Name: Advanced CFD for Francis Turbines: Hill Chart Prediction and Part-Load Analysis
Uploaded: 2026-04-23T16:30:06.555Z
Duration: 39 min 8 s
Description: Advanced CFD for Francis Turbines: Hill Chart Prediction and Part-Load Analysis

Transcript for "Advanced CFD for Francis Turbines: Hill Chart Prediction and Part-Load Analysis": Welcome everyone to today's webinar. We will begin in a few minutes. Thank you for joining today's webinar. We will begin in a few minutes. On behalf of Cadence, I'd like to welcome you to today's webinar, advanced CFD for Francis turbines, hill chart prediction and part load analysis. My name is Cale Hawkins. I'm the field marketing manager here at Cadence. Before I introduce today's presenter, I'd like to go over a few housekeeping items. All attendee lines are muted to prevent background noise. There will be a question and answer segment at the conclusion of today's webinar. If you have any questions during the presentation, you may type it in the ask a question section at the bottom right hand side of the screen. We will attempt to answer all questions at the end of the webinar. However, if we run out of time, questions may be answered via email within the next few days. Please ask your questions early so that we can get them queued up and ready to go for the q and a session. Today's webinar is being recorded, and a link to the recording will be emailed to you in the coming days. Final note, we will control the movement of the slides. Some slides include graphics and may take some time to advance depending on the speed of your connection. At this time, I will turn the event over to our engineering team. Hi. Hello. Thanks thanks for the introduction, and, welcome everybody to to this webinar. My name is Domenico Medici, and I'm a product engineer here at Cadence. This webinar is part of a larger series of talks on product features and application that we find interesting and we want to share and discuss with our customers and users. And to this conversation about transits turbines aligns with Cadence ambitious goals and strategy for turbo machinery applications. We are dedicated to making substantial effort in turbo machinery simulation, through three key directions. On one side, we are we are keep being working on our solver fine turbo, and this is our fully structured solver. It's fully available on GPU, and it's a very rapid tool that can be used from the very beginning of the design, especially thanks to its light generator auto grid. On the on the opposite side, LES is our higher fidelity solution. When we need to go beyond the runs your runs approach, we run our LES server, which is also on GPU. This makes simulations that once were extremely expensive, affordable both in time and and in cost. Today, however, we'll focus on the detailed design direction, where the level of complexity can extend along two big lines. One one is the physics. So when we are talking about phenomena where the off the off design hydraulics or aerodynamics can dominate, this means partial loading or, high swirl draft tube instabilities, efficiency loss. On the other hand, we have, geometrical, complexity. This means that the structure grid are too complicated to make, and the reliability of a structure grid is necessary. The workflow that we are going to use for today's webinar, goes beyond the it's too much complexity that can be handled by the structural server and extends to a higher level of complexity also in terms of physics. So the core of this approach is our pressure based solver. It's a coupled solver that can produce results fast, can handle mixer grids, and is ready to run on GPU. One key advantage is of the solver is that it's suitable for all type of applications or all speeds on our Mac. It can run applications that vary from the low speed fan, to the high Mac turbine. Today has however, it's about hydraulic turbine, so I'll link the stage to Margarita. Margarita is another product engineer here at Cadence. She holds a master's degree in mechanic engineering from Politecnico di Torino and, degree in mechanical engineering from Institute of Technology called the Buenos Aires. So thanks, Margarita. Go ahead. Thank you, Emenico, for the interaction. So let's dive in into a Francis turbine simulation for HIL chart and off design analysis. In this webinar, I will, present how to set up the case, how to calculate the HIL diagram, and what happens at part load or part load analysis. The case that we're tackling in today is the Francis 99 case. It's a very well known case for in our industry. There were a series of workshops done in the water powered laboratory where the experimental results are available as well as the computational domains. The turbine is actually a scale model of a turbine operating at the TOCA power plant in Norway, and it's a relatively high head, low specific speed turbine for a Francis turbine. The computational domains, we can see them here in different colors. The 14 state vanes are included in the inlet spiral. There are as well 28 guide vanes in a different domain, the runner with 15 blades and 15 splitters, and finally, the outlet draft tube. For meshing this case, I have decided to use two separate tools. First, for the guide vanes and the runner, I'm doing a structured grid. This is done with the Fidelity auto grid that can handle a lot of cases that are common, configurations for hydraulic turbines. It can handle with high quality and high fidelity, fill it meshing, clearances, all types of leading and trailing edges, bulbs, etcetera. Here for the runner, you can see I have a splitter and a main blade, and the periodic are done automatically by AutoGrid. The generation time for these meshes is extremely fast. Both of them are done in under half a million, and the number of cells is around 2,000,000 in total for both meshes. The the quality is well, with under the criteria that we require for a Fidelity flow solver. On the other hand, for the draft tube and the spire inlet spire, I have used an unstructured method with ANSA