Large Scale Computation of the Rotating Stall in a PumpTurbine using an Overset Finite Element Large Eddy

Simulation Numerical Code

THSE NO 6183 (2014)

PRSENTE LE 25 AVRIL 2014
LA FACULT DES SCIENCES ET TECHNIQUES DE L'INGNIEUR
LABORATOIRE DE MACHINES HYDRAULIQUES
PROGRAMME DOCTORAL EN MCANIQUE

COLE POLYTECHNIQUE FDRALE DE LAUSANNE

POUR L'OBTENTION DU GRADE DE DOCTEUR S SCIENCES

PAR

Olivier PACOT

accepte sur proposition du jury:

Prof. I. Botsis, prsident du jury
Prof. F. Avellan, Prof. C. Kato, directeurs de thse
Dr J.-B. Houdeline, rapporteur
Prof. O. Mtais, rapporteur
Prof. L. Villard, rapporteur

Suisse
2014

iii

If you cant fly then run,

if you cant run then walk,
if you cant walk then crawl,
but whatever you do,
you have to keep moving forward.
Martin Luther King Jr.

EPFL - Laboratory for Hydraulic Machines

iv

EPFL - Laboratory for Hydraulic Machines

Acknowledgements
This research has been carried out during the past four years between Lausanne and
Tokyo, thanks to Prof. Avellan at EPFL-LMH, Prof. Kato at The University of Tokyo
and the Swiss National Science Foundation which provided the financial support (grant
number 200021-130605).
First and most important, I would like to express my sincere gratitude to my thesis
directors Prof. Francois Avellan and Prof. Chisachi Kato for their motivation, enthusiasm,
patience and immense knowledge. Their guidance helped me all along my research work
and the writing of the thesis.
Besides my advisors, I would like to thank the rest of my thesis committee: Prof.
Ioannis Botsis, Dr. Jean-Bernard Houdeline, Prof. Olivier Metais and Prof. Laurent
Villard, for their encouragement, insightful comments, and hard questions.
For some people, the choice to do a PhD is made quite early into their studies. For
others, it is the result of personal experiences, meetings and opportunities. In my case, the
latter is true, firstly thanks to coniglietto (Steven) who introduced me to the laboratory,
located in the downtown of Lausanne. Secondly, I should thank Cecile who introduced
me to the Large Eddy Simulation and both of them motivated me to go further. Even if
during the last year I wondered several time why I decided to do a PhD, now I can be
nothing else than grateful, because it gave me the opportunity to meet many interesting
people, to learn a lot of exciting things and finally to explore a part of the world which
was totally unknown to me!
In view of the importance of computing ressources in my numerical research, and
the many trips between Geneva and Tokyo, I would like to thank Philippe Cerrutti and
Isabelle for their availability and constant support with my technical and administrative
needs.
My research wouldnt have been possible without all the work done by my predecessor
Olivier Braun. I would like to thank him for all his support and the advise he gave me in
order for me to pursue the investigation under the best conditions.
A PhD is also a long adventure shared among many people and not only between
PhD students. At LMH we have the opportunity to interact with other groups as the
mechanical group at the workshop, the design group or the test bed group. All together,
it was great to share raclettes and bbq parties, weekends at Ovronnaz and volley during
the EPFL championship. A special thank to Viktor, for making me suffer during our
jogging and to Max for helping me with all my big problems: shoes, door,... and for the
nice time we had during our weekly apero on Friday evenings.

vi

The length of the list of former and present PhD students is quite frightening. I
hope I will not forget anybody! Olivier: you were the first to leave and I wont forget
the nice samba show after your public presentation. Philippe Ausoni: you showed me
the effect that the writing has on PhD students and I wont forget the wonderful and
hilarious lunch together, during that period. Ruch: even if you are a snowboarder and
a windsurfer (nobody is perfect hehe) our ski trip by night and kite/windsurf sessions
were very nice. Sebounet, Cecile and your two wonderful daughters: Your are the best
family and a model for me!! Thanks for all the time we spent together and Im already
looking forward to our next kitesurf and ski sessions. What about Brazil 2015? Vlad:
how to forget our enjoyable discussions drinking Munichs beer after our nice experience of
the German highway. Coniglieto: the list is too long! From our Thursday nights during
the master to now, passing by our flat in Renens, there is too much to say and most
cannot be written here. Lets remember what good viewers we were during the winter
Olympic games 2010... Ebrahim: I wont forget your impressive knowledge and thanks
to introduce to me the SPH method. Vessaz: for sure I will miss your incredible local
stories you told us. Good luck with your remarkable SPH videos. Big body our preferred
scapegoat: For sure being from the low-Wallis is not easy..., but thanks to your constant
and appreciable good mood. Tthieuma: too much to say but I will remember how you
loved the glouglouglou at the ABL lecture! Recla the king: you were not supposed to
finish your PhD that early! But thanks for our telemark trips! It was awesome, so next
time in Hokkaido. Arthur: be the force with you. You recalled me some stuff of my past.
Good luck and do better than me!! Dr. Tinguely: still now Im amazed by the Marc
during the day and the Marc by night. Dont change! Boby: your name is always present
at the lab. We try to maintain your spirit and to honor the beer as you do. Finally, I
would like to make a special thank to the known Herr M
uller (HM)! Thanks for all the
support during these past months and for welcoming me at your place once Im back in
Lausanne. For sure, our trips to Okinawa and Hokkaido while tasting nice whiskies will
stay in my mind for long.
Two years ago, I went to Tokyo for the first time to work at the laboratory of Prof.
Kato. I would like to express my gratitude to Yamade-san, Guo-san, Nishimura-san,
Suzuki-san, Tsubuki-san, Takekoshi-san and all the others students for their help and
their warm welcome. They were always available and ready to help me either with my
research, or with my acclimation to the lifestyle in Japan. These two years were wonderful!
Besides the lab, I should thank my dinner mates Adrian, Aldo and Mike who supported
me especially these past months when I was writing. Also, a special thank to Nathalie
and Pauline who helped me when I had no place to go and to Pierre, Hel`ene, Nico, Mehul,
Alia, Koko-san, Jean-Pierre and Albert for all the nice activities we had together. Again
thanks to HM, Tthieuma, Mr. Tinguely, Max and the M&Ms for coming to visit me. It
was great to see you there.
During my stay in Tokyo, several people arrived at the laboratory to do a PhD or a
postDoc. I would like then to wish you Elena, Simon, Keita, Quentin, Philippe Kobel
and Loc all the best with your research. I hope you will appreciate the lifestyle of the
lab and of Lausanne. Good luck!
Finally, I would like to thank my family, who always supported me and encouraged
me to go further.

EPFL - Laboratory for Hydraulic Machines

R
esum
e
Suite aux nouveaux besoins energetiques et economiques, de plus en plus de pompeturbines sont utilisees dans les centrales de pompage turbinage, utilisees pour garantir
la stabilite du reseau electrique dune part et pour stocker lexcedent denergie electrique
dautre part. Pour cela, les pompe-turbines doivent pouvoir etre manoeuvrees efficacement dans une large plage de fonctionnement, cest-`a-dire pouvoir oeuvrer sans probl`eme
significatif `a des points de fonctionnement eloignes de celui pour lequel elles sont concues.
Cependant, lorsque lon seloigne trop du point de fonctionnement optimal, lecoulement
peut devenir instable ce qui engendre du bruit, des vibrations et une chute du rendement. Un exemple est lutilisation dune pompe-turbine en mode pompe `a charge partielle o`
u une instabilite du nom decollement tournant apparat. Dans ce cas de figure,
lecoulement passant autour de certaines directrices decolle et perturbe lecoulement du
diffuseur. Lendroit o`
u lecoulement decolle change avec le temps, ce qui engendre une perturbation locale et temporelle de lecoulement, do`
u la propagation de cette perturbation
dune directrice `a sa voisine.
Letude de ce phenom`ene a` ce jour est restee tr`es limitee due `a sa localisation dans
la machine qui est difficilement accessible. Pour une approche experimentale, il faut
trouver des solutions pour saffranchir des differentes parois qui obstruent la visualisation de lecoulement. Pour une approche numerique, il faut en trouver de nouvelles, car
lapproche RANS - traditionnellement utilisee - a montre ses limites. Une alternative
numerique serait lapproche par la simulation des grandes echelles (SGE) qui est connue
pour sa capacite `a calculer precisement de tels ecoulements, mais dont les besoins en
puissance de calcul ont ete trop contraignants dans le passe, pour que cette approche soit
utilisee. Aujourdhui, suite au developpement constant des supercalculateurs, la puissance
de calcul accessible offre de nouvelles perspectives pour etudier des phenom`enes physiques
par lapproche SGE. Lobjectif de ce travail de recherche est donc de demontrer que le
decollement tournant peut etre reproduit precisement par une simulation des grandes
echelles, afin delargir les connaissances de ce phenom`ene. Pour cela, la pompe-turbine
HYDRODYNA a ete choisie et lecoulement calcule est celui en mode pompe a` charge
partielle, =0.026 (76% BEP). En effet, une etude experimentale a demontre la presence
du decollement tournant pour ce point de fonctionnement.
La complexite dans la simulation des grandes echelles reside principalement dans la
generation du domaine de calcul que lon appelle maillage. Ce maillage est predominant,
car il gouverne la separation entre les grandes et les petites structures presentes dans
lecoulement. Sachant que lapproche SGE calcule explicitement les grandes structures
et modelise linfluence des petites, il est donc essentiel que le maillage soit suffisamment
fin afin de capturer toutes les structures principales de lecoulement pour assurer un
resultat de bonne precision. Dans le cadre de cette recherche, une etude de faisabilite

viii

a demontre que le decollement tournant peut etre simule en utilisant un maillage de

600 milliards delements pour un nombre de Reynolds identique a` letude experimentale,
ce qui nest pas envisageable. Cependant, si le nombre de Reynolds est reduit dun
facteur 5 compare a` lexperimental (Re/5), le maillage necessaire se reduit `a 7 milliards
delements, ce qui devient acceptable. De la meme facon, si le nombre de Reynolds est
reduit dun facteur 25, le maillage necessaire se reduit a` 80 millions delements (Re/25).
En consequence, un maillage de 85 millions delements a ete genere en vue dinitialiser
lecoulement et pour cela, le supercalculateur PRIMEHPC-FX10 de luniversite de Tokyo
- qui poss`ede un peu moins de 80,000 coeurs de calcul - a ete utilise. Concernant le
software, le code FrontFlow/blue, developpe par la meme institution, a ete selectionne.
Ce code est parallelise ce qui a permis deffectuer les calculs en utilisant 2,048 coeurs de
calcul, soit le calcul dun tour de roue en moins de 7 heures. Par ailleurs, ce code est base
sur la methode des elements finis pour lintegration des equations regissant le mouvement
des fluides Newtoniens (equations de Navier-Stokes), il utilise le mod`ele de Smagorinsky
dynamique pour fermer le syst`eme dequations propre a` lapproche SGE et il utilise une
methode de recouvrement des maillages comme interface entre la partie tournante et les
parties fixes.
Linitialisation de lecoulement pour le cas Re/5 avec le maillage de 85 millions delements a montre que le phenom`ene est dej`a capture apr`es le calcul de 10 tours de roue.
Ainsi, le decollement de lecoulement en 4 regions du diffuseur, appele ulterieurement
cellule de decollement, est observable, correspondant a` ce qui a ete rapporte experimentalement. De plus, lenergie calculee est proche de la valeur experimentale, soit 3.35%
superieure. Cependant, la vitesse de propagation des cellules est environ 60% plus lente
que celle escomptee, ce qui sexplique par le manque de resolution de lecoulement etant
donne quun maillage de 7 milliards delements est necessaire pour un tel nombre de
Reynolds.
Le calcul de lecoulement pour le cas Re/25, avec le meme maillage de 85 millions
delements, a permis de capturer precisement le phenom`ene du decollement tournant dans
la pompe-turbine. Lenergie calculee est 0.36% plus elevee que la valeur experimentale,
le nombre de cellules de decollement est identique et la vitesse de propagation des cellules
de decollement est seulement 3.4% plus lente que la valeur experimentale. Ceci confirme
la qualite du calcul et demontre la possibilite de traiter des probl`emes de type industriel
par cette approche.
En utilisant lecoulement obtenu dans le cas Re/25, le mecanisme de propagation des
cellules a pu etre etudie. Pour le point de fonctionnement selectionne, certains passages
du diffuseur sont affectes par un decollement de lecoulement sur la directrice amont, ce
qui engendre un blocage du passage. De plus, lendroit o`
u le decollement sop`ere change
avec le temps, ce qui est observable par la succession de phases croissantes et decroissantes
du blocage dun passage. La consequence dun tel blocage dun passage est la deviation de
lecoulement vers le passage voisin en aval. En fonction de lintensite de cette deviation,
lecoulement dans le passage aval va egalement decoller et bloquer ainsi le passage, do`
u
la rotation de ces cellules de decollement.
Mots-cl
es: SGE, Pompe-turbine, Decollement Tournant, Mecanisme de Propagation,
Mode de Pompage, Charge Partielle

EPFL - Laboratory for Hydraulic Machines

Abstract
The increasing need of pump-turbines in todays and tomorrows electrical power storage and grid stability requires new approaches to develop the knowledge of the structure
of the internal flow of pump-turbines to improve their efficiency and to widen their range
of operation. The main issue encountered when operating the pump-turbine in pumping
mode at part load condition is the development of an instability in the diffuser known
as the rotating stall. This instability consists in the development of recirculation zones,
called stall cell, which result from a stall of the flow on a guide vane. Such a stall cell
propagates from one guide vane channel to the neighboring downstream channel, which
locally can be seen as the grow and decay of a recirculation zone.
As the stall cells are located in the diffuser, their investigation proves to be a complex
task. Therefore, only a few studies exist up to this day. Experimentally, the visualization
of the flow field is difficult because of the components (guide vanes, stay vanes,...) constituting the diffuser. Numerically, because of the high Reynolds number involved and
the low computing power ressources, only the RANS approach could be used and showed
the limits to accurately compute such a phenomenon with this approach. Thanks to the
constant development of supercomputers, new perspectives are opened up and more elaborated numerical approach as the Large Eddy Simulation (LES) can be selected. Thus,
the objective of this thesis is to perform a resolved Large Eddy Simulation to demonstrate
that the rotating stall can be accurately computed by the LES approach to shed light on
the physical mechanism behind the rotating stall phenomenon. To do this, the HYDRODYNA pump-turbine was selected and the computed flow is the one in pumping mode at
part load condition, =0.026 (76% BEP). At such an operating condition, experimental
investigation reported the presence of the rotating stall phenomenon.
To perform a resolved LES of a full pump-turbine, a sufficiently fine mesh has to
be generated, as the size of the mesh governs the split between the large and the small
structures in the flow. Since the LES approach computes explicitly the large structures
and models the effect of the small ones, it is essential to capture all the main structures
of the flow in order to accurately simulate the rotating stall phenomenon. In the present
study, a feasibility study showed that the phenomenon could be captured using a mesh
composed of 600 billions elements for a computation similar to the experiment, which
is not feasible. However, if the Reynolds number is reduced by a factor 5 (Re/5), the
required mesh is composed of 7 billions elements, which is today acceptable. If the
Reynolds number is further reduced by a factor 5 (Re/25) the required mesh size decreases
to 80 millions elements. Therefore, a mesh featuring 85 millions hexahedral elements was
generated and used to initiate the flow field in the pump-turbine. The computation of the
rotating stall requires the computation of several revolutions of the impeller and the fine
spatial and temporal discretizations induce time consuming computations. Therefore, it

is mandatory to use a powerful supercomputer featuring a large number of cores. For the
present computation, the PRIMEHPC-FX10 supercomputer of the University of Tokyo,
which features approximately 80,000 cores, is used with an overset finite element large
eddy simulation numerical code named FrontFlow/blue, which is developed at the same
institution. Using 2,048 cores and the dynamic Smagorinsky model to close the system of
equations, the computation of one impeller revolution can be performed within 7 hours.
The initialization of the flow field for the Re/5 case using the 85 millions elements
mesh showed that the phenomenon could be captured after the computation of 10 impeller
revolutions, i.e. four propagating stall cells are present in the diffuser. The computed
energy coefficient is close to the experiment and is 3.35% higher compared to the
experimental data. However, the computed propagation speed was approximately 60%
slower compared to that of the experiment. This difference is due to insufficient resolution
of the structure of the flow, as a 7 billions elements mesh would be required.
The flow computation for the Re/25 case, using the same 85 millions elements mesh,
accurately reproduced the flow field in the pump-turbine. The resulting energy coefficient
is only 0.36% higher compared to the experimental data. The same number of stall cells
is predicted and the propagation speed of the stall cells is 3.4% slower than that of the
experiment. Therefore, it is concluded that the rotating stall phenomenon is accurately
reproduced.
Based on the accurate computation, the stall mechanism could be studied. It was
shown that this instability involves the development of a recirculation zone near the
trailing edge of the guide vanes on the suction side. The size of this recirculation zone
almost periodically fluctuates and influences the surrounding flow field. Because of the
geometry of the guide vane diffuser, it appeared that the recirculation possesses a strong
influence and has the capability to nearly block the guide vane channel, which forces
the flow to deviate to the neighboring channel. This deviation induces a modification of
the flow in the downstream channel and results in favorable conditions for the flow to
separate.
Keywords: LES, Pump-turbine, Rotating Stall, Propagation Mechanism, Pumping
Mode, Part-load Condition

EPFL - Laboratory for Hydraulic Machines

Contents
I

Introduction

1 Problem Overview
1.1 Pumped-storage Power Plants . . . . . . .
1.2 Rotating Stall . . . . . . . . . . . . . . . .
1.3 High Performance Computing . . . . . . .
1.4 Numerical Simulation Specification . . . .
1.4.1 RANS, LES and DNS Tradeoffs . .
1.4.2 The smallest eddies to be computed
1.4.3 Mesh Requirements . . . . . . . . .
1.4.4 Computing Resources Requirement
1.5 Thesis Objective . . . . . . . . . . . . . .
1.5.1 Document Organization . . . . . .

II

1
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

Methodology

15

2 Large Eddy Simulation

2.1 Large and Small Scales Separation . . . . . . . . . . . . .
2.2 Low-Pass Filtered Incompressible Navier-Stokes Equations
2.3 Subgrid-scale Modeling . . . . . . . . . . . . . . . . . . . .
2.3.1 Smagorinsky Model . . . . . . . . . . . . . . . . . .
3 Numerical Tools
3.1 FrontFlow/blue . . . . . . . . . . . . . . . . . . . . .
3.1.1 Numerical Code Structure . . . . . . . . . . .
3.1.2 Numerical Code Performance on BlueGene/P
3.1.3 Stationary and Rotating Interface . . . . . . .
3.2 Computing Resources . . . . . . . . . . . . . . . . . .
3.2.1 IBM BladeCenter . . . . . . . . . . . . . . . .
3.2.2 Blue Gene/P . . . . . . . . . . . . . . . . . .
3.2.3 FX10 - K Supercomputer . . . . . . . . . . .
3.3 VisIt . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 File Format . . . . . . . . . . . . . . . . . . .
3.3.2 Parallelized Visualization . . . . . . . . . . . .
EPFL - Laboratory for Hydraulic Machines

3
3
5
6
7
7
9
10
12
12
13

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

17
17
17
19
21

.
.
.
.
.
.
.
.
.
.
.

23
23
23
23
26
27
27
28
29
30
30
30

xii

CONTENTS

4 HYDRODYNA Pump-Turbine
4.1 Pump-Turbine Characteristics . . . . . . . . .
4.1.1 Former Research on the Pump-Turbine
4.2 Flow Domain Discretization . . . . . . . . . .
4.2.1 Process . . . . . . . . . . . . . . . . .
4.2.2 Blocking . . . . . . . . . . . . . . . . .
4.2.3 Overset Discretization . . . . . . . . .
4.3 Computation Parameters & Flow initialization
4.3.1 Numerical Setup . . . . . . . . . . . .
4.3.2 Computation Convergence . . . . . . .
4.3.3 Propagation Speed Discovery . . . . .

III

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

Results

33
33
33
35
35
36
43
44
44
46
48

51

5 Reduced Reynolds Number Computation

5.1 Numerical Setup . . . . . . . . . . . . . .
5.2 Comparison with Experimental Data . . .
5.2.1 Propagation Speed Discovery . . .
5.2.2 Flow Patterns at Specific Locations
5.3 Discussion . . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

6 Rotating Stall Analysis

6.1 Stall Onset . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Observation of Stalled Channel Behavior . . . . . . . . . .
6.3 Propagation Mechanism of Stalled Cells . . . . . . . . . .
6.3.1 Pressure Evolution in Diffuser Channel . . . . . . .
6.3.2 Flow Incidence of Vanes . . . . . . . . . . . . . . .
6.3.3 Summary of the Stall Cells Propagation Mechanism
6.4 Rotating Stall Impact . . . . . . . . . . . . . . . . . . . . .
6.4.1 Impeller . . . . . . . . . . . . . . . . . . . . . . . .
6.4.2 Guide Vane . . . . . . . . . . . . . . . . . . . . . .
6.4.3 Stay Vanes . . . . . . . . . . . . . . . . . . . . . .
6.5 Grid Resolution Effect . . . . . . . . . . . . . . . . . . . .
6.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . .

IV

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

53
53
53
53
55
60

.
.
.
.
.
.
.
.
.
.
.
.

67
67
68
69
73
75
79
80
80
81
82
83
86

Conclusions and Perspectives

89

7 Conclusions and Perspectives

7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

91
91
92

Bibliography

97

Curriculum Vitae

105
EPFL - Laboratory for Hydraulic Machines

List of Figures
1.1

1.3

Evolution of the energy consumption in the world: a) per region, b) per

resources. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
Microprocessor transistor counts from 1971 to 2011. The straight line
shows the Moores law. . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
Flat plate wall skin friction coefficient as a function of the Reynolds number 11

2.1

Schematic view of the energy spectrum showing the scales separation. .

18

3.1
3.2

Flow computation algorithm. . . . . . . . . . . . . . . . . . . . . . . . .

Pie chart showing the normalized time spent in each main part composing
the computation of one time step. . . . . . . . . . . . . . . . . . . . . .
FFB7 weak-scaling benchmark up to 8,192 cores for a load per a core of
0.1 million elements on BlueGene/P. . . . . . . . . . . . . . . . . . . .
Overset scheme of two neighboring meshes. . . . . . . . . . . . . . . . .
Illustration of the half Pump-turbine domain decomposed into 512 subdomains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

24

1.2

3.3
3.4
3.5
4.1
4.2

4.3
4.4

4.5

4.6

Pump-turbine main components. . . . . . . . . . . . . . . . . . . . . . .

Specific energy and efficiency scaled by the efficiency at BEP versus
discharge of the HYDRODYNA pump-turbine. Instrumented impeller
(2006) and plain impeller (2008), N =600 rpm and N =900 rpm. Data
obtained from [3] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
EPFL LMH PF2 test rig. . . . . . . . . . . . . . . . . . . . . . . . . . .
a) Wire-frame representation of the cone geometry. The cyan curves
represent the physical limits of the cone and the grey curves represent
the overset limits. The dashed grey lines represent the outer limits of the
overset. b) Cone blocking. Using the rotational symmetry only 1/9 of
the volume is blocked. . . . . . . . . . . . . . . . . . . . . . . . . . . .
a) Wire-frame representation of the impeller geometry. The cyan curves
represent the physical limits of the impeller and the grey curves represent
the overset limits. The dashed grey lines represent the outer limits of the
overset. b) Impeller single channel blocking. . . . . . . . . . . . . . . .
a) Wire-frame representation of the diffuser channel geometry (symmetric
part). The cyan curves represent the physical limits of the diffuser channel and the grey curves represent the overset limits. The dashed grey
lines represent the outer limits of the overset. b) Diffuser single channel
blocking. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

EPFL - Laboratory for Hydraulic Machines

25
27
28
30
34

34
35

36

37

38

xiv

4.7

LIST OF FIGURES

a) Wire-frame representation of the diffuser channels geometry around

the tongue (nonsymmetric part). The cyan curves represent the physical
limits of the diffuser and the grey curves represent the overset limits.
The dashed grey lines represent the outer limits of the overset. b) 3
nonsymmetric diffuser channels blocking. . . . . . . . . . . . . . . . . .

