessed the performance of PANS when complex shapes and unideal
relatively too large to the resolution requirements of turbulent struc­ meshes are involved. Therefore, whether PANS is economical enough for
tures in LES. The TKE results in the horizontal plane near the bottom building simulation or not is still an open question until more scenarios
boundary also demonstrate this point, where PANS_C30 is closer to the like the building canyon and block array benchmarks are checked.
WTE. Even for the peak caused by the side separation, PANS_C30 almost
has the same value as the LES_C30. 5. Conclusions
However, it has to admit that the advantage of PANS over LES in the
coarse mesh here is not as obvious as that shown in the previous research This study applied a PANS model based on k − ω SST RANS to the
[24,27]. There are some possible reasons for that. First, in this research unsteady simulation of turbulent flow past an isolated building model
PANS and LES used the same mesh, time step interval, and same dis­ with a 1:1:2 shape and evaluated its performance by comparing the
cretization scheme, which have been testified to be the optimal selec­ results to WTE measurements, LES, and URANS (k − ω SST) results. To
tions to ensure the performance of LES [13]. Thereby, LES still has simulate the building that emerged in a high level of turbulence, DFM
enough robustness even with the coarse mesh. Future research is was employed to generalize the turbulent inflow with different levels of
necessary to investigate the optimal combination of acceptable grid explicit fluctuations to meet different resolution requirements of PANS.
resolution and discretization schemes for PANS. Besides, fk was imposed Our main conclusions can be summarized as follows.
as a spatiotemporal constant in this research. Considering that it

Fig. 12. Comparisons of mean streamwise velocity and total TKE kt predicted by LES_C30, PANS_C30, and the measurements of WTE. Top: vertical plane(y = 0),
bottom: horizontal plane (z = H /16), left: streamwise velocity, right: total TKE.

12
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

1. It was confirmed that PANS’s performance can be adjusted between This study represents the early stages of applications of PANS in the
LES and URANS by changing the value of fk . When the mesh is fine urban wind environment area. Therefore, there are some limitations and
enough, smaller fk makes PANS explicitly resolve more turbulence unclear issues, leading to additional research questions. For example,
and yields better prediction. what are the optimal settings for fk with tempo-spatial distribution?
2. By comparing PANS (fk = 0.3) to LES and URANS, we found that What are the best advection schemes for the PANS simulation of flow
PANS reaches the same level of accuracy as LES in the time-averaged past buildings? It should also be investigated whether PANS can provide
velocity field. URANS exhibits the largest deviation from WTE satisfactory simulation results with economical mesh for different
measurements because it fails to reproduce the vortex shedding building simulation benchmarks. Finally, it has been demonstrated that
around buildings correctly. Consequently, the predictions of the the partially averaged filtering process can be adopted in almost all
mean streamline and reattachment behaviors of the flows were RANS models to develop different PANS models. These PANS models
significantly improved by PANS compared to URANS. The separation would inherit the characteristics of the basic RANS models. It is neces­
bubble was completely reproduced by PANS and the deviation in the sary to determine the optimal PANS models for different scenarios. In
estimation of the reattachment position in the wake region relative to our opinion, future research should consider these problems.
the WTE results was reduced by approximately 88% compared to
URANS. The averaged flow field in the vertical plane of PANS is CRediT authorship contribution statement
almost the same as that of LES.
3. In general, PANS provides a better prediction of TKE than URANS Hongyuan Jia: Writing – original draft, Visualization, Validation,
because it can resolve more turbulence around the building. How­ Methodology, Data curation, Conceptualization. Hideki Kikumoto:
ever, its accuracy is not as good as that of LES in the vicinity of the Writing – review & editing, Supervision, Funding acquisition,
bottom boundary because fk = 0.3 is not small enough for PANS to Conceptualization.
resolve the small-scaled turbulence in this region. The quantitative
analysis reinforced the results of graphical comparisons. Declaration of competing interest
4. According to the PSD results for four representative points, PANS can
reproduce the critical vortex shedding caused by buildings at specific The authors declare that they have no known competing financial
frequencies. There are deficiencies in the high frequency domain interests or personal relationships that could have appeared to influence
because small-scaled turbulence is modeled in the PANS. Despite the work reported in this paper.
that, at the point in the wake region, PANS resolved 99% of sto­
chastic fluctuations in the relatively low-frequency domain Data availability
compared to LES, which is a recognizable advantage over URANS.
Data will be made available on request.
In summary, by introducing the resolution adjustment coefficients
into the equations of URANS, PANS significantly improves the simula­ Acknowledgements
tion performance for flows past an isolated building and is comparable
to LES in some aspects. It has the potential to be one of the efficient This work was supported by the Japan Society for the Promotion of
models to simulate urban airflow in industrial applications in the future. Science (JSPS) KAKENHI Grant Number 20H02308.