that is a powerful preprocessor that has one environment for CFD applications. Here, the geometry creation and editing was not necessary, but it's possible, And as well, it handles, well, the meshing, morphing, etcetera. These, two domains require a little bit more cells. So in total, I have around 7,000,000 more. The quality is also extremely good and well aligned with what we require for unstructured meshes, and, it's it's very, very high quality as well. For the setup of the structured case, we have many operating points with different guide vane angles and 15 rotational speeds per each guide vane angle. We can see in the picture in different colors the the different boundary conditions. So in blue is the inlet, and it's mass flow imposed. We have mixing planes for the rotor stator interfaces here displayed in green. We have as well the matching periodic boundaries that were generated automatically by AutoGrid. And at the outlet, we have an opening. Here, the option was to put an outlet, but an opening was chosen because there are several operating points where backflow is expected. And, therefore, the opening can handle these coming and going of, I mean, entering and exiting of mass flow through the surface. For the convergence, we can see that, the best efficiency point, the convergence is achieved in less than 100 iterations. We have the residuals decreasing well over minus four 10 to the power of minus four at even 50 iterations, and that the global quantities are also well stabilized. We can check as well the partial load convergence. Since the physics of this point is more unsteady, we have some oscillations as well. So more iterations were were done, but as we can see, the the results oscillate a bit around these points. Anyways, the it it is stable enough even at 100 iterations. And this is the point that I will, later on analyze with a urine simulation. For the performance assessment, we can see here the hill diagram obtained, of all the operating points. The best efficiency point is location is well captured as well as the efficiency isochurve shape. These results are very well aligned with what we see with the experimental values provided by the workshops. There is a slight overestimation of the efficiency that could be explained by the silicate losses that were not considered in this case because they were not also provided in the domains. And this could be dominant in a low specific speed turbine such as this one. For the off design analysis, the case setup is modified a bit. So here, we we mesh the full wheel and also all of the guide vanes to be able to analyze the rotor stutter inter interaction. And this is done also automatically by AutoGrid, and the mesh size, of course, is increased to almost 37,000,000 cells in total. Between the runner domain and under other domains, we use a sliding grid. And for between the savings and the guide vanes, we use instead a non matching boundary. For the turbulence models, there are three turbulence models that are analyzed. First, the shear stress transport, which is quite, standard in our industry, and it's also what was used for the steady state simulations. As well, the ERSM model was used, and it represents better the turbulence and anisotropy that we might expect in a case such as this one. Finally, as well, the SAS, sim turbulence model was used, which is an extension of the SST. It's based on the runs model, but it can capture large scale unsteadiness, like the LAS, if the grid and the flow allows. So to to run 40 full runner revolutions with GPU, it takes ten hours. So we can actually do two urine simulations in a day if we use GPU. This is six times faster than the CPU run that takes around sixty hours. Whereas for the runs, simulation, it in GPU, in in only one core of the GPU, it takes five minutes. And the good thing is that we can run eight simulations at the same time. So in five minutes, we actually get eight operating points, which is very efficient and, very cost effective. For the convergence system assessment, I will analyze the phase repeatability, the unsteady fluctuations in each revolution, and the mean values. As well, the web the workshop provides, experimental pressure probes that we can see the location here. And, therefore, for the simulation, a lot of control points were added also in these locations. There is one in the vaneless space between the guide vane and the runner. There's three on the on the runner blade, but one of them, we can see the p 71, is quite near the outlet of the draft tube. And then there is two control points as well in the draft tube. So here we are seeing the different time step. We're exploring what's the difference between using different time steps in our simulation. So for five degrees, 2.4 degrees, and one degree. For since we only have 12 degrees between two different blades of the runner, the five degrees is quite a course time step. Anyways, for for the convergence of this case, we can see that the phase repeatability is quite good. We see that the cycle average mean is bounded, and we can see here on the top on the right. Even, we see some small, low frequency fluctuations, and this could could be due to some low frequency unsteadiness. And, in the chart below, particularly at the static pressure difference, we see that the one degree, unsteadiness is higher than the two point four and five degrees. And this could mean that there is more unsteady phenomena being resolved for this one degree time step. For the comparison between them in the in the control points, we can first see the on the chart on the on the right at the top, we can first see the in the valet space, the control point, and we can see two peaks at fifteen and thirty, not normalized frequency. And these are actually the first and second harmonic of the runner. So it's the the blade passing frequency. And this is captured by by all of the time steps. However, with the five degree time step, we can see, an artificial frequency at 12, and this is actually the aliasing of the runner's fourth