Вы находитесь на странице: 1из 14

EACWE4 The Fourth European & African Conference on Wind Engineering

J. Naprstek & C. Fischer (eds); ITAM AS CR, Prague, 11-15 July, 2005, Paper #138

Numerical simulation of the flow across an asymmetric street intersection




J. Franke, W. Frank

ABSTRACT: A numerical simulation of the flow across an asymmetric street junction formed by four rings
of buildings has been performed with the commercial flow solver FLUENT V6.1 using the Reynolds averaged
Navier-Stokes equations with six different turbulence models. The turbulence models comprised four linear
two-equation models and a differential Reynolds stress model with two formulations for the pressure-strain.
The simulations were done for two directions of the approach flow using three systematically refined grids for
each case. The results on the finest grid were shown to differ only slightly from the next coarser grid. These
results were then compared with velocity and turbulence measurements that are available from the CEDVAL
model while the
database. The best results for the velocity are obtained with the realizable and standard
Reynolds stress models perform best for the Reynolds stresses and slightly worse for velocity.


much better agreement for the LES with the measuerments than for the RANS with standard
model. LES was also used by [Hanna et al. (2002)]
for the flow over a matrix of cubes with reasonable agreement between simulation results and experiments. While LES leads in general to better predictions than RANS its costs are currently too high for
engineering applications. Therefore industrial scenarios are normally treated with the RANS approach,
e.g. [Ferreira et al. (2002), Richards et al. (2002),
Ketzel et al. (2002), Westbury et al. (2002)]. None of
the cited works uses higher-order turbulence models
and many did not analyse the influence of the grid resolution on the results.

1 INTRODUCTION

The detailed knowledge of the interaction of the atmospheric boundary layer flow with clusters of obstacles
like buildings, trees or other obstructions is important for several reasons. Understanding and predicting
the forces and loadings on the obstructions requires
knowledge of the complex flow patterns around the
obstacles and within the streets. This information is
also needed to assess the mechanical wind effects on
pedestrians and to determine the dispersion of pollutants or contaminants.
Besides the experimental investigation of the
flow within the built environment in boundary
layer wind tunnels, Computational Fluid Dynamics (CFD) is increasingly used for the numerical
simulation of the flow around clusters of buildings. [Lien & Yee (2004)] computed the flow over
a matrix of cubes solving the Reynolds Averaged
Navier Stokes (RANS) equations with the standard
and [Kato & Launder (1993)]
turbulence model.
While there was a good agreement between measured and computed velocities, the turbulent kinetic
energy was underpredicted. A similar case has been
studied by [Cheng et al. (2003)]. Besides the RANS
equations with the standard
turbulence model
they used Large Eddy Simulation (LES) with different subgrid scale models. While RANS computations solve for the statistically steady flow field directly, LES computes the unsteady flow field formed
by the larger eddies and only models the influence of
the unresolved eddies. [Cheng et al. (2003)] report


In the present work we therefore analyse the performance of the RANS approach with six turbulence
models, including two differential Reynolds stress
models, for the prediction of the flow over a moderately complex cluster of four rings of buildings
forming an asymmetric street junction. For this case
velocity and turbulence measurements are available
for three directions of the approach flow from the
C EDVAL database (www.mi.uni-hamburg.de/cedval)
[Leitl (2000)]. The simulations are performed with
the commercial flow solver F LUENT V6.1 for two directions of the approach flow, using three systematically refined grids in each case. The grid dependance
of the solutions is analysed only qualitatively. The
computational results for the mean velocities and the
turbulence on the finest grid are then compared with
the experimental data.

Dr.-Ing. Jorg Franke, Department of Fluid- and Thermodynamics, University of Siegen, e-mail franke@ift.uni-siegen.de
Prof. Dr.-Ing. Wolfram Frank, Department of Fluid- and Thermodynamics, University of Siegen, e-mail frank@ift.unisiegen.de

The Fourth European & African Conference on Wind Engineering, Paper #138

2 NUMERICAL METHOD

The modelled equation for the turbulent kinetic


energy is
!

2.1 Basic equations


The basic equations describing a turbulent, neutrally
stratified flow are the Navier-Stokes equations with
constant density and dynamic viscosity . When
these equations are averaged over time one arrives at
the RANS equations for a statistically steady turbulent flow. These are mass conservation,





!

(4)


is the production of
*

and defined as


(5)

(1)

For all two-equation models that are used in this


work,
is computed from

and momentum conservation,











,


where
tensor,

denotes the velocity components of the


Here,
mean, i.e. time averaged velocity vector and
is
the mean pressure.
are the Reynolds stresses
containing the time averaged correlations of velocity
fluctuations . The Reynolds stresses have to be determined to close the system of equations (1) and (2).
In this work we use the linear eddy viscosity approximation for the Reynolds stresses and differential
second-moment modelling, where additional equations are solved for each of the Reynolds stresses.
Contrary to that the deviatoric part of the Reynolds
stresses is approximated by analogy with the molecular stresses in linear eddy viscosity modelling,

(6)

(2)

is the modulus of the mean rate-of-strain


3







(7)

3
3

3


In Equation (4)
denotes the diffusion coefits dissipation. These quantificient of and
ties have different definitions for the different twoequation turbulence models that are used. Altogether
four models are used of which three belong to the
class of
models and the fourth is the shear stress
transport
model of [Menter (1994)]. All
models solve a modelled transport equation for the
dissipation in addition to Equation (4).
)






!




(3)









8

(8)

'

Here,
is the turbulent kinetic energy,
the Kronecker delta function and the turbulent
or eddy viscosity. The turbulent viscosity is computed
from the turbulent kinetic energy and a measure of
its dissipation, for each of which an additional transport equation is solved. Therefore these models are
called two-equation turbulence models.