Appendix. A

Fig. A.1. Comparisons of mean streamwise velocity fields in a vertical plane (y = 0) between PANS cases with different fk values and WTE.

13
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

Fig. A.2. Comparisons of mean velocity fields in a horizontal plane (z = H/16) between PANS cases with different fk values and WTE. Top: streamwise; Bot­
tom: spanwise.

References [6] B. Zhang, R. Ooka, H. Kikumoto, Analysis of turbulent structures around a

rectangular prism building model using spectral proper orthogonal decomposition,
J. Wind Eng. Ind. Aerod. 206 (2020), 104213, https://doi.org/10.1016/j.
[1] B. Blocken, T. Stathopoulos, J.P.A.J. van Beeck, Pedestrian-level wind conditions
jweia.2020.104213.
around buildings: review of wind-tunnel and CFD techniques and their accuracy for
[7] Y. Zhang, X. Yang, H. Yang, K. Zhang, X. Wang, Z. Luo, J. Hang, S. Zhou, Numerical
wind comfort assessment, Build. Environ. 100 (2016) 50–81, https://doi.org/
investigations of reactive pollutant dispersion and personal exposure in 3D urban-
10.1016/j.buildenv.2016.02.004.
like models, Build. Environ. 169 (2020), 106569, https://doi.org/10.1016/j.
[2] N. Ikegaya, T. Kawaminami, T. Okaze, A. Hagishima, Evaluation of exceeding wind
buildenv.2019.106569.
speed at a pedestrian level around a 1:1:2 isolated block model, J. Wind Eng. Ind.
[8] S. Hassan, MdM. Molla, P. Nag, N. Akhter, A. Khan, Unsteady RANS simulation of
Aerod. 201 (2020), https://doi.org/10.1016/j.jweia.2020.104193,
wind flow around a building shape obstacle, Build. Simulat. 15 (2022) 291–312,
104193–104193.
https://doi.org/10.1007/s12273-021-0785-8.
[3] H. Kikumoto, R. Ooka, A numerical study of air pollutant dispersion with
[9] T. Kono, T. Tamura, Y. Ashie, Numerical investigations of mean winds within
bimolecular chemical reactions in an urban street canyon using large-eddy
canopies of regularly arrayed cubical buildings under neutral stability conditions,
simulation, Atmos. Environ. 54 (2012) 456–464, https://doi.org/10.1016/j.
Boundary-Layer Meteorol. 134 (2010) 131–155, https://doi.org/10.1007/s10546-
atmosenv.2012.02.039.
009-9434-y.
[4] F. Li, J. Liu, J. Ren, X. Cao, Predicting contaminant dispersion using modified
[10] B. Blocken, LES over RANS in building simulation for outdoor and indoor
turbulent Schmidt numbers from different vortex structures, Build. Environ. 130
applications: a foregone conclusion? Build. Simulat. 11 (2018) 821–870, https://
(2018) 120–127, https://doi.org/10.1016/j.buildenv.2017.12.023.
doi.org/10.1007/s12273-018-0459-3.
[5] X. Liu, J. Niu, K.C.S. Kwok, Evaluation of RANS turbulence models for simulating
[11] Y. Tominaga, Flow around a high-rise building using steady and unsteady RANS
wind-induced mean pressures and dispersions around a complex-shaped high-rise
CFD: effect of large-scale fluctuations on the velocity statistics, J. Wind Eng. Ind.
building, Build. Simulat. 6 (2013) 151–164, https://doi.org/10.1007/s12273-012-
Aerod. 142 (2015) 93–103, https://doi.org/10.1016/j.jweia.2015.03.013.
0097-0.