harmonic, which should be at 60. As well, if we look at the at the at the pressure at the control point on the pressure side of the blade, we can see the the peak of frequency at 28 normalized frequency, which corresponds to the guide main passing frequency. So this is also well captured for all of the of the time steps. Finally, if we look at the left, we can see the draft tube signal, and we can see that all of these, time step different time steps can model, capture a a very small frequency around 0.2, natural natural frequencies. And these are captured by all of these, by by all of these time stamps. So what's important here is to analyze whether we want to capture the retro saturate interaction or not. Because if if it's not interesting, then we can go with a coarser time step and a mixing plane that will kill, this interaction, and then we can still, analyze the the low frequency the low frequency phenomena. But if we want to capture the the adjusted reinteraction, then, of course, only two, time steps per blade passage is, too few. So, all of these simulations are done with one degree because we want to capture everything, and these are the the convergence assessment of of all the turbulence models that I mentioned before. So here we can see that the the phase repeatability is also very, very good for these are the last 20 revolutions of the simulation. And as well, we can see the cycle average mean and the RMS per revolution. If we look at the cycle average mean, both in the static pressure difference and in the axial thrust at the top right of the slide, we can see that the SAS model seems to have some low frequency phenomena that is not correctly captured by the other turbulence models. However, all of these unsteady simulations are are well converged because the the mean differences and variations, as well as the RMS, are are well bounded. So if we take a look at the at the control points, we can see that in the vaneless space at the left, we capture both, first and second harmonic of the runner's blade passing frequency. On the right, we can see as well the guide vane passing frequency also on all of the turbulence models. If we continue, and check the pressure side of the blade, we as well can see the the guide vane passing frequency again at 28, And we can see as well, and we saw before in in the section side, that there is some modulation of of this frequency at 28, and this could indicate that there is some unsteady vertical flow or unsteadiness in the guide ring trailing edge. However, if we look as well to the lower frequencies of p 71, which was the one that was near the draft tube, we can see at at the chart below that there is a peak of low frequency only for the SAS model that is not correctly captured for the other turbulence models. And if we check on the draft tube, as we can see here, we see that this effect is repeated. So we only see the draft tube in we are you only see in the draft tube the low frequencies in the SAS model. And this low frequency, which is at around 0.2, could be identified as the vertex rope frequency or the rain gun's frequency. And it is only captured, I as I mentioned before, with the SAS model. With the other turbulence models, we don't see any any frequency there. As well, if we look at the at at the low higher frequencies in this probe, we can see that the runner passing frequency at 30 is well identified by all of the turbulence models and even lower, but is there still, is a a guideline passing frequency. And this is normal because we are looking at the draft tube, so we are seeing both both of the frequencies. If we compare with the experimental results, we can see that these frequencies were correctly captured. So if we look at the first, chart on the left, we see that the other turbulence models capture correctly these runner passing frequencies even though the SAS seems to capture the amplitude better. On the other hand, if we look at the draft tube, two two different probes, we see that these low frequency phenomena is well captured, but only on the SAS model. So if we if we check the the three d results and we do a cue criterion of the velocity, we can see with the SAS the vertex rope structure, and this is colored with the swirling swirling speed. So we can see that there is the the vertex rope that we're seeing with the probes. Whereas with the ERSM, this is not formed. We just see a a blob that that could explain a little bit of the of the very low peaks that we had, but no vertex rope is, distinguished. Again, if we if we show instead with an isosurface of the swirling velocity, we see that the SAS is much better aligned to what we are expecting with for for this, partial load case, where in the ERSM, we we cannot see we see a very axisymmetrical flow that does not show this effect. So to conclude, the Affinity Flow solver was able to accurately, measure the this case, the hill chart and also the partial load analysis. CPU is is very, very convenient because we can do even two urine simulations per day. There is good accuracy, and it's cost effective. Also, the SAS seems to be more suitable to to simulate this case than other surveillance models due to this unsteady phenomena that is only identified with the SAS surveillance model. And, finally, the for further investigations, we will conduct as well the same simulation using the Reynolds stress model. So we can wrap it up here. Thank you very much for your attention. Thanks, Margarita. Thanks. Thanks for the presentation. That was great. Yeah. Please, send us all your questions. We are going to stay here available for a few minutes. So if you have any anything you want us to, reply to, please send us a question in the in the q and a box. We'll do our best to to reply. Alright. Seems like we probably have a question to lead off, Domenico. It looks like you may have one or two already. Yeah. And it's, I think it's about yeah. I think that the last sentence, I think, was of your presentation, Marietta, was was interesting. So, what is the RSM model and why didn't you