Unsteady Flow Analysis in Hydraulic Turbomachinery

Albert Ruprecht Institute of Fluid Mechanics and Hydraulic Machinery University of Stuttgart, Germany ABSTRACT
In the field of hydraulic machinery Computational Fluid Dynamics (CFD) is routinely used today in research and development as well as in design. At that nearly always steady state simulations are applied. In this paper, however, unsteady simulations are shown for different examples. The presented examples contain applications with self excited unsteadiness, e. g. vortex shedding or vortex rope in the draft tube, as well as applications with externally forced unsteadiness by changing or moving geometries, e. g. rotor-stator interactions. For these examples the requirements, potential and limitations of unsteady flow analysis assessed. Particularly the demands on the turbulence models and the necessary computational efforts are discussed.

INTRODUCTION For more than a decade Computational Fluid Dynamics (CFD) is used in the field of hydraulic machinery in research and development as well as in the daily design business. Early successful demonstrations are given e. g. in the GAMM workshop [1]. The applications are steadily increasing. This is expressed in fig. 1, where the percentage of papers dealing with CFD is shown, which were presented at the IAHR Symposium on Hydraulic Machinery and Cavitation. Starting with Q3D-Euler and 3DEuler today usually the Reynolds averaged Navier-Stokes equations together with a robust model of turbulence (usually the k- model) is used. It is common practice to apply steady state simulations, the unsteadiness in consequence of the rotor-stator interactions is addressed by averaging procedures. By this method accurate results are obtained for many questions in the design of components. Fig. 1: Percentage of papers at the IAHRHowever, different problems in Symposium dealing with CFD turbomachinery arise from unsteady flow phenomena. In order to get information on this phenomena or solutions to the problems an unsteady flow analysis is necessary. This requires a much higher computational effort, roughly a factor 5-10 compared to steady state, depending of the problem and of the degree of modeling assumptions. With todays computers and software, however, unsteady problems can be solved.

Two major groups of unsteady problems can be distinguished. The first group are flows with an externally forced unsteadiness. This can be caused by unsteady boundary conditions or by changing of the geometry with time. Examples are the closure of a valve, the change of the flow domain in a piston pump, or the rotor-stator interactions. The second group are flows with self excited unsteadiness, which are e. g. turbulent motion, vortex shedding (Karman vortex street) or unsteady vortex behavior (e. g. vortex rope in a draft tube). Here the unsteadiness is obtained without any change of the boundary conditions or of the geometry. There can also occur a combination of both groups (e. g. flow induced vibrations, change of geometry caused by vortex shedding). All these phenomena can take place in a turbine or pump and require different solution procedures. BASIC EQUATIONS AND NUMERICAL PROCEDURES In hydraulic turbomachinery today usually the Reynolds averaged Navier-Stokes equations for an incompressible flow are applied. Compared to the steady state the momentum equations contain an additional term prescribing the unsteady change: U j Ui Ui 1 P Ui + ij ' = 0 + Uj + (1) x j x i t x j x i x j ij are the Reynolds stresses, which are calculated from the turbulence model. The continuity equation for incompressible flow reads Ui (2) =0 x i and does not contain a time depending term. It has to be emphasized that the equations (1) and (2) behaves different in time and in space. In space they show elliptic behavior, therefore they require boundary conditions on all surfaces. In time, however, they are of parabolic nature, which mean that there is no feed-back from the future to the present or past. Because of that no boundary conditions are required in the future. This is schematically shown in fig. 2. This is the reason, why the time discretization is generally carried out in a different way than the spatial discretization. For spatial discretization usually a Finite Volume or a Finite Element approximation is applied. For time discretization, however, mostly the Finite Difference method is used. A few of the most popular finite Fig. 2: Boundary and initial conditions difference approximations are shown in fig. 3. In addition explicit multi-point schemes of RungeKutta-type or predictor-corrector schemes are often applied.