14
H. Jia and H. Kikumoto Building and Environment 223 (2022) 109506

[12] J.-O. Mo, A. Choudhry, M. Arjomandi, Y.-H. Lee, Large eddy simulation of the wind [32] X. Luo, R. Huang, B. Ji, Transient cavitating vortical flows around a hydrofoil using
turbine wake characteristics in the numerical wind tunnel model, J. Wind Eng. Ind. k-ω partially averaged Navier-Stokes model, Mod. Phys. Lett. B 30 (2016) 1–10,
Aerod. 112 (2013) 11–24, https://doi.org/10.1016/j.jweia.2012.09.002. https://doi.org/10.1142/S0217984915502620.
[13] T. Okaze, H. Kikumoto, H. Ono, M. Imano, N. Ikegaya, T. Hasama, K. Nakao, [33] N. Ikegaya, T. Okaze, H. Kikumoto, M. Imano, H. Ono, Y. Tominaga, Effect of the
T. Kishida, Y. Tabata, K. Nakajima, R. Yoshie, Y. Tominaga, Large-eddy simulation numerical viscosity on reproduction of mean and turbulent flow fields in the case
of flow around an isolated building: a step-by-step analysis of influencing factors of a 1:1:2 single block model, J. Wind Eng. Ind. Aerod. 191 (2019) 279–296,
on turbulent statistics, Build. Environ. (2021), 108021, https://doi.org/10.1016/j. https://doi.org/10.1016/J.JWEIA.2019.06.013.
buildenv.2021.108021. [34] T. Okaze, H. Kikumoto, H. Ono, M. Imano, T. Hasama, T. Kishida, K. Nakao,
[14] Z. Xie, I.P. Castro, LES and RANS for turbulent flow over arrays of wall-mounted N. Ikegaya, Y. Tabata, Y. Tominaga, Large-eddy simulations of flow around a high-
obstacles, Flow, Turbul. Combust. 76 (2006) 291–312, https://doi.org/10.1007/ rise building: validation and sensitivity analysis on turbulent statistics, in:
s10494-006-9018-6. Presented at the Proceedings of the 7th European and African Conference on Wind
[15] V.B.L. Boppana, Z.T. Xie, I.P. Castro, Large-eddy simulation of dispersion from Engineering, 2017, pp. 3–6. Liege, Belgium.
surface sources in arrays of obstacles, Boundary-Layer Meteorol. 135 (2010) [35] Y. Tominaga, A. Mochida, R. Yoshie, H. Kataoka, T. Nozu, M. Yoshikawa,
433–454, https://doi.org/10.1007/s10546-010-9489-9. T. Shirasawa, AIJ guidelines for practical applications of CFD to pedestrian wind
[16] H. Jia, H. Kikumoto, Construction of urban turbulent flow database with wavelet- environment around buildings, J. Wind Eng. Ind. Aerod. 96 (2008) 1749–1761,
based compression: a study with large-eddy simulation of flow and dispersion in https://doi.org/10.1016/j.jweia.2008.02.058.
block-arrayed building group model, J. Wind Eng. Ind. Aerod. (2020), https://doi. [36] Y. Tominaga, A. Mochida, S. Murakami, S. Sawaki, Comparison of various revised
org/10.1016/j.jweia.2020.104433, 104433–104433. k-ε models and LES applied to flow around a high-rise building model with 1:1:2
[17] Y. Tominaga, T. Stathopoulos, Steady and unsteady RANS simulations of pollutant shape placed within the surface boundary layer, J. Wind Eng. Ind. Aerod. 96
dispersion around isolated cubical buildings: effect of large-scale fluctuations on (2008) 389–411, https://doi.org/10.1016/j.jweia.2008.01.004.
the concentration field, J. Wind Eng. Ind. Aerod. 165 (2017) 23–33, https://doi. [37] H.F. Wang, Y. Zhou, The finite-length square cylinder near wake, J. Fluid Mech.
org/10.1016/J.JWEIA.2017.02.001. 638 (2009) 453–490, https://doi.org/10.1017/S0022112009990693.