use it? That's a that's a good question indeed. So the RSM is is a full renal stress model, which means that, basically, we, use all the renal stressants, and we use one at least one equate at least we use one equation for renal stress plus a closure equation. This is a is an extremely expensive and, let's say, accurate model, and, we want to run it. At the moment, this model is not yet, available on GPU, so we we I think we will run it in the future. And, yeah, stay tuned for another power possibly for another webinar, when, when this model will be available on GPU. Generally, it's something that extremely appreciated and extremely, as I said, important when we have anisometry, when there is rotation curvature. So it's it's actually the type of application we would like to to to tackle with this model. Great. Yeah. There are really a lot of question. And do you have a comparison of the cost SAS with respect to SST or ERSM? I think the cost is pretty much the same. I don't think we have as slides with that, but, generally, the SAS doesn't cost much more than an SST or ERSM model. And, yeah, as far as it allows us to capture more physics, as long as the mesh, of course, allows it for it. So the requirements in terms of meshing are slightly different, but it's a model that goes a bit beyond the limits of the of the limitations of the of the pure of the pure runs. We we do have another question related to cavitation. So how cavitation can be modeled in the solver and if it is supported on GPU? Cavitation is indeed possible to to to to be exploited in the solver. And we have basically three, different ways of, simulating this kind of phenomena. So the cheapest, I would say, will be the barotropic model where we prescribe a law for the variation of, of density with pressure. This is a model that is not suitable for all type of application, but it's quite fast, quite robust, and on which we have a lot of experience with all our solvers. We have the option of using tabulated fluids. So, what this is a position that is not usually water or as is more for cryogenic fluids, where basically we give we provide the the solar with the tabulated set of data, regarding the phase. And we have, let's say, a transport equation model, that is a more general, let's say, approach where we transport either the bubbles or or the the the mass fraction, let's say, the volume fraction of of vapor, through throughout the domain. And, yes, they all support it. The the all the three models are all the three, let's say, three approaches because the transport equation model has several variations. They are supported on GPU. Yeah. We have another question about the deep part load conditions. So we didn't study deep part load conditions, at least for this project. I I think it would be an interesting interesting study to run, especially with the with the other side model. So so it's a it's a it's a nice question and inspiration for, for for the next webinar. Yeah. We we use ANSA and not the fidelity structural measures. Indeed, we use a combination of measures, for spiral, for draft tube. So our cadence, we have really a lot of tools that we can use we we we can use for reprocessing. And, I believe the draft tube was matched with, with Ansel. So we tried, let's call it a surface to volume approach where the mesh is generated on the side on the surfaces and then, let's see, project it into the domain, whereas the spiral, if I remember one, was generated with Haxpres, and I think I I believe we use the the opposite approach. We generate the, the mesh in the volume, and then we project the mesh onto the surfaces. It's, yeah, it's just a matter of convenience, and that that that doesn't matter. I did did we consider in the hill chart the comparison between model test and CT? For the ASD results, we didn't consider it, but we didn't have enough time to, let's say, we didn't have enough GPU resources to to run that. So I can actually have a lot of GPUs, but we have also a lot of fewer one that want GPUs. So it's a it's a race for for for hardware. And, again, indeed, it will be it will be something interesting to do once once we can convince our colleague to leave our to leave us some hardware. Yeah. We have another question about target web blasting the simulation. Perhaps, Margarita, you can answer this. Yeah. Of course. So the target wipe loss for the guide vein, the runner, and the draft tube was one. But for the inlet spiral, we we chose to do it high Reynolds, so 30, mainly not to increase, very, very much the the cell count, of course. Then I see another question on how were the guide vanes rotated. So the guide vanes were rotated as well, automatically with auto grid, and, and then they were remeshed. So we just rotate them and remesh them. Alright. We we have a question about if we consider small walls, boundaries, or or rough walls. I believe it's small. Right? Yeah. It's smooth wall boundaries. Okay. Then. I Yeah. we. we have a last question about the geometry that was provided in the workshop. So this is actually the workshop one geometry, and it has, like, a well, in this question, it's mentioned cone, between the runner and the draft tube. And, yes, it's meshed with the with the with the with the meshes I showed is there. Of course, It's very easy to mesh with AutoGrid and ANSA as well. Right. If you want, get get in touch with us. I believe there is some way of getting in touch with us. So, yeah, get in touch with us. We can show you how how we mesh it more in detail. Yep. Alright. I think we can we can wrap it up. If. you have still questions, please get in touch with us in any way or another, and we'll be happy to to answer. That's fantastic. Thank you so much for a great presentation. Ladies and gentlemen, that's gonna conclude the webinar today, but I'm gonna launch a quick survey. It's only a couple of questions. It'll take you a few seconds. If you wouldn't mind, answering that for us, it'll really let us know how we're doing, and, we hope to see you next time. Thank you so much for attending.