>

>

>

Here,
is the diffusion coefficient of and
and
its production and dissipation, respectively.
Their definitions are listed in Table 1 for the twoequation turbulence models which are briefly introduced in the following.
)

>

>

Table 1 Definitions for


?

turbulence models
C

model
SKE
RNG
RKE

>

>

>

>

'


8




'

!
!

Q


D
D

O
G





8

>


G

X
G

1.44
1.42


1.92
1.68
1.9

0.09
0.0845
U
K

turbulence models
C

'

>

Table 2 Constants for

model
SKE
1.3 1.0
RNG

RKE 1.0 1.2

'

G


U
8

0.4187 9.793
0.4187 9.793
0.4187 9.793

>

J. Franke, W. Frank


is computed from Equation 9. The remaining constants are given in Table 2.

Standard
(SKE) model
The SKE model of [Launder & Spalding (1972)]
is still widely used in CWE despite its stagnation
point anomaly, which leads to an overprediction of
in stagnant flow regions. Despite this shortcoming it yields reasonable results for the flow within
clusters of obstacles, e.g. [Lien & Yee (2004)].
The constants used in this model are given in Table 2.
ReNormalization Group (RNG)
model
The RNG model of [Yakhot & Orszag (1986)]
mainly differs from the SKE model in the equation
for . The parameter in the dissipation term of
is defined as


As has been said before, the fourth two-equation


model also solves the transport equation for , Equation 4. The second equation is now for the specific
.
dissipation


with

(9)

from Equation (7). The constants are


and
. This modification leads to a
in regions of large strain rate and theresmaller
fore to lower values of , thus alleviating the stagnation point anomaly.
in the diffusion coeffiFinally the coefficient
and , see Table 1, has the value
cients for
.
Realizable
(RKE) model
The RKE model of [Shih et al. (1995)] takes two of
the three realizability conditions for the Reynolds
stresses into account. As the Reynolds stresses
form a real valued tensor all its invariants must be
non negative, e.g. [Schumann (1977)]. The RKE
model fulfills the condition that the normal stresses
are non negative and that the Schwarz inequality
holds. The first condition leads to a limitation of
the production of at high rate of strain and therefore reduces the stagnation point anomaly. To fulis not constant
fill the realizability conditions
is a
but defined as given in Table 2.
constant and is defined as


Shear Stress Transport (SST)


model
The SST model of [Menter (1994)] combines the
model of [Wilcox (1988)] with
standard
the SKE model and also takes the transport effects of the principal turbulent shear stress into account through a modified turbulent viscosity formulation. It has been shown to perform very well
for adverse pressure gradient aerodynamic flows
but was not often applied for building aerodynamics, see e.g. [Westbury et al. (2002)]. The blendand the SKE model leads
ing of the standard
to many equations that are necessary for a detailed
description. Due to space limitations we do not
provide the detailed model here but refer to the
original work of [Menter (1994)] and the manual
of [F LUENT (2003)].

All linear two-equation models are known to


have shortcomings and deficiencies wherever stress
transport and normal stresses are important, leading to poor predictions in separating and reattaching flows [Hanjalic (1999)]. While non linear twoequation models are able to render improved results,
full second-moment modelling is the most general
approach for the closure of Equations (1) and (2).
Within that approach a transport equation is solved
for each of the Reynolds stresses. These modelled
equations are

"

'

(
1

/
R

(
/

P
D

B
P

Y
P

,
2

(13)
P

@
@

'

where

(10)

is the mean rate-of-rotation tensor,


.

(
1

@
,

*
1

*


(11)


/
(

is defined as

(


P
P

D


@
%

8
7

<

:
=

<

>

(12)

where the model of [Lien & Leschziner (1994)]


for the turbulent diffusive transport has been already
taken into account, as well as the approximation of the
dissipation tensor with the scalar dissipation .
denotes the convective term and
the shear stress

The dissipation of is also changed, cf. Table 1.


is the kinematic viscosity. Additionally,
is no longer constant but a function of which
A

"

"

The Fourth European & African Conference on Wind Engineering, Paper #138

generation. Both terms do not need further modelling


as does the pressure-strain term . We use two different models for , a linear and a quadratic model.


!






(18)
















Linear Pressure-Strain Model (LRR-IP)


This model is based on the proposal of Launder, Reece and Rodi [Launder et al. (1975)]. The
pressure-strain term is decomposed as



#

1







/







>
#

#
#

1
=




(14)






(15)




Both, the LRR-IP and the SSG model use Equation (8) to compute the scalar dissipation . The production and dissipation term are computed as in the
SKE model with the same constants. However, the
production term
is now computed exactly from
as
. In this form it is also used in the
transport equation for , Equation (4), where the coefficient , which is also used in the diffusion term of
Equation (14), is now
. is computed as in
the SKE model with the same value for . The equation for is only necessary to compute in the wall
adjacent cells. In the rest of the domain is computed
directly from the normal stresses. Near the wall is
needed to compute the Reynolds stresses. Assuming
a logarithmic distribution of the velocity and equilibrium of the turbulence and neglecting convection and
diffusion in Equation (14), the stresses are

is the slow pressure-strain term,


where
the rapid pressure-strain term and
the wallis modelled according to
reflection term.
[Rotta (1951)],


where
is the Reynolds stress anisotropy tensor
. The rapid part
defined as
is modelled as proposed by [Fu et al. (1987)],


"

!



!