Fig. 3: Time discretization schemes

It has to be mentioned that the explicit methods require a restriction of the time step according to stability criteria (CFL-criteria), which depend on the local velocities and the local grid size. The implicit methods, in contrary, are always stable, there is no restriction of the time step. It can be chosen only according to the physical requirements. In order to obtain accurate solutions the time discretization should be at least of 2nd order, similar to the spatial discretization. Otherwise extremely small time steps would be required. The above description of the flow in the Eulerian coordinates can be applied for unsteady boundary condition problems as well as for self excited unsteadiness. However, to express problems with moving geometries in Eulerian coordinates is more difficult. At the moving boundary a Lagrangian description can be applied very easily since the fluid particles can be traced by this method. Combining these two methods an Arbitrary Lagrangian Eulerian (ALE) method can be utilized. This method is suitable for the solution of problems with moving boundaries. In the ALE method the reference coordinates can be chosen arbitrary. In this referential coordinate system the material derivative can be described as f x L f x R f x E i ,t i ,t i ,t = + (u j w j ) (3) t t x j with the coordinates R E xL i ... Lagrangian cooddinates; x i ... referential cooddinates; x i ... Eulerian cooddinates and W i ... reference velocity. The momentum equations in the ALE formulation can be written as follows U j ~ Ui U 1 P Ui + ij = 0 + (U j W j ) i + (4) x t x j x i x j x j i The moving of the reference system W i can be chosen arbitrary. If W i is equal to zero one gets the Eulerian description, on the other hand, if wi is equal to the velocity of the fluid particle the Lagrangian formulation is obtained. The convective term in the transport equations for scalar quantities changes in the same way than in the momentum equations. This applies also to the k- and -equations. The numerical realization of moving or changing grids can either be obtained by deformation of an existing mesh in each time step. For large deformations this requires an automatic grid smoothing algorithm or even an automatic remeshing after a few time steps. An other method is the use of different embedded grids, which can move

( )

against each other. In this case a sliding interface between the non-matching grids is required. This procedure is schematically shown in fig. 4 for two different problems, namely rotor-stator interaction and vibration of a cylinder in a fluid. In FENFLOSS, the computer code developed at our institute at University of Stuttgart, the second approach is applied. The interface between the grids is realized by means of dynamic boundary conditions, where downstream the node values (velocities and turbulence quantities) are prescribed and upstream pressure and fluxes are introduced as surface conditions. A brief overview on the numerical procedures is given in [2], for more details the reader is referred to [3,4]. One point has to be emphasized. Since the unsteady simulations require a severe increase of computational effort compared to steady state Fig. 4: Moving grid examples solutions, parallel procedures are necessary. In this case the ALE formulation with moving grids leads to a dynamic change of communication because the location of exchange boundaries varies with time and can therefore change the computational domain of the processors, see [2]. In FENFLOSS an implicit solution algorithm is applied. As already mentioned this has the advantage that there is no stability limitation for the time step. The overall solution procedure including the fluid-structure interaction is shown in fig. 5. If the movement of the grid does not depend on the flow situation the fluid-structure loop vanishes.

Fig. 5: Flow chart of FENFLOSS including fluid-structure interaction

APPLICATIONS In the following selected applications are shown and the specific problems for this examples are discussed. Firstly some cases with self excited unsteadiness are presented. Vortex shedding at the inlet of a power plant Problem description: The first example shows the flow behavior at the inlet of a lowhead power plant. It is an existing plant with two identical bulb turbines. During operation the inner turbine showed severe bearing problems whereas the outer turbine operates smoothly. The reason was expected to be vortex shedding at the inlet. By numerical analysis the problem was investigated and it was tried to find a solution to the problem. In fig. 6 the geometry is shown. The calculation has been carried out in 2D as well as in 3D. Firstly it was tried to carry out a steady state simulation, however, no converged solution could be obtained. Therefore an unsteady simulation was undertaken. The results indicate a strong unsteady motion. In fig. 7 the velocity distribution at a certain time step is presented. Clearly visible are the vortices, shedding from the inlet and moving Fig. 6: Geometry of power plant inlet downstream into the inner turbine. This is the reason of the destruction of the bearings. In order to improve the flow behavior a modified geometry was suggested. This geometry, shown in fig. 8, has been built in the meantime. There are no longer Fig. 7: Instantaneous velocity vectors, vortex shedding at the inlet pier problems with vortex shedding. Further details about this application can be found in [5,6].