39

a) Wire-frame representation of the volute geometry. The cyan curves

represent the physical limits of the volute and the grey curves represent
the overset limits. The dashed grey lines represent the outer limits of the
overset. b) Volute blocking. . . . . . . . . . . . . . . . . . . . . . . . .

39

4.9

Spatial discretization of the pump-turbine. . . . . . . . . . . . . . . . .

40

4.10

Mesh quality quantification using the equiangle skewness criterion. The

bars give the criteria distribution for each mesh in percentage. . . . . .

41

Mesh quality quantification using the volume change criterion. The bars
give the criteria distributions for each mesh in percentage. . . . . . . .

42

Mesh quality quantification using the aspect ratio criterion. The bars
give the criteria distributions for each mesh in percentage. . . . . . . .

43

(a) Time history of the pressure coefficient at SP1 during the flow initialization. (b) Time history of the pressure coefficient during the 10th
impeller revolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

Instantaneous normalized velocity on the diffuser symmetry plane after

the computation of 10 impeller revolutions. . . . . . . . . . . . . . . . .

48

(a) Time history of the pressure coefficient fluctuation at SP1 (black

curve). The red curve is the low-pass filtered pressure coefficient. (b)
Discrete Fourier Transform of the instantaneous pressure coefficient. . .

48

(a) Time history of the pressure coefficient fluctuation at SP1.The red

curve is the pressure coefficient fluctuation that is low-pass filtered using
a cutoff frequency equal to fn . (b) Discrete Fourier Transform of the
instantaneous pressure coefficient fluctuation. . . . . . . . . . . . . . . .

54

2D location of the velocity sampling points represented by circles. The

black equidistant circles compose the circular surface with a radius of
277 mm. The hyperbolically distributed green circles compose the throat
surface. The hyperbolically distributed blue circles compose the surface
0.5 mm apart from the guide vane. The two black circle filled in red are
the nearest sampling points to SP1 and the guide vane suction side. The
crosses SP1 and SP2 are two pressure sampling points. . . . . . . . . .

55

Time histories of the flow rate through 4 guide vane throats separated by
5 channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

Analytic signal of the low-pass filtered pressure fluctuation signal shown

on Fig. 5.1 (a): a) analytic/stall phase, b) analytic amplitude. . . . . .

57

Stall-phase-averaged normal (streamwise direction) velocity in the guide

vane throat section measured by LDV at = 0.026. a-f) Velocity contour
plots along different phase angles from 0 to 3/2. The bold line represents
the level of zero velocity. Figures obtained from [3]. . . . . . . . . . . .

58

4.8

4.11
4.12
4.13

4.14
4.15

5.1

5.2

5.3
5.4
5.5

EPFL - Laboratory for Hydraulic Machines

LIST OF FIGURES

5.6

5.7

5.8

5.9

5.10

6.1

6.2
6.3

6.4

6.5

6.6
6.7

6.8
6.9

Stall-phase-averaged normal (streamwise direction) velocity in the guide

vane throat section computed by LES at = 0.026. a-f) Velocity contour
plots along different phase angles from 0 to 3/2. The bold line represents
the level of zero velocity. . . . . . . . . . . . . . . . . . . . . . . . . . .
Stall-phase-averaged tangential velocity at 0.5 mm apart from guide vane
4, measured by LDV at = 0.026. The bold line represents the level of
zero velocity. Figures obtained from [3]. . . . . . . . . . . . . . . . . . .
Stall-phase-averaged tangential velocity at 0.5 mm normal distance from
guide vane 4, computed by LES at = 0.026. The bold line represents
the level of zero velocity. . . . . . . . . . . . . . . . . . . . . . . . . . .
Stall-phase-averaged velocity on cylindrical section at guide vane, measured by LDV at = 0.026. a, b and c) radial velocity, d, e and f) circumferential velocity. The bold lines represent the level of zero velocity
for the radial velocity and the level of 0.5 velocity for the circumferential
velocity. Figures obtained from [3]. . . . . . . . . . . . . . . . . . . . .
Stall-phase-averaged velocity on cylindrical section at guide vane, computed by LES at = 0.026. a, b and c) radial velocity, d, e and f) circumferential velocity. The bold lines represent the level of zero velocity
for the radial velocity and the level of 0.5 velocity for the circumferential
velocity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Guide vane profile. The grey area shows the region favorable for flow
separation in pumping mode and the dashed lines the two horizontal
tangents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Velocity triangle at the impeller exit for the BEP and the part load conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instantaneous spanwise-averaged velocity magnitude in six diffuser channels showing the spatial organization of a stall cell. Channels a-b and f-g
exhibit the highest flow rate passing through the channel. Channel b-c is
stalling, channel c-d is stalled and channels d-e and e-f are recovering. .
Spanwise-averaged instantaneous velocity magnitude passing through the
diffuser for 5 different stall phases: a) late recovery phase, b) highest flow
rate phase, c) stalling phase, d) stalled phase and e) early recovery phase.
f) evolution of the flow rate passing through the section shown by the red
line in figure a). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Spanwise-averaged instantaneous Cp in the diffuser for 5 different stall
phases: a) late recovery phase, b) highest flow rate phase, c) stalling
phase, d) stalled phase and e) early recovery phase. f) evolution of the
flow rate passing through the section shown by the red line in figure a).
The interval of the contour line (shown in black) is Cp = 0.015. . . .
Location of the diffuser sampling points (SP). . . . . . . . . . . . . . .
a) Time history of Cp located at SP1, represented by the black line.
The blue line is an interpolation. b) Interpolations representing the time
history of the pressure coefficient along a diffuser channel. . . . . . . . .
Time history of interpolated Cp in the area between the guide vane and
the stay vanes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Difference of the pressure coefficient at SP1 and SP2. . . . . . . . . . .

EPFL - Laboratory for Hydraulic Machines

xv

59

62

63

64

65

67
68

69

70

71
74

76
76
76

xvi

6.10
6.11
6.12
6.13

6.14

6.15

6.16
6.17

6.18

6.19
6.20

6.21

LIST OF FIGURES

Evolution of the AOA on GVA and GVB for three locations: 5%, 10%
and 15% of the guide vane chord length upstream of the leading edge. .
Evolution of the low-pass filtered radial component Cr at two locations,
see Fig. 5.2 on page 55. a) pt22, b) pt37. . . . . . . . . . . . . . . . . .
Evolution of the AOA on SVA and SVB for three locations: 5%, 10% and
15% of the guide vane chord length upstream of the leading edge. . . .
Comparison between the evolution of a quantity in the impeller for one
impeller revolution with the corresponding quantity on the stationary
part, i.e. the diffuser. a) pressure coefficient, b) flow rate (LF: Low Flow
rate, HF: High Flow rate). . . . . . . . . . . . . . . . . . . . . . . . . .
(a) Lift coefficient variation on a guide vane during one quarter stall
cycle. The black line is the instantaneous lift coefficient and the red line
the low-pass-filtered lift coefficient with a cutoff frequency set to fn . (b)
The green and the blue lines are respectively the normalized flow rate
through the upstream and downstream channels. The black line is the
flow rate difference between the two channels. . . . . . . . . . . . . . .
(a) Lift coefficient variation on a stay vane during one quarter stall cycle.
The black line is the instantaneous lift coefficient and the red line the
low-pass-filtered lift coefficient with a cutoff frequency set to fn . (b)
The green and the blue lines are, respectively, the normalized flow rate
through the upstream and downstream channels. The black line is the
flow rate difference between the two channels. . . . . . . . . . . . . . .
Comparison of the computed Head for the initial computation (Re/5) and
the reduced Reynolds number computation (Re/25). . . . . . . . . . . .
Comparison of the computed flow rate variation through the guide vane
throat for the initial computation (Re/5) and the reduced Reynolds number computation (Re/25). . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of the normal (streamwise direction) velocity in the guide
vane throat section at the minimum flow rate: a) initial computation
(Re/5), b) reduced Reynolds number computation (Re/25). . . . . . . .
Comparison of the kinetic energy spectrum near the pressure side and the
leading edge of the guide vane at the mid-spanwise. . . . . . . . . . . .
Comparison of the computed flow fields in the guide vane channels at
maximum (HF) and minimum (LF) flow rate for the initial computation
(Re/5) and the reduced Reynolds number computation (Re/25). . . . .
Schematic of the variation of the flow within a guide vane channel. . . .

77
78
79

81

82

82
83

84

85
86

87
88

EPFL - Laboratory for Hydraulic Machines

List of Tables
1.1
1.2
1.3

3.1

3.2
3.3
3.4
4.1
4.2
4.3

4.4
4.5

5.1
5.2

Supercomputer top 10 of the Top500 list, as of November 2013. . . . . .

Sizes of streamwise vortices expressed in wall unit: y + =yC /. . . . . .
Physical sizes of the streamwise vortices together with the corresponding grid discretization expressed in millimeter. The number of elements
estimation is done for a velocity reference C1 and for a velocity 5 times
smaller, C1 /5, and 25 times smaller, C1 /25. The physical domain, consisting in the volume of fluid between two guide vanes, is represented by
a rectangular box of 100x75x35 mm. The total number of elements is obtained assuming that the pump-turbine can be represented by 87 similar
volumes of fluid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

Investigation of the time spent in the five main parts composing the
computation of the flow field as well as the relative performance of the
code with regards to the peak performance of the machine, which is 13.6
GFLOPS for a computing node. . . . . . . . . . . . . . . . . . . . . . .
FFB characteristics on BlueGene/P. . . . . . . . . . . . . . . . . . . . .
Blue Gene/P specification. . . . . . . . . . . . . . . . . . . . . . . . . .
PRIMEHPC FX10 and K supercomputers specifications. . . . . . . . .

26
26
29
29

Specifications of the HYDRODYNA pump-turbine scale model. . . . .

Number of nodes and elements for each component of the pump-turbine.
Number of layers and number of elements for each overset. The third
column gives the ratio expressed in % of the overset mesh size regarding
the component mesh size, which include the overset. . . . . . . . . . . .
Summary of the computation parameters. . . . . . . . . . . . . . . . . .
Volume of each component of the pump-turbine scale model with an estimation of the number of impeller revolutions for all initial fluid particles
to leave each component. Q and T respectively represent the pumpturbine flow rate and the impeller period and are equal, respectively, to
2.304102 m3 /s and 0.4 s. . . . . . . . . . . . . . . . . . . . . . . . . .
Surface discretization for comparison with LDV measurements. The physical location of these surfaces is shown in Fig. 5.2. . . . . . . . . . . . .
Comparison of the experimental and numerical values for the flow coefficient , the energy coefficient and the propagation speed s . . . . . .

EPFL - Laboratory for Hydraulic Machines

8
10

33
40

44
46

47
57
61

Notations

Latin

b
fn
g
k
p
s
t
x, y, z
y+
z
A
C
C
Cm
Cn
Co
Cr
Cs
Ct
Cu
C

channel width
rotational frequency
gravitational acceleration: g '9.81
turbulent kinetic energy
pressure
curvilinear coordinate
time
Cartesian coordinates
non-dimensional wall distance: y + = C y
number of blades, number of channels
surface area
absolute velocity
bulk velocity
meridional velocity component
velocity component normal to surface
characteristic velocity
radial velocity component
Smagorinsky constant
velocity component tangential to surface
circumferential velocityq
component
friction velocity: C = w

[m]
[s1 ]
[ms2 ]
[m2 s2 ]
[Pa]
[-]
[s]
[m]
[-]
[-]
[m2 ]
[ms1 ]
[ms1 ]
[ms1 ]
[ms1 ]
[ms1 ]
[ms1 ]
[-]
1
[ms ]
[ms1 ]
[ms1 ]

E
H
L
Lo
N
Q
R
U

pump specific hydraulic energy: E = gHI gHI

Head: H = E/g
length
characteristic length
rotational speed
flow rate
radius
peripheral velocity

[Jkg1 ]
[m]
[m]
[m]
[min1 ]
[m3 s1 ]
[m]
[ms1 ]

EPFL - Laboratory for Hydraulic Machines

xx

LIST OF TABLES

Greek

absolute flow angle

relative flow angle
turbulent dissipation rate
stress tensor
angular coordinate
viscosity
kinematic viscosity (=/)
stall phase
density
impeller angular velocity
specific dissipation rate
stall propagation velocity

[-]
[-]
2 3
[m s ]
[Pa]
[-]
[Pas]
[m2 s1 ]
[-]
[kgm3 ]
[rads1 ]
[s1 ]
[rads1 ]

Subscripts
1
2
e
n
o
r
ref
s
t
u
w

reference to the circumference formed by the outer impeller diameter

reference to the circumference formed by the guide vane leading
edge
reference to the shroud or element
normal to section
reference for non-dimensional numbers or nominal value
radial
reference value
stall
throat
peripheral
wall
shear

Superscripts

b
bb
eb

normalized quantity
fluctuating or modeled quantity
resolved quantity
test-filtered quantity
low-pass filtered quantity
EPFL - Laboratory for Hydraulic Machines

LIST OF TABLES

xxi

Dimensionless Numbers
cf =

cf

skin friction coefficient

Cp

pressure coefficient

Re

Reynolds number

Re =

Co Lo

flow rate coefficient (Pump-turbine)

Q
R3

specific energy coefficient

2E
2 R2

Abbreviations
AOA
EPFL
BEP
BPF
CFD
CPU
DDR
DFT
DNS
FFB
GV
LDV
LES
LMH
MIC
RANS
RSI
SGS
SP
SV

Angle Of Attack

Ecole
polytechnique federale de Lausanne
Best Efficiency Point
Blade Passing Frequency
Computational Fluid Dynamics
Central Processing Units
Double Data Rate
Discrete Fourier Transform
Direct Numerical Simulation
FrontFlow/blue
Guide Vane
Laser Doppler Velocimetry
Large Eddy Simulation
Laboratoire de Machines Hydrauliques
Mass Imbalance Correction
Reynolds Averaged Navier-Stokes
Rotor Stator Interaction
Subgrid Scale
Sampling Point
Stay Vane

EPFL - Laboratory for Hydraulic Machines

Cp =

1
2
U1e
2

p pref
1
2
U1e
2

Part I
Introduction

EPFL - Laboratory for Hydraulic Machines

Chapter 1
Problem Overview
1.1

Pumped-storage Power Plants

During these past decades, the energy demand has increased drastically requiring to
improve the existing power stations and to find new sources of energy. Indeed, the world
consumption of energy has increased by 180% during these past 30 years and has reached
in 2010 150 quadrillion (1015 ) Watt hour, see Fig. 1.1 a). Besides, the environmental
concerns, regarding the present usage of the energy resources, push to reduce or even to
abandon the use of some existing non-renewable resources such as nuclear, coal, oil or natural gas. However, a glance at Fig. 1.1 b) clearly shows the energetics dependence on the
oil, coal and gas resources, as they are by far the main energetics contribution. Therefore,
to reduce this energetics dependence, new resources with high electrical capacity should
be found and hydropower is one of them. The hydropower generation can provide a large
quantity of electricity and at the same time it has a lower impact on the environment,
which makes it very popular nowadays. Furthermore, thanks to different technologies, a
hydropower plant can be built in different places with different geographical topology to
generate electricity: along a river, thanks to the high flow rate, or in mountainous area,
thanks to the difference of height.
As mentioned above, the configuration of a hydropower plant is not unique and different approaches exist. The common approach is to use a dam to store and accumulate
water. In terms of energy, this consists in storing water as potential energy. In a lower
altitude than the dam, one or several water turbines are connected to the dam through a
penstock. Released from the dam, the potential energy of the water is converted to kinematic energy, which is finally converted into electrical energy through the water turbine
connected to a generator. Therefore, the difference of altitude and the flow rate passing
through the water turbine determine the amount of power that can be generated. The
drawback of this approach is that the available resources depend on the volume of water
retained from the dam. Another approach is quite similar as the previous one as it also
uses a dam, but with a low or no reservoir capacity. Such technology is known as runof-the-river and it uses the flow rate of the river to generate electricity. The advantage
of this approach is that it produces an almost constant amount of electricity provided
by a constant river flow. There are other approaches that exist and which are under
development, such as the ones using the tide (tidal power), the ocean currents (marine
current power) or some small river (micro hydro).
EPFL - Laboratory for Hydraulic Machines

1.1. PUMPED-STORAGE POWER PLANTS

150

Source: www.eia.gov

50

a)

Source: www.bp.com

b)

100

50

PWh (1015 Watt hour )

PWh (1015 Watt hour )

40
North America
Europe
Asia & Oceania
Eurasia
Middle East
Center and South America
Africa
World

30
20

Nuclear
Hydro
Oil
Gas
Coal
Renewable

10
0
1980

1985

1990

1995
year

2000

2005

2010

1970

1980

1990
year

2000

2010

Figure 1.1: Evolution of the energy consumption in the world: a) per region, b) per
resources.
Another energy issue encountered nowadays concerns the stability of the grids used to
transport the electricity. On the one hand, these present grids are used near their limits
and the variation between the supply and the demand induces electrical fluctuation on
the grid, which cannot be avoided. Indeed, typical examples of the drop of the supply
arise when the clouds decrease the radiant light and heat from the sun or when the
intensity of the wind decreases and makes the wind turbines stop. On the other hand,
the main variation of the demand is identifiable by the industry operating time, as well
as by the short and large residential demand early in the morning and around noon.
Therefore, it is primordial to have the technology able to respond to these variations of
the supply and the demand. This means that the technology has to provide energy when
the supply is lower than the demand, and that the technology has to store the energy
when the demand is lower than the supply. In that context, the pumped-storage power
plants are the best technology. Indeed, only a few minutes are required to start the
machine and to synchronise it with the grid. Furthermore, by adjusting the flow rate, it is
possible to control precisely the electrical power generated, making that technology very
attractive. To respond to the supply excess, pumps are installed to pump water from a
lower reservoir to an upper reservoir. Such pumped water will be used later to generate
electricity when the demand exceeds the supply. Finally, another reason that makes this
approach attractive is that it is economically viable. The price of the kilowatt hour (kWh)
during the day is not constant and it fluctuates regarding the demand. Therefore, the
pumped-storage power plants generate electricity during the peak of demand and sell the
electricity when the price of the kWh is high. Inversely, the plants buy the electricity
to pump the water when, as already mentioned, the demand is lower than the supply
and, therefore, the price of the kWh is low. Originally, such a plant was equipped with a
water turbine, a pump and an electrical engine or generator, and it runs the engine or the
generator depending on whether the plant is in pumping or generating mode. However,
for economical reason, it is advantageous to merge the water turbine and the pump which
results in the so called reversible pump-turbine.
The reversible pump-turbines are devoted to a specific implementation and are, therefore, designed for a fixed flow rate Q and a fixed net hydraulic head H. However, the
EPFL - Laboratory for Hydraulic Machines

Chapter 1. Problem Overview

present requirements make the pump-turbines to be used at off-design conditions. Therefore, there is a need to expand the operating range of such a machine and at the same
time to ensure the safety of the machine for that range. So far, two main problems have
been reported: at full load condition (Q > QBEP ) cavitation may occur, and at part load
condition (Q < QBEP ) the rotating stall phenomenon may occur. These phenomena have
to be studied because they can have a dramatic issue on the machine. In a weak occurrence of these phenomena, noise and vibration are perceptible and the efficiency of the
machine decreases. However, if the intensity of these phenomena are high, it can result in
severe damage to the machine, which in the worst case, can destroy the machine within a
few hours of operation. Therefore, there is a necessity to understand these phenomena in
order to expand the operating range of the machine. So far, the rotating stall is still an
unclear phenomenon and needs to be more detailed, especially for the hydraulic machine
for which almost no study exists.

1.2

Rotating Stall

The rotating stall is a generic name to describe a particular phenomenon appearing in

a rotating machine. This phenomenon can be seen as a local modification of the flow by
the presence of a single or multiple recirculation zones. As these zones are located near
walls and they imply a reverse flow, such zones are often referred to as stall or as flow
separation [29]. Besides, depending on the location of the recirculation zones, the stall
phenomenon can be referred to as forward/backward rotating stall, alternate blade stall or
asymmetric stall.
The problem of the presence of the so called rotating or alternate stall phenomenon in
a centrifugal pump is now well known. However, the onset and the mechanism of this phenomenon are still poorly understood. The first time that this phenomenon was reported
was by Emmons in 1959 when he explained the phenomena using cascade theory, [14],
and was intensively studied numerically and experimentally in the case of compressors,
[18], [9]. However, much longer time was required to get similar studies for centrifugal
pump experiencing rotating stall. Indeed, one of the first experimental studies on radial machine was performed by Krain, [33], and Inoue et al., [23], where focus was put on
impeller blade pressure and flow field investigation in centrifugal impeller for radial turbomachinery. Deeper investigations followed with the studies of the rotor stator interaction
performed by Arndt et al., [2], [1]. They showed that the pressure fluctuation appears
to be larger on the vane suction side near the trailing edge and that the magnitude of
the pressure fluctuation is strongly dependent on the radial gap (distance between the
impeller discharge and the diffuser vane leading edge). The importance of such gap will
be later emphasized by Sano et al., [51].
Finally, the first experimental investigation of the so called rotating stall in centrifugal
pump took place only during these last ten years. Indeed, Sinha et al., [54], [55], [56] were
the first to investigate this phenomenon in a vaned diffuser of a centrifugal pump with the
Particle Image Velocimetry (PIV). In the same time, new details on the importance of the
narrow clearance and its impact on the rotating stall were demonstrated by Sano et al.,
[51]. The clearance, also refered to as gap, is the distance between the trailing edge of the
impeller blade and the leading edge of the diffuser vane. Finally, almost all experimental
studies were performed with the PIV method to investigate the structure of the flow or
EPFL - Laboratory for Hydraulic Machines

1.3. HIGH PERFORMANCE COMPUTING

the influence of the flow rate on the rotating stall, [64], [45], [34]. Numerical research on
the rotating stall phenomenon started once again for an axial type vaned diffuser coupled
with a centrifugal impeller performed by Torbergsen [59]. This simulation was performed
using the Reynolds Averaged Navier Stokes (RANS) method. Another study was carried
out by Sano, [52] where the interest was on the onset of the rotating stall and to assess the
accuracy of the computation by comparison with their experimental data [51]. Finally,
3D numerical simulation was performed by Braun, [3] for a complete centrifugal pump.
Recently, other simulations on the rotating stall were performed by Lucius [38] in order
to compare the performance of the SAS and the SST turbulence models.
Concerning the Large Eddy Simulation (LES) method the first simulation was performed by Kato, [27] in the case of a mixed-flow pump and the first LES simulation in a
centrifugal pump with the interest in the rotating stall was performed by Byskov, [4].

