You are on page 1of 8

Conference on Modelling Fluid Flow (CMFF’03)

th
The 12 International Conference on Fluid Flow Technologies
Budapest, Hungary, September 3 - 6, 2003

VERY LARGE EDDY SIMULATION FOR THE PREDICTION OF UNSTEADY


VORTEX MOTION

Albert RUPRECHT Thomas HELMRICH Ivana BUNTIC


Head of Fluid Mechanics Group* Research assistant PhD Student

Institute of Fluid Mechanics and Hydraulic Machinery


University of Stuttgart

*Corresponding author: Pfaffenwaldring 10, D-70550 Stuttgart, Germany


Tel.: (+49) 711 685 3256, Fax: (+49) 711 685 3255, Email: ruprecht@ihs.uni-stuttgart.de

ABSTRACT
An new turbulence model for Very Large Eddy 1. INTRODUCTION
Simulation, based on the extended k-ε model of Industrial calculations are usually based on the
Kim and Chen is developed. Introducing an Reynolds-averaged Navier-Stokes (RANS)
adaptive filtering technique the model can equations. This means that the complete turbulence
distinguish between numerically resolved and behavior is expressed by means of an appropriate
unresolved parts of flow. The model is applied to turbulence model. The necessary turbulence model
the unstable vortex motion in a pipe trifurcation. has to take into account all the turbulent scales
This flow phenomenon could not be predicted with ranging from the largest turbulent eddies down to
classical RANS methods and usually used the Kolmogorov scale. Consequently the model has
turbulence models. By applying the VLES method to be very sophisticated and it is impossible to
with the new turbulence model the phenomenon define a model suitable for all flow phenomena,
was well predicted and the results agree quite especially for unsteady vortex motions.
reasonable with measurement data. A direct numerical simulation (DNS), where all
Key Words: Adaptive turbulence model, Pipe scales down to the Kolmogorov scale are resolved
trifurcation, Very Large Eddy Simulation. by the computation is impossible. As the smallest
scales strongly decrease with increasing Reynolds
NOMENCLATURE number, extremely high computational effort (i.e.
f filter function large grid size) is required for high Reynolds
hmax local grid size number flows. Therefore DNS cannot be applied for
k turbulent kinetic energy flow of practical relevance in the foreseeable future.
L Kolmogorov length scale A promising compromise is the Large Eddy
Pk production term Simulation (LES). In a “real” LES (from the
U local velocity turbulence research point of view) all anisotropic
α model constant
turbulence structures are resolved in the
∆ resolved length scale
computation and only the smallest isotropic scales
∆t time step
are modeled. Therefore the used turbulence models
ε dissipation rate
can be simpler compared to RANS models, since
ν kinematic viscosity
they only have to describe the influence of the
νt turbulent viscosity
isotropic scales on the resolved anisotropic flow
field. Unfortunately with increasing Reynolds
numbers the anisotropic scales decreases and cannot the turbulent spectrum. It has an adaptive
be resolved in the computation of flow of practical characteristic, in such a way that it can be applied
relevance, although there are many “LES” for the whole range of approaches from RANS up
applications in the literature. However from the to DNS.
turbulence research point of view these simulations The presented application of the new adaptive
are mostly unsteady RANS (URANS), since they turbulence model is the flow in a pipe trifurcation of
only resolve the unsteady mean flow but do not a water power plant. In the spherical geometry an
resolve any turbulence structures. unsteady not periodic vortex motion has been
In order to apply a “classical” URANS there observed. This phenomenon is predicted by the
must be a gap in the turbulence spectrum between adaptive model.
the unsteady mean flow and the turbulent flow.
Only then the classical turbulence models (e.g. k-ε) 2. SIMULATION METHOD
can be applied, as they are developed for modeling
the whole range of turbulent scales. If it is not 2.1 Very large Eddy Simulation (VLES)
possible to distinguish between mean flow and “Real” Large Eddy Simulation (LES) from the
turbulence these models cannot be applied properly. turbulence research point of view require an
In this case a Very Large Eddy Simulation enormous computational effort since all anisotropic
(VLES) can be applied. Contrary to URANS there turbulence structures have to be resolved in the
is a requirement for the turbulence model, that it can computation and only the smallest isotropic scales
distinguish between numerically resolved unsteady are modeled. Consequently this method also can not
movement and not resolved turbulent fluctuations be applied for industrial problems today.
which have to be modeled. VLES is similar to LES, Today’s calculations of flows of practical
only that a smaller part of the turbulence spectrum relevance (characterized by complex geometry and
is resolved in the unsteady simulation. The model high Reynolds number) are usually based on the
must express the influence of a greater part of the Reynolds-averaged Navier-Stokes (RANS)
spectrum and therefore it has to be more equations. This means that the influence of the
sophisticated. complete turbulence behavior is expressed by
As mentioned above DNS and LES requires a means of an appropriate turbulence model. To find a
much too high computational effort for complex turbulence model, which is able to capture a wide
industrial problems. RANS is not suitable for range of complex flow effects quite accurate is
unsteady vortex phenomena. For this type of impossible. Especially for unsteady flow behavior
problem today VLES seems to be a promising way. this method often leads to rather poor results. The
In Table 1 the availability of the different RANS and LES approach can schematically be seen
approaches for the simulation of the flow around an in Figure 1, where a typical turbulent spectrum and
aircraft is summarized. This data apply similarly for its division in resolved and modeled parts is shown.
other complex flow problems.