#




"

#


(16)


"

is approximated as proposed by
Finally,
[Gibson & Launder (1978)]





'
(

#



(17)

"


/




(19)




5


>

>

>

>







#


#
#

3
1

in a local coordinate system with the tangential,


the normal and the binormal coordinate.
The logarithmic velocity distribution is also used
to compute the tangential velocity
in the wall adjacent cells with all the presented turblence models.
As part of the walls in the computational domain are
rough the formulation of the standard wall function
with roughness is used for all turbulence models.
V

Here, are the components of the unit normal vector to the wall, is the wall normal distance and
the von Karman constant. Its value and the one of
are the same as for the SKE model, cf. Table 2.
This model for the wall-reflection term is
known to be inadequate for impinging flows
[Hanjalic (1999), Murakami (1998)]. Therefore
simulations are performed with and without inclu. The grid convergence study of the
sion of
. One simulaLRR-IP model is done without
, denoted LRR-IP+, is perormed on
tion with
the fine grid for comparison.
Quadratic Pressure-Strain Model (SSG):
The model of Speziale, Sarkar and Gatski (SSG)
[Speziale et al. (1991)] uses an approximation for
the slow part of the pressure strain which is
quadratic in , and for the rapid part a quasi-linear
model. The entire pressure-strain follows from

'

"

^
N

_



(20)


`

'

L


Here,
is the wall shear stress,
the normal distance from the wall and
is
the friction velocity in the case of local equilibdepends on the non dimensional roughrium.
ness height
, where
is the dimensional roughness height.
is computed from the
following formulas which have been derived from
measurements in sand-grain roughened pipe flow
[Cebeci & Bradshaw (1977)].
\

"

J. Franke, W. Frank

is smaller in this case but still large enough to have a


small blockage ratio. In fact the distance to the inflow,
outflow and top boundary can be chosen smaller but
we want to have only a small influence of the boundary conditions on the solution.




'




&


5


2



-


9

5
0


:

4


&




(21)
For the constant
the default value
is
used in the present work. Finally, the standard treatment for , and with the wall function approach
is used [F LUENT (2003)]. Thus the boundary conditions at rough and smooth solid walls are described.
The other boundary conditions applied are described
in subsection 2.3.

&

&

250

250

130

>

90

2.2 Numerical approximations


F LUENT uses a finite volume method to solve the
transport equations described in the previous subsection. The fluxes through the surfaces of the control
volumes are numerically integrated with the midpoint
rule. The local values of flow variables at the surface
centres are approximated with a second order upwind
scheme. Gradients at the surface centres are approximated with central differencing. For the coupling of
pressure and velocity the S IMPLE algorithm is used.
The stopping criterion for the scaled
norm of
the residuals of all equations is
. In F LUENT scaling of the continuity residual is done with the largest
absolute value of the continuity residual in the first
five iterations. For the other equations the
norm
of the residual is divided by the sum of
over all
cells, where is the value of the solution variable
in cell and its coefficient in the algebraic equation
that results from discretizing the differential equation.
The stopping criterion is reached for all simulations
only on the coarsest of the three grids, except for the
SSG model which diverged. On all meshes iterative
convergence is therefore also checked with the aid of
pressure, velocity and turbulence quantity histories at
two monitoring positions within the junction.

60

H = 60

Figure 1 Geometry of the buildings and the street canyons.


All sizes are in
.
N

(a)

15H

10H

10H

10H

30H

(b)
15H

8.3H

9.2H

2.3 Computational domains, boundary conditions


and grids
In Figure 1 the geometry of the street interscetion is
shown. Two of the ring shaped buidlings partly have
a slanted roof while all other roofs are flat. The height
of the building with flat roof is
. The cross
section of the streets is quadratic with edge length .
This cluster of buildings is located in the box shaped
computational domain shown in Figure 2(a) for the
and in Figure 2(b) for the
approach flow. For
the
direction the buildings are rotated by that angle in clockwise direction around the z-axis. Thus
the minimum distance from the domain boundaries

9.3H

27.8H

Figure 2 Computational domain with distance of the built


area from the boundaries. (a) , (b)
approach flow.

Symmetry boundary conditions are used at the


top and at the sides of the domain. At the outflow
constant static pressure is prescribed. The walls of
the buidlings and the streets between the buidlings
are treated as smooth walls while the ground around

The Fourth European & African Conference on Wind Engineering, Paper #138

the intersection is treated as rough wall with a hy, indrodynamic roughness length of
ferred from the measurements of the approach flow.
The measured approach flow corresponds to a neutrally stratified boundary layer flow over a rough wall
with the cited roughness length and a friction veloc. In addition to these data the
ity
exponent in the power law approximation of the velocity profile is provided by the C EDVAL database as
. In Figure 3(a) the two resulting velocity profiles
are shown together with the measurements. While the
power law approximates the measured velocity profile
very well, the logarithmic distribution leads to much
too high velocities. Therefore the power law is chosen for the prescription of the velocity at the inflow
boundary.


kinetic energy is smaller than the measured values, cf.


Figure 3(b). In case of the Reynolds stress model the
assumption of isotropic turbulence is used to compute
the individual stress components from the constant
. While this contradicts the well known anisotropy
of the normal stresses and the relation
in boundary layer flows, the influence on the velocity distribution in the junction is small as is shown
in subsection 3.1. But for the normal stresses a better agreement with the measurements is obtained in
the junction when measured values are used at the inlet. Then also the turbulent kinetic energy matches the
from the inlet plane above
experiments at
. The influence of the higher at the inlet
on the results of the two-equation models is currently
investigated.
The given hydrodynamic roughness length has
to be converted into the roughness height to be used
with the wall function, Equation (21). By equating
(20) in the fully rough regime (
) with the logarithmic velocity distribution in terms of one finds
the relation
. Thus we use
on the ground around the built area, while treating all
other walls as smooth. The large roughness height
would require a cell height of
at the wall to place the first computational location
higher than the roughness height. This contradicts the
need for a good resolution of the boundary layer even
when the wall function approach is applied. Therefore we decided to use lower cell heights at the wall,
on the fine
the smallest being
grid. For the medium and the coarse grid this height
has always been doubled. We are currently investigating the effect of using
.