[18] P.R. Spalart, W.H. Jou, M. Strelets, S.R. Allmaras, Comments on the feasibility of [38] B. Zhang, R. Ooka, H. Kikumoto, Identification of three-dimensional flow features
LES for wings, and on a hybrid RANS/LES approach, in: Presented at the Ist AFSOR around a square-section building model via spectral proper orthogonal
Int. Conf. On DNS/LES, 1997, 1997. decomposition, Phys. Fluids 33 (2021), 035151, https://doi.org/10.1063/
[19] J. Liu, J. Niu, CFD simulation of the wind environment around an isolated high-rise 5.0041395.
building: an evaluation of SRANS, LES and DES models, Build. Environ. 96 (2016) [39] H.K. Versteeg, W. Malalasekera, An Introduction to Computational Fluid
91–106, https://doi.org/10.1016/j.buildenv.2015.11.007. Dynamics: the Finite Volume Method, second ed., Pearson Education Ltd, Harlow,
[20] J. Liu, J. Niu, C.M. Mak, Q. Xia, Detached eddy simulation of pedestrian-level wind England ; New York, 2007.
and gust around an elevated building, Build. Environ. 125 (2017) 168–179, [40] F. Menter, T. Esch, Elements of industrial heat transfer prediction, in: 16th
https://doi.org/10.1016/j.buildenv.2017.08.031. Brazilian Congress of Mechanical Engineering, 2001.
[21] J. Liu, J. Niu, Delayed detached eddy simulation of pedestrian-level wind around a [41] F.S. Pereira, G. Vaz, L. Eça, S.S. Girimaji, Simulation of the flow around a circular
building array – the potential to save computing resources, Build. Environ. 152 cylinder at Re =3900 with Partially-Averaged Navier-Stokes equations, Int. J. Heat
(2019) 28–38, https://doi.org/10.1016/j.buildenv.2019.02.011. Fluid Flow 69 (2018) 234–246, https://doi.org/10.1016/j.
[22] B.W. Yan, Q.S. Li, Detached-eddy and large-eddy simulations of wind effects on a ijheatfluidflow.2017.11.001.
high-rise structure, Comput. Fluids 150 (2017) 74–83, https://doi.org/10.1016/j. [42] M. Germano, Turbulence: the filtering approach, J. Fluid Mech. 238 (1992)
compfluid.2017.02.009. 325–336, https://doi.org/10.1017/S0022112092001733.
[23] T. Zhou, Q. Yang, B. Yan, X. Deng, Y. Yuan, Detached eddy simulation of turbulent [43] S.S. Girimaji, S. Wallin, Closure modeling in bridging regions of variable-resolution
flow fields over steep hilly terrain, J. Wind Eng. Ind. Aerod. 221 (2022), 104906, (VR) turbulence computations, J. Turbul. 14 (2013) 72–98, https://doi.org/
https://doi.org/10.1016/j.jweia.2022.104906. 10.1080/14685248.2012.754893.
[24] S. Krajnović, G. Minelli, B. Basara, Partially-averaged Navier–Stokes simulations of [44] H.G. Weller, G. Tabor, H. Jasak, C. Fureby, A tensorial approach to computational
two bluff body flows, Appl. Math. Comput. 272 (2016) 692–706, https://doi.org/ continuum mechanics using object-oriented techniques, Comput. Phys. 12 (1998),
10.1016/J.AMC.2015.03.136. https://doi.org/10.1063/1.168744, 620–620.
[25] F.R. Menter, M. Kuntz, Adaptation of eddy-viscosity turbulence models to unsteady [45] J. Smagorinsky, General circulation experiments with the primitive equations: I.
separated flow behind vehicles, in: R. McCallen, F. Browand, J. Ross (Eds.), The The basic experiment, Mon. Weather Rev. 91 (1963) 99–164, https://doi.org/
Aerodynamics of Heavy Vehicles: Trucks, Buses, and Trains, Lecture Notes in 10.1175/1520-0493(1963)091<0099:GCEWTP>2.3.CO;2.
Applied and Computational Mechanics, Springer Berlin Heidelberg, Berlin, [46] H. Ono, H. Takimoto, T. Michioka, A. Sato, Convection term discretization for large