1.3

High Performance Computing

The term High Performance Computing (HPC) refers to the use of supercomputers or
clusters of computers to solve difficult computational problems. Concretely, it consists in
assembling a large amount of processors to increase the size of the memory and to decrease
the operating time by using a large number of cores. Looking at the last fifty years, one
can see that a drastic change has occurred. Indeed, although the first supercomputer
featured only a few number of cores, the latest supercomputers feature up to million of
cores. This change comes from the capacity to reduce the size of the transistors, to make
them faster and to assemble them in a way to produce massively parallel computers.
Interestingly, this change in the computing power was approximately predicted in the
1960s by Gordon Moore, who is the designer of one of the first supercomputers. At that
time, while working on integrated circuit, which is the basis for computers, he made the
estimation that the number of transistors per square inch on integrated circuits would
roughly double every year [42]. However, in 1975, this estimation had to be changed and
the new estimation moved from 12 to 24 months [43]. Following this previous estimation,
we can see that it got verified for the evolution of the computing power. Figure 1.2 shows
the computing power evolution from 1971 to 2011 in a logarithmic representation. The
straight line corresponds to the trend for doubling the number of transistors every two
years, which corresponds well to the computing power evolution. To achieve this power
evolution, the size of the transistor was constantly scaled down following the Dennard
scaling down to 130 nm [11]. This allowed to increase the power density and meanwhile
to increase the performance. However, to decrease the transistor size below 130 nm,
a new approach had to be found as the Dennard scaling was not sufficient enough to
increase the performance. To reach from the 90 nm transistors to the 32 nm transistors,
work was done on the material, which permitted the size decrease with the performance
increase. However, the transistor size will not become infinitely small and, therefore, one
can wonder what will be the size limit of a transistor?
Since 1993, a project named Top500 1 has compiled twice a year (June and November) a list indexing the 500 fastest supercomputers based on the standardized LINPACK
benchmark developed by Jack Dongarra [13]. Basically, the LINPACK benchmark solves a
1. http://www.top500.org/

EPFL - Laboratory for Hydraulic Machines

Chapter 1. Problem Overview

7
16-core SPARC T3
Six-core i7
10-core Xeon Westmere-EX
Six-core Xeon 7400
8-core POWER7
Quad-core z196
Dual-core Itanium 2
Quad-core Itanium Tukwila
AMD K10
POWER6
8-core Xeon Nehalem-EX
Itanium 2 with 9MB cache
Six-core Opteron 2400
AMD K10
Core i7 (Quad)
Core 2 Duo
Cell
Itanium 2

2,600,000,000
1,000,000,000

Transistor count

100,000,000

AMD K8

The curve shows transistor

count doubling every
two years

10,000,000

Pentium 4

Pentium

Barton

Atom

AMD K7
AMD K6-III
AMD K6
Pentium III
Pentium II
AMD K5

80486

1,000,000
80386
80286

100,000

68000
8086

10,000
2,300

8080

8085
6800
Z80

80186
8088
6809

8008
MOS 6502
4004
RCA 1802

Source: Wiki Commons

1971

1980

1990

2000

2011

Date of introduction

Figure 1.2: Microprocessor transistor counts from 1971 to 2011. The straight line shows
the Moores law.
dense n by n system of linear equations Ax=b to measure the number of floating point operations per second. The authors of these lists are Hans Meuer (University of Mannheim,
Germany), Erich Strohmaier (currently at Lawrence Berkeley National University, but
previously at the University of Mannheim), Jack Dongarra (University of Tennessee and
ORNL) and since 2000, Horst Simon (Lawrence Berkeley National University). These
lists also permit to index the location of the supercomputers and to describe the kind
of applications for which a supercomputer is used. Table 1.1 gives the top 10 supercomputer elaborated in November 2013. As it can be seen, the fastest computer (Tianhe-2)
is located in China and features 3.12 millions of cores. Furthermore, Japans K computer was the first in November 2011 to achieve a performance level of 10 PFLOPS (1015
calculations per second). Since then, all the effort is done to reach the EFLOPS (1018
calculations per second) level by 2020. Last december, RIKEN (in Japanese: Rikagaku
Kenkyuujo and in English: Institute of Physical and Chemical Research) was selected to
develop an exascale supercomputer, which should start operating by 2020.

1.4
1.4.1

Numerical Simulation Specification

RANS, LES and DNS Tradeoffs

Since the development of computers and CFD software, numerical simulation has been
extensively used, both in academia and industries. Such an approach has the advantage
that it can be a cheaper approach compared to the experiment and can, for a specific case
study, provide more accurate and more detailed data. However, this approach also faces
EPFL - Laboratory for Hydraulic Machines

1.4. NUMERICAL SIMULATION SPECIFICATION

Table 1.1: Supercomputer top 10 of the Top500 list, as of November 2013.

Rank

Country

Name

Manufacturer

1
2
3
4
5
6
7
8
9
10

China
United States
United States
Japan
United States
Switzerland
United States
Germany
United States
Germany

Tianhe-2
Titan
Sequoia
K computer
Mira
Piz Daint
Stampede
JUQUEEN
Vulcan
SuperMUC

NUDT
Cray Inc.
IBM
Fujitsu
IBM
Cray Inc
Dell
IBM
IBM
IBM

Nb of cores
3,120,000
560,640
1,572,864
705,024
786,432
115,984
462,462
458,752
393,216
147,456

Linpack Perf.
TFLOPS
33,862.7
17,590.0
17,173.2
10,510.0
8,586.7
6,271.0
5,168.1
5,008.9
4,293.3
2,897.0

its inherent limitations, which induce different ways in performing numerical simulation
known as Reynolds Averaged Navier-Stokes (RANS), Large Eddy Simulation (LES) and
Direct Numerical Simulation (DNS).
DNS is the most constraining approach as it explicitly computes all the turbulence
scales using the unsteady Navier-Stokes equations, but it provides the most accurate
solution as long as the space-time discretization is sufficiently fine. To estimate the spatial
discretization, one can use the work performed by Kolmogorov [31], [47], which provides
an estimation on the ratio of the biggest to the smallest eddies present in a flow with regard
to the Reynolds number. Using such an estimation, the space discretization requirement
is of the order L/=O(Re3/4 ), where L characterizes the largest scales and the smallest
ones. Re is the Reynolds number, which expresses the ratio of the inertial force to the
viscous force. The definition of the Reynolds number is as follows: Re=C0 L0 /, where C0
is a characteristic velocity, L0 a characteristic length and the kinematic viscosity. For 3D
cases, such space requirement becomes O(Re9/4 ) in terms of number of the computational
grids, which is considerable. Knowing that the typical range of the Reynolds number for
industrial applications is 106 -108 , it implies that the number of the computational grids is
of the order of 1013 -1018 , which is not feasible today and neither is for the next decades.
Furthermore, accurate computation also requires a fine temporal discretization, which
makes DNS a difficult approach. Therefore, DNS is mainly used for computation of a low
Re flow with a simple geometry.
RANS can be seen as the opposite of DNS as it is the most economical and the most
popular approach for engineering applications. RANS approach computes flows using
the Navier-Stokes equation, where the pressure and the velocity are decomposed into the
statistical average quantity and the fluctuating quantity. Therefore, RANS computation
provides only statistical average solution, and no information regarding the fluctuation
is available. The fluctuation term arising from the decomposition is thus unknown and
has to be modeled. Furthermore, such a fluctuation is flow dependent and is therefore
complicated to model. Since the above decomposition was introduced by O. Reynolds,
many models were developed aiming at reproducing correctly the influence of the fluctuation on the statistical average quantities. Nowadays, one of the most popular RANS
turbulence models used is the one developed by Menter [40],[41], called k SST. This
EPFL - Laboratory for Hydraulic Machines

Chapter 1. Problem Overview

model combines the two known models for k , [24], [35] and k , [62], [63], in order to
take advantage of both models, and gives satisfactory results most of the time. However,
regarding the phenomena of interest, it was shown by the study of O. Braun [3] that the
RANS approach with the k SST model faces difficulties in accurately simulating the
rotating stall. Therefore, another approach should be considered.
LES stands between the RANS and DNS approaches. As DNS does, LES explicitly
computes the different scales, but only to a specific size dictated by a filter width. Similarly to RANS, the influence of the small scales on the big ones is modeled through the
so called subgrid scale (SGS) model. The challenge with this approach is to find the appropriate balance between time-space resolution and accuracy, or in other words to have
a sufficiently fine grid resolution in order to dissipate all the energy produced. The turbulent production arises from the large scales, whereas the dissipation occurs in the small
scales. Therefore, if the grid is too coarse to accurately represent the small scales, the
energy accumulates within the flow leading to nonphysical solutions. As the LES grid is
not fine enough to resolve all the turbulence scales, the task of the subgrid scale model is
to mimic the effect of the scales smaller than the filter width on the scales bigger than the
filter width. However, such a modeling is not trivial because it requires the knowledge of
the turbulence structures, which is flow dependant. Furthermore, another main difficulty
encountered by the LES approach is the computation of boundary layers. This requires
a fine discretization near walls in order to represent accurately the boundary layer up
to the viscous sublayer, increasing drastically the need in spatial resolution [46]. These
resolution requirements are thus the limiting reason to use such an approach. However,
the constant increase of the computing power during these past decades makes the LES
approach more and more feasible. Meanwhile, the work performed on the development
of numerical codes in order to have a good scalability to accommodate tens of thousand
cores also increased the opportunity to perform LES computation. The use of such an
amount of cores provides a large amount of available memory with a relatively weak core
load, resulting in computations that can be carried out in a relatively short amount of
time. Nevertheless, despite all these improvements some industrial investigations can not
be numerically performed yet. Therefore, a feasibility study on the mesh size and the
computing resources requirements is mandatory.
The process of the feasibility study can be decomposed into three parts. The first part
consists in a short analysis to estimate what are the physics involved in the research of
interest. Such estimation is primordial, as it indicates the size of the smallest structures
present in the flow to be computed. The second part consists in estimating the size of
the spatial discretization requirement with regards to the small scales to be computed.
Finally, the third part consists in the estimation of the computing resources requirement
to perform the computation of interest. In the following, a feasibility study is carried out
to determine the requirement to compute the rotating stall in a full pump-turbine scaled
model by the LES approach.

1.4.2

The smallest eddies to be computed

The present research focuses on the rotating stall appearing in a pump-turbine being
operated in pumping mode and at part load condition (76% BEP). The rotating stall
phenomena is an instability recognizable by the presence of stall cells in the pump-turbine
diffuser, which propagate, at a low ratio of the impeller frequency, from one diffuser
EPFL - Laboratory for Hydraulic Machines

10

1.4. NUMERICAL SIMULATION SPECIFICATION

Table 1.2: Sizes of streamwise vortices expressed in wall unit: y + =yC /.

Length
300
Diameter
30
Spanwise spacing 150
channel to the neighboring diffuser channel in the same direction as the impeller rotation.
Their presence is the result of part load condition, which induces a misalignment of
the incoming flow to the diffuser guide vanes and induces favorable condition for flow
separation. The physics to be computed involve mainly the flow exiting the impeller
and the flow around the guide vanes, which can be regarded as a hydrofoil. Because the
phenomena of interest is unsteady, the time and spatial evolutions of the boundary layer
around these hydrofoils have to be computed. Regarding the Reynolds number considered,
the nature of the boundary layer is turbulent. Based on extensive former researches
[21],[30],[48],[53], it is known that a turbulent boundary layer can be characterized by
the presence of coherent structures, which are long streamwise vortices. For the present
research, it is assumed that capturing such structures should be sufficient to compute the
phenomena of interest, that is, the rotating stall. The physical size of such vortices can
be estimated using their dimensionless size, see Table 1.2, and such a physical size will be
used to estimate the spatial discretization requirement.

1.4.3

Mesh Requirements

To estimate the physical size of the streamwise vortices, the viscous length l= /C ,
which links the dimensionless wall unit to the dimension one, has to be determined.
However, the complexity of such an evaluation is that the friction velocity, which is
unknown and is dependent on time and space, also has to be determined. Based on the flat
plate boundary layer theory, an approximation of the friction velocity can be computed
using the following formulas developed initially by Chapman [5], and a correction for
high Reynolds number suggested by Choi et al. [6]. Such formulas relate the mean flow
velocity to the friction velocity using the wall skin friction coefficient cf and the Reynolds
number. The variation of the wall skin friction coefficient with regards to the Reynolds
number is depicted in Fig. 1.3.
r
cf
C
=
C
2
cf = 0.0577Re1/5
for Rex 106
(1.1)
x
1/7
6
9
cf = 0.027Rex
for 10 Rex 10
However, the Reynolds number is dependent on the choice of a characteristic length and a
characteristic velocity, which can vary significantly. As the velocity of the flow field in the
diffuser is driven by the impeller revolution, it is assumed that the absolute velocity C1 at
the impeller outlet is the characteristic velocity. Furthermore, the flow in the diffuser will
first evolve by passing around the guide vanes. Then, the characteristic length selected is
the guide vane chord length, which results in a Reynolds number of 1.9 millions.
For a Reynolds number of 1.9 million, the ratio between the friction velocity and the
absolute velocity is 4% giving a viscous length of 1.24 m. The expected dimensions
EPFL - Laboratory for Hydraulic Machines

Chapter 1. Problem Overview

11

0.014

cf

0.0577Re x

0.012

0.027Re

1
7
x

0.010
0.008
0.006
0.004
0.002
0.000
3
10

10

10

10

10

10

Re x

10

Figure 1.3: Flat plate wall skin friction coefficient as a function of the Reynolds number
in mm of these streamwise vortices are thus given in Table 1.3. To estimate the total
number of elements, it is assumed that the volume of fluid between two adjacent diffuser
guide vanes can be represented by a rectangular box with the following dimension: 100 x
75 x 35 mm representing respectively the streamwise, transverse and spanwise directions.
Furthermore, to estimate the total number of elements for the full pump-turbine, it is
assumed that the pump-turbine has 87 similar volumes of fluid: 3 for each impeller channel, 1 for each guide vane channel and 2 for each stay vane channel. Finally, to compute
the number of elements per direction, the following spatial discretization was selected:
x+ =50, y+ = z+ =20. This estimation has resulted in that the present computation will
need 600 billions of elements. This is of course not feasible and the following change was
adopted: based on the Reynolds-number similarity, the Reynolds number was reduced by
a factor 5 in order to reach a realistic spatial discretization. The new required mesh is
approximately 7 billions, which is nowadays feasible.
Table 1.3: Physical sizes of the streamwise vortices together with the corresponding grid
discretization expressed in millimeter. The number of elements estimation is done for
a velocity reference C1 and for a velocity 5 times smaller, C1 /5, and 25 times smaller,
C1 /25. The physical domain, consisting in the volume of fluid between two guide vanes,
is represented by a rectangular box of 100x75x35 mm. The total number of elements is
obtained assuming that the pump-turbine can be represented by 87 similar volumes of
fluid.
C1
C1 /5 C1 /25
Vortex length [mm]
0.37
1.6
7.43
Vortex diameter [mm]
0.037
0.16
0.74
Vortex spacing [mm]
0.19
0.8
3.7
Streamwise grid resolution [mm]
0.06
0.27
1.24
Transverse and spanwise grid resolution [mm] 0.025
0.11
0.5
9
6
Number of elements per passage [-]
710
8010
1106
9
9
Number of elements for the pump-turbine [-] 60010
710
80106

EPFL - Laboratory for Hydraulic Machines

12

1.4.4

1.5. THESIS OBJECTIVE

Computing Resources Requirement

Since the estimation on the mesh size is carried out, one can estimate the computing
resources in terms of memory size, amount of computing nodes and resources allocation requirement. To perform this estimation, some characteristics of the CFD software and the
computation are necessary. In the present research, computation will be performed using
FrontFlow/blue (FFB) software developed at the University of Tokyo. This software has
already been successfully ported to several supercomputer architectures, and time measurements have shown that the sustained peak performance of the code is approximately
5% of the peak performance up to several ten thousands cores, that it requires about
20k floating point operation (FLOP) per element per time step and that the appropriate
load balancing is 0.1 million elements per processor core. As introduced previously, the
estimated size of the mesh is 7 billion elements. Furthermore, the time discretization
is 10,000 time steps per impeller revolution and the number of impeller revolutions to
compute to perform statistics is 40.
The computation will be performed on the supercomputer K, belonging to RIKEN.
This supercomputer features 8 cores per computing node with a total number of computing
nodes over 80,000 for a peak performance of 128 GFLOPS per computing node. Using the
above information, the computation requires 10,000 computing nodes and one impeller
revolution lasts approximately 7 hours. This results in the computation of 40 impeller
revolutions within 12 days.
In conclusion, the feasibility study showed that such a computation is nowadays possible and that thanks to important resources, results can be obtained in a reasonable
time.

1.5

Thesis Objective

Nowadays, the rotating stall is still a common issue when pumps are operated at part
load condition. To better understand this phenomenon and to predict its impact for
a specific machine, experimental or numerical research is usually performed. However,
the location of the instability in the machine makes the experimental study difficult.
Therefore, the numerical approach is often preferred, as the study of internal flows is
easier. So far, numerical investigations were mainly done using the RANS approach or a
resolved LES approach. Because of the constraining requirements of LES, some studies
simplified the geometry by using symmetry assumption. Such simplification reduces the
size of the computational mesh and make the study possible. However, as the rotating stall
phenomenon modifies the flow in the upstream area of the diffuser in the circumferential
direction, the flow field in the impeller may be affected and, therefore, the use of symmetry
assumption should be avoided. From the knowledge of the author, there is presently
no LES computation performed to study the rotating stall phenomenon without using
any geometrical simplification. Therefore, thanks to new computational opportunities,
the objective of this research is to perform a resolved LES, by resolving the streamwise
vortices, and to demonstrate that the rotating stall can be accurately computed by the
LES approach without the use of geometric symmetry. The selected case study is the
HYDRODYNA pump-turbine, where the rotating stall phenomenon was experimentally
observed [3], when operating the pump at 76% of the design point. This experimental
EPFL - Laboratory for Hydraulic Machines

Chapter 1. Problem Overview

13

research will be used to assess the accuracy of the computation, which will allow us to
demonstrate the mechanism of the rotating stall occurring in the present pump-turbine.

1.5.1

Document Organization

The thesis has been divided into four parts:

Part I regroups the introducing chapters regarding the rotating stall phenomenon.
It shows that the increasing needs to extend the operating range of a pump-turbine
may lead to the appearance of the rotating stall phenomenon. Meanwhile, the
increase of the computing power offers the possibility to study numerically this
phenomenon using the LES approach, which has been often seen as an unfeasible
approach for industrial studies. A feasibility study is then carried out and assesses
that the LES approach is suitable to compute the flow field of the complete scale
model pump-turbine.
Part II is dedicated to the numerical investigation methodology. It introduces the
LES approach with the turbulence model used. Then the different numerical tools
required for this research are described: the numerical code, the computing resources
and the visualization software. Part II concludes with the presentation of the pumpturbine characteristics, its spatial discretization and the flow initialization.
Part III presents the results of the unsteady computation of the pump-turbine part
load flow. First the computation is compared to the experimental data, and secondly
the propagation mechanism of the rotating stall is discussed.
Part IV draws the concluding remarks.

EPFL - Laboratory for Hydraulic Machines

Part II
Methodology

EPFL - Laboratory for Hydraulic Machines

Chapter 2
Large Eddy Simulation
2.1

Large and Small Scales Separation

The LES approach relies on the separation of the large and small scales from the solution. This separation is performed by applying a spatial filter to the governing equations
of the fluid motion and is illustrated in Fig. 2.1, where a schematic representation of
the kinetic energy spectrum is split in two parts. The left part represents the energy
contained in the large scales, whereas the right part represents the energy contained in
the small scales. The scale size at which the separation takes place depends on the filter
width, which is a parameter when using an explicit filter or is the computational grid size
when using an implicit filter. The LES approach computes thus explicitly all the scales
larger than the filter width and computes the effect of the small ones through the use
of a subgrid scale model. The filtering process and the subgrid modeling are introduced
respectively in the following two sections.

2.2

Low-Pass Filtered Incompressible Navier-Stokes

Equations

The governing equations of an incompressible fluid arise from two conservation laws,
which are the momentum (based on Newtons second law) and mass conservation laws.
The momentum equation expresses the link between the acceleration of a fluid particle to
the surface forces and body forces experienced by the fluid. In the case where the body
force is gravity, the force can be expressed through the gravitational potential , defined
as =gZ for a constant gravitational force. Therefore, the body force per unit mass is
defined as g=. Using the previous definition for the body force, the two laws can
be mathematically written as follow:
DCi
=

Dt

Ci
Ci
+ Cj
t
xj


=

ij

xj
xi

+
(Ci ) = 0
t xi
EPFL - Laboratory for Hydraulic Machines

(2.1)

(2.2)

18

2.2. LOW-PASS FILTERED INCOMPRESSIBLE NAVIER-STOKES

EQUATIONS

log(E(k))
-5/3

filter width
Resolved scales

Modeled scales
kc

log(k)

Figure 2.1: Schematic view of the energy spectrum showing the scales separation.
where ij is the stress tensor, Ci components of the absolute flow velocity and the density.
In the case of an incompressible Newtonian fluid, the stress tensor can be expresses as:


Ci Cj
+
(2.3)
ij = P ij +
xj
xi
where P is the pressure, ij the Kroneckers delta and the viscosity. In the case of an
incompressible fluid, Eq. (2.2) reduces to Ci /xi =0 and the stress can be seen as the
sum of the isotropic tensor (P ij ) and the deviatoric contributions by using the last
expression and assuming that the density is time and space independent, we obtain the
Navier-Stokes (NS) equations for an incompressible fluid:


2 Ci

Ci
Ci
P
+

(2.4)

+ Cj
=
t
xj
xi
xj xj
xi
Defining a modified pressure p, as
p = P +

(2.5)

the governing equations of a Newtonian incompressible flow becomes:

Ci
Ci
1 p
2 Ci
+ Cj
=
+
t
xj
xi
xj xj

(2.6)

Ci
=0
xi

(2.7)

where is the kinematic viscosity of the fluid.

The modified pressure shows that the effect of the isotropic stress and the body forces
have the same effect. Hence, the body force has no effect on the velocity field and on
the modified pressure. Henceforth, the modified pressure p may be simply referred to as
pressure [47]. Besides, Eqs. (2.6) and (2.7) are commonly expressed in the dimensionless
form using the following dimensionless variables xi = xi /L, Ci = Ci /C, t = t/(L/C)
and p = p/(C 2 ), where L and C are respectively a characteristic length and velocity.
EPFL - Laboratory for Hydraulic Machines

Chapter 2. Large Eddy Simulation

19

This dimensionless form clearly highlights the importance of the non-linear term for flows
with high Reynolds number.

Ci
p
1 2 Ci
Ci
=

+
+
C
j
t
xj
xi
Re xj xj

(2.8)

Ci
=0
xi

(2.9)

The LES approach relies on a filtering process to damp out the turbulence scales
smaller than the filter width. Therefore, every instantaneous quantity is decomposed into
a resolved quantity (f ) and a modeled quantity (f 0 ):
Ci (x, t) = C i (x, t) + Ci0 (x, t)

p(x, t) = p(x, t) + p0 (x, t)

(2.10)

By applying these decompositions to the NS equations, one obtains the incompressible

low-pass filtered NS equations:
 



C i
C i C j
1 p
+
+
+
C iC j =

Bij
(2.11)
t
xj
xi xj
xj
xi
where Bij represents the subgrid-scale (SGS) tensor, which can be expressed as:
Bij = Ci Cj C i C j

(2.12)

Using the statistical property, one can rewrite the SGS as:


Bij = C i C j C i C j + C i Cj0 + C j Ci0 + Ci0 Cj0
{z
} |
|
{z
} | {z }
L

(2.13)

where L represents the Leonard term, C the cross term and R the Reynolds-stress
like term. The Leonard term is the only one which can be calculated explicitly. Depending on the case studied it could be better to take out the Leonard term from the
previous equation, which led to a new kind of SGS. However, in most SGS model this
decomposition is not done and the SGS tensor Bij is modeled as a whole.