(a)

(b)

Power law

7
6

z/H

z/H

Exp.
SKE-F
RNG-F
RKE-F
SST-F
LRR-IP-F
SSG-F
Inlet

"

8
7

Log law

0 1 2 3 4 5 6 7

0.25 0.5 0.75


2

U [m/s]

k [m /s ]

Figure 3 Inflow boundary conditions for velocity (a) and


turbulent kinetic energy (b). The coloured lines are the
respective profiles on the fine grid at
away from
the inlet.


(a)

z/H

Also shown in Figure 3(a) are the computed velocity profiles after
from the inlet plane for
approach flow case which are the same as for
the
the
case. The profiles are nearly identical for all
turbulence models. Below
the computations have smaller velocities than the measurements
and the power law at the inflow. This momentum loss
is attributed to the usage of a cell height at the wall
which is too low for the applied value of the roughness height , see discussion below.
A similar behaviour can be observed for the turbulent kinetic energy , see Figure 3(b). There the computed also decreases considerably below
from the constant value prescribed at the inlet.
The inflow profile for (and for or ) is computed from
and
according to the proposal of
[Richards & Hoxey (1993)]. The resulting turbulent
"

0
-5

'

-4

-3

-2

-1

x/H

'

(b)

z/H

0
-5

-1

The grids are all blockstructured consisting only


of hexahedral cells. The fine (F) grid has 2 335 360
cells. The other two grids have been generated by
omitting every second node, leading to 291 920 cells

-2

-3

Figure 4 Resolution of the buildings height with (a) the


medium grid and (b) the fine grid.

-4

x/H

J. Franke, W. Frank

for the medium (M) and 36 490 cells for the coarse
(C) grid. While the cell volume change is kept below 2.5 for the fine grid, its maximum increases to
3.5 on the medium and 5.5 on the coarse grid. This is
one reason why no Richarson extrapolation is used to
examine the grid convergence. The resolution of the
buildings height with the medium and the fine grid is
shown on the previous page in Figure 4(a) and 4(b),
respectively.

flow directions. For the


approach flow only the
topology of the grid is changed, as can be seen in Figure 5(b). To better resolve the boundary layers on the
building walls the grid lines are perpendicular to the
walls up to a normal distance of .


3 RESULTS

(a) 6

The computational results are linearly interpolated on


the measurement locations which are located in six
planes within the crossing. Two planes have constant
, namely
and
, cf.
Figure 4. The other planes are for constant
or
, respectively. The primed coordinates (and vector and tensor components) are used for the
case
and are aligned with the streets. They are obtained
from the and coordinate by clockwise rotation of
around the -axis. The respective positions are
, see Figure 5(a).

y/H

-2

3.1 Grid dependance and hit rate of the solutions


-4
-6
-6

First a qualitative comparison of the results on the different grids is presented. As there are in total 1706
measurement locations for the
case and 1528 for
the
case the computed results are plotted against
the measured ones at all positions where data are
available. As 2D laser doppler anemometry has been
used in the experiments, only the velocity and the
stress
are available at all positions. In Figure 6
the corresponding results are shown for the RKE turbulence model. For the other models similar results
are obtained. The computed velocities are normalised
with
at
from the inlet, cf. Figure (3)(a). For the experimental velocities the corresponding value from the approach flow measurements
is used. It can be seen that the values on the medium
and fine grid differ only little, with a few exceptions.
For the turbulence quantities similar results are obtained.
As already stated in subsection 2.3, a grid convergence study by Richardson extrapolation is not used
as the expansion ratio differs considerably between
the grids and only linear interpolation to the measurement locations is used. Therefore the quantitative assessment of grid convergence is based on the hit rate
which is recommended by the [VDI (2003)] for code
validation. The hit rate is defined as


-4

-2

x/H

(b) 8

"

4
2

"

y/H

0
-2
-4
-6
-8
-8

-6

-4

-2

x/H
Figure 5 Resolution of the crossing with medium grid. (a)
, (b)
approach flow. Red colored is the surface mesh
on the rough ground around the crossing, black colored the
surface mesh on the smooth ground between the buildings
and blue colored the surface mesh on the smooth building
roofs.


&

With the medium grid 10 cells are used to resolve


and with the fine grid 20. The resolution as well as
the total number of cells is the same for both approach

&

)
'

(22)

where

'

is the total number of measurements and

The Fourth European & African Conference on Wind Engineering, Paper #138

where the values at the monitoring points have large


oscillations. This is especially the case for the
approach flow. The hit rate of the normal Reynolds
stress components displays a similar behaviour with
only small changes between the medium and fine grid.
The results on the fine grid can therefore be regarded
as grid independent.

(a) 1.5

RKE-C
RKE-M
RKE-F

U SIM / U SIM,REF

0.5

Table 3 Hit rate (in %) of velocity and normal Reynolds


stress components for the approach flow.
5

0
-0.5
-0.5

0.5

U EXP / U EXP,REF

1.5

(b) 1.5
RKE-C
RKE-M
RKE-F

USIM / U SIM,REF

turb. model
SKE-C
SKE-M
SKE-F
RNG-C
RNG-M
RNG-F
RKE-C
RKE-M
RKE-F
SST-C
SST-M
SST-F
SSG-C
SSG-M
SSG-F
LRR-IP-C
LRR-IP-M
LRR-IP-F
LRR-IP+F
LRR-IP+F, ES

0.5

29
-0.5
-0.5

0.5

1.5

50
58
59
43
54
51
51
59
61
43
51
48

58
58
51
62
68
60
61

38
56
57
36
53
53
39
56
55
36
54
45

59
59
38
60
63
57
60

14
14
15
16
16
18
14
16
15
14
15
18

16
16
19
17
18
21
21

72
77
78
37
55
60
63
68
64
29
36
42

70
71
51
72
73
72
73

62
69
69
30
40
40
53
52
51
24
30
38

57
58
41
60
64
83
81

42
34
39
59
65
60
42
46
49
54
56
51

72
74
65
59
60
64
75

UEXP / U EXP,REF
Table 4 Hit rate (in %) of velocity and normal Reynolds
stress components for the
approach flow.