Heidelberg, 2004, pp. 339–352, https://doi.org/10.1007/978-3-540-44419-0_30. eddy simulations based on the finite volume method: effect of thermal
[26] S.S. Girimaji, Partially-averaged Navier-Stokes model for turbulence: a Reynolds- stratification on the dispersion characteristics of rooftop exhaust, Part 2,
averaged Navier-Stokes to direct numerical simulation bridging method, Journal of J. Environ. Eng. 80 (2015) 1143–1151, https://doi.org/10.3130/aije.80.1143.
Applied Mechanics, Transactions ASME 73 (2006) 413–421, https://doi.org/ [47] P. Ranjan, A. Dewan, Partially Averaged Navier Stokes simulation of turbulent heat
10.1115/1.2151207. transfer from a square cylinder, Int. J. Heat Mass Tran. 89 (2015) 251–266,
[27] E. Jeong, S.S. Girimaji, Partially averaged Navier–Stokes (PANS) method for https://doi.org/10.1016/j.ijheatmasstransfer.2015.05.029.
turbulence simulations—flow past a square cylinder, J. Fluid Eng. 132 (2010) [48] M. Klein, A. Sadiki, J. Janicka, A digital filter based generation of inflow data for
1–11, https://doi.org/10.1115/1.4003153. spatially developing direct numerical or large eddy simulations, J. Comput. Phys.
[28] S.S. Girimaji, K.S. Abdol-Hamid, Partially-averaged Navier Stokes model for 186 (2003) 652–665, https://doi.org/10.1016/S0021-9991(03)00090-1.
turbulence: implementation and validation, in: 43rd AIAA Aerospace Sciences [49] P. Domingo, L. Vervisch, D. Veynante, Large-eddy simulation of a lifted methane
Meeting and Exhibit - Meeting Papers, 2005, pp. 12887–12900, https://doi.org/ jet flame in a vitiated coflow, Combust. Flame 152 (2008) 415–432, https://doi.
10.2514/6.2005-502. org/10.1016/j.combustflame.2007.09.002.
[29] G. Minelli, E.A. Hartono, V. Chernoray, L. Hjelm, B. Basara, S. Krajnović, [50] G.R. Tabor, M.H. Baba-Ahmadi, Inlet conditions for large eddy simulation: a
Validation of PANS and active flow control for a generic truck cabin, J. Wind Eng. review, Comput. Fluids 39 (2010) 553–567, https://doi.org/10.1016/J.
Ind. Aerod. 171 (2017) 148–160, https://doi.org/10.1016/j.jweia.2017.10.001. COMPFLUID.2009.10.007.
[30] C. yue Wang, F. jun Wang, B. hong Wang, Y. Tang, H. ru Zhao, A novel Omega- [51] Y. Meng, K. Hibi, Turbulent measurments of the flow field around a high-rise
driven dynamic PANS model, J. Hydrodyn. 32 (2020) 710–716, https://doi.org/ building, Wind Engineers, JAWE (1998) 55–64, https://doi.org/10.5359/
10.1007/s42241-020-0052-y. jawe.1998.76_55, 1998.
[31] B. Akula, P. Roy, P. Razi, S. Anderson, S. Girimaji, Partially-averaged Navier-Stokes [52] M. Schatzmann, H. Olesen, J. Franke, COST 732 MODEL EVALUATION CASE
(PANS) simulations of lid-driven cavity flow—Part 1: comparison with URANS and STUDIES: APPROACH AND RESULTS, Meteorological Inst, 2010.
LES, in: S. Girimaji, W. Haase, S.-H. Peng, D. Schwamborn (Eds.), Notes on [53] S. Murakami, A. Mochida, On turbulent vortex shedding flow past 2D square
Numerical Fluid Mechanics and Multidisciplinary Design, Springer International cylinder predicted by CFD, Journal of Wind Engineering and Industrial
Publishing, Cham, 2015, pp. 359–369, https://doi.org/10.1007/978-3-319-15141- Aerodynamics, Third Asian-Pacific Symposium on Wind Engineering 54–55 (1995)
0_29. 191–211, https://doi.org/10.1016/0167-6105(94)00043-D.

15