Discussion: The physical unsteadiness of the flow has been indicated by the inability to achieve a converged steady state solution. This is very often the case with flows showing vortex shedding in reality.

Fig. 8: Modified geometry

A necessary condition for that is, that the numerical scheme does not contain serious artificial diffusion, which would suppress the unsteady motion. The same applies to the used turbulence model. The standard k- model usually produces a too high eddy viscosity, especially in swirling flows, and therefore it very often suppresses the unsteady motion. This will be discussed again in other applications. For many cases at least a streamline curvature correction or even a non-linear eddy viscosity formulation is necessary in order to avoid a too high turbulence production. Another point in turbulence modeling is the treatment of the near wall flow. It is well known that the use of wall functions usually tends to predict a flow separation too late. In case of vortex shedding this can cause a severe reduction of the vortex sizes or even a complete suppression of the vortices. More accurate results can be obtained by solving the flow up to the wall (if possible) by a low-Reynolds- or a twolayer model. The results shown above are achieved by an algebraic turbulence model (Baldwin-Lomax-type) where the flow is resolved up to the wall.

Vortex rope in a draft tube Problem description: As an other self excited unsteady flow example the simulation of a vortex rope in a draft tube is shown. Here a straight axisymmetrical diffuser is considered. The inflow conditions to the diffuser are chosen according to the part load operation of a Francis turbine. This means that the flow shows a strong swirl component. The inlet velocity distribution and the geometry are presented in fig. 9. The instantaneous flow for a certain time step is given in fig. 10, where an isopressure surface as well as the secondary velocity vectors in three cross-sections are plotted. Clearly the cork-screw type flow with an unsymmetrical form is visible, although the geometry and the boundary conditions are completely axisymmetrical.

Fig. 9: Geometry and inlet conditions

Fig. 10: Iso-pressure and secondary flow of a vortex rope In fig. 11 the secondary velocity and the low pressure region, which represents the vortex center, is shown in the cross-section S, indicated in fig. 9, for certain time steps. Clearly the revolution of the vortex center can be observed. This, of course, causes pressure fluctuations and therefore dynamical forces on the draft tube surface.

Fig. 11: Secondary motion and low pressure region for different time steps

Discussion: Concerning the numerical scheme and the turbulence models the discussion above also applies here, e. g. application of the standard k- model leads to a steady state, symmetrical solution. This is also reported in [7]. The results shown above are achieved by applying the multi-scale k- model of Kim [8] together with a streamline curvature correction. This model shows a much lower eddy viscosity than the standard model, especially in swirling flows. The application of wall functions does not give any problems here, since the flow instability has its origin in the center and is not affected by the prediction of the near-wall region.

Vortex instability in a pipe trifurcation Problem description: In the following another problem caused by a vortex instability is shown. It is a pipe trifurcation, which is established in a power plant in Nepal. The trifurcation distributes the water from the penstock to the three turbine units. The problem in this plant arises from severe fluctuations of the power output of the both outer turbines. By field measurements the trifurcation was discovered as the reason for the fluctuations. By means of CFD and by model tests, carried out at ASTROE in Graz, Fig. 12: Geometry of the trifurcation the flow behavior should be analyzed and a cure of the problem should be found. The geometry of the trifurcation is shown in fig. 12. It has a spherical shape. The fluctuation in the trifurcation is caused by a strong vortex, which tends to be unstable. It skips between the two situations, sketched in fig.13. In the model tests the secondary velocity of the vortex could be found to be 30 times higher than the transport velocity. The reason is that at the top of the sphere there is enough space for a huge vortex to form. This vortex concentrates in the side branches and there- Fig. 13: Vortex structure fore increases the swirl intensity. Because of this strong secondary motion there are strong losses at the inlet of the branch, which reduces the head of the turbine and therefore causes the reduction of power output. During the project it was tried to obtain the unsteady behavior by a k- simulation on relatively coarse grids (200-300.000 nodes). However, these calculations did not show the vortex instability. Merely a vortex forms which extends from one side branch to the other. The swirl intensity was underpredicted by more than a factor five. Because of the low swirl rate the vortex is completely stable and has no tendency of skipping between different stations. Even by a dynamical excitation caused by changes of the outlet boundary condition of one branch the predicted vortex did not change its position. Only when applying finer grids and another turbulence model the predicted swirl intensity could be increased. Here an algebraic turbulence model with a limitation of the eddy viscosity is applied. The used grids consists of about 500 000 nodes. As a consequence this leads to an instability of the vortex. In the prediction the vortex skips between the two structures shown in fig. 14. One of these structures corresponds quite well with the structure observed in the model tests. In the second situation the vortex expends from one side branch to the other. This complies with the above mentioned stable results. The calculated swirl intensity is still more than two times lower compared to the results of the model tests. Therefore further investiga-