Figure 6 Grid convergence of normalised and


velocity for the RKE model. (a) , (b)
approach flow.

for
otherwise

&

turb. model
SKE-C
SKE-M
SKE-F
RNG-C
RNG-M
RNG-F
RKE-C
RKE-M
RKE-F
SST-C
SST-M
SST-F
SSG-C
SSG-M
SSG-F
LRR-IP-C
LRR-IP-M
LRR-IP-F

"

(23)

Here,
denotes the normalised results of the
computation and the normalised experimental results.
is the absolute value of the relative error and
the allowed deviation which has different values
for the examined quantities. For the velocity components we use
and for the Reynolds stresses
. The resulting hit rates for the velocity and
normal Reynolds stress components are listed in Table 3 and Table 4 for the
and
approach flow,
respectively.
For both cases the hit rate of the velocity components converges in general with grid refinement.
Some models display oscillating behaviour which can
be attributed to the bad convergence on the fine grid
+

47
57
59
50
51
49
48
57
59
43
49
50
44
52
54
45
51
50

29
53
54
43
45
47
27
55
56
30
46
39
32
46
50
33
51
41

21
23
24
16
20
17
23
21
20
20
19
16
19
14
15
18
16
16

55
49
56
66
66
64
59
63
65
52
54
53
68
65
62
67
67
63

7
3
3
25
26
20
17
8
10
52
44
37
38
14
13
39
15
14

25
20
25
61
61
52
34
37
39
52
42
34
69
65
64
66
60
64

J. Franke, W. Frank

The quantitative analysis of the hit rates shows


case the best results are obtained
that for the
with the differential Reynolds stress models. Here,
LRR-IP+F, ES refers to the LRR-IP model with wallreflection term, Equation (17), and the Experimental Stresses as inflow condition. Comparing the results of this model with the ones of the LRR-IP+ and
the LRR-IP model, the best results for the velocities
are obtained with the LRR-IP model while the normal stresses are best predicted with the LRR-IP+, ES
model. The inclusion of the wall-reflection therefore
leads to a better prediction of the turbulence at the
cost of the velocity prediction. The best results for
the normal stresses are obtained when using the measured stresses as inflow condition.
The results with the SSG model are similar to the
where
ones with the LRR-IP+ model except for
worse agreement is obtained and for
which is
close to the one of LRR-IP+, ES. Concerning the results for
all models have bad results with a maximum of 21 % of the computed values agreeing with
the measurements within 25 %. The reason for the
bad agreement is visualised in subsection 3.2.
Comparing the two-equation models, the SKE and
the RKE model perform equally well with results for
the velocities only slightly worse than with the differential Reynolds stress models. The SKE model also
has remarkably good results for the normal stresses,
which is best predicted with the RNG
except for
model. But this model performs worse for the velocities. The least accurate results are obtained with
the SST model. This is also true for the velocities
in the
case but not for the normal stresses where
the SKE model performs worst, see Table 4. But for
the velocities the best results are obtained with the
SKE model, followed by the RKE and then the SSG
model. The LRR-IP model is even outperformed by
the RNG model. For the normal stresses the situation
is inversed as the differential Reynolds stress models have the best results. All models predict
very
badly, as well as
again. The reason for the latter
is displayed in subsection 3.3.
Resuming the results we can state that all models
perform reasonably well in predicting the
and
velocities, especially in the case, but fail for
the
velocity. When averaging the results for all
velocities for the and
case we get the following
ranking of the models: the SKE model performs best,
slightly better than the RKE model. Then the LRRIP, SSG, RNG and SST model follow. Performing the
same averaging for the normal stresses we have the
following order: SSG, LRR-IP, RNG, RKE, SKE and
SST. Thus the best results for the velocities are ob-

tained with linear two-equation models, namely the


SKE and the RKE model.

3.2 Mean velocities and turbulence for


flow

After presenting the main quantitative results of the


simulations in the previous subsection we here want
to present the flow field in the junction at two of the
six measurement planes. These measurement planes
are chosen as the worst results based on the previously introduced hit rate are obtained there. For the
two measurement planes parallel to the ground this is
. In Figure 7 the experimental
the one at
velocity vectors from and are dispplayed. The
flow enters the intersection through the street facing
the approach flow and the crossroad from the positive
direction. At the entrance into the crossroad a
recirculation region is formed. The fluid then follows
the crossroad until it is blocked by the fluid coming
from the street entering the junction. Therefore the
flow from the crossroad changes direction and leaves
the intersection by the street parallel to the
-axis
which faces the outlet plane of the computational domain. In this street a weak recirculation region can be
observed behind the corner of the upper right building. Though this is merely a 2D picture of the 3D
flow within the intersection, the flow from the street
entering the intersection has high enough momentum
to deflect the flow from the crossroad. This high modirecmentum flow is itself deflected in negative
tion in the crossroad, with a small recirculation zone
at the corner of the lower left building. Another recirwhere
culation region can be observed at
the flow leaves the crossroad and is deflected by the
external flow around the buildings.

approach


6
5

EXP.
U EXP,REF

Y/H

1
0

-1

-2

-3

-4
-5
-3 -2 -1 0

X/H
Figure 7 Velocity vectors from
Experimental results.

and


at

"