Table 1: Application of the different approaches to


the flow around an aircraft [1].

In this paper the development of a VLES


turbulence model is presented. The model is based Figure 1: Modelling approach for RANS and LES.
on the extended k-ε model of Chen and Kim [2]. By
adding an appropriate filtering techniques - which The recently new established approach of Very
depends on the local grid spacing and the Large Eddy Simulation can lead to quite promising
computational time step - the new turbulence model results, especially for unsteady vortex motion.
distinguishes between resolved and modeled part of Contrary to URANS there is a requirement to the
applied turbulence model, that it can distinguish After the solution of the momentum and
between resolved unsteady motion and not resolved continuity equations the turbulence quantities are
turbulent motion which must be included in the calculated and a new turbulence viscosity is
model. It is similar to LES, only that a minor part of obtained. The k- and ε-equations are also linearized
the turbulence spectrum is resolved (schematically by successive substitution and the linear systems are
shown in Figure 2). VLES is also found in the solved by the BICGSTAB2 algorithm with ILU
literature under different other names: preconditioning. The whole procedure is carried out
- Semi-Deterministic Simulation (SDS), in a global iteration until convergence is obtained.
- Coherent Structure Capturing (CSC), For unsteady simulations the global iteration has to
- Detached Eddy Simulation (DES), be carried out in each time step.
- Hybrid RANS/LES, The parallelization of the code is introduced by
- Limited Numerical Scales (LNS). domain decomposition using double overlapping
grids. The linear equation solver BICGSTAB2 is
carried out in parallel and the data exchange
between the domains is organized on the level of
the matrix-vector multiplication in the BICGSTAB2
solver. The preconditioning is carried out locally on
each domain. The data exchange is carried out using
MPI (Message Passing Interface) on machines with
Figure 2: Turbulence treatment in VLES. distributed memory. On shared-memory-computers
the code runs also parallel by applying OpenMP.
For details on the numerical procedures and
2.2 Numerical Method parallelization the reader is referred to [7,8].
The calculations are carried out using the
program FENFLOSS which has been developed at 3. ADAPTIVE TURBULENCE MODEL
the institute for more than a decade [3,4]. Classical turbulence models, which are usually
The partial differential equations are solved by applied in engineering flow predictions, contain the
a Galerkin Finite Element Method. The spatial whole turbulent spectrum. They usually show a too
discretization of the domain is performed by 8-node viscous behavior and very often damp out unsteady
hexahedral elements. For the velocity components motion to early. As discussed above the turbulence
and the turbulence quantities a tri-linear model for the VLES have to distinguish between the
approximation is applied. The pressure is assumed resolved and unresolved part of the turbulent
to be constant within each element. For advection spectrum (Figure 2). Therefore an adaptive model
dominated flow a Petrov-Galerkin formulation of is developed, which adjusts its behavior according
2nd order with skewed upwind orientated weighting to the approach (schematically shown in Figure 3).
functions is applied. The time discretization is done This means that this model can be applied for all
by a three-level fully implicit finite difference approaches.
approximation of 2nd order.
For the solution of the momentum and
continuity equation a segregated solution algorithm
is used. Each momentum equation is handled
independently. The momentum equations are
linearized by successive substitution. The linear
systems are solved by the BICGSTAB2 algorithm
of van der Vorst [5] with an incomplete LU Figure 3: Adjustment for adaptive model.
decomposition (ILU) for preconditioning. The
The advantage of the adaptive model is, that
pressure is treated by a modified Uzawa type
with increasing computer power the resolved part of
pressure correction scheme [6]. The pressure
the turbulent spectrum increases and the modeled
correction is carried out in a local iteration loop
part decreases and consequently the accuracy of the
without reassembling the system matrices until the
calculations improves.
continuity error is reduced by a given order (usually
There are several filtering techniques in the
6-10 iterations needed).
literature [e. g. 9-11]. Here a filtering techniques
similar to Willems [12] is applied. In the following
the new adaptive turbulence is presented. The  u ⋅ ∆t  ∆V for 2D
nomenclature of resolved and modeled parts can be ∆ = α ⋅ max  with h max =  3 (6)
 h max  ∆V for 3D
seen in Figure 4.
with a model constant (α= 2 to 5). ∆V is the volume
of the local element, u is the local velocity and ∆t is
the time step. The Kolmogorov scale L is given as
k3/2
L= . (7)
ε
The filtering procedure leads to the final
equations
 )
∂k ∂k ∂ ν t  ∂k  )
+ Uj =  ν +   + Pk − ε (8)
∂t ∂x j ∂x j  σ k  ∂x j 
and
 )
Figure 4: Distinguishing of turbulence spectrum for ∂ε ∂ε ∂ ν t  ∂ε  ε)
VLES. + Uj =  ν +   + c 1ε Pk
∂t ∂x j ∂x j  σ k  ∂x j  k
) (9)
The basis of this adaptive model is the k-ε ε2 P  )
− c 2ε + c 3 ε  k  ⋅ Pk
model of Chen and Kim [2]. This model has been k  k 
chosen, because it is quite simple and its results are with the production term
much better – especially for unsteady flows –
) )  ∂U ∂U j  ∂U i
compared to the standard k-ε model. The transport Pk = ν t  i +  . (10)
equations for k and ε are given as  ∂x ∂xi  ∂x
 j  j

∂k ∂k ∂  ν t  ∂k  For details on the model and its development


+ Uj =  ν +   + Pk − ε (1)
∂t ∂x j ∂x j  σ k  ∂x j  the reader is referred to Ruprecht [4].
The simulation of vortex shedding, which can
∂ε ∂ε ∂  ν t  ∂ε  ε
+ Uj =  ν +   + c 1ε Pk be considered as a very convenient test case for
∂t ∂x j ∂x j  σ k  ∂x j  k CFD computations, shows very often great
ε2 P  difficulties when applying URANS. For example
− c 2ε + c 3 ε  k  ⋅ Pk (2) applying URANS with the standard k-ε model to
k
14 42k 
44 3 the flow behind a bluff trailing edge (Figure 5)
additional term
leads to a steady state solution. The vortex shedding
is suppressed by the too diffusive turbulence model.
These equations prescribe the whole turbulence Applying a more sophisticated turbulence model, e.
spectrum and therefore a filtering procedure has to g. the extended k-ε model of Kim and Chen [2], the
be implemented. vortex shedding is obtained. The comparison is
According to Kolmogorov theory it can be shown in Figure 6, where the streamlines at a
assumed that the dissipation rate is equal for all certain time step are plotted.
scales. This leads to
)
ε = ε. (3)
However the turbulent kinetic energy needs a
filtering
)   ∆ 
k = k ⋅ 1 − f   . (4)
  L 
A suitable filter can be obtained to
0 for ∆ ≥L
 2/3
f = ∆ (5)
1− L>∆
  L 
for Figure 5: Flow behind bluff trailing edge.

where
The trifurcation distributes the water from the
penstock into the three branches to the turbines. It
has a spherical shape because of structural reasons.
The water passage is partly shown in Figure 8.
In the power plant severe power oscillations
were encountered at the outer turbines 1 and 3, in
the range of +/-10% of nominal power. A vortex
instability could be discovered as a reason for that.
A vortex is forming in the sphere, starting at the top
and extending into a side branch. After a certain
Figure 6: Streamlines of the vortex shedding period it changes its behavior and extends to the
behind bluff trailing edge. opposite side branch, again for a certain time
period. Then it jumps back again. The two vortex
positions are schematically shown in Figure 9. This
unstable vortex motion, which is not periodic,
causes the power fluctuations, since the vortex
produces very high losses in the branch, in which it
is actually located, because of strong swirling inlet
flow to the branch. These losses reduce the head of
the turbine and consequently the power output. In
the following the simulation of this vortex
movement is presented and the results are compared
with measurement in a model test.

Figure 7: Vortex shedding behind a bluff trailing


edge, pressure distribution, comparison of Kim &
Chen model and adaptive VLES model.

Furthermore by more detailed observation of


the obtained results, a relatively strong damping of
vortex shedding in the downstream flow is noticed Figure 8: Water passage with trifurcation.
applying the Kim and Chen model.
Using the VLES method the results severely
improve. In Figure 7 the pressure distribution
behind the bluff trailing edge clearly shows a
reduced damping of the vortex motion, which
coincided better with experimental observations.

4. APPLICATION

4.1 Problem description


The application presented is the flow in a pipe
trifurcation of a water power plant. The water Figure 9: Unsteady vortex structures.
passage consists of the upper reservoir, channel,
surge tank, penstock, trifurcation and three turbines.
4.2 Unsteady vortex simulation
For the simulation the geometry is discretized
by approximately 500.000 elements. The
computational grid, shown in Figure 10, is
decomposed into 32 partitions and the computations
are run on a CRAY T3E.
Applying URANS with the standard k-ε model
leads to a steady state solution. The obtained vortex
structure is shown in Figure 11. The vortex ranges
from one side branch to the other. It is completely
stable and the swirl component is severely
Figure 12: Flow in the trifurcation, vortex position
underpredicted. This also leads to an
in branch 1.
underprediction of the losses in the side branches.
The URANS is not able to predict the unsteady flow
phenomenon.

Figure 13: Flow in the trifurcation, vortex position


in branch 2.

Figure 10: Computational grid. Due to the strong swirl at the inlet to the side
branch, in which the vortex is located, the inlet
Applying VLES with the adaptive turbulence losses into this branch are much higher compared to
model the unstable vortex movement can be the others. Consequently the discharge through this
predicted. In Figure 12 the flow at a certain time branch is reduced. In Figure 14 the discharge
step is presented. In this time step the vortex characteristic is shown. It is obvious, that the
extends into the first side branch. It starts at the top discharge varies alternatively between the two side
of the sphere. The vortex is prescribed by an iso- branches. Low discharge corresponds with the
pressure surface and by instantaneous streamlines. location of the vortex in this branch. If the vortex is
Some time later, Figure 13, the vortex has located in the other side branch the discharge is
“jumped” to the opposite side branch. Since the high. The discharge in the middle branch only
geometry is not completely symmetric the vortex shows much smaller oscillations.
stays longer in branch 3 than in branch 1. This is
shown in the simulation as well as in model tests. Left
Middle
Discharge in %

Right

Time in s
Figure 11: Vortex structure obtained by URANS Figure 14: Discharge through the different
with k-ε turbulence model. branches.
Loss coefficient ζ
right

left

middle
Time in s

Figure 15: Loss coefficients of the three branches Figure 16: Test rig at ASTRÖ, Graz.

In the simulation a free outflow boundary


condition is applied at the end of the branches. In
reality, however, the turbine is located there.
Consequently the discharge variation is rather small
since the flow rate through the different branches is
prescribed by the turbines. For comparison with
experimental data therefore the loss coefficients are
calculated and compared with the model test data.
In Figure 15 the loss coefficients for the three
branches are presented. It can be observed, that the Figure 17: Vortex in the trifurcation, measurements
branch on the right hand side shows the highest of ASTRÖ.
values and has a strong variation. The loss
coefficient of the middle branch is nearly constant. Comparing the measured loss coefficients with
the ones from the simulation (Figure 15) one can
4.3 Model tests observe, that the maximum values are still
Model tests were carried out at ASTRÖ in Graz, underpredicted. For the right hand side branch the
Austria. For details on the measurements the reader numerical simulation shows a maximum of 6
is referred to [13]. In Figure 16 the test rig is whereas in the measurements the value of 8 is
shown, looking down from the penstock to the obtained. The values for the left hand side branch
trifurcation. The trifurcation was build in acrylic are clearly lower (approximately 3 in the simulation
glass so that the vortex motion could be observed. and 4 in the measurements). The overall tendency
In Figure 17 the view from a top of the of the flow, however, is well predicted. Also the
trifurcation is shown. One can see the vortex quantitative prediction is quite reasonable.
starting at the top and expanding into the side
branch. The vortex is made visible by reducing the
pressure level, that small cavitation bubbles appear
in the vortex center but without disturbing the flow
structure severely. In the model tests the vortex also
moves from one side to the other, as described
above.
From the pressure and discharge measurement
the loss coefficients of the different branches were
calculated. In Figure 18 the variations are shown. Figure 18: Loss coefficients for the different
branches, measurements by ASTRÖ.