tions with other turbulence models and with finer grids are necessary and will be carried out in future.

Fig. 14: Predicted vortex structures For completeness the solution to the problem is shown. It consists of the installation of two plates in the upper and lower part of the sphere. This is shown in fig. 15. Hence no free space is available, where the vortex can form. Consequently the intensity of the vortex is dramatically reduced and the vortex is completely stable. In the meantime the reconstruction was carried out and the fluctuation of the power output vanished. As a by-product the losses in the trifurcation are severely reduced, which results in an increase of power output of approximately 5%. Further details of this problem can be found in [9,10].

Fig. 15: Modified geometry

Discussion: As already mentioned the calculations using the k- model were not successful. It is well known that this model is not able to predict highly swirling flows accurately. The unsteady motion of the vortices (especially of very slim vortices), however, very much depends on the swirl intensity. In order to prescribe such types of flow with sufficient accuracy it is necessary to have highly sophisticated turbulence models and very fine grids, maybe the only way to achieve it is the application of large eddy simulation. Rotor-stator interaction in an axial tubine The following example belongs to the second group, the unsteadiness is forced by moving geometries. The problem in question is the

Fig. 16: Geometry of the investigated axial turbine

flow in an axial turbine. The speciality of this turbine is its relatively low specific speed. It has been designed for pressure recuperation in piping systems. The advantage is that the discharge is nearly independent of the speed, because of that the turbine cannot introduce waterhammers in the system. The geometry of the turbine is shown in fig. 16. It consists of the inlet confuser, 12 fixed guide vanes, 15 runner blades and the draft tube. The stator and rotor part is shown in more detail in fig. 17. For the simulation the complete turbine is considered including all flow channels in the guide vanes and in the runner, although a symmetry condition of 120 could be used. The reason is, that also a variant with unsymmetrical outlet has been investigated.

Draft tube

runner

guide vanes

Fig. 17: Geometry of rotor and stator

Fig. 18: Part of the computational mesh

The computational mesh consists of more than 2 million grid nodes, part of the grid is shown in fig. 18. These are roughly 60000 nodes per flow channel. It is a rather coarse grid, considering that the clearance between runner blades and casing has to be included in the model, which is necessary since the clearance flow very much affects the channel flow because of the short runner blades. The calculations are carried out using the standard k- model. In the following some results of the calculation will be shown. In fig. 19 the instantaneous flow in the runner is presented. The figure shows the pressure distribution of the runner surface as well as streamlines started at different locations. Looking at the pressure one clearly sees the stagnation point at the leading edge. The location of the

Fig. 19: Instantaneous flow in the runner

stagnation point varies slightly with the runner position. Generally the inlet flow angle seems to be slightly too flat. Therefore the stagnation point is shifted towards the suction side. Considering the flow in the tip clearance one can observe that at the inlet the shear forces dominate. The flow tends to go from the suction to the pressure side. In the second half of the blade the pressure forces dominate. The flow in the clearance goes from the pressure to the suction side. It can already be seen by this results that the design of the runner is not optimal. This is a first version, in the meantime a much better runner has been designed. However this geometry is numerically investigated since extensive measurements have been carried out for this configuration and the numerical results can be validated.

Fig. 21: Cavitation observation in the runner Fig. 20: Calculated pressure distribution for a certain runner position In fig. 20 again the instantaneous pressure for a certain time step is shown. One can observe the low pressure region on the suction side at the top of the runner blades. Clearly visible is the variation of the pressure with the position. The low pressure region corresponds quite well with the cavitation observation at the test rig, see fig. 21. There one also can observe the variation of the cavitation bubbles according to the runner position. As a quantitative comparison the pressure at two locations is shown in fig. 22. Position 1 is located in front of the guide vanes and the second position lies between the guide vanes and the runner. At both locations the measured and the calculated pressure corresponds quite well. One can see that even in front of the guide vanes pressure fluctuations can be observed. Between the stator

Fig. 22: Pressure distribution at two spot points

and the rotor fluctuations of nearly 25% of the head of the turbine can be seen. This, of course, leads to dynamical forces on the blades. In fig. 23 the torque on one runner blade as well as the torque of the complete runner is shown. The calculated torque fluctuation on a single blade are nearly 30% of the averaged torque. This is a dynamical force on the blading. The total torque, however, is nearly constant due to the great number of blades and due to different phases of the fluctuations. Further details concerning the turbine and the measurements are published in [11], details on the calculations are given in [12]. Discussion: Since the unsteadiness of the flow is forced by the changing of the geometry this problem is easier to attack than the examples shown above. Here sufficient results are obtained applying Fig. 23: Torque on a runner blades the same models than in steady state simulations and similar criteria apply for accuracy than for the steady state simulations. As seen in the comparison with the measurements the prediction of pressure (but also of velocities which are not shown here, see [12]) is quite accurate. Therefore this kind of calculation is suitable to predict dynamical forces. POTENTIAL, LIMITATIONS, REQUIRED RESOURCES The applications show that many unsteady problems can be investigated by CFD and many phenomena can be studied, even rather complicated ones. Here we will discuss again the potential and the limitations as well as the required resources, which are necessary for an unsteady simulation. Firstly the rotor-stator problem is discussed. It can be said, that for this type of flow the unsteady computations behave similar to steady state. The accuracy principally depends on the grid size and on the turbulence model used as in steady state. The reasons for inaccuracy also correspond very close to that of steady state simulations, e. g. wake flow, swirling flow etc. It has to be pointed out, that the requirements of computational effort for unsteady flows is much higher than for steady state. Looking at a single component an unsteady simulation needs at least 3-5 times more computing time. But due to the absence of any periodicity the complete turbine including all stator and rotor channels have to be considered, in opposition to steady state, where periodicity can be applied by circumferential averaging. Depending on the type of machine and on the number of guide vanes and runner blades the necessary grid nodes can be increased by a factor of 20-30, to achieve a similar accuracy. Another problem for unsteady computations of a complete turbine or pump is, that the flow contains a large range of frequencies. In order to resolve the high frequen-