The Fourth European & African Conference on Wind Engineering, Paper #138

-1

-2

-2

-3

-3

-4

-4
1

-1

-5
-3 -2 -1 0

U SIM,REF

Y/H

6
5

-5
-3 -2 -1 0

X/H
6
5

RKE

SST
U SIM,REF

U SIM,REF

-1

-2

-2

-3

-3

-4

-4

-5
-3 -2 -1 0

-5
-3 -2 -1 0

Y/H

Y/H

U SIM,REF

SSG

U SIM,REF

X/H

LRR-IP

-1

-1

-2

-2

-3

-3

-4

-4

-5
-3 -2 -1 0

-5
-3 -2 -1 0

6
5
1

X/H
at


Y/H

and

X/H
6

LRR-IP+, ES

U SIM,REF

U SIM,REF

-1

-1

-2

-2

-3

-3

-4

-4

-5
-3 -2 -1 0

-5
-3 -2 -1 0

LRR-IP+

The corresponding numerical results with the twoequation models are displayed in Figure 8. None of
the models predicts the recirculation zone when the
flow enters the crossroad from the positive
direction. The deflection of this flow towards the street
in positive
direction is only captured by the SKE
and RKE model, where the last one is the only model
that shows the experimentally observed flow reversal
above
. The RNG and the SST model predict that the flow goes through the entire crossroad
and does not enter the street in positive
direction.
This is due to the fact that the flow into the junction
from the street facing the inlet plane has lower momentum than in the experiment. This is already obvious from the first velocity vectors at the entrance into
this street. While in the experiment these have nearly
equal length, representative of a block profile, all simulations have smaller velocities close to the walls, the
smallest for the SST model. Thus the prediction of
the flow near the walls seems to be responsible for the
low momentum that then leads to a flow field which

X/H

X/H
Figure 8 Velocity vectors from
Two-equation model results.

-1

Y/H

Y/H

U SIM,REF

is qualitatively different from the one observed in the


experiments.
Due to the bad prediction of the flow within the
crossroad all models predict too large a recirculation
region behind the corner of the upper right building.
Another common feature of the simulations is that the
is undervelocity of the external flow at
predicted. This is also the case for the simulations
with the differential Reynolds stress models which are
displayed in Figure 9.

RNG

Y/H

SKE

Y/H

X/H

Figure 9 Velocity vectors from and


at
Differential Reynolds stress model results.


X/H


An excellent prediction of the flow is obtained


with the LRR-IP and the SSG model which both show
the flow reversal above
. But the LRRIP model better predicts several other features of the
flow than the SSG model. While both models do not
compute the recirculation region at the entrance of the
flow into the crossroad from the positive
direction, the LRR-IP model better predicts the strength
and direction of the flow. The same is true for the flow
in the crossroad below
. When using the
wall-reflection term, Equation (17), with the LRR-IP

10

J. Franke, W. Frank

model worse results are obtained that compare well


with the ones of the SKE model, cf. Figure 8. The
use of the experimental normal stresses as boundary
condition has obviously no influence on the velocities
in this plane.

U / U REF

SKE
RNG
RKE
SST
LRR-IP
SSG
EXP

0.5

Reynolds stress models. All models predict a larger


into the crossroad and a
flow from above
recirculation region, roughly in the lower right part of
the crossroads cross section. This recirculation zone
is largest and strongest for the RKE, SSG and LRRIP model which showed the best agreement with the
, see Figure
measurements in the plane at
7, 8 and 9. The inclusion of the wall-reflection term
leads to a smaller recirculation zone with again no visible influence of the boundary condition of the normal
Reynolds stresses. The smallest recirculation zone is
obtained with the SST model. These results indicate
a correlation between the prediction of the flow in the
and
. The better the
planes
results are for the
plane the worse they are
for the
plane.


z/H=0.5
x/H=0.17

-0.5
-5

-4

-3

-2

-1

Y/H
0.1

z/H=0.5
x/H=0.17

-4

-3

-2

-1

2
1.5
1
0
-3

velocity and
.


normal

0
-3

SKE
y/H=-1

U SIM,REF
-2

-1

2.5
2
1.5
1
0.5

0
-3

RNG
y/H=-1

U SIM,REF
-2

-1

X/H

2.5

Z/H

X/H

Z/H

-1

1.5
0.5

-2

Next, a comparison between the computed and


measured
velocity and the
normal Reynolds
stress at the position
within the measurement plane at
is shown in Figure 10.
For the SKE, RKE, SSG and LRR-IP model we obtain an excellent agreement with the measurements
which deteriorates towards the ends of the crossroad.
The RNG and SST model predict too low velocities
around
which is a result of the low momentum of the flow entering the junction from the negative
direction. These models also significantly
underpredict . For the other models the results are
better, largely capturing the qualitative distribution of
. The magnitude of the maximum value is best predicted with the SKE model.
The velocity vectors from and
in the measurement plane at
are shown in Figure
11 for the experiment and the two-equation models.
This is the measurement plane where with all models
the worst results for the
velocity have been obtained. It cuts the recirculation region in the crossroad at the corner of the lower left building. From
the experiments one sees that the entire flow is directed away from the bottom of the crossroad. There
is also hardly any flow from above
into the
crossroad. These features are not captured by any of
the models used, cf. Figure 12 for the differential

y/H=-1

U EXP,REF

2.5

Z/H

EXP

X/H

Y/H
Figure 10 Comparison of
Reynolds stress at

2.5

Z/H

uu / U REF

SKE
RNG
RKE
SST
LRR-IP
SSG
EXP

0
-5

1.5
1
0.5

0
-3

RKE
y/H=-1

U SIM,REF
-2

-1

X/H
2.5
2

Z/H

0.5

0.05

1.5
1
0.5
0
-3

SST
y/H=-1

U SIM,REF
-2

-1

X/H

Figure 11 Velocity vectors from and


at
Experimental and two-equation model results.

11

The Fourth European & African Conference on Wind Engineering, Paper #138

model, good agreement for the SKE, RKE and LRRIP+ model and no agreement for the RNG and SST
which is not
model. For the plane
shown here, all models yield good agreement. On the
other hand all models performance is weak for the
const. planes.

2.5

Z/H

2
1.5

SSG

1
0.5

y/H=-1

U SIM,REF

0
-3

-2

-1

X/H

2.5

2.5

2.5

y/H=-1
x/H=0

y/H=-1
x/H=0

1.5

1.5

1.5
0.5

y/H=-1

U SIM,REF

0
-3

-2

-1

Z/H

LRR-IP

Z/H

Z/H

0.5

0.5

X/H
2.5

Z/H

2
1.5

LRR-IP+

1
0.5
0
-3

y/H=-1

U SIM,REF
-2

-1

-0.2

0.2

0.02

0.04

0.06

ww / U REF2

Figure 13 Comparison of
Reynolds stress at
legend see Figure 10.

2.5

Z/H

W / U REF

X/H

velocity and


normal
. For colour


1.5

LRR-IP+, ES
U SIM,REF

1
0.5

0
-3

-2

3.3 Mean velocities and turbulence for


approach
flow
In the
case the same measurement planes are
used but now in the primed coordinates. For the two
planes
const. which are not shown here, the
qualitative and quantitative results of all models are
case. The lower hit rate prebetter than in the
sented in Table 4 is therefore due to the results in the
const. planes. There again
has
the worst results in terms of the hit rate. The experimental and numerical velocity vectors in this plane
are shown in Figure 14. In the experiments the flow
and into
is directed towards the roofs for
the crossroad for
up to
. Another feature of the flow inside the crossroad is that
the
component is nearly always negative. This is
not captured by any of the models which all predict
positive
velocities for
. Also all models except the SKE and RKE model display flow in
negative
direction for
in contrast to
the experiments. These characteristics of the simulations explain the bad agreement with the experimentally observed components.
component the results are better. EspeFor the
cially the RKE and SSG model predict upward flow
for
like in the experiment. Also the flow
direction above
is fairly good predicted.
This can be also seen in Figure 15 where the
velocity and
normal Reynolds stress are displayed


y/H=-1

-1

X/H

Figure 12 Velocity vectors from and


at
Differential Reynolds stress model results.


Finally, a comparison between the computed and


measured
velocity and the
normal Reynolds
stress at the position
within the measurement plane at
is shown in Figure 13.
Above
all models have negative
velocities due to the aforementioned flow into the crossroad whereas the experiments show
. Inside
the crossroad the simulation results show the recirculation region while the experiment has only positive
we obtain the best agreement with
velocities. For
the SKE model. While the RKE and even the RNG
model also produce reasonable results, the differential
Reynolds stress models and the SST model compute
without even reproducing the distinct
too low a
peak at
. So while in general the differential
Reynolds stress models show a better agreement for
with the experimental results, as has been shown
with the hit rate in Table 3, they perform less good
than most of the other models at this position.
For the approach flow in general we obtain an
excellent qualitative agreement with the experiments
in the plane
for the SSG and LRR-IP


"

&

"

0.4

12

J. Franke, W. Frank

at


within the plane


2.5

2.5

y/H=-1
x/H=0

2.5
2

1.5

1.5

0.5
0
-3

EXP
y/H=-1

U EXP,REF
-2

-1

z/H

z/H

z/H

2
1.5

y/H=-1
x/H=0

0.5

0.5

x/H
2.5

z/H

2
1.5
1
0.5
0
-3

SKE

0
-0.2

y/H=-1

U SIM,REF
-2

-1

z/H

0.5
0
-3

y/H=-1

U SIM,REF
-1

x/H

z/H

2
1
0.5
0
-3

RKE
y/H=-1

U SIM,REF
-2

-1

x/H

z/H

2
1
0.5
0
-3

velocity and


"

&

'

y/H=-1

U SIM,REF

&

x/H

0.1

normal
. For colour


'

-1

SST

-2

2.5
1.5

0.05

ww / U REF

2.5
1.5

Up to
all models predict too low
velocities with the SST and RNG model even having negative velocities close to the ground. The best
agreement is obtained with the SSG, LRR-IP and
the agreement with
RKE model. Above
the experiments is good although a stronger downward movement is observed experimentally.
For the normal stress all models display too low
values within the crossroad. There the SKE model is
closest to the experiments but predicts too high a peak
value. This peak is reproduced by all models except
than in the
for the SSG model albeit at a larger
experiments.
So for the
approach flow case we find in gencase. While the
eral the same results as in the
qualitative and quantitative prediction of the flow in
const. planes is very good for all modthe
els, even better than in the case, the results for the
const. planes are worse.
%

RNG

-2

0.4

2.5

0.2

Figure 15 Comparison of
Reynolds stress at
legend see Figure 10.

x/H

1.5

W / U REF

2.5

z/H

2
1.5
1
0.5
0
-3

4 CONCLUSIONS
LRR-IP
y/H=-1

U SIM,REF
-2

Six turbulence models comprising four linear twoequation models and two differential Reynolds stress
models were used to simulate the flow from two approach flow directions across an asymmetric street intersection formed by four rings of buildings which
partly have a slanted roof. Three systematically refined grids were used in each case to assess the grid
dependance of the results and the simulations on the
finest grid were found to be nearly grid independent.
The predictions were compared with available wind
tunnel data from the C EDVAL database. The agreement between the predicted and measured velocity
components in the measurement planes parallel to the
ground is in general very good. There the standard