The underprediction of the loss coefficient is


assumed to be due to the rather coarse grid on one
side. On the other side it is due to the strong
anisotropic turbulence behavior, which can not be
predicted accurately by a turbulence model based
on the eddy viscosity assumption. Therefore it is
intended to develop an adaptive, algebraic [3] Ruprecht, A. (1989): Finite Elemente zur Be-
Reynolds-stress model in future. rechnung dreidimensionaler turbulenter Strömungen
in komplexen Geometrien, Doctorate Thesis, Uni-
4.4 Problem solution versity of Stuttgart.
In order to solve the oscillation problem in the
[4] Ruprecht A. (2003): Numerische Strömungs-
hydro power plant, the trifurcation was changed. To
simulation am Beispiel hydraulischer Strömungs-
avoid the forming of the vortex the upper and lower
maschinen. Habilitationsschrift, Universität Stutt-
region of the sphere is cut off by welding in flat
gart.
blades. The modified shape is shown in Figure 19.
In the meantime this modification has been carried [5] Van der Vorst, H. A., (1994): Recent
out and the power oscillations disappeared. As a Developments in Hybrid CG Methods, Proc. High
side effect the power output increased severely Performance Computing & Networking, München,
since the losses in the trifurcation decreased without 1994.
the existing of the vortices.
[6] Zienkiewicz, O.C., Vilotte, J. P., Toyoshima, S.,
Nakazawa, S. (1985): Iterative method for
constrained and mixed finite approximation. An
inexpensive improvement of FEM performance,
Comp. Meth. Appl. Mech. Eng., 51.
[7] Maihöfer, M. (2002): Effiziente Verfahren zur
Berechnung dreidimensionaler Strömungen mit
Figure 19: Modified trifurcation to avoid the power nichtpassenden Gittern, Dissertation Universität
oscillations. Stuttgart.
[8] Maihöfer, M., Ruprecht, A. (2003): A Local
Grid Refinement Algorithm on Modern High-
5. CONCLUSIONS Performance Computers, Parallel CFD, Moskau.
An adaptive turbulence model for Very Large
Eddy Simulation is presented. The model is based [9] Spalart, P. R., Jou, W. H. Strelets M, Allmaras,
on the extended k-ε model of Kim and Chen. By S. R. (1997): Comments on the Feasibility of LES
introducing a filtering technique the model can for Wings, and on Hybrid RANS/LES Approach,
distinguish between numerically resolved and 1st AFOSR International Conference on DNS/LES
unresolved parts. Rouston.
Applying this new model the unstable vortex [10] Constantinescu, G. S., Squires, K. D. (2000):
motion in a pipe trifurcation is calculated. This flow LES and DES Investigations of Turbulent flow over
phenomenon could not be predicted with classical a Sphere, AIAA-2000-0540.
RANS methods and usually used turbulence
models. By applying the VLES method with a new [11] Magnato, F., Gabi, M., “A new adaptive turbu-
turbulence model the phenomenon can be obtained lence model for unsteady flow fields in rotating ma-
and the results agree quite reasonable with chinery”, ISROMAC 8, 2000.
measurement data. The vortex swirl, however, is [12] Willems, W., Peters, N. (1997): Large Eddy
still underpredicted. A further improvement is Simulation of a Turbulent Mixing Layer Using a
expected by applying an adaptive Reynolds-stress New Two-Level-Turbulence Model, 11th Symp. On
model to account for the anisotropy of the Turb. Shear Flows, Grenoble.
turbulence in swirling flows.
[13] Hoffmann, H., Roswora, R. R., Egger, A.
(2000): Rectification of Marsyangdi Trifurcation,
6. REFERENCES Hydro Vision 2000, Charlotte.
[1] De Langhe C., Dick E. Very Large Eddy
Simulation of Turbulent Flows.
http://allserv.rug.ac.be/~edick/
[2] Chen, Y. S., Kim, S. W. (1987): Computation of
turbulent flows using an extended k-ε turbulence
closure model, NASA CR-179204.

You might also like