cies sufficiently accurate a small computational step has to be chosen. This, however, results in very long computational times when also low frequency phenomena have to be resolved, see [2]. Concerning the self excited vortex flow qualitative predictions can be obtained. The phenomenon of vortex shedding can be calculated quite accurate. Even if the detailed flow behavior may not be kept completely correct the frequencies and amplitudes of integral quantities (e. g. forces) can be predicted with sufficient accuracy for most of the problems. Flow instabilities and correct vortex movements, however, depend very much on the detailed flow situation. Even small changes of velocity can have a great response in the flow structure. As an example the flow in the spherical trifurcation is mentioned. There the vortex instability strongly depends on the swirling rate. This means it is essential to predict the highly swirling flow very accurate, but that is a severe problem for all the turbulence models used in practice today. At least it is necessary to apply non-linear models or Reynolds-stress models, since swirling flows are dominated by anisotropic effects. Maybe sufficient results can only be achieved by Large Eddy Simulations. Because the vortices are often very concentrated and consequently show very steep gradients their prediction require very fine computational grids. If the vortices move with time a self adaptive mesh refinement would be desirable. This, however, is rather complicated. Since the required computational effort is very high, parallel computing must be applied in order to obtain reasonable response times. An adaptive grid generation then leads to a dynamical load distribution. Both of the two approaches, either using extremely fine grids in total geometry or using adaptive meshes with dynamical load balancing, leads to a computational effort, which is at least 20 times higher than a similar steady state solution. CONCLUSIONS Unsteady simulations for the different applications have been shown, among them are applications dominated by vortex shedding or vortex instabilities as well as applications with forced unsteadiness which is e. g. rotor-stator interactions. All simulations have in common a quite large requirement of computational resources. Especially for rotor-stator interactions the complete turbine has to be considered and all flow channels in the stator as well as in the rotor have to be included. This leads to many grid nodes and an enormous computational effort. On the other hand the flow with vortex instability, e. g. vortex rope in a draft tube, represents a great challenge for an unsteady simulation. There the swirl intensity very much affects the overall movement of the vortex. But just strong swirling flows are extremely difficult to be calculated and the turbulence models applied today cannot capture this behavior sufficiently accurate. For this type of flow it is necessary to use improved models of turbulence. Even if there are problems with the computational accuracy for certain applications generally many information can be obtained from unsteady simulations. For example the dynamical forces in the axial turbine could be predicted with sufficient accuracy in order to allow an assessment of the durability. On the other side flow phenomena can be predicted and potential of influencing them can be assessed.

References
[1] [2] GAMM-Workshop on 3D Computation of incompressible internal flows, Lausanne, 1989, NNFM, Vieweg, 1993. Ruprecht, A., Heitele, M., Helmrich, T., Moser, W., Aschenbrenner, T., Numerical Simulation of a Complete Francis Turbine including unsteady rotor/stator interactions, 20th IAHR Symposium on Hydraulic Machinery and Cavitation, Charlotte, 2000. Ruprecht, A., Bauer, C., Gentner, C., Lein, G., Parallel Computation of Stator-Rotor Interaction in an Axial Turbine", ASME PVP Conference, CFD Symposium, Boston, 1999. Heitele, M., Helmrich, T., Maihfer, M., Ruprecht, A., "New Insight into an Old Product by High th Performance Computing", 5 European SGI/CRAY MPP Workshop, Bologna, 1999. Ruprecht, A., Maihfer, M., Gde, E., Flow analysis for the intake of low-head hydro power plants, 18th IAHR Symposium on Hydraulic Machinery and Cavitation, Valencia, 1996. Ruprecht, A., Maihfer, M., Gde, E., Untersuchung der Strmung in Kraftwerkseinlufen, 9. Int. Seminar Wasserkraftanlagen, Wien, 1996. Skotak, A., Draft tube swirl flow modelling, IAHR WG "The Behaviour of Hydraulic Machinery under Steady Oscillatory Conditions", Brno, 1999 Kim, S.-W., Chen, C.-P., A multiple-time-scale turbulence model based on variable partitioning of the turbulent kinetic energy spectrum, Numerical Heat Transfer 16(B), 1989. Ruprecht, A., Bauer, C., Gde, E., Janetzky, B., Llosa, P., Maihfer, M., Welzel, B., Strmungseffekte in einem Dreifach-Abzweiger (Trifurkation) - Numerische Analyse und deren Grenzen, 10. Internationales Seminar Wasserkraftanlagen, Wien, 1998.

[3] [4] [5] [6] [7] [8] [9]

[10] Hoffman, H., Egger, A., Riener, J., "Rectification of the Marsyangdi trifurcation", Hydropower into the Next Century, Gmunden, Austria, 1999. [11] Gentner, Ch., "Experimentelle und numerische Untersuchung der instationren Strmung in einer Axialturbine", Doktorarbeit, Universitt Stuttgart, 2000. [12] Bauer, C., "Instationre Berechnung einer hydraulischen Axialturbine unter Bercksichtigung der Interaktion zwischen Leit- und Laufrad", Doktorarbeit, Universitt Stuttgart, 2000.