-1

x/H
2.5

z/H

2
1.5
1
0.5
0
-3

SSG
y/H=-1

U SIM,REF
-2

-1

x/H

Figure 14 Velocity vectors from

and

at

13

The Fourth European & African Conference on Wind Engineering, Paper #138
lence models for engineering applications. AIAA Journal, 32, 8, pp. 1598-1605.
Launder, B. E. & Spalding, D. B. (1972) Lectures in Mathematical Models of Turbulence. Academic Press, London.
Yakhot, V. & Orszag, S. A. (1986) Renormalization group
analysis of turbulence: I. Basic theory. Journal of Scientific Computing, 1, 1, pp. 1-51.
eddy-viscosity model
Shih, T.-H. et al. (1995) A new
for high Reynolds number turbulent flows Model development and validation. Computers & Fluids, 24, 3,
pp. 227-238.
Schumann, U. (1977) Realizability of Reynolds-stress turbulence models. Phys. Fluids, 20, 5, pp. 721-725.
Wilcox, D. C. (1988) Reassessment of the scaledetermining equation for advanced turbulence models.
AIAA Journal, 26, 11, pp. 1299-1310.
F LUENT (2003) FLUENT V6.1 Users guide, Fluent Inc.,
Lebanon, New Hampshire, USA.
Hanjalic, K. (1999) Second-moment turbulence closures
for CFD: Needs and prospects. International Journal of
Computational Fluid Dynamics, 12, 11, pp. 67-97.
Lien, F. S. & Leschziner, M. A. (1994) Assessment of
turbulent transport models including non-linear RNG
eddy-viscosity formulation and second-moment closure.
Computers & Fluids, 23, 8, pp. 983-1004.
Launder, B. E. et al. (1975) Progress in the development
of Reynolds-stress turbulence closure. Journal of Fluid
Mechanics, 68, pp. 537-566.
Rotta, J. (1951) Statistische Theorie nichthomogener Turbulenz. 1. Mitteilung. Zeitschrift fur Physik, 129, pp.
547-572.
Fu, S. et al. (1987) Accomodating the effects of high strain
rates in modelling the pressure-strain correlation. Thermofluids report TFD/87/5, UMIST, Manchester.
Gibson, M. M. & Launder, B. E. (1978) Ground effects on
pressure fluctuations in the atmospheric boundary layer.
Journal of Fluid Mechanics, 86, pp. 491-511.
Murakami, S. (1998) Overview of turbulence models applied in CWE-1997. Journal of Wind Engineering and
Industrial Applications, 74-76, pp. 1-24.
Speziale, C. G. et al. (1991) Modelling the pressure-strain
correlation of turbulence: an invariant dynamical systems approach. Journal of Fluid Mechanics, 227, pp.
245-272.
Cebeci, T. & Bradshaw, P. (1977) Momentum Transfer in
Boundary Layers, Hemisphere Publishing Corporation,
New York.
Richards, P. J. & Hoxey, R. P. (1993) Appropriate boundary
conditions for computational wind engineering models
model. Journal of Wind Engineering and
using the
Industrial Aerodynamics, 46 & 47, pp. 145-153.
VDI (2003) Environmental meteorology. Prognostic microscale windfield models Evaluation for flow around
buildings and obstacles. VDI-Richtlinie 3783, Blatt 9
(Entwurf), in German.

and the realizable


model perform slightly better than the differential Reynolds stress models when
both approach flow directions are taken into account.
The worst results are obtained with the renormalization group
and the shear stress transport
model. Greater discrepancy is obtained for the velocity components in planes perpendicular to the crossroad with all models.
The normal turbulent stresses are best predicted
with the two differential Reynolds stress models. But
the standard and the realizable
model yield again
reasonable agreement with the measurements. The
results therefore indicate that these two linear twoequation models are accurate enough for the prediction of at least the mean velocities in a cluster of buildings with moderate complexity.


5 REFERENCES
Lien, F.-S. & Yee, E. (2004) Numerical modelling of the
turbulent flow developing within and over a 3-D building array, Part I: A high-resolution Reynolds-Averaged
Navier-Stokes approach. Boundary-Layer Meteorology,
112, pp. 427-466.
Kato, M. & Launder, B. E. (1993) The modeling of turbulent flow around stationary and vibrating square cylinders. Proceedings of the 9th Symposium on Turbulent
Shear Flows, Kyoto, Japan.
Cheng, Y. et al. (2003) A comparison of large eddy simuReynolds-averaged Navierlations with a standard
Stokes model for the prediction of a fully developed turbulent flow over a matrix of cubes. Journal of Wind Engineering and Industrial Aerodynamics, 91, pp. 13011328.
Hanna, S. R. et al. (2002) Comparisons of model simulations with observations of mean flow and turbulence
within simple obstacle arrays. Atmospheric Environment, 36, pp. 5067-5079.
Ferreira, A. D. et al. (2002) Prediction of building interference effects on pedestrian level comfort. Journal of
Wind Engineering and Industrial Aerodynamics, 90, pp.
305-319.
Richards, P. J. et al. (2002) Pedestrian level wind speeds
in downtown Auckland. Wind & Structures, 5, 2-4, pp.
151-164.
Ketzel, M. et al. (2002) Intercomparison of numerical urban dispersion models - Part II: Street canyon in Hannover, Germany. Water, Air, and Soil Pollution: Focus,
2, 5-6, pp. 603-613.
Westbury, P. et al. (2002) CFD application on the evaluation of pedestrian-level winds. Proceedings of the Workshop Impact of Wind and Storm on City life and Built
Environment, Nantes, France, pp. 172-181.
Leitl, B. (2000) Validation data for microscale dispersion
modeling. EUROTRAC Newsletter, 22, pp. 28-32.
Menter, F. R. (1994) Two-equation eddy-viscosity turbu

14

Вам также может понравиться