2.3

Subgrid-scale Modeling

As introduced previously, the filtering process separates the large (resolved) and the
small (modeled) scales and generates new unknowns gathered into the subgrid-scale tensor
Bij , which has to be modeled. The process of this modeling is first to represent the structure of the small scales of motion and secondly to represent the interactions between the
modeled and the resolved scales. As LES is performed using a coarse computational mesh,
the filter width is obviously located somewhere in between the largest and the smallest
scales and deals with the inter-scales flow structures. It is evident that no universal tensor
can be provided, as the physical behavior of the inter-scales and the small scales are flow
dependent. Therefore, two strategies are adopted to incorporate the action of the small
scales: functional modeling and structural modeling [50]. Looking at Eq. (2.11), one can
EPFL - Laboratory for Hydraulic Machines

20

2.3. SUBGRID-SCALE MODELING

see that the subgrid scale tensor in the NS equations is in the form B. Hence, one can
model the term B as a whole or model the tensor B itself. The first approach is used
for the functional modeling, whereas the second one is used for the structural modeling.
Since the modeling process requires the knowledge of the flow structure, the study case
of the fully developed isotropic homogeneous turbulence was extensively used, as it is the
only case that can be accessible by theoretical analysis. Therefore, the initial modeling
were only dedicated for the computation of isotropic flow. The DNS research dedicated
to find the location where the subgrid energy transfer occurs and what correlation exists
between the different scales, permitted to show that the interaction between the small and
the large scales is driven by two mechanisms. The first, which is the dominant mechanism,
is assimilate to an energy drainage from the large scales to the small ones and is referred to
as forward energy cascade phenomenon. The second mechanism is a weak energy feedback
to the large scales and is referred to as backward energy cascade [50]. Furthermore, it
was shown that the cascade mechanisms are associated to specific features of the velocity
and vorticity field in the physical space [8], [28].
Functional modeling
The functional modeling in the physical space explicitly model the forward energy
cascade mechanism to the subgrid scales. It relies on the hypothesis that the transfer
mechanism from the resolved to subgrid scales is analogous to the molecular mechanisms
represented by the diffusion term, in which the viscosity appears. Therefore, the mathematical form of the subgrid scale model, as Boussinesq proposed, is written as:
B d = sgs C + T C

(2.14)

where B d is the deviatoric of B defined as:

1
Bijd = Bij Bkk ij
3

(2.15)

The spherical tensor 13 Bkk ij is added to the filtered static pressure and does not need
to be modeled. Such

decomposition is required, because for incompressible flows the
T
tensor C + C has a zero trace and thus the model has to have also a zero trace.
Therefore, the model reduces in determining a relation of the form:
sgs = F (C)

(2.16)

Several examples of functional models can be found in [50] and the one from Smagorinsky is introduced later in section 2.3.1.
Structural modeling
The structural modeling in opposite to the functional modeling does not rely on the
knowledge of the nature of the interaction between the large and the small scales, but
aims to approximate directly the subgrid tensor Bij from the resolved velocity field or
from a formal series expansion [50].
EPFL - Laboratory for Hydraulic Machines

Chapter 2. Large Eddy Simulation

2.3.1

21

Smagorinsky Model

The Smagorinsky model [57] is a subgrid viscosity model of the form of Eq. (2.16),
which relies on the large scales. This model is expressed as:

2

1/2
(2.17)
sgs (x, t) = Cs 2|S(x, t)|2
where Cs is the Smagorinsky constant, the filter width and S the resolved strain rate.
The two main drawbacks of this model is firstly that it depends on the constant Cs that
needs to be adjusted and secondly that it depends on the strain rate S of the large scales.
Therefore, as long as the flow exhibits spatial variations, the subgrid viscosity is non-zero
even if all scales are resolved as for example with a laminar flow. Such a model should
thus be used for fully turbulent and under-resolved flows. Regarding the constant, it
should be tuned in order that the ensemble-average subgrid kinetic energy dissipation
rate is equal to , with  being the kinetic energy dissipation rate. The first LES using the
Smagorinsky model was done by Deardorff [10], where he used Cs =0.1 for a plane channel
flow. Another study by Clark et al. [7] used a Cs =0.2 for an isotropic homogeneous
turbulence computation. Several other studies suggested a constant between 0.1 and 0.12
[39], [60], [44]. Therefore, no universal constant is available. An approach to avoid this
constant is to evaluate its value locally in space and time with regards to the local flow
field. This approach is introduced in the following.
Dynamic Smagorinsky Model
The dynamic Smagorinsky model is similar to the Smagorinsky model except that the
coefficient Cs is no longer constant but changes in space and time during the simulation.
The purpose is to dynamically adapt this coefficient to the state of the flow to increase
the accuracy of the solution. The first dynamic approach to evaluate the Smagorinsky
coefficient was presented by Germano et al. [16]. This approach is subject to a source of
singularity, which has led to a modification of the Germano subgrid-scale closure method
by Lilly [37]. These two dynamic models rely on a second filtering of the Navier-Stokes
b is often set
equations. This second filter is usually called test filter and its width ()
twice the original filter width. The idea from this two filtering is known as the Germano
identity, which shows that two tensors corresponding to two different filtering levels can
be related by an exact relation.
The first tensor, is simply the subgrid tensor from the initial filtering, with the
Smagorinsky model:
1
2
(2.18)
Bij ij Bkk = 2Cs |S|S ij
3
The second tensor, is obtained by applying the test filter to the filtered equations:
1
bS
b
b 2 |S|
Tij ij Tkk = 2Cs
(2.19)
ij
3
Germano showed on the following identity that the consistency between the two previous tensors depends on the choice of Cs . Hence, the relation known as the Germano
identity:
b b
bij = Cd
Lij = Tij B
iC j + C iC j
EPFL - Laboratory for Hydraulic Machines

(2.20)

22

2.3. SUBGRID-SCALE MODELING

The tensor Lij represents the stress components of the scale motion between the test
scale and the grid scale. These scales are referred to as the test window and can be
explicitly evaluated. Furthermore, this evaluation can be compared to the Smagorinsky
closure approximation by subtracting (2.18), where the test filter was applied from (2.19):
1
Lij ij Lkk = 2Cs Mij
3

(2.21)

where
d 2 |S|
bS
b ,
Mij = |S|S
ij
ij

=2

(2.22)

Equation (2.21) provides the value for Cs to insert in Eq. (2.18), but Eq. (2.21)
represents five independent equations for a single unknown. Therefore Cs cannot be
uniquely determined and an error minimization process is selected by using a least squares
approach. Cs is thus obtained by the following equation:
Cs =

1 hLij Mij i
2 hMij Mij i

(2.23)

where
 1 

d
b
b
d
b
b
Lij = C i C j C i C j ij C k C k C k C k
3


(2.24)

Equation (2.23) differs in two manners from the original Germano model introduced
below:
Cs =

1 Lij Sij
2 Mij Sij

(2.25)

The first is that Germano multiplied Eq. (2.21) by Sij to obtain the above equation.
The issue is that the term Mij Sij can vanish or becomes very small leading to unstable
value of Cs . The second difference is the averaging process mandatory to ensure the
computational stability. This avoids negative coefficient and prevents it from changing
sharply.

EPFL - Laboratory for Hydraulic Machines

Chapter 3
Numerical Tools
3.1
3.1.1

FrontFlow/blue
Numerical Code Structure

FrontFlow/Blue 1 is an open source Overset Finite Element code developed by the

University of Tokyo, under the supervision of Professor Chisachi Kato, featuring Direct Numerical Simulation, DNS, Large Eddy Simulation, LES, and Reynolds Averaged
Navier-Stokes, RANS. For LES computations, the static Smagorinsky turbulence model
[57] as well as the dynamic version developed by Germano [16], where the Smagorinsky
constant is determined locally in time and space using the modification of Lilly [37] are
implemented. The computation of incompressible flows is performed using the fractional
step method to solve the pressure Poisson equation with the Crank-Nicolson implicit time
integration scheme. The resulting linear systems of equations are solved by the Biconjugate Gradient Stabilized method (Bi-CGSTAB) [61] incorporated with the residual cutting
method [58]. Furthermore, the spatial and temporal discretization are both second-order
accurate. To be able to handle large computational mesh, FFB is parallelized using the
Message Passing Interface (MPI) library [17] for inter-domains communication and uses
the METIS library [25] to decompose the computational domain into several sub-domains.
The code structure of the flow computation is given in Fig. 3.1. FFB has also a multiple
frame of reference implemented and uses an overset grid as interface. Regarding the grid,
the solver version of FFB used for the present computations only computes flow using
hexahedral element.

3.1.2

Numerical Code Performance on BlueGene/P

The performance of the numerical code on the supercomputer BlueGene/P has been
investigated using the Hardware Performance Monitoring library provided by IBM. Such a
library allows one to monitor, among others, the memory usage, the number of FLOP, the
quantity of data read or written of the whole or parts of the numerical code. To monitor
what parts of the code are the time consuming ones, the numerical code to compute the
flow field for one time step can be decomposed as follow:
Initial flow field development;
1. http://www.ciss.iis.u-tokyo.ac.jp/english/dl/

EPFL - Laboratory for Hydraulic Machines

24

3.1. FRONTFLOW/BLUE

Computation Initialization

Foregoing Computation

Calculate fractional velocity

FALSE

RES EPS
OR
NITR NMAX

TRUE

Solve pressure equation

FALSE

RES EPS
OR
NITR NMAX

TRUE

Velocity Field Correction

Subsequent computation

FALSE

DIVMAX DIVESC
OR
ITIME NTIME

TRUE

Computation Finalization

Figure 3.1: Flow computation algorithm.

Overset grid compilation for parallel mode;

Wall shear stress computation;
Smagorinsky constant computation;
Effective viscosity computation;
Velocity computation;
EPFL - Laboratory for Hydraulic Machines

Chapter 3. Numerical Tools

25

Pressure computation;
Velocity field correction;
Subsequent computations.
As the primary interest is in the computation of the velocity and the pressure fields,
the analysis of the performance of the computation of the flow development, the overset
grid computation, the wall shear stress computation, the element Smagorinsky constant
computation and the element effective viscosity computation are merged into one analysis
block called foregoing computation. The computation of the velocity, the pressure, the
velocity correction and the subsequent computations compose each of the analysis block.
The subsequent computations involve miscellaneous sub-block such as: the computation
of the force acting on a specified body, the computation of the flow statistics and to save
the data. As for the initial computation, these sub-blocks are merged into a single block.
It results that the numerical code for the flow computation is decomposed into five distinct
parts.
The analysis of the sustained performance of the numerical code is carried out by
performing LES computations with the dynamic Smagorinsky turbulent model for the
lid driven cavity flow [32]. The Reynolds number is 1,000 and the computational grid is
decomposed into 512 subdomains with 0.1 million elements per subdomain. A subdomain
is computed by a core. The lid driven cavity flow benchmark corresponds to a confined
flow in a squared box with the lid moving at constant velocity in its plane. As the number
of iterations required for the inner iterative solvers to converge changes, the convergence
tolerance were set low enough to always impose 5 iterations for the velocity computation
and 50 iterations for the pressure computation in this benchmark test. The computation
of the pressure field requires more iterations than for the velocity field as the convergence
rate is lower. The computation was performed for 100 time steps as it is estimated to be
sufficient to obtain a statistically good representation of the time spent and the number
of FLOP performed in each of these 5 blocks. The result of such a measurement is shown
in Fig. 3.2 illustrating the time required for each block.
Foregoing computations: 9%

Calculate fractional velocity: 22%

Subsequent computations: 3%
Velocity field correction: <1%

Solve pressure equation: 65%

Figure 3.2: Pie chart showing the normalized time spent in each main part composing
the computation of one time step.
As expected, the most time consuming block is the computation of the pressure,
because 50 iterations were imposed whereas the number of iterations imposed for the
EPFL - Laboratory for Hydraulic Machines

26

3.1. FRONTFLOW/BLUE

Table 3.1: Investigation of the time spent in the five main parts composing the computation of the flow field as well as the relative performance of the code with regards to the
peak performance of the machine, which is 13.6 GFLOPS for a computing node.

Elapsed
time [s]
% of peak
performance

foregoing
computation
50.836

Velocity
computation
152.484

Pressure
computation
404.524

Velocity
correction
3.306

Subsequent
computations
12.427

5.69

5.82

5.34

4.82

1.01

computation of the velocity field was only 5 iterations as previously mentioned. However,
such numbers of iterations are typical values for incompressible flow computation and
only for the velocity computation a decrease to 2-3 iterations may be expected. During
the computation, the number of FLOPS in each part were recorded, which permitted to
compute the sustained performance presented in Table 3.1. As it can be seen, the relative
performance is near 6% to the peak. This result is corresponding to the performances
of other numerical codes optimized on massively parallel supercomputer and it is found
satisfactory. Finally, this monitoring also permitted to evaluate the characteristics of
FFB for the BlueGene/P, in term of the number of FLOP per element per time step,
the memory requirement per element and the average sustained performance. These
characteristics are given in Table 3.2.
To evaluate the behavior of FFB with regard to the number of cores used, a weak
scaling analysis is performed using the lid driven cavity benchmark. The weak-scaling
analysis increases the size of the computational domain in proportion to the number of
cores so as to set a constant load for each core. This weak-scaling is performed for 1 to
8,192 cores (2,048 computing nodes) for the typical load of 0.1 million elements per core,
requiring 74% of the main memory of one core. The averaged elapsed total time per a
time step required for each computation is given in Figure 3.3; total elapsed time being
defined as the sum of the computational time and the communication time. It should be
noticed that with a half of the machine it is possible to have a computational domain near
one billion elements. Furthermore, we can see that the total elapsed time per a time step
slightly increases with respect to the number of cores. However, this increase is only 1.2
seconds when comparing the results for 1 and 8,192 cores which can be considered totally
acceptable from a practical point of view.

3.1.3

Stationary and Rotating Interface

The interface between the rotating and the stationary meshes are done using an overset
mesh [26]. As illustrated in Fig. 3.4, the idea is that each mesh includes an appropriate

Table 3.2: FFB characteristics on BlueGene/P.

Number of FLOP per Memory requirement
element per time step
per element
22,600
7.4kB

Sustained
performance
5.3%

EPFL - Laboratory for Hydraulic Machines

Chapter 3. Numerical Tools

27

10
t / time step
9
8
7
6
5
4
3
2
1
8192

4096

2048

1024

512

256

128

64

32

16

Number of cores

Figure 3.3: FFB7 weak-scaling benchmark up to 8,192 cores for a load per a core of 0.1
million elements on BlueGene/P.
margin, which overlaps with its neighboring mesh upstream and downstream. At each
time step, the velocity components and the static pressure in the overlap region are
interpolated from the neighboring mesh with element tri-linear interpolation function
[19]. For the specific case of turbomachinery, a dual frame of reference is used to compute
flows in the stationary and the rotating mesh. Therefore, when the velocity field between
the rotating and the stationary meshes are overset, a transformation is applied in order to
take into account the difference of the two different frames of reference. In the case of the
fractional-step method, where the velocity and the pressure are computed successively,
the velocity field is first overset. Secondly, the pressure gradient is overset in the inner
iteration of matrix solver poisson equation. Finally, the static pressure is overset [19].
As the mass balance through the interface is not exactly guaranteed due to the interpolation in the overset interface, it may lead to nonphysical oscillation of the pressure
field. To handle this issue, two approaches are used and implemented. The first approach
is to accept small mass imbalance at the overset by using the low Mach number assumption. This assumption let an incompressible flow have a small compressibility, which then
become compatible with the small mass imbalance. The second approach is to impose the
mass balance at the overset by using the mass imbalance correction (hereafter denoted
as MIC) [65]. Therefore, with the inconsistency generated by the mass imbalance being
removed, the pressure solver does not exhibit nonphysical oscillation [19]. Furthermore,
it is worth to mention that the overset does not affect the pressure field.

3.2
3.2.1

Computing Resources
IBM BladeCenter

The first type of computing architecture used is the local cluster implemented in the
Laboratory for Hydraulic Machines, composed of two IBM BladeCenter [12]. The first
EPFL - Laboratory for Hydraulic Machines

28

3.2. COMPUTING RESOURCES

Computational Domains
Interface
Grid 1
6x11 elements

Grid 2
3x5 elements

Overset from grid 2 on grid 1

3x1 elements

Overset from grid 1 on grid 2

6x3 elements

Figure 3.4: Overset scheme of two neighboring meshes.

BladeCenter is composed of 8 computing nodes, each of which has a bi-processor quad
core Intel Xeon CPU X5570 (Gainstown) operated at 2.93 GHz (64 cores in total). The
memory cache is 8 MB of L3 and each node processes 32 GB of DDR3 memory operated
at 1333 MHz.
The second BladeCenter is composed of 12 computing nodes, each of which has a
bi-processor quad core Intel Xeon CPU E5450 (Harpertown) operated at 3 GHz (96 cores
in total). The memory cache is 12 MB of L2 and each node processes 8 GB of DDR3
memory operated at 1066 MHz.
The interconnection is composed of a Gigabit Ethernet, that is, a bandwidth up to 1
Gbps.

3.2.2

Blue Gene/P

The second type of computing architecture used is the CADMOS IBM Blue Gene/P
supercomputer [22]. This supercomputer is a specific machine designed for massively
parallel simulations, featuring a total of 16,384 cores. A unit of such a machine is called
a rack. The packaging hierarchy of a BlueGene/P from the rack size to a single core is
as follow. A rack accommodates two Midplanes containing 16 node cards. These node
cards consist in the assembly of 32 nodes containing two CPUs (two processors) with two
cores. That results in a total of 4 cores per computing node, i.e. a total of 4,096 cores per
rack. One core corresponds to a PowerPC 450, running at 850 MHz and thanks to the
dual-pipeline FPU, floating-point unit, it can simultaneously execute two fused multiplyadd instructions per a machine cycle. This gives a peak performance of 4 floating-point
operations (FLOP) per a machine cycle per core. Thus, the peak performance of a node is
13.6 billion floating-point operations pre second (GFLOPS). The CADMOS BlueGene/P
is composed of 4 racks providing a peak performance of 56 TFLOPS with 16 TB of
memory available.
The computing nodes are interconnected by a 3D torus network with a node-to-node
bandwidth of 2 times (bidirectional) 3.4 GB/s in each direction.
EPFL - Laboratory for Hydraulic Machines

Chapter 3. Numerical Tools

29

Table 3.3: Blue Gene/P specification.

CPU Name
Cores/Node
Performance
Cache
Node configuration
Memory capacity
Number of nodes/rack
Number of rack
Rack peak performance

3.2.3

BG/P
PowerPC 450
4 cores (@0.85GHz)
13.6 GFLOPS
2xL3 4MB
1 CPU/Node
4 GB
1,024
4
14 TFLOPS

FX10 - K Supercomputer

The third and fourth types of computing architectures are the PRIMEHPC FX10
and the K supercomputers. These supercomputers share the same architecture except
that the PRIMEHPC FX10 is equipped with the new CPU generation of SPARC64 IXfx,
whereas the K supercomputer is equipped with the SPARC64 VIIIfx. The SPARC64
IXfx is composed of 16 cores running at 1.848 GHz with 12MB of L2 cache, whereas
the SPARC64 VIIIfx is composed of 8 cores running at 2 GHz with 6MB of L2 cache.
Furthermore, these two CPU can perform 8 FLOP per machine cycle [15]. Regarding the
packaging, the approach is similar to the one previously introduced. A rack is composed
of 24 system boards, each with 4 nodes. Therefore, a rack features 96 nodes. The peak
performance of a rack is 22.7 TFLOPS for the PRIMEHPC FX10 supercomputer and
12.3 TFLOPS for the K supercomputer.
The computing nodes are interconnected by a 6D torus network with a node-to-node
bandwidth of 2 times (bidirectional) 5 GB/s in each direction.

Table 3.4: PRIMEHPC FX10 and K supercomputers specifications.

FX10
CPU Name
SPARC64 IXfx
Cores/Node
16 cores (@1.848GHz)
Performance
23.5 GFLOPS
Cache
L2 12MB
Node configuration
1 CPU/Node
Memory capacity
32 GB
Number of nodes/rack 96
Number of rack
50
Rack peak performance 22.7 TFLOPS

EPFL - Laboratory for Hydraulic Machines

K
SPARC64 VIIIfx
8 cores (@2GHz)
128 GFLOPS
L2 6MB
1 CPU/Node
16 GB
96
864
12.3 TFLOPS

30

3.3
3.3.1

3.3. VISIT

VisIt
File Format

VisIt is an open source software developed by the Lawrence Livermore National Laboratory (LLNL) 2 to interactively visualize and analyse scientific data in serial or parallel
mode. Furthermore, VisIt offers the advantage to leverage both the use of a remote parallel computer and the use of a local workstation, which has graphics acceleration hardware.
Besides, such an approach is more natural since in most cases the scientific data are stored
on a remote machine. Therefore, there is no need to transfer the data for analyses.
Regarding the file format, VisIt has a wide range of readers implemented to read data
coming from different database. The default file format name of VisIt is Silo. Silo is a Clanguage library with a well-defined application programming interface (API) allowing the
user to write all kind of data to a Silo file. The Silo library is architecturally divided into
an upper-level API and a lower-level I/O implementation called driver. There are several
choices regarding these drivers, but the two main ones are the HDF5 (Hierarchical Data
Format 5) and the PDB (Portable Data Base) drivers. For the utility program converting
the data from a FFB database to a Silo database, the PDB driver was selected. The
advantages using this file format are that it allows to store the data in binary format and
that it is platform independent.

3.3.2

Parallelized Visualization

As the computational mesh increases, the data to load for visualization increases
accordingly. Therefore, the visualization using a single CPU is no longer feasible and
requires a parallel visualization to distribute the data among several CPUs. This is easily
achieved with VisIt by compiling a parallel executable. Regarding how to handle the data
to load in VisIt, two approaches may be used.

Figure 3.5: Illustration of the half Pump-turbine domain decomposed into 512 subdomains.
2. https://wci.llnl.gov/codes/visit/

EPFL - Laboratory for Hydraulic Machines

Chapter 3. Numerical Tools

31

The first consists in converting each partitioned flow data from the simulation to a Silo
file format and to generate a specific file, called master file. The master file is generated
using also the Silo library and works like an assembly file. It means that VisIt will
know which data to read in in order to visualize the full computational domain. Such an
assembly is illustrated in Fig. 3.5, where the computational domain of the pump-turbine
is decomposed into 512 subdomains. Furthermore, for scalability purposes, this master
file contains the spatial and the variables extents, which allows, when plotting a slice, an
iso-surface or anything else to load only the subdomains containing the data to visualize.
Therefore, avoiding to load all the data to retrieve only the one of interest.
The second approach consists in merging the partitioned flow data to built a single
file containing the flow data of the full computational domain. This merge is nevertheless
sometimes not feasible, when the data size is too large. But, on the other hand this
approach may be appropriate when the data are analyzed on a machine with several
CPUs and a sufficient amount of RAM memory. In that case, it is preferable to reduce
the number of files to the minimum, as the demand for the file I/O degrades drastically
the efficiency of VisIt.

EPFL - Laboratory for Hydraulic Machines

Chapter 4
HYDRODYNA Pump-Turbine
4.1

Pump-Turbine Characteristics

The reduced scale model HYDRODYNA pump-turbine is depicted in Fig. 4.1 with
its specifications given in Table 4.1.
Table 4.1: Specifications of the HYDRODYNA pump-turbine scale model.

zb
zo
D1e
D1e
b
gv
R2e R1e
(R2e R1e )/R1e

Specific speed
Number of impeller blades
Number of guide vanes
Outer impeller diameter
Inner impeller diameter
Guide vane width
Guide vane opening angle
Impeller-diffuser radial gap
Impeller-diffuser relative gap

0.19
9
20
523.5 mm
250 mm
36 mm
20
25 mm
9.5%

Two measurement campaigns were carried out in 2006 and 2008 to establish the pumpturbine performances in pumping and generating modes. For the first campaign, an
instrumented impeller was used with 30 pressure sensors mounted. For the second campaign, a plain impeller was used to avoid any damage to the instrumented impeller during
off-design investigations in generating mode. The resulting performance measurements
are shown in Fig. 4.2, where the specific energy and the scaled efficiency are plotted
against the discharge. It can be seen that the specific energy curves do not exhibit the
expected monotonic behavior. Near a discharge coefficient of =0.021 and =0.026 the
slopes become positive, which indicates the presence of instabilities within the machine.
For this specific machine it was shown that this instability is associated with the rotating
stall phenomenon.

4.1.1

Former Research on the Pump-Turbine

The internal flow of the HYDRODYNA pump-turbine reduced scaled model has been
investigated, both in generating and pumping modes. This pump-turbine was installed in
one of the three EPFL-LMH hydraulic machines testing facilities. Such a test facility is
EPFL - Laboratory for Hydraulic Machines

34

4.1. PUMP-TURBINE CHARACTERISTICS

Volute

Impeller

Pump
mode

Stay vanes
Guide vanes

Cone

Draft tube

Figure 4.1: Pump-turbine main components.

illustrated in Fig. 4.3 showing the PF2 test rig. The maximum specific energy achievable
is E=1,250 Jkg1 , the maximum rotational velocity is N =2,500 rpm and the maximum
discharge is Q=1.4 m3 s1 . The relative error of the discharge, the shaft torque and
the differential pressure measurements are within 0.2%. The flow within the closed loop
is driven by a single circulating pump coupled to a 1,000 kW electrical motor and the
pump-turbine is coupled to a 300 kW electrical generator.

Figure 4.2: Specific energy and efficiency scaled by the efficiency at BEP versus discharge
of the HYDRODYNA pump-turbine. Instrumented impeller (2006) and plain impeller
(2008), N =600 rpm and N =900 rpm. Data obtained from [3]

EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

35

Figure 4.3: EPFL LMH PF2 test rig.

Four former studies were dedicated to study different phenomena related to this specific
pump-turbine. One study was dedicated to the investigation of the pump-turbine in
pumping mode and the other three studies are dedicated to the investigations of the
pump-turbine in generating mode.
The first research was performed by Dr. A. Zobeiri [66]. He studied numerically and
experimentally the Rotor Stator Interaction (RSI) in generating mode for three different
conditions: part load, best efficiency and full load. This research identified the origin of
the high pressure fluctuations and their impact on the structure of the different component
of the machine.
The second research was performed by Dr. O. Braun [3]. He studied experimentally
and numerically the internal flow in the diffuser in pumping mode at part load condition.
This research quantitatively and qualitatively described the rotating stall occurring in the
vaned diffuser.
The third research was performed by Dr. Hasmatuchi [20]. He studied experimentally
and numerically the internal flow field in generating mode. This research describes how
the flow field evolves within the impeller when operating the pump in the so-called Sshape.
The fourth research was performed by Dr. Roth [49]. He studied experimentally the
high periodic excitation due to the RSI on the pump-turbine guide vanes. This research
elaborates a model to predict the dynamic behaviour of the entire guide vane cascade.

4.2
4.2.1

Flow Domain Discretization

Process

The spatial discretization of the pump-turbine was performed using the commercial
software ANSYS ICEM CFD (ANalysis SYStem; Integrated Computer-aided Engineering
and Manufacturing; Computational Fluid Dynamics). This software is very powerful as
it gives a means to fully control the build of the mesh down to the very small detail. This
EPFL - Laboratory for Hydraulic Machines

36

4.2. FLOW DOMAIN DISCRETIZATION

a)

b)

Figure 4.4: a) Wire-frame representation of the cone geometry. The cyan curves represent
the physical limits of the cone and the grey curves represent the overset limits. The dashed
grey lines represent the outer limits of the overset. b) Cone blocking. Using the rotational
symmetry only 1/9 of the volume is blocked.
full control is achieved by the generation of the so called blocking. This blocking can
be seen as a decomposition of the geometrical volume (i.e, here the volume of water in
the machine) into multiple smaller volumes, called block. These blocks are primordial,
as they, on one side, represent the geometry and on the other side represent the mesh
topology. To represent the geometry, each block has to be associated to the geometry.
Typically, a block is represented by vertices, edges and faces, and they are associated
with the geometry represented by points, curves and surfaces. Furthermore, this process
of generating small blocks and mapping them to the geometry, gives the liberty to choose
the kind of mesh topology, e.g., C-mesh, H-mesh or O-mesh. Once the topology is mapped,
it remains to set the number of nodes on edges as well as the nodes distributions along
each edge. This is very powerful, as it allows to have, for example, a high node density
near walls and decrease the node density away from the wall, reducing the total amount
of the nodes. Once all these operations are preformed, the computational mesh can be
generated and exported.

4.2.2

Blocking

As shortly introduced previously, the mesh generation requires several steps, which
can be time consuming and lead to extremely complex blocking. In order to simplify
the blocking process, the geometry of the full pump-turbine is first decomposed into four
distinctive parts: the cone, the impeller, the diffuser and the volute. Each of these parts
are meshed separately and are assembled together using the overset capability introduced
in section 3.1.3.

EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

37

Cone
The cone is the upstream most component of the pump-turbine in pumping mode
as it connects the inflow to the impeller. Its wire-frame geometry is depicted in Fig.
4.4 a). The physical limits of the cone correspond to the cyan curves, whereas the grey
curves represent the overset limits. The overset surfaces here are the interface between the
stationary cone and the rotating impeller. As the geometry has a rotational symmetry, the
total volume to mesh could be simplified and only 1/9 of the cone was manually meshed.
The corresponding blocking is depicted in Fig. 4.4 b). The resulting mesh features 6.7
million elements and 7 million nodes (including the overset).
Impeller
The impeller is the second (rotating) component of the pump-turbine. Its wire-frame
geometry is depicted in Fig. 4.5 a). The physical limits of the impeller correspond to the
cyan curves, whereas the grey curves represent the overset limits. The overset surfaces here
are the interface between the stationary cone and the rotating impeller and the interface
between the stationary diffuser and the rotating impeller. Since the impeller possess a
rotational symmetry, only one impeller channel is blocked and is depicted in Fig. 4.5
b). The resulting mesh features 41 million elements and 42 million nodes (including the
overset).

a)

b)

Figure 4.5: a) Wire-frame representation of the impeller geometry. The cyan curves
represent the physical limits of the impeller and the grey curves represent the overset
limits. The dashed grey lines represent the outer limits of the overset. b) Impeller single
channel blocking.

Diffuser
The diffuser is the third component of the pump-turbine. As known, such a component
has no rotational symmetry because of the presence of the tongue. Downstream of the
tongue, which is in the counterclockwise direction, see Fig. 4.1, there are a smaller stay
EPFL - Laboratory for Hydraulic Machines

38

4.2. FLOW DOMAIN DISCRETIZATION

vane and the tip part of the volute, both of which contribute to the non-symmetry of the
diffuser. The choice to integrate the tip part of the volute into the diffuser part was made
firstly due to the fact that the mesh has to be fully composed of hexahedral elements. Secondly, without this integration, the blocking in the tip of the volute would be catastrophic
and lead to totaly deteriorated elements, which is not affordable. Consequently, in order
to simplify the diffuser blocking, three diffuser channels were isolated as shown in Fig.
4.7 a), which include the three non-symmetric elements (tongue, smaller stay vane and
volute tip). For the remaining 17 diffuser channels, the rotational symmetry was used,
resulting in the blocking of only one single channel as shown in Fig. 4.6 b). In this case,
the assembly of the two meshes will not be performed via an overset interface. Thus,
the blocking and the node distribution has to be identical between those two meshes.
The areas in common are depicted in Fig. 4.6 b) and Fig. 4.7 b) by red surfaces. As
previously mentioned, the physical limits of the diffuser surfaces correspond to the cyan
curves, whereas the grey curves represent the overset surfaces limits. The overset surfaces
here are the interface between the the rotating impeller and the stationary diffuser and
the interface between the stationary diffuser and the stationary volute.
After the generation and the merge of the two meshes, the resulting diffuser mesh
features 73 million elements and 76 million nodes (including the overset).

a)

b)

Figure 4.6: a) Wire-frame representation of the diffuser channel geometry (symmetric

part). The cyan curves represent the physical limits of the diffuser channel and the grey
curves represent the overset limits. The dashed grey lines represent the outer limits of
the overset. b) Diffuser single channel blocking.

Volute
The last component of the pump-turbine is the volute, whose symmetrical wire-frame
geometry is depicted in Fig. 4.8 a). The physical limit of the volute correspond to the
cyan curves, whereas the grey curves represent the overset limits. The overset surfaces
EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

39

a)

b)

Figure 4.7: a) Wire-frame representation of the diffuser channels geometry around the
tongue (nonsymmetric part). The cyan curves represent the physical limits of the diffuser
and the grey curves represent the overset limits. The dashed grey lines represent the outer
limits of the overset. b) 3 nonsymmetric diffuser channels blocking.
here are the interface between the stationary diffuser and the stationary volute. As the
geometry has a plane symmetry, the total volume to mesh can be simplified and only half
the volute was manually meshed. The corresponding blocking is depicted in Fig. 4.8 b).
The resulting volute mesh features 7.8 million elements and 8.1 million nodes (including
the overset).

a)

b)

Figure 4.8: a) Wire-frame representation of the volute geometry. The cyan curves represent the physical limits of the volute and the grey curves represent the overset limits.
The dashed grey lines represent the outer limits of the overset. b) Volute blocking.

EPFL - Laboratory for Hydraulic Machines

40

4.2. FLOW DOMAIN DISCRETIZATION

Table 4.2: Number of nodes and elements for each component of the pump-turbine.
Pump-Turbine Part
Cone
Impeller
Diffuser
Volute
Total

Number of nodes
106
7.00
42.45
76.15
8.11
133.73

Number of elements
106
6.76
41.44
73.36
7.82
129.40

Ratio of the global mesh

%
5.23
32.02
56.70
6.05

Pump-turbine
Table 4.2 summarizes the size of the mesh for each different component. The total
number of the elements reached 129 million elements, which fits the 130 millions estimated
in the previous chapter. The generated mesh is shown in Fig. 4.9.
Mesh Quality
There exist several criteria to check the quality of a mesh and ICEM offers more than
35 criteria. However, it is estimated that only three criteria are sufficient to quantify the
mesh quality. These criteria are the equiangle skewness, the volume change and the aspect
ratio. The equiangle skewness is a measure regarding how an element is distorted. The
volume change indicates how the volume evolves between an element and its neighboring
elements. Finally, the aspect ratio measures how elongated an element is.

Figure 4.9: Spatial discretization of the pump-turbine.

EPFL - Laboratory for Hydraulic Machines

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Chapter 4. HYDRODYNA Pump-Turbine

35
30
25
20

Equiangle Skewness
min ave
Cone
0.42 0.81
Impeller 0.34 0.79
Diffuser 0.16 0.83
Volute 0.27 0.85

41

max
0.99
0.99
0.99
0.99

15
10
5
0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Figure 4.10: Mesh quality quantification using the equiangle skewness criterion. The bars
give the criteria distribution for each mesh in percentage.
Equiangle Skewness
The equiangle skewness criterion takes a value between 0 and 1 where 0 represents a
completely flat hexahedral element and 1 represents the perfect element with all angles
equal to 90 degrees. The definition of the equiangle skewness criterion is as follow:
skew = 1.0 max((Qmax Qe )/(180 Qe ), (Qe Qmin )/Qe )
where Qmax is the largest angle between two neighboring sides in the element, Qmin is
the smallest angle in the element and Qe is the angle of an equiangular element (e.g., 90
degrees for a square).
Figure 4.10 shows the distributions of the equiangle skewness criterion for the different
pump-turbine components. The ordinate gives the number of elements, expressed in
percentage regarding the total number of elements in the individual component. As it
can be seen, the majority of the elements possess a value higher than 0.5. This tendency
is confirmed by high averaged values, which are approximately 0.8, confirming the good
quality of the meshes. It can also be noted that the criterion distribution for the impeller
does not follow the same tendency as the three other meshes. This can be explained by
the characteristic complexity of the impeller geometry with high curvatures. Because the
mesh has to fit closely the geometry, it induces mesh curvatures in both the streamwise
and the spanwise direction, explaining the above mentioned difference.
The worst equiangle skewness qualities are located in the diffuser and the volute components. These components also possess complex geometry to be discretized, consisting
mainly in sharp edges. The consequences are that the blocking has to be built in accordance to these sharp edges, inducing non regular blocking with bad shaped elements.
Volume Change
This quality metric is calculated for the individual element by finding the maximum
volume of its neighboring elements and dividing it by the volume of the element itself. In
EPFL - Laboratory for Hydraulic Machines

42

4.2. FLOW DOMAIN DISCRETIZATION

an area with important flow structure this metric should not exceed 1.2. Elsewhere where
the flow does not change significantly, this metric could be relaxed to values higher than
1.2. However, it is preferable to keep this metric within the range 1.1-1.2.
Figure 4.11 shows the distribution of the volume change criterion for each mesh. For
visualization purpose, only the range between 1 to 2 is shown as it contained the majority
of the elements. Near walls, the expansion ratio is set to 1.1, whereas it is set to 1.2
elsewhere. However, it is not possible to meet the prescribed expansion ratio because it
resulted in the increase of the number of nodes, which was sometimes not affordable. This
constraint explains why the average values are slightly above 1.2.
The maxima shows that in some cases the expansion ratio reaches 25 or even 37, which
should be avoided. Because all the settings are set manually, some of them may have been
missed and may explain such a high ratio.

70

Volume Change
min ave
Cone
0.98 1.26
Impeller 0.97 1.27
Diffuser 0.97 1.27
Volute 0.96 1.30

60
50
40

max
1.86
8.42
25.8
37.0

30
20
10
0
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0

Figure 4.11: Mesh quality quantification using the volume change criterion. The bars give
the criteria distributions for each mesh in percentage.

Aspect Ratio

70
60
50
40
30
20
10

For hexahedral elements, the aspect ratio is defined as the size of the maximum element
edge divided by the size of the minimum element edge.
Figure 4.12 depicts the distributions of the aspect ratio for each mesh in the range
1-20. Except the volute which contains the majority of its elements in the range 1-3, other
components contain only aboutCone
30% of their elements in this range. This is confirmed by
Impeller
an average value for the volute, which is 2-3 times smaller than in the other components.
Diffuser
One reason is that, for the volute
component, the discretization near walls to capture the
Volute
boundary layer was not taken into account. A mesh refinement was still carried out near
walls in this component, but it is not as fine as in other components.
The necessity to refine the mesh near the wall always induces high aspect ratio elements
as it is not affordable to have an equivalent size in the wall normal direction, spanwise
and streamwise directions. For this reason, due to the particular discretization of the
EPFL - Laboratory for Hydraulic Machines

0
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0

Chapter 4. HYDRODYNA Pump-Turbine

43

boundary layer, elements with high aspect ratio up to 100 are present in the impeller and
diffuser components.
60

Aspect Ratio
min ave
Cone
1.00 4.28
Impeller 1.00 7.00
Diffuser 1.00 4.89
Volute 1.00 2.33

%
50
40

max
25.9
98.5
43.3
32.8

30
20
10
0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Figure 4.12: Mesh quality quantification using the aspect ratio criterion. The bars give
the criteria distributions for each mesh in percentage.

4.2.3

60
%
50
40
30
20
10
0

Overset Discretization

As presented in section 3.1.3, the interface between meshes is done using an oversetting
approach. Thus, for each mesh to be connected to another mesh, the geometry was
modified in order to incorporate the geometry of the neighboring component to build
the topology of the overset volume. Figures 4.4 a), 4.5 a), 4.6 a) and 4.8 a) (see pages
36-39) showed the result of such expansion by the grey surfaces. Table 4.3 gives the size
of each overset volume in term of number of elements. In the present mesh, there are
three interfaces, which are the interface between the cone and the impeller, between the
impeller and the diffuser and finally between the diffuser and the volute. Since the overset
process is bidirectional, it results that for each interface, two overset area are required,
and hence Table 4.3 has six entries. As it can be seen, each overset volume are composed
of several layers between six and ten. For an optimal use of the overset approach, it is
often required to set a ten layers overset. However, because of geometrical constraint,
e.g. the presence of a wall, it was
not always possible to extend the overset as far as it
Cone
should be. Furthermore, another Impeller
determining constraint is the size of each element. One
Diffuser
example is the overset between the diffuser and the volute. The interface between the
two components is located near Volute
the trailing edge of the stay vanes, where the mesh is
fine in order to capture the physics near the wall. However, the volute exhibits the larger
volume to be discretized, thus large elements are required in order for the number of
elements to be acceptable. It results then, that the two overset area will present different
element density and that according to the limitation in elements volume change, it will not
be possible to achieve ten elements within the overset volume. This constraint explains
the difference in mesh density for the interfaces between the cone and the impeller and
between the diffuser and the volute. In both cases, the ratio is about one-third.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

EPFL - Laboratory for Hydraulic Machines

44

4.3. COMPUTATION PARAMETERS & FLOW INITIALIZATION

Table 4.3: Number of layers and number of elements for each overset. The third column
gives the ratio expressed in % of the overset mesh size regarding the component mesh
size, which include the overset.
Pump-Turbine Component
Cone on Impeller
Impeller on Cone
Impeller on Diffuser
Diffuser on Impeller
Diffuser on Volute
Volute on Diffuser
Total

Number of layers
10
9
7
7
6
6
-

Number of elements
414,000
1,595,000
1,341,000
1,710,000
1,502,000
547,000
7,112,000

% to global mesh
6.12%
3.85%
3.24%
2.33%
2.05%
7.00%
5.50%

The third column in Table 4.3 gives the proportion of the overset in term of the
number of the component mesh elements, which includes the overset elements. As it
can be seen, the overset features 7 million elements, which corresponds to 5.5% of the full
pump-turbine mesh. This quantity could be considered as a significant increase. However,
when performing large-scale computations involving a large number of cores, i.e. 2048,
such a quantity of added elements is acceptable and it corresponds to only a few thousand
elements added per sub-domain.

4.3

Computation Parameters & Flow initialization

The flow computation of a full pump-turbine reduced scale model by the Large Eddy
Simulation for a high Reynolds number requires a fine temporal and spatial discretization
to capture all the relevant turbulent structures within the flow. For these discretizations,
the necessary computing power resources can be very high and the allocation of these
resources can be long. In order to save time and computing resources, the flow initialization is performed using a coarse mesh, featuring 85 million elements by down scaling
the mesh introduced in section 4.2, and one impeller revolution discretized by 10,000
time steps. With such a spatial and temporal discretization, the computation of one impeller revolution using 2,048 cores can be performed in less than 7 hours on the FX10
supercomputer.

4.3.1

Numerical Setup

The execution of the unsteady flow computations requires several settings, such as
the choice of a turbulence model, the numerical scheme and some settings specific to the
start of a computation in order to avoid numerical instabilities. FFB gives two choices
for the turbulence model: the static and the dynamic Smagorinsky turbulence models.
As it is known that the Smagorinsky model tends to be too dissipative [36], the dynamic
Smagorinsky model is chosen with the no-slip condition at the walls. The presence of
the overset interfaces in the mesh requires specific settings to avoid the above mentioned
issues regarding the mass imbalance at the interface. Therefore, the Fractional Step
method with the low Mach number assumption is selected. The setting of the low Mach
number is done via a parameter called FSMACH which is the inverse of the numerical
EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

45

speed of sound: 1/a. Using the definition of the Mach number: Ma=C/a, where C is a
characteristic velocity, the parameter FSMACH becomes the ration of M a/C and M a and
C are respectively set to 0.1 and 4.1135 m/s which is the peripheral speed of the impeller.
However, in some computations the low Mach assumption may not be sufficient, especially
at the early phase of the computation, and it brings about the mass imbalance issue. Such
a situation was encountered during the start of the present computation and was solved
by using a specific function (MIC) imposing the mass balance at the overset interfaces.
Such a function has been used during the computation of the first 33 impeller revolutions.
Finally, the last setting concerns the overset between the rotating and the stationary
components. In this situation, the overset area attached to the impeller rotates with time
and the corresponding overlapping region between the rotating and stationary parts has
to be updated. In order to save time without loosing the accuracy, the overlapping is
updated at every 4 time steps.
FFB requires the choice of a temporal numerical scheme and weighting functions for
the elements. For an accurate computation, the Crank-Nicolson scheme with the standard
Galerkin method are chosen, which results in a second order accuracy both in time and
space.
Finally, to compute the flow corresponding to the part load condition =0.026, the
inlet uniform velocity and the impeller rotational speed have been set respectively to 2.57
m/s and 15.708 rad/s.

EPFL - Laboratory for Hydraulic Machines

46

4.3. COMPUTATION PARAMETERS & FLOW INITIALIZATION

Table 4.4: Summary of the computation parameters.

Re

n
Cin

Reynolds number
1.1 106
Kinematic viscosity
106 m2 /s
Flow coefficient
0.026
Impeller rotational speed 15.708 rad/s
Inlet velocity
2.57101 m/s

Computation Initialization
The start of a computation can lead to numerical instabilities as the flow undergoes
a sudden change at boundaries. To prevent such an instabilities from occurring, it is
common to use functions which gradually increase the boundary values. Therefore, the
inlet uniform velocity and the impeller rotational speed were exponentially increased
from 0 to their steady values during the computation of the first impeller revolution.
Moreover, to prevent unphysical pressure fluctuation from taking place associated with
the RSI, no overset update were carried out during the first 5 impeller revolutions (i.e.
static oversetting mode was set).
In the following, all the results shown were obtained using the University of Tokyo
Supercomputer PRIMEHPC FX10.

4.3.2

Computation Convergence

The convergence of a computation to a steady, a cyclic or a statistically steady solution

can be demanding, especially for large computational domains and a small time step. To
estimate the required number of impeller revolutions to obtain a fully developed flow field,
one can compute the number of impeller revolutions necessary to renew the flow within
the full pump-turbine. Table 4.5 gives the volume of each components of the pumpturbine with the estimation of the number of impeller revolutions required for all initial
fluid particle present in the individual pump-turbine component to exit from it. As it
can be seen, the computation of 25 impeller revolutions would be required to completely
renew the initial fluid present in the machine, which is too huge for an initialization as it
would need the use of 358,400 CPU hours. However, the components of interest are the
impeller and the diffuser, which are the two smallest volumetric components of the pumpturbine, respectively 5% and 7%. Therefore, the required number of impeller revolutions
to renew the flow within these components is lower than 2. However, such an estimation
is just an indication of the duration for the required initial computation, thus such an
estimation is generally taken several times its value to get a safer assumption. For the
present computation, it is assumed that the computation of 10 impeller revolutions (5
times the estimation) should be sufficient to obtain a fully developed flow field in the
impeller and the diffuser.
The pressure coefficient at the sampling point SP1, see Fig. 5.2 on page 55, was
monitored since the start of the computation for evaluating the computation convergence.
The time history of this signal is shown in Fig. 4.13(a). The pressure coefficient is
ref
computed as follow: Cp = pp
1
2 , where pref is located in the cone (upstream the impeller)
U1e
2
at [-0.1190,0.0687,-0.2302]. As it can be seen, the start of the computation induces a
EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

47

Table 4.5: Volume of each component of the pump-turbine scale model with an estimation
of the number of impeller revolutions for all initial fluid particles to leave each component.
Q and T respectively represent the pump-turbine flow rate and the impeller period and
are equal, respectively, to 2.304102 m3 /s and 0.4 s.
Component
Cone
Impeller
Diffuser
Volute
Total

Volume
Volume/Q/T
3
[m ]
[Nb of revolutions]
3.04 102 (13%)
3.3
2
1.09 10 (5%)
1.2
2
1.49 10 (7%)
1.6
1.72 101 (75%)
18.7
2.28 101
25

high-amplitude pressure wave, which is dampened within 3 impeller revolutions. After 5

impeller revolutions, the overset mode was switched from the static mode to the dynamic
mode and the fluctuations due to the Blade Passing Frequency (BPF) appeared in the
signal. A better presentation of those fluctuations is given in Fig. 4.13(b), which shows
the pressure coefficient evolution during one impeller revolution. It can be seen that the
main fluctuation is occurring at twice the BPF. A glance at the evolution of the pressure
during the following 5 impeller revolutions does not show significant variation. This
suggests that the flow field in this area of the machine already reached a fully developed
state and confirms that only a few impeller revolutions were required to initiate the flow
field against the estimation given in Table 4.5. However, this short pressure time series
does not necessarily show the presence of the instability within the diffuser. To clarify the
occurrence of the rotating stall, Fig. 4.14 shows the instantaneous normalized velocity
magnitude in the diffuser symmetry plane. As it can be seen, some diffuser channels
exhibit higher velocity than their neighboring channels. Moreover, these high velocity
zones can be assembled into 4 groups which are separated by approximately 90 degrees.
This confirms the presence of the instability in the present computation.
2.5

1.00

(a)

Cp

Cp

2.0

(b)

0.95

1.5
0.90
1.0
0.85

0.5
0.0

t/T

10

0.80
9.0

9.2

9.4

9.6

9.8

t/T

10.0

Figure 4.13: (a) Time history of the pressure coefficient at SP1 during the flow initialization. (b) Time history of the pressure coefficient during the 10th impeller revolution.

EPFL - Laboratory for Hydraulic Machines

48

4.3. COMPUTATION PARAMETERS & FLOW INITIALIZATION

C/U1

1.00
0.75

0.50
0.25
0.00

Figure 4.14: Instantaneous normalized velocity on the diffuser symmetry plane after the
computation of 10 impeller revolutions.

4.3.3

Propagation Speed Discovery

The previous section showed the existence of the instability within the diffuser since the
early computation stage. Therefore, as the experiment reported that such an instability
(a)

0.3

Cp

|Cp|

0.10

(b)

18.0

0.08

0.2

0.07

0.1

0.06
0.05

0.0

0.04

0.1

0.03

0.2

0.02

0.3

0.01

10

0.039

0.09

15

20

25

30

35

40

t/T

45

0.00

9.0
0

10

15

20

25

f/fn

30

Figure 4.15: (a) Time history of the pressure coefficient fluctuation at SP1 (black curve).
The red curve is the low-pass filtered pressure coefficient. (b) Discrete Fourier Transform
of the instantaneous pressure coefficient.

EPFL - Laboratory for Hydraulic Machines

Chapter 4. HYDRODYNA Pump-Turbine

49

propagates at 2% of the impeller revolution speed [3], significant pressure variation in the
gap between the impeller trailing edge and the guide vane leading edge is expected. Figure
4.15(a) shows the time history of the fluctuation (black curve) and the low-pass filtered
fluctuation (red curve) of the pressure coefficient at SP1, during 38 impeller revolutions.
The instantaneous fluctuation exhibits a peak at t/T =33, which is the results of turning
off the MIC option at that instant. The low-pass filtered pressure fluctuation was obtained
by applying a zero-phase digital filter, using a low-pass 3rd order Butterworth filter with
a normalized, by the impeller frequency, cutoff frequency set to fn . Such filter will be
used in the remaining of the document for any instantaneous signal unless otherwise
explicitly stated. The pressure signal clearly exhibits a cyclic pattern recognized by the
two pressure drops at t/T approximately equal to 16 and 38. The time lag between
these two drops indicates that the individual channel is encountered with the passage of
a stall cell at every 22 impeller revolutions which is about twice longer period than what
was reported by the experiment, i.e. approximately 12 impeller revolutions. Such a low
propagation speed is confirmed by the Discrete Fourier Transform (DFT) shown in Fig.
4.15(b) with a normalized frequency resolution of 0.019. Three main peaks are visible
at the normalized frequencies 0.038, 9 and 18. The latter two normalized frequencies
correspond to the BPF and its first harmonics, respectively. The first one corresponds to
the rotating stall. Since there are 4 stall cells, the normalized frequency of the first peak
has to be divided accordingly, which results in a normalized propagation speed of 0.95%
(26 impeller revolutions for one stall cell passage), i.e. approximately twice smaller than
the experiment. The counter measure that the present research has taken against this
discrepancy will be explained in the next chapter.

EPFL - Laboratory for Hydraulic Machines

Part III
Results

EPFL - Laboratory for Hydraulic Machines

Chapter 5
Reduced Reynolds Number
Computation
5.1

Numerical Setup

The initial computation of the rotating stall performed for a Reynolds number 5 time
smaller than the experiment showed that the four rotating stall cells can be captured, but
the computed propagation speed was approximately twice smaller compared to that of the
experiment. It is assumed that the cause of this difference is due to an under resolution of
the physics, as the mesh used is too coarse for that Reynolds number. Therefore, to verify
this hypothesis, the Reynolds number is again reduced by a factor 5, resulting in a total
reduction by a factor of 25 compared to the experiment. The numerical setup is strictly
identical to the previous computation (use of the same mesh of 85 million elements) except
for the kinematic viscosity, which is multiplied by a factor 5. Furthermore, the initial flow
field was set to the latest flow field obtained with the initial computation and 5 impeller
revolutions computation were carried out for the flow to converge to the new state for the
further reduced Reynolds number.

5.2
5.2.1

Comparison with Experimental Data

Propagation Speed Discovery

Local Pressure Fluctuation

The presence of the rotating stall is clearly visible by the pressure variation in the
area between the trailing edge of the impeller blades and the leading edge of the guide
vanes, as shown in the former computation. Therefore, the pressure at SP1 location is
monitored during the computation of 35 impeller revolutions. Such a time history fluctuation is depicted in Fig. 5.1(a), where three pressure drops can be identified. Compared
to the initial computation, which showed that the passage of one stall cell required approximately 26 impeller revolutions, see Fig. 4.15(a), the passage of one stall cell requires
approximately 12 impeller revolutions. This propagation speed is confirmed by the frequency spectrum of the pressure coefficient fluctuation as shown in Fig. 5.1(b), where
three main peaks at a normalized frequency of 0.076, 9 and 18 can be seen. The two latter
EPFL - Laboratory for Hydraulic Machines

54
0.3
Cp
0.2

5.2. COMPARISON WITH EXPERIMENTAL DATA

0.12
|Cp|
0.10

(a)

0.1

0.08

0.0

0.06

0.1

0.04

0.2

0.02

0.3

10

15

20

25

30 t/T 35

0.00

(b)

18.0

0.076

9.0
0

10

15

20

25 f/f 30
n

Figure 5.1: (a) Time history of the pressure coefficient fluctuation at SP1.The red curve is
the pressure coefficient fluctuation that is low-pass filtered using a cutoff frequency equal
to fn . (b) Discrete Fourier Transform of the instantaneous pressure coefficient fluctuation.
frequencies correspond respectively to the BPF and its first harmonic, whereas the first
frequency corresponds to the stall rotation. The resolution of the spectrum is similar to
the former computation and is 0.019. From the investigation of the local pressure it can
be concluded that the propagation speed is 9% slower compared to the experiment, i.e.
s / = 1.9%. However, the present spectrum resolution is not sufficient to accurately
assess the propagation speed, as the two neighboring frequencies would give a propagation
31% lower and 13% faster than the experiment.
Flow Rate Fluctuation at the Guide Vane Throat
Another approach to quantify the propagation speed of the stall cells is to monitor the
flow rate variation through each guide vane channel because the presence of a stall cell
in a channel reduces the amount of flow passing through it. Using the FFB capability to
save the value of any quantity at each time step, a discrete surface representing the guide
vane throat was generated to record the velocity field for each time step. A 2D view of
the guide vane throat is shown in Fig. 5.2 by the green circles and they are composed of
27 sampling points hyperbolically distributed in order to have a finer discretization near
the guide vane walls. In the spanwise direction 29 sampling points were also distributed
hyperbolically, which resulted in a total of 783 sampling points per guide vane throat.
The previous visualization of the flow, see Fig. 4.14, showed that the stall cells exhibit
a double spatial symmetry, which suggests a propagation speed approximately similar for
each stall cells. Figure 5.3 shows the time histories of flow rate through 4 guide vane
throats separated by 5 guide vane channels with each other, i.e. throats 4, 9 14 and 19. It
can be confirmed that these four channels exhibit a similar variation. Furthermore, using
the four flow rates, a spatial average was computed and is depicted by the black line.
Such an average has been used to estimate the number of impeller revolutions required
to reach back to the initial condition for the individual channel. It takes 12.3 impeller
revolutions for a channel to see the full passage of one stall cell. Using this evaluation,
49.3 impeller revolutions are required for one full stall cycle, which results in a normalized
propagation speed of s / = 2.03%. This results is very satisfactory as it confirmed the
EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

55

SP2
s

SS

SP1
pt22

GV
PS

y
pt37

Figure 5.2: 2D location of the velocity sampling points represented by circles. The black
equidistant circles compose the circular surface with a radius of 277 mm. The hyperbolically distributed green circles compose the throat surface. The hyperbolically distributed
blue circles compose the surface 0.5 mm apart from the guide vane. The two black circle
filled in red are the nearest sampling points to SP1 and the guide vane suction side. The
crosses SP1 and SP2 are two pressure sampling points.
previous estimation of the propagation speed using the pressure signal and also because
it means that the computed stall cells propagate approximately the same speed as in the
experiment. In fact, the computed propagation speed is only 3.4% slower than that of
the measurement.

5.2.2

Flow Patterns at Specific Locations

During the experimental campaign, Laser Doppler Velocimetry (LDV) measurements

were carried out to investigate the local flow patterns at three different locations. These
locations are the guide vane throat surface where the velocity normal to the surface was
measured; a surface 0.5 mm apart from a guide vane where the velocity tangential to the
surface was measured and a cylindrical surface with a radius of 277 mm (15.2 mm apart
from the impeller trailing edge) where both the radial and the tangential velocities were
measured. Furthermore, the data were averaged regarding the phase of the stall. Such a
phase is obtained using the time history of a pressure signal located at the same location
at SP1 for 100 impeller revolutions to make an analytic signal as follow:
P(t) = p0 (t) + iH(p0 (t))

(5.1)

where p0 is the fluctuation of the low-pass filtered pressure and H(p0 ) its Hilbert transform.
The analytic amplitude and analytic phase are respectively a(t) = |P(t)| and (t) =
EPFL - Laboratory for Hydraulic Machines

56

5.2. COMPARISON WITH EXPERIMENTAL DATA

1.8
Q*t
1.6

12.3 revolutions

1.4
1.2
1.0
0.8
Throat 4
Throat 9
Throat 14
Throat 19

0.6
0.4
0.2

10

t/T

15

Figure 5.3: Time histories of the flow rate through 4 guide vane throats separated by 5
channels.
arg(P(t)). This decomposition is reversible and the original signal can be reconstructed
as follow:
p0 (t) = a(t)cos((t))

(5.2)

Figures 5.4 a) and b) show respectively the analytic phase and the analytic amplitude
of the pressure signal computed at SP1 shown in Fig. 5.1 (a). The three cycles that can
be seen in Fig. 5.1 (b) correspond to three passages of a stall cell. Therefore, the analytic
phase will be hereafter referred to as stall phase.
For comparison purposes, the flow through the guide vane throat and the cylindrical
section have been recorded along all the diffuser circumference. However, with the number
of the sampling points being large, the flow near the guide vanes could be recorded only for
five guide vanes. The resulting number of the sampling points is 160,000 to save the three
instantaneous velocity components. To the contrary, as the rotating stall phenomena
propagates at a low speed, the present computation computes only three passages of a
stall cell through a guide vane channel which makes the number of samples for stall phase
average low. Therefore, to increase the number of samples, a space average is performed,
which takes into account all the samples in the diffuser experiencing a similar stall phase.
Doing so gives around 50 samples for the throat and the cylindrical surfaces and 12
samples for the guide vane surfaces.
As introduced previously, the numerical data were obtained by storing the velocity
vector on a discrete surface at each time step, see Fig. 5.2, using hyperbolic distribution
to better discretize the flow near walls. For all the surfaces, the spanwise side is discretized
by 29 sampling points. The resulting discrete surfaces are given in Table 5.1 and the flow
patterns in these surface will be described below.

EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

a)

(t)

a(t)

57

b)

0.06
0.04
0.02
0.00

0.02
0.04
0.06
0.08

0.10
0

10

15

20

25

30

t/T

35

10

15

20

25

30 t/T 35

Figure 5.4: Analytic signal of the low-pass filtered pressure fluctuation signal shown on
Fig. 5.1 (a): a) analytic/stall phase, b) analytic amplitude.
Flow Patterns at the Throat Surface
The stall-phase-averaged velocity normal to the throat surface obtained by the measurement and the computation, are depicted respectively in Fig. 5.5 and Fig. 5.6. The
stall-phase-averaged flow is shown for 6 different phases, i.e. =[0, /2, 3/4, , 5/4,
3/2], with =0 representing the phase when the guide vane channel experiences the
maximum flow rate, i.e. the minimum pressure. The comparison between the experiment
and the computation shows that some discrepancies exist between them.
The experiment reported a reverse flow between the stall phases = 3/4 to = 5/4.
The reverse flow spans all the vane-to-vane channel pitch near the hub and occupies
approximately 20% of the throat surface area. The computation also shows a reverse
flow, but occupies a smaller area than the experiment. The reverse flow is mainly located
at the guide vane suction side near the hub. A reverse flow is also visible at the guide
vane pressure side near the hub, but no decisive reverse flow is observed at the mid-pitch
in the computation.
Regarding the magnitude of the flow velocity, it can be seen that the computation
indicates that the velocity extrema are higher than the experiment. Indeed, the computed
velocity is lower than the measured velocity when the channel is stalled, see Fig. 5.5 (c)
and Fig. 5.6 (c). In the same manner it can be seen that the computed velocity is higher
than the measured velocity, when the channel experiences a high flow rate, see Fig. 5.5
(a) and Fig. 5.6 (a).
These differences may be partially attributed to the stall phase averaging, which picks

Table 5.1: Surface discretization for comparison with LDV measurements. The physical
location of these surfaces is shown in Fig. 5.2.
Surface
Discretization Total
Throat
27x29
783
Guide vane 2x50x29
2,900
R277
40x29
1,160

EPFL - Laboratory for Hydraulic Machines

58

5.2. COMPARISON WITH EXPERIMENTAL DATA

a) =0

b) =/2

0.5

c) =3/4
0.5

0.5

z/b 2

PS

0.4

hub

z/b 2

SS

PS

0.4

hub

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

-0.1

-0.1

-0.1

-0.2

-0.2

-0.3

-0.3

0.50

0.55

-0.3
-0.4
-0.5

0.2

0.4

0.6

0.8

-0.5

shroud
0

d) =

0.2

0.4

0.6

0.8

z/b2

PS

0.4

hub

0.75

0.25

0.40
0.00

-0.5

-0.25

shroud
0

0.2

0.4

0.6

0.8

f) =3/2
0.5

z/b2

SS

Cn /U1e

SS

0.50

e) =5/4
0.5

0.5

hub

-0.4

-0.4

shroud

PS

0.4

0.3

-0.2

z/b 2

SS

PS

0.4

hub

z/b2

SS

PS

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

-0.1

-0.1

hub

Cn/U1e

SS

0.75

0.50

0.40

-0.2
0.35

-0.3
-0.4
-0.5

0.40

-0.2

-0.2

-0.3

-0.3

0.2

0.4

0.6

0.8

-0.5

0.00

-0.4

-0.4

shroud

0.25

-0.1

0.2

shroud

0.4

0.6

0.8

-0.5

-0.25

shroud
0

0.2

0.4

0.6

0.8

Figure 5.5: Stall-phase-averaged normal (streamwise direction) velocity in the guide vane
throat section measured by LDV at = 0.026. a-f) Velocity contour plots along different
phase angles from 0 to 3/2. The bold line represents the level of zero velocity. Figures
obtained from [3].
a single phase to average the data. Namely, the flow field fluctuates rapidly and thus
changes in a short time. Therefore, a small difference in the time of picks up may result
in a noticeable difference in the average flow pattern. Furthermore, in both cases a discrete
grid was used and the corresponding data were interpolated. The grid for the computation
is approximately twice finer than the experiment and refinement is performed near walls
for the computation, which may introduce some differences. However, these possible
source of errors cannot by themselves explain the differences between the computed and
the measured flow velocities. These differences are most likely attributed to the difference
of the Reynolds numbers in the computation and in the experiment.
Flow Patterns at the Guide Vane Surface
In a similar manner, the measured and computed stall-phase-averaged tangential flows,
EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

z/b2

0.5

(c) =3/4

(b) =/2

(a) =0
PS

hub

SS

0.4

z/b2 0.5

PS

hub

SS

0.4

z/b2 0.5

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.0

0.35

0.0

0.0

0.1

0.1

0.2

0.2

0.2

0.3

0.3

0.3

0.4

0.4

0.4

0.5
0.0

0.5
0.0

0.4 0.6
shroud

0.8

s 1.0

(d) =
z/b2

0.5

0.2

0.4 0.6
shroud

0.8

s 1.0

hub

SS

0.4

z/b2 0.5

PS

SS

z/b2

0.5

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0.0

0.0

0.0

0.1
0.2

0.2

0.3

0.3

0.3

0.4

0.4

0.4

0.5
0.0

0.5
0.0

0.4 0.6
shroud

0.8

s 1.0

0.2

0.50

0.25

0.00

0.30

0.2

0.4 0.6
shroud

0.4 0.6
shroud

PS

0.8

hub

s 1.0

0.25

0.8

s 1.0

SS

0.5
0.0

0.75
^

Cn/U1e
0.50

0.25

0.1

0.40

0.2

0.2

0.4

0.3

0.30

0.75

(f) =3/2

hub

0.4

0.1

SS

Cn/U1e

0.5
0.0

(e) =5/4

PS

hub

0.1

0.1

0.2

PS

0.4

0.3

0.55

59

0.50
0.00

0.2

0.4 0.6
shroud

0.8

s 1.0

0.25

Figure 5.6: Stall-phase-averaged normal (streamwise direction) velocity in the guide vane
throat section computed by LES at = 0.026. a-f) Velocity contour plots along different
phase angles from 0 to 3/2. The bold line represents the level of zero velocity.
in the streamwise direction, on a surface apart from 0.5 mm of a guide vane are shown
in Fig. 5.7 and Fig. 5.8, respectively. The surface around the guide vane is decomposed
into the suction side (SS) part and the pressure side (PS) part. These figures show the
stall-phase-averaged flow patterns for the 6 different phases as in Fig. 5.5 and 5.6. Regarding the computed flow pattern, it is noticeable that recirculations are always present
at the guide vane trailing edge (s=1) all along the span. The reason is that the blunt
trailing edge always induces flow separation. This recirculation is not visible from the
experiment, as this area was not accessible. In this representation of the flow field, the
computation delivers again a more symmetric (with respect to the spanwise direction)
flow patterns compared to the experiment. The experiment shows a recirculation located
near the hub at the suction side. Furthermore, the velocity gradients are less pronounced
in the experiment compared to the computation. In the experiment, the flow decelerates
gradually from the leading edge toward the trailing edge near the shroud, whereas the
computation exhibits higher gradients located near the leading edge. However, it should
EPFL - Laboratory for Hydraulic Machines

60

5.3. DISCUSSION

be recalled that only a few samples (approximately 12) were available for this phase average in the computation. Hence, these patterns in the computation may not represent
the phase-averaged flow and may just show that of an instantaneous flow.
In addition to that, at most of the phases (Fig. 5.8 (a) to (f)), low tangential velocities
are observed near the guide vane leading edge both near the shroud and the hub in the
computation. These low velocity zones are associated with the presence of two horseshoe
vortices, which are the result of the interaction of the impinging boundary layer at the top
and the bottom ring on the guide vane. The effect of these horseshoe vortices is mainly
seen in the computation on the suction side at =/2, thus before the channel is fully
stalled.
Flow Pattern at the Cylindrical Section
The pattern of the flow field entering the guide vane is very important as it is a key parameter to determine the favorable conditions for the stall to appear. Thus, the radial and
tangential components of a cylindrical surface with a radius R=277 mm are monitored.
The resulting stall-phase averaged flows are shown in Fig. 5.9 for the experimental data
and in Fig. 5.10 for the computational data. In both cases, figures (a) to (c) show the
normalized radial velocity and figures (e) to (f) show the normalized tangential velocity
for =[0, /2, 3/4].
The comparison between the experiment and the computation shows that there are
good qualitative and quantitative agreements for the radial velocity. The computed magnitude of the velocity as well as the location of the reverse flow agree well with the
measurements. However, the sizes of the computed reverse flow are smaller compared
to the experiment. For the tangential velocity, one can see that the computed velocity
is about 10% higher compared to the measurements. One possibility is that the relative
velocity W exiting the pump is slightly slower in the computation than in the experiment,
which would be the typical result of an under resolution in the boundary layer thickness.
If the impeller boundary layers are under-resolved, the thickness of the boundary layers
are also under-predicted, which increases the effective channel pitch and reduces the flow
velocity. Therefore, the absolute angle is reduced, increasing the tangential velocity.

5.3

Discussion

This chapter compared the computed flow with the experiment. It confirmed that
the Reynolds number plays a significant role in the computation of the rotating stall
phenomenon. Reducing the Reynolds number compared to the initial computation resulted in the capture of the phenomenon reported by the experimental investigation. The
consequence of the reduction of the Reynolds number was an increase of the size of the
turbulence scales. Therefore, the ability of the present grid to compute the turbulence
scales was enhanced.
However, it was shown that the patterns of the flow through the throat, around a
guide vane and through a circular surface near the guide vane inlet are different in the
velocity magnitude and in the velocity distribution. These differences may partially be
attributed to the difference of the surfaces discretization and by the difference of the
number of samples for the stall averaging. However, these cannot explain the observed
EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

61

difference between the computation and the experiment by themselves. The observed
difference is most likely due to the difference in the Reynolds numbers in the experiment
and the computation. Effects of the Reynolds number will further be discussed in the
next chapter.
Despite discrepancies of the flow patterns, the computed specific energy coefficient
are consistent with the experiment. Indeed, the relative difference between the computed
and the experiment is only of 0.36%. Meanwhile, the propagation speed computed
was only 3.4% slower than the experiment. Table 5.2 shows the comparison between the
experimental and computed flow coefficient , energy coefficient and the propagation
speed s . The results of the RANS computation were obtained from [3]. Such a comparison enhances the accuracy of the Re/25 computation. Therefore, it is believed that,
despite the flow pattern discrepancies, the present computation has produced essential
features of the flow field and can thus be used to study the rotating stall phenomenon in
detail.
Table 5.2: Comparison of the experimental and numerical values for the flow coefficient
, the energy coefficient and the propagation speed s .
Exp. RANS LES(Re/5)

0.026 0.032
0.026
Rel. difference
23%
0%

1.079 0.962
1.115
Rel. difference
-10.8%
3.35%
s /
0.021 0.029
0.0008
Rel. difference
38%
-60%

EPFL - Laboratory for Hydraulic Machines

LES(Re/25)
0.026
0%
1.082
0.36%
0.0203
-3.4%

62

5.3. DISCUSSION

0.5

a) =0

0.5

hub

z/b2

hub

z/b2

0.5

PS 0.0

b) =/2

PS 0.0
shroud

-0.5
0.5

hub

z/b2

0.6

0.4
shroud

-0.5
0.5

hub

z/b2
0.7

SS 0.0

SS 0.0

0.5

shroud

0.2

0.4

0.6

0.8

1.0

c) =3/4

0.5

hub

z/b2

-0.5
0.0

shroud

0.2

0.4

0.6

0.8

1.0

d) =
hub

z/b2

PS 0.0

PS 0.0

0.3

shroud

-0.5
0.5

hub

z/b2

SS 0.0

SS 0.0

shroud
hub

0.5

0.5

shroud

0.2

0.4

0.6

0.8

1.0

e) =5/4

shroud
hub

z/b2
0.5.5

0.6

0.8

0.4

1.0
hub

shroud

-0.5
0.5

hub

0.5

SS 0.0

0.6

0.8

1.0

-0.5
0.0

0.8
^
Ct /U1e

0.4
0.2

-0.2

0.8
^
Ct /U1e
0.6

0.4

z/b2

shroud

0.2

0.4

f) =3/2

PS 0.0

-0.5
0.5
SS 0.0

shroud

0.2

z/b2

0.4

PS 0.0

-0.5
0.0

0.5

hub

z/b2

-0.5
0.0

-0.2

0.0

0.6
-0.5
0.0

0.2

0.6

0.4

-0.5
0.5

z/b2

0.4

0.0

0.7
-0.5
0.0

0.8
^
Ct /U1e

0.4
0.2
0.0

shroud

0.2

0.4

0.6

0.8

1.0

-0.2

Figure 5.7: Stall-phase-averaged tangential velocity at 0.5 mm apart from guide vane
4, measured by LDV at = 0.026. The bold line represents the level of zero velocity.
Figures obtained from [3].

EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

(a) =0

(b) =/2

0.50

0.50

hub

z/b2

0.25

shroud

shroud
hub

z/b2

0.25

0.25

0.7

SS 0.00

0.25

0.25
shroud

0.50
0.0

0.2

0.4

0.6

0.8

(c) =3/4

0.2

0.4

0.6

0.8

0.50

hub

s 1.0

hub

z/b2

0.00

PS 0.00

0.25

0.25
shroud

0.50
0.50

hub

z/b2

0.6

shroud

0.50
0.50

hub

0.25 0.6

0.00

SS 0.00

0.25

0.25

0.7
shroud

0.2

0.4

0.6

0.8

s 1.0

z/b2

0.25

0.3

0.8

s 1.0

hub

hub

0.5

-0.2

0.8
^

0.6

shroud

0.50
0.50

z/b2

0.25

hub

0.6

SS 0.00

0.25

0.2

Ct/U1e

0.25
shroud

SS 0.00

0.50
0.0

0.6

z/b2

PS 0.00

0.50
0.50
0.25

0.4

0.50
0.25

0.25

z/b2

0.2

0.4

PS 0.00

shroud

(f) =3/2
hub

0.4

0.0

0.50
0.0

(e) =5/4

0.8

Ct/U1e

z/b2

0.25

0.50
0.0

-0.2

0.25

0.3

0.30

0.25

0.2

(d) =

0.50

z/b2

shroud

0.50
0.0

s 1.0

0.4

0.0

0.80

SS 0.00

0.50

0.6

0.40

0.50
0.50

hub

z/b2

SS

0.25

0.50
0.50

0.8

Ct/U1e

PS 0.00

0.25

PS

hub

z/b2

0.25

0.5

PS 0.00

63

0.4
0.2
0.0

0.25
shroud

0.2

0.4

0.6

0.8

1.0

0.50
0.0

shroud

0.2

0.4

0.6

0.8

s 1.0

-0.2

Figure 5.8: Stall-phase-averaged tangential velocity at 0.5 mm normal distance from guide
vane 4, computed by LES at = 0.026. The bold line represents the level of zero velocity.

EPFL - Laboratory for Hydraulic Machines

64

5.3. DISCUSSION

0.5
z/b2
0.4

a) =0

0.5
z/b2
0.4

hub

b) =/2

0.5
z/b2
0.4

hub

c) =3/4

0.25

hub

0.20

0.3

0.3

0.3

0.15

0.2

0.2

0.2

0.10

0.1

0.1

0.1

0.05

0.0

0.0
0.15

-0.1

0.00

0.0
0.10

-0.1

0.10

-0.1

-0.05

-0.2

-0.2

-0.2

-0.10

-0.3

-0.3

-0.3

-0.15

-0.4

-0.4

-0.5
30

0.5
z/b2
0.4

shroud
24

18

12

d) =0

-0.5
30

0.5
z/b2
0.4

hub

-0.20

-0.4
shroud
24

18

12

e) =/2

-0.5
30

0.5
z/b2
0.4

hub

shroud
24

18

12

f) =3/4

-0.25

0.75

hub

0.70

0.3

0.3

0.3

0.65

0.2

0.2

0.2

0.60

0.1

0.1

0.1

0.0

0.0
0.60

-0.1

-0.1

0.60

0.0

0.50

-0.1

0.45

-0.2

-0.2

0.40

-0.3

-0.3

-0.3

0.35

-0.5
30

-0.4
shroud
24

18

12

-0.5
30

0.30

-0.4
shroud
24

18

12

-0.5
30

^
Cu/U1e

0.55

0.50

-0.2

-0.4

^
Cr/U1e

shroud
24

18

12

0.25

Figure 5.9: Stall-phase-averaged velocity on cylindrical section at guide vane, measured

by LDV at = 0.026. a, b and c) radial velocity, d, e and f) circumferential velocity.
The bold lines represent the level of zero velocity for the radial velocity and the level of
0.5 velocity for the circumferential velocity. Figures obtained from [3].

EPFL - Laboratory for Hydraulic Machines

Chapter 5. Reduced Reynolds Number Computation

(a) =0

(b) =/2
hub

0.5

z/b2

z/b2

0.4

(c) =3/4
hub

0.5

65

z/b2

0.4

hub

0.5

0.25 ^

Cr/U1e

0.4

0.20

0.3

0.3

0.3

0.15

0.2

0.2

0.2

0.10

0.1

0.1

0.1

0.05

0.0

0.0

0.0

0.00

0.1

0.1

0.1

0.05

0.2

0.2

0.2

0.20

0.3

0.3

0.4

0.4

0.5
30

24

18

shroud

0.5
30

12

(d) =0
z/b2

0.15

z/b2

0.4
0.3

12

0.15

0.4

0.20

0.5
30

0.5

24

18

shroud

12

0.25

(f) =3/4
hub

z/b2

0.4
0.3

0.70

0.2

18

shroud

0.3

(e) =/2
hub

0.5

24

0.10

0.10

0.2

hub

0.5

0.75

0.4

0.70

0.3

0.65

0.2
0.65

0.60
0.60

0.55

0.1

0.1

0.0

0.0

0.0

0.50

0.1

0.1

0.1

0.45

0.2

0.2

0.2

0.40

0.3

0.3

0.3

0.35

0.4

0.4

0.4

0.30

0.5
30

0.5
30

0.5
30

0.25

24

18

shroud

12

0.1

24

18

shroud

12

Cu/U1e

24

18

shroud

12

Figure 5.10: Stall-phase-averaged velocity on cylindrical section at guide vane, computed

by LES at = 0.026. a, b and c) radial velocity, d, e and f) circumferential velocity. The
bold lines represent the level of zero velocity for the radial velocity and the level of 0.5
velocity for the circumferential velocity.

EPFL - Laboratory for Hydraulic Machines

Chapter 6
Rotating Stall Analysis
6.1

Stall Onset

The previous chapter revealed the existence of four stall cells in the diffuser, propagating in the sense of the impeller rotation, at a speed of s /=2.05%. Furthermore,
comparisons with the experimental data produced satisfactory results, which enable the
investigation of the rotating stall phenomenon using the previously computed flow fields.
Therefore, the present chapter focuses on the evolution of the flow within the diffuser, in
order to obtain a better understanding of the propagation mechanisms of the stall cells
in the HYDRODYNA pump-turbine.
A stall cell can be regarded as a flow recirculation region. The main reason for such a
recirculation to exist is because of the part load operating conditions. A decrease in the
flow rate passing through the impeller results in a reduction of the exit flow angle of the
impeller as shown in Fig. 6.2. This decrease of causes an increase of the angle of attack
(AOA) on the diffuser guide vane, which leads to favorable conditions for flow separation,
which results in a recirculation region at the guide vane trailing edge on the suction side.
Furthermore, it is also worth mentioning that the present pump-turbine has, regardless
of its operating conditions, a recirculation region at the trailing edge of the guide vanes.
The guide vanes are dedicated for turbine operation and feature a standard hydrofoil
shape with a thicker leading edge and a thinner trailing edge to avoid flow recirculation.
However, when operating the pump-turbine in pumping mode, the guide vane profile is
used in the opposite way, resulting in a thick trailing edge, which always generates a
recirculation region. This region is shown by the grey area in Fig. 6.1, bound by the two
horizontal tangents, shown as dash lines.

Figure 6.1: Guide vane profile. The grey area shows the region favorable for flow separation in pumping mode and the dashed lines the two horizontal tangents.

EPFL - Laboratory for Hydraulic Machines

68

6.2. OBSERVATION OF STALLED CHANNEL BEHAVIOR

Part Load

W1

U1

C1,m

BEP

C1

W1

U1

C1,u

C1,m

C1

C1,u

Figure 6.2: Velocity triangle at the impeller exit for the BEP and the part load conditions.

6.2

Observation of Stalled Channel Behavior

The present diffuser features 20 guide vanes and 20 stay vanes, that results in 20 vane to
vane channels for the flow to reach the volute. Incompressibility imposes that at any time,
the flow rate at the diffuser entry matches the one at the exit of the impeller. It results
that any channel exhibiting a lower flow rate will have a counter channel experiencing
a higher flow rate to keep the overall flow rate constant. In a space perspective of the
rotating stall, this incompressibility constrain is the key parameter for the stall pattern
into the diffuser.
The operation of the pump-turbine at 76% BEP induces the presence of four stall cells.
These cells locally perturb the flow by reducing the discharge passing through a channel,
as already seen in Fig. 4.14. However, continuous pressure and velocity variations between
each of those cells necessitate a similar spatial organization. As the instability involves
four cells at this specific operating condition, each cell has to cover 5 channels. This spatial organisation is shown in Fig. 6.3, which shows the spanwise-averaged instantaneous
velocity in six channels. Channels a-b and f-g are experiencing the highest flow rate.
The flow rate passing through these guide vane channels is approximately Qt /Qo =1.6,
with Qo being the nominal flow rate passing through the channel. As channels a-b and
f-g are exhibiting a similar flow pattern, they represent the two bounds of a stall cell
and the channels in between reflect the spatial organization of a stall cell. Taking the
direction of the impeller rotation as reference, the upstream neighboring channel, channel
b-c, exhibits a lower flow rate because of the presence of a recirculation region. This
recirculation region occupies almost half the guide-vane channel pitch and the flow rate
is close to Qt /Qo =1.0. This channel is in so called stalling process. The next upstream
neighboring channel, channel c-d, is stalled corresponding to a flow rate approximately
Qt /Qo =0.4. The recirculation region occupies almost the entire channel pitch and only
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

69

some flow can pass near the pressure side of the guide vane c. The channel with maximum
and minimum flow rates are thus close as they are separated only by a single channel.
Finally, to reach a cyclic pattern, the two remaining channels, channels d-e and e-f, have
to represent a situation where the flow rate is increasing and the recirculation region is
decreasing. Such a situation is indeed reached, as a greater flow rate is passing through
channels d-e and e-f. The combination of the increase of flow rate and the decrease of the
recirculation region results in an important wake spreading until the stay vane channel.
Such channels have respectively a flow rate approximately Qt /Qo =0.8 and Qt /Qo =1.2.
From this first observation it can be concluded that the stall cells have a strong impact
on the flow rate distribution within the diffuser. Indeed, it was shown that the presence
of a stall cell induces a relative difference of the flow rate compared to the nominal flow
rate up to 60%.

6.3

Propagation Mechanism of Stalled Cells

As the stall phenomenon is unsteady, each guide-vane channel experiences a flow rate
fluctuation as shown previously in Fig. 5.3. This requires a continuous variation of the
local flow field to switch from conditions favorable for the stall cell to grow to conditions
favorable for the stall cell to decay and vice versa. To illustrate this evolution, Fig. 6.4
shows the flow field passing through a guide-vane channel (between GVA and GVB), at five
successive different phases representing conditions encountered in the spatial organization
presented in Fig. 6.4. Furthermore, each phase is temporally located in Fig. 6.4 f), where
the black line shows the evolution of the flow rate passing through the channel shown in
Fig. 6.4 a) by the red line. Except for the flow field shown in Fig. 6.4 c), the phase of
the impeller is identical. In the case of Fig. 6.4 c) the phase of the impeller is rotated 2
in the counterclockwise direction.
To study how the flow evolves temporally within a single channel, one can set the
initial phase of the rotating stall cycle when the channel does not have any significant
C/U1

1.00
0.75
0.50

channel a-b

0.25
0.00

b
d

Figure 6.3: Instantaneous spanwise-averaged velocity magnitude in six diffuser channels

showing the spatial organization of a stall cell. Channels a-b and f-g exhibit the highest
flow rate passing through the channel. Channel b-c is stalling, channel c-d is stalled and
channels d-e and e-f are recovering.

EPFL - Laboratory for Hydraulic Machines

70

6.3. PROPAGATION MECHANISM OF STALLED CELLS

C/U1

1.00

C/U1

1.00

a)

0.75

0.75

0.50

0.50
GVA

0.25

SV

b)

GVA

0.25

0.00

SV

0.00

GVB

GVB
SV

C/U1

1.00

SV

C/U1

c)

1.00

0.75

d)

0.75

0.50

0.50
GVA

0.25

SV

GVA

0.25

0.00

SV

0.00

GVB

GVB
SV

C/U1

1.00

e)

Q*
t

0.75
0.50
0.25

SV

GVA

SV

1.8

f)

1.6

Number of impeller revolutions:

2

1.45

1.4
Fig. b)

1.2

0.00

1.55

1.0

Fig. a)
Fig. c)

0.8
0.6
GVB

Fig. e)

0.4
SV

0.2

Fig. d)
0

10

12

t/T

14

Figure 6.4: Spanwise-averaged instantaneous velocity magnitude passing through the diffuser for 5 different stall phases: a) late recovery phase, b) highest flow rate phase, c)
stalling phase, d) stalled phase and e) early recovery phase. f) evolution of the flow rate
passing through the section shown by the red line in figure a).
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

Cp

1.30

71
Cp

a)

1.30

1.00

1.00

0.69

0.69
GVA

0.39

SV

b)

GVA

0.39

0.08

SV

0.08

GVB

GVB
SV

Cp

1.30

SV

Cp

c)

1.00

1.30

d)

1.00

0.69

0.69
GVA

0.39

SV

GVA

0.39

0.08

SV

0.08

GVB

GVB
SV

Cp

1.30

e)

Q*
t

1.00
0.69
0.39

SV

GVA

SV

1.8

f)

1.6

Number of impeller revolutions:

2

1.45

1.4
Fig. b)

1.2

0.08

1.55

1.0

Fig. a)
Fig. c)

0.8
0.6
GVB

Fig. e)

0.4
SV

0.2

Fig. d)
0

10

12

t/T

14

Figure 6.5: Spanwise-averaged instantaneous Cp in the diffuser for 5 different stall phases:
a) late recovery phase, b) highest flow rate phase, c) stalling phase, d) stalled phase and
e) early recovery phase. f) evolution of the flow rate passing through the section shown
by the red line in figure a). The interval of the contour line (shown in black) is Cp =
0.015.
EPFL - Laboratory for Hydraulic Machines

72

6.3. PROPAGATION MECHANISM OF STALLED CELLS

recirculation. Such a phase is shown in Fig. 6.4 a), where the guide-vane channel experiences a flow rate higher than the nominal flow rate Qo , i.e. Qt =Qt /Qo =1.15. The
particularity of this phase is characterized by a flow incidence on GVA and GVB with a
near zero angle of attack. Furthermore, the high velocity passing around GVB, leads to
the flow separation, both on the suction and the pressure sides, because of the the guide
vane geometry. The resulting wake is elongated and aligned with the guide vane chord.
Two impeller revolutions later, the channel reaches its maximum flow rate of Qt =1.42,
see Fig. 6.4 b). As previously, the flow incident on GVB is nearly parallel to the guide
vane chord, but GVA exhibits a flow deviation at the leading edge, which is visible by
the flow not aligned with the guide vane suction side near the leading edge, even if the
flow velocity nearly reaches its maximum intensity. Regarding the recirculation zone, it
can be seen that its size has reduced compared to the previous phase a). More precisely,
it reaches its minimum size and tends to be less aligned with the guide vane chord. The
reason is that the flow rate near the the pressure side of GVB and between GVB and
the Stay Vane (SV) is increased. Because of this increase, the flow is less favorable to
separate and stays attached to the GVB pressure side.
After, the channel reaches the maximum flow rate, it enters the stalling phase. Therefore, the flow rate in the passage is decreasing. Such a phase is shown in Fig. 6.4 c), which
is 1.55 impeller revolutions later and has a flow rate of Qt =1.02. Clearly, the recirculation
zone has grown and occupies about the downstream 30% of the GVB suction side. The
zone extends toward the center of the channel and deviates the flow passing through the
channel toward the GVA pressure side, which explains why during the previous phase the
flow rate near the GVB pressure side increased. With this increase of the recirculation
zone toward the center of the channel, it becomes difficult for the upstream flow to pass
through the channel and therefore circumvent the channel by deviating toward the downstream channel. Regarding the flow passing through the gap formed by GVB and SV, the
flow rate is decreasing but still large.
Finally the flow rate trough the channel reaches its minimum and corresponds to the
stalled phase, Qt =0.44. As we can see in Fig. 6.4 d), the flow can barely go through
the guide vane channel because the recirculation occupies almost all the guide vane pitch.
Namely, the channel is almost blocked. This situation implies a significant rise of the
pressure along the guide-vane channel and the flow that can pass trough the guide-vane
channel is deviated towards the gap between GVA and SV as the pressure is lower in this
area.
When the flow rate reaches its minima, the channel starts to recover, as shown in Fig.
6.4 e), Qt =0.58. This is characterized by an increase of the flow rate near the pressure
side of GVA. Furthermore, the flow incidence on GVB is decreasing. As a result, the
stall cell size decreases as well. This phase is characterized by the presence of an large
wake, which elongates far downstream the flow. Furthermore, as the flow velocity near
the pressure side of GVB is increasing, the flow starts again to separate.
The distribution of the pressure coefficient in the diffuser is shown in Fig. 6.5. In a
similar manner to the velocity fields, the distribution of the pressure coefficient is shown
for 6 different phases. Figure 6.5 a) shows the pressure coefficient when the flow is near
to reach its maximum flow rate through the channel. The spherical pattern of the isopressure at the leading edge of GVB reconfirms that the flow is almost aligned with the
guide vane chord length. Therefore, no significant acceleration of the flow is seen near the
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

73

leading edge on the suction side. But, from approximately 40% to 90% of the chord length,
the contour lines are concentrated on the suction side of GVB, showing the deceleration
of the flow. But, once the flow passes the guide vane throat no significant decrease is
visible.
Once the flow rate through the channel reaches its maximum, see Fig. 6.5 b), the
pressure gradients on the suction side of GVB exhibits iso-pressure line starting from the
leading edge to some points on the suction side (up to 25% of the chord length). Therefore,
the flow first accelerates when it passes the leading edge and secondly decelerates where
stronger adverse pressure gradients are presented. Again, once the flow passed the guide
vane throat, no significant decrease is visible.
The phase shown in Fig. 6.5 c) exhibits the same pattern as previously except that
the acceleration of the flow near the leading edge of GVB on the suction side is stronger.
The flow is still experiencing a strong deceleration shown by the concentrated iso-pressure
lines. The origin of such adverse gradients is the greater pressure drop occurring in the
clearance area formed by the tailing edge of the impeller and the leading edge of the guide
vane.
The next phase represents the one where the channel is stalled, see Fig. 6.5 d). In
that situation, one can see that the flow acceleration on the suction side near the leading
edge decreased considerably. Furthermore, the pressure gradient along the suction side
also decreased considerably. Therefore, the flow can again go through the channel.
Finally, since the adverse pressure becomes weaker, the condition are more favorable
for the flow to pass through the channel and will gradually increase. With this gradually
increase, the velocity increases again and the pressure in the clearance will decrease, which
will generate stronger adverse pressure gradients.

6.3.1

Pressure Evolution in Diffuser Channel

Under a normal condition, the pressure distribution along a diffuser channel should
be constant in time. However, when stalled, because of the temporal and spatial non
uniformity of the flow passing through a diffuser channel, the pressure along a channel
varies according to the flow speed. The relation between the pressure and the velocity in
the clearance area can be described with the Bernoullis principle as in this area the flow
is still accelerating with low vorticity and that the viscous force can be neglected. The
following relation can be then applied to any arbitrary point along a streamline in the
clearance area:
p
C2
+ gZ + = cst
2

(6.1)

Therefore, if the velocity of a fluid particle on an iso-altitude streamline accelerates

it results that the pressure has to decrease accordingly to maintain the same level of the
specific energy. To analyse how the pressure varies in a diffuser channel, the pressure at
four sampling points (SP), see Fig. 6.6, is recorded. A typical evolution of the pressure
coefficient, at the sampling point SP1, is shown for 14 revolutions in Fig. 6.7 a), where
the black line is the instantaneous signal, the blue line is a polynomial interpolation of
the signal and the letters, with their corresponding dashed lines, represent the stall phase
introduced previously in Fig. 6.3. As it can be seen, the instantaneous signal exhibits
high frequency and amplitude fluctuation up to 8%. The frequencies present in the signal
EPFL - Laboratory for Hydraulic Machines

74

6.3. PROPAGATION MECHANISM OF STALLED CELLS

SP4

SPa SP3

Im

pel

ler

SP2

rad

ius

SP1

SPb

y
x

Figure 6.6: Location of the diffuser sampling points (SP).

are the blade passing frequency (BPF) and its first harmonic. For comparison purposes,
only the polynomial interpolation of the four pressure signals are shown in Fig. 6.7 b).
Such signals also show fluctuation at the BPF and its first harmonic, but the amplitude
is decreasing along the diffuser channel flow path. The oscillation amplitudes for SP2,
SP3 and SP4 are respectively 6.5%, 4% and 1%. Except the pressure near the volute
(SP4) all signals show the sudden pressure decrease after 6 impeller revolutions. This
confirms the presence of a stall cell, where as the flow rate increases the pressure level
decreases in the whole channel to reach its minimum. However, the minimum among the
signals are not reached at the same time. SP2 and SP3 reach their minimum when the
channel experiences a high flow rate as shown in Fig. 6.3 f). But the minimum at SP1 is
reached a little later, because of the location of the sampling point. Indeed, its location is
in the gap between the impeller and the guide vane channel, where the flow field evolves
differently compared to the one passing through the diffuser channel. In that gap, the
variation of the tangential flow has to be taken into account and it will be shown later
that the velocity passing near the sampling point reaches its maximum at the same time
corresponding to that pressure minimum.
After reaching its minimum, the pressure starts to increase as the flow rate decreases.
Once the stall cell is well developed, the pressure in the stay vane channel reaches the
same pressure level as the volute, corresponding to the stalled phase. This phase lasts
approximately during one impeller revolution and then since the flow starts again to pass
through the channel, the pressure level in the channel decreases.
A closer view of the pressure coefficient evolution in the area between the guide vanes
and the stay vanes is introduced in Fig. 6.8. Furthermore, the pressure coefficient evolutions (SPa and SPb) in the gap between the guide vane and the stay vane are also taken
into account. As explained previously, the data is interpolated for a better visualization.
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

75

It can be seen that the evolution of the pressure located at SPb follows closely the evolution of the pressure at SP2. Therefore, it is expected that SPa is representative of the
local pressure level in the downstream channel. Interestingly, we can see that the pressure
at SPa is lower than all others pressures for approximately during 3.5 impeller revolutions
(from t/T=8 to t/T=11), which has an impact on the flow field. As already discussed, the
flow field approaching the stay vane channel is sometime deviated toward the gap formed
by the guide vane and the stay vane when the channel is in the stalling phase and in the
stalled phase. The reason is the difference of the flow rate passing through the two channels until that the downstream channel stalls. During that period, the difference of the
flow rates induces a difference of the pressure level between the two channels. Therefore,
the flows entering the stalled channel from the guide vane channel or from the upstream
gap, are more favorable to exit the diffuser via the downstream gap than using the stay
vane channel.
From the previous observations, the main impact of the rotating stall on the pressure
level takes place in the clearance area, i.e. at SP1. It results that the difference of
pressures between SP1 and SP2 is not constant and such a difference is shown in Fig. 6.9.
It can be seen that the pressure difference increases when the channel experiences high
flow rate and also during the first half of the stalling phase. Then, the difference decreases
which corresponds to the second half of the stalling phase until the early recovery phase.
Therefore, the flow is subject to a varying intensity of the adverse pressure gradient which
explains one of the favorable conditions for the stall cell to grow. Since the boundary
layer is decelerating, separation can take place if the momentum is not strong enough
to overcome the adverse pressure gradient. In the present situation, as the flow rate
through a channel increases, the pressure difference increases. However, this increase in
the pressure difference does not stop the flow rate to decrease during the early stage of the
stalling phase as the approaching flow has a high velocity. This approaching flow is the
flow deviated from the upstream channel and this deviation is maximum at the instant c),
see Fig. 6.9, because the upstream channel is fully stalled. Therefore, it can be concluded
that a channel experiences its highest adverse pressure gradient once its upstream channel
is fully stalled.

6.3.2

Flow Incidence of Vanes

Guide Vane
Figure 6.10 shows the evolution of the AOA on guide vanes GVA and GVB at three
different distances upstream of the leading edge, i.e. 5%, 10% and 15% of the guide vane
chord length. The AOA corresponds to the angle formed by the velocity vector with the
orientation vector of the guide vane chord. As expected, the AOA becomes higher when
approaching the leading edge, but the same trend is observed on each curve. Furthermore,
both graphs show a similar pattern, but with a temporal shift assessing the disruption
generated by a stall cell and this shows that the stall cell is propagating.
A strong correlation is visible between the flow incidence on GVA and the evolution of
the flow rate passing through the channel shown in Fig. 6.4 f). Indeed, when the channel
is in the phase of the increasing flow rate (Fig. 6.4 a) and b)), the AOA is decreasing and
the flow tends to be aligned with the guide vane. However, when the flow rate through
the channel is decreasing, i.e in the stalling phase Fig. 6.4 c), a clear increase of the AOA
EPFL - Laboratory for Hydraulic Machines

76

6.3. PROPAGATION MECHANISM OF STALLED CELLS

1.1
Cp

1.10
Cp
1.05

1.0

1.00
0.9

a)

0.95

0.8

b)

c)

d)

0.90
0.85

0.7

Cp at SP4
Cp at SP3
Cp at SP2
Cp at SP1

0.80
0.6
0.5

e)

a)
0

b)

c)

d)

0.75

e)

10

0.70

12 t/T 14

10

12 t/T 14

Figure 6.7: a) Time history of Cp located at SP1, represented by the black line. The blue
line is an interpolation. b) Interpolations representing the time history of the pressure
coefficient along a diffuser channel.
1.10
Cp
1.08

a)

b)

c)

d)

e)

1.06
1.04
1.02

Cp at SPb
Cp at SPa
Cp at SP4
Cp at SP3
Cp at SP2

1.00
0.98
0

10

12

t/T

14

Figure 6.8: Time history of interpolated Cp in the area between the guide vane and the
stay vanes.
Cp(SP2)Cp(SP1)

0.25
Cp
0.20

0.15

0.10
0

a)
2

b)
6

c)
8

d)

e)
10

12

t/T

14

Figure 6.9: Difference of the pressure coefficient at SP1 and SP2.

EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

77

is visible. This increase lasts during the entire phase of the decreasing flow rate (Fig. 6.4
b) to d)) and the maximum AOA is reached once the flow rate through the channel starts
again to increase.
As mentioned above, the same trend is visible in Fig. 6.10 b), but with a forward
temporal shift. Observing the relation between the AOA on GVB and the channel flow
rate, some general comments can be made. First of all, the AOA reaches its minimum
about 2 impeller revolutions before that on GVA reaches its minimum, which is consistent
with the propagation speed of 2.4 impeller revolutions for the propagation of one stall cell
from one diffuser channel to the downstream channel. The early increase of AOA (from
a) to b)) does not impact on the flow passing through the channel as it is still increasing.
However, we can see that the start of decrease in the flow rate can be associated with a
high incident of the flow on GVB. Then, the maximum AOA (phase c) corresponds to the
middle phase of the stalling process. Beyond that phase, the oncoming flow on GVB still
has a high AOA, but is decreasing. Once the decrease is sufficient, i.e. the flow is almost
aligned with the guide vane orientation, the flow rate into the channel starts to increase
again.
Since it was shown that the orientation of the flow near the leading edge of the guide
vane changes in a stall cycle, one can expect a variation of the flow orientation all along
the pitch at the inlet of the guide vane channel. To confirm this, the evolutions of the
radial component of the velocity on the cylindrical section at two different locations, see
Fig. 5.2, are shown in Fig. 6.11. Furthermore, when the flow velocity in the clearance
area is decomposed into its tangential and radial components, with respect to the cylindrical section, the main parameter determining the orientation of the flow is the radial
component as the tangential component does not change significantly. Therefore, to avoid
misunderstanding with the angle describing the AOA of the flow on the guide vane, the
radial component is shown as it dictates the orientation of the flow.
As it can be expected, the evolution of the flow near the center of the pitch and near
the leading edge of the suction side are different. Indeed, the orientation of the flow near
the leading edge is mainly dictated by the physics taking place in the upstream channel,
whereas the orientation of the flow in the center of the guide vane pitch is dictated by
80
AOA

80
AOA

GVA

60

60

40

40

20

20

GVB

0.05C
0.10C
0.15C

0
0.05C
0.10C
0.15C

20
0

a)
4

b)
6

c)
8

d)

20

e)
10

12 t/T 14

a)
0

b)
6

c)
8

d)
10

e)
12 t/T 14

Figure 6.10: Evolution of the AOA on GVA and GVB for three locations: 5%, 10% and
15% of the guide vane chord length upstream of the leading edge.

EPFL - Laboratory for Hydraulic Machines

78

6.3. PROPAGATION MECHANISM OF STALLED CELLS

0.06

0.13
a)
Cr
0.12

Cr
0.04

0.11
0.10

0.02

0.09

0.00

0.08
0.07

b)

a)
4

b)
6

c) d)
e)
8
10
12

t/T

14

0.02

a)

b)

c)
8

d)

e)
10

12

t/T

14

Figure 6.11: Evolution of the low-pass filtered radial component Cr at two locations, see
Fig. 5.2 on page 55. a) pt22, b) pt37.
the physics taking place in the channel. It was already shown that the AOA of the flow
incidence on GVA increases from the instant a) to c). Therefore, the corresponding change
occurs for the flow near the leading edge, see Fig. 6.11 b). The deviation of the flow in
the upstream channel is such that the radial component becomes negative, which means
that the deviation induces a reverse flow near the leading edge. The consequences of
this change are that the flow is not directed toward the radial direction as it is under
normal condition and that the momentum of the flow near the suction side is decreasing.
Regarding the center of the channel, no reverse flow is observed. However, the evolution
is different as this area is mainly affected by the recirculation zone taking place in the
channel. When the channel experiences high flow rate, the orientation of the flow is close
to the orientation of the guide vane, which results in high radial velocity. However, since
the flow near the leading edge on the suction side is more deviated and the recirculation
zone grows, the flow at the center of the channel is more favorable to deviate toward the
downstream channel resulting in the decrease of the radial velocity.
Stay Vane
Figure 6.12 also shows the evolution of the AOA, but for the corresponding stay vane.
Furthermore, for comparison purposes, the distances from the leading edge are normalized
using the guide vane chord length, hence the notation CGV . As it can be seen, the flow
incidence on the stay vanes undergoes a fast transition from negative incidences to positive
ones and then reduces back to negative incidences. The main characteristic driving the
incidence of the flow on the stay vane is the size of the wake generated by the guide vane
and the recirculation region. In situations of high flow rates within the channel, only a
small recirculation region is present and is located on both pressure side and suction side
of the trailing edge of the guide vane. Therefore, it does not perturb the flow and the
latter approaches the stay vane with negative incidence. However, since the stall cell is
growing, the flow is deviated toward the guide vane pressure side and generate a large
wake there. As it can be seen in Fig. 6.3 b) such a wake reaches the leading edge of
the stay vane which explains such a fast transition. Furthermore, the maximum AOA is
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

100
AOA
75

a)

b)

c)

d)

79
100
AOA
75

e)

50

50

25

25

25

0.05Cgv
0.10Cgv
0.15Cgv

50
75

SVA
0

10

12 t/T 14

0.05Cgv
0.10Cgv
0.15Cgv

a)

b)

c)

d)

e)

25
50
75

SVB
0

10

12 t/T 14

Figure 6.12: Evolution of the AOA on SVA and SVB for three locations: 5%, 10% and
15% of the guide vane chord length upstream of the leading edge.
reached when the diffuser channel is fully stalled. In that specific instance, the pressure
in the stay vane channel reaches the high pressure of the volute, as previously explained
in Fig. 6.8. Meanwhile, the pressure in the downstream neighboring diffuser channel has
a lower pressure as the flow is near the nominal flow rate. Therefore, some part of the
flow approaching the stay vane channel is deviated toward the gap formed be the guide
vane trailing edge and the stay vane leading edge, see Fig. 6.3 c) and d). Such a deviation
results in these very high AOA.

6.3.3

Summary of the Stall Cells Propagation Mechanism

Step 0 The flow rate within a channel is increasing and higher than the nominal flow
rate, see instant a) in Fig. 6.4 f). Resulting from this increase, the pressure decreases
within the channel (Fig. 6.7). However, because of the high velocity taking place in
the clearance area, the pressure decrease is higher than downstream in the diffuser
(Fig. 6.9). Therefore, the increase of flow rate contributes to the increase of the
adverse pressure gradient within the guide vane channel.
Meanwhile, the flow near the leading edge on the upstream guide vane suction
side starts to be deviated toward its downstream channel (Fig. 6.10 GVB). This
deviation on the upstream channel modifies the orientation of the flow near the
leading edge on the suction side of the guide vane (Fig. 6.11 b). The flow is less
aligned with the guide vane and deviates toward the downstream channel.
Step 1 The deviation of the flow near the leading edge on the suction side becomes
sufficiently large and the flow at the inlet of the channel all along the pitch is
modified (Fig. 6.11 a). This instant corresponds to the instant with the highest
flow rate (instant b). It results that the AOA on the downstream guide vane start
to increase (Fig. 6.10 GVA) and that the flow rate passing through the channel
starts to decrease (Fig. 6.4 f). However, since the flow velocity in the clearance
is still high and that the flow rate decreases the pressure downstream the diffuser
(SP2) increases, resulting in increasing further the pressure difference (Fig. 6.9).
Furthermore, the combination of the decrease of the flow rate and the high pressure
EPFL - Laboratory for Hydraulic Machines

80

6.4. ROTATING STALL IMPACT

gradient give favorable conditions for a growth of the recirculation area near the
trailing edge on the suction side to start. Besides, because of the guide vanes
geometry configuration, the shortest distance between the two guide vanes is from
point x/C=2.2% on the downstream guide vane (GVA) pressure side and point
x/C=84.9% on the upstream guide vane (GVB) suction side. Any, perturbation of
the flow near the trailing edge of the upstream guide vane has therefore an impact
on the flow near the upstream guide vane.
Step 2 The deviation taking place near the leading edge on the suction side reaches its
maximum (Fig. 6.11 b) instant c). The flow rate through the channel reduced
considerably and the adverse pressure gradient reached its maximum (Fig. 6.5 c)
and Fig. 6.9). It results that the recirculation zone becomes more significant (Fig.
6.4 c). The flow within the channel is redirected to the downstream guide vane due
to the recirculation zone.
Step 3 Since the conditions stay favorable, the recirculation zone grows and occupies
almost all the guide vane pitch (Fig. 6.4 d), which blocks the flow and forces
the flow to pass through the downstream channel, which results in high AOA on
the downstream guide vane (Fig. 6.10). However, the deviation of the flow near
the leading edge on the upstream guide vane suction side reduces (Fig. 6.11 b)
and since that the flow velocity decreased the pressure difference decreases as well
(Fig. 6.9). Therefore, the conditions are less favorable for the recirculations zone
to maintain, i.e. the flow is less subject to separate and the size of the recirculation
zone decreases.

6.4
6.4.1

Rotating Stall Impact

Impeller

The flow exiting the impeller is subject to changes imposed by the stalled channels.
Furthermore, it was shown that the main variation of the pressure in the diffuser takes
place in the clearance area, namely downstream of the impeller exit. Figure 6.13 a) shows
the instantaneous pressure coefficient located on the impeller represented by the red curve
for one impeller revolution. On the diffuser, the pressure at SP1 is monitored for each
channel. Using such signals, the local pressure in the diffuser near a specific blade channel
can be built. This results in the black line and it can be confirmed that the pressure at
the impeller exit is similar to the pressure field in the stationary part. The amplitude and
the phase agree well. Furthermore, with regards to the pressure field in the clearance,
the interaction with some guide vanes are missing. Indeed, the diffuser is composed of
20 guide vanes, but the pressure signal on the impeller shows only 16 main oscillations.
The fluctuation of the pressure in the impeller is slightly smaller than in the diffuser. The
main frequency of the fluctuation is the first harmonic of the BPF and the amplitude is
approximately 8%. But, since the impeller is rotating at a much faster speed compared
to the propagation of the stall, there is no similar variation as the pressure drop seen in
the diffuser. Finally, the presence of the stall cell confirms that the frequency of the RSI
changes under rotating stall and therefore may affect or even worse damage the machine.
The evolution of the flow rate through a single impeller channel is shown in Fig. 6.13
b) for one impeller revolution. Since the instantaneous flow fields were saved at 100 time
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

81

a)
Cp

1.05
1.00

1.06
Q*t
1.04

Stationnary
Rotating

0.95

LF HF

1.02

0.90
0.85

1.00

0.80

0.98

0.75
0.70

0.96

0.65
0.0

b)

0.2

0.4

0.6

0.8

t/T

1.0

0.94
0.0

0.2

0.4

0.6

0.8

t/T

1.0

Figure 6.13: Comparison between the evolution of a quantity in the impeller for one impeller revolution with the corresponding quantity on the stationary part, i.e. the diffuser.
a) pressure coefficient, b) flow rate (LF: Low Flow rate, HF: High Flow rate).

steps, which corresponds to the revolution of the impeller with 3.6 degrees, the above flow
rate was also computed at every 3.6 degrees. Furthermore, the approximate duration when
the impeller channel is close to a high/low flow rate guide vane channels is represented
by the grey and pink strips. It can be seen that the flow rate in the impeller changes
in accordance with the change occurring in the diffuser. However, the magnitude of the
fluctuations is one order smaller compared that in the diffuser. Indeed, it was shown
that the fluctuation in the diffuser are 60%, whereas in the impeller the fluctuation are
6%. Such difference can be explained by the fact that the flow within the impeller is
very energetic and is driven by the mechanical rotation of the impeller. Therefore, the
rotating stall only has relatively weak impacts on the flow within the impeller.

6.4.2

Guide Vane

It was shown that the presence of the stall cells modifies the pressure field and the
velocity field. These changes induce a temporal and spatial local variation of the stress
applied to different components of the machine. To elucidate this behavior, the evolution
of a guide vane lift coefficient (black line) during 1/4 stall cycle, i.e. one passage of a stall
cell, is depicted in Fig. 6.14 (a). The red line is the low-pass filtered lift coefficient. In
the previous section, we see the close relation between the pressure field and the velocity
field. It can then be expected that the lift of a guide vane is correlate with the flow
rate passing through the channel surrounding the guide vane. Figure 6.14 (b) shows thus
the instantaneous flow rates of the two channels (here upstream and downstream denote
direction with regards to the impeller rotation), as well as their difference shown by the
black line. It clearly shows that the high lift occurred when the upstream channel is in
the stalling process. As soon as the downstream channel starts to stall, and thus the flow
rate difference start to decrease, the lift coefficient starts to decrease accordingly. The lift
coefficient changes by a factor of 3, when a stall cell is passing.
EPFL - Laboratory for Hydraulic Machines

82

Cl

6.4. ROTATING STALL IMPACT

0.18

(a)

0.16
0.14

1.5
Q*t

0.12

1.0

(b)

0.10
0.5

0.08
0.06

0.0

0.04
0.02
0.00

10

12 t/T 14

0.5

10

12 t/T 14

Figure 6.14: (a) Lift coefficient variation on a guide vane during one quarter stall cycle.
The black line is the instantaneous lift coefficient and the red line the low-pass-filtered
lift coefficient with a cutoff frequency set to fn . (b) The green and the blue lines are
respectively the normalized flow rate through the upstream and downstream channels.
The black line is the flow rate difference between the two channels.

6.4.3

Stay Vanes

An identical approach was adopted for the effect of the stall cell on the stay vane
load, see Fig. 6.15 (a). The low-pass filtered lift coefficient (red curve) is near zero.
Furthermore, the instantaneous lift coefficient (black curve) periodically takes negative
value, indicating that the stay vane pressure side is acting as a suction side and viceversa.
As previously, it can be seen that the lift coefficient is function of the flow rate difference
between the two channels surrounding a particular stay vane.
0.06

(a)

Cl

(b)
1.5
Q*t
1.0

0.04
0.02

0.5
0.00
0.0
0.02
0.5
0.04

10

12 t/T 14

10

12 t/T 14

Figure 6.15: (a) Lift coefficient variation on a stay vane during one quarter stall cycle.
The black line is the instantaneous lift coefficient and the red line the low-pass-filtered
lift coefficient with a cutoff frequency set to fn . (b) The green and the blue lines are,
respectively, the normalized flow rate through the upstream and downstream channels.
The black line is the flow rate difference between the two channels.
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

1.18

1.16

83

exp. = 1.079
Rel. diff. (Re/5) = 3.35%
Rel. diff. (Re/25) = 0.36%

Re/25
Re/5

1.14
1.12
1.10
1.08
1.06
1.04

t/T

Figure 6.16: Comparison of the computed Head for the initial computation (Re/5) and
the reduced Reynolds number computation (Re/25).

6.5

Grid Resolution Effect

Energy coefficient
The latest computation performed for a Reynolds number reduced by a factor 25
agrees well with the measurements. Therefore, the question regarding the grid resolution
influence arises. To begin with, one can compare the specific energy coefficient computed. The computed energy coefficients during 5 impeller revolutions are shown in Fig.
6.16. There is no significant change, as already the specific energy coefficient computed
in the initial computation is already close to the experiment with a relative difference
of 3.35%. For the second computation, the relative difference drops to 0.36%, which fits
precisely the experiment. Furthermore, the specific energy coefficient is a good measure
to ensure that the appropriate recovery of the flow is achieved. In the initial computation,
the flow is slightly too much decelerated compared to the experiment. However, such a
variable does not provide any information regarding the physics taking place in the guide
vane channels, but only assesses if the flow is first sufficiently accelerated and secondly
sufficiently decelerated.
Flow rate through a guide vane channel
To evaluate the difference taking place in the diffuser, one can monitor, as previously
introduced, the temporal evolution of the flow rate passing through each guide vane
channels. An example of such an evolution is shown in Fig. 6.17. For each comparison
between the initial and the second computation, the same guide vane channel is used.
Such representations allow to confirm that a significant difference is taking place between
the initial and the second computations. In Fig. 6.17 a), one can see that the flow rate
passing through a guide vane channel in the case of the second computation (Re/25)
experiences a entire passage of one stall cell. Regarding the Re/5 case, only the recovery
EPFL - Laboratory for Hydraulic Machines

84

6.5. GRID RESOLUTION EFFECT

1.8
Q*t a)
1.6

1.8
b)
Q*t
1.6

1.4

1.4

1.2

1.2

1.0

1.0

0.8

0.8
2 rev.

0.6
0.4
0

Re/25
Re/5
6

10

t/T

12

Re/25
Re/5

4 rev.

0.6
0.4
0

2 rev.
8
10

t/T

12

Figure 6.17: Comparison of the computed flow rate variation through the guide vane
throat for the initial computation (Re/5) and the reduced Reynolds number computation
(Re/25).
and the early phase of the stall are visible in the same duration. Fig. 6.17 b) shows
the stalled phase experienced by another guide vane channel. From this observation, it
can be concluded, that one of the main differences between the two computations and
therefore responsible for the difference of propagation speed, is taking place during the
stalled phase. Indeed, in the case of Re/5, a guide vane channel is in the stalled phase for
more than 4 impeller revolutions, whereas for the Re/25 case this phase lasts less than 2
impeller revolutions.
The comparison between the rate of increase and decrease of the flow rate shows a
difference but not as significant as for the stalled phase. The phase where the flow rate
is maximal is similar for the two cases, but in the case of Re/5 the maximum flow rate
is approximately Qt =1.7. Finally, using the two graphs, one can estimate the number
of impeller revolutions needed for one stall cell to propagate from one channel to the
downstream channel. It requires approximately 20 impeller revolutions, which results in
a propagation speed 60% slower for the Re/5 case.
Flow pattern at the throat section
The normal velocity in the downstream direction through the guide vane throat is
shown in Fig. 6.18 at an instant when the channel experiences the minimum flow rate for
the two different computations. It can be seen that a significant difference exists between
the two computations. The pattern shown in Fig. 6.18 b) is obtained from the flow
for a reduced Reynolds number as already introduced in Fig. 5.6 on page 59. For this
computation, the reverse flow is mainly located near the hub on the suction side. Figure
6.18 a) represents the flow pattern obtained with the initial computation. It results in a
better agreement with the experiment, as the reverse flow also occupies the whole pitch
of the guide vane channel near the hub. Therefore, further computations of the initial
calculation should be performed to identify the origin of the non-symmetry in the flow in
a symmetrical diffuser.
EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

26
05
0.1

Cn/U1e

0.3

053

0.2

632

0.052

0.2
1

52
63
2

0.26316
57 0.2
1053
8
0.10 9
0.15789
526
0.10526
00 0.052632
0.052632
.05263 2
0

0.0

0.0

0.1

0.10526

53

89

0.2

0.1

0
0

789

10

32

52
6

.0

0.5

57

0.5
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
s

63

0.15

0.1

0.4

0.6

0
52

0.1

0.26316
0.21053
0.2
0
0 .0526320.10520.615789
9
78
15
0.
0 0526320 10526
0 052632

89
57

1
0.

0.0

0.3
26
05
0.1

0.0

0.4

53
10

526

0.4

0.1

0.15789

0.3

0.26316

0.2

0.15789

0.2

0.10526

0.1

0.3

0.0526032
.10

52

0.0

0.4

b)

0
0.1

0.1

z/b2

0.5

0.2

0.2

0.2
63
16

0.3

0.20.210
0.052632
53 89 0.10526
6310.157
6
0.15789 0.10526
0.052632
0.0
5263
0
2
0
32
6
0.05
0.052 0526
2
6
32
0.1
0.1052
0
0.15789
6
0.05
0.15
2632
789
0.21053
0.2
0.1
10
05
53
2

0.21053

0.4

a)

0.15789

z/b2

0.5

85

0
0.5
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
s

0.1
0.2
0.3
0.4

Figure 6.18: Comparison of the normal (streamwise direction) velocity in the guide vane
throat section at the minimum flow rate: a) initial computation (Re/5), b) reduced
Reynolds number computation (Re/25).
Kinetic Energy Spectrum
To estimate how the turbulence is capture by the computation, one can draw the
kinetic energy spectra using the time history of the velocity at a specific location. The
location selected is the one near the leading edge of the guide vane on the pressure side
at the mid-span of the guide vane. The spectra obtained from the two computations are
shown in Fig. 6.19 and three distinct ranges with a slope of -1, -5/3 and -7 can be seen.
These ranges are, respectively, the production range, the inertial range and the dissipation
range. The comparison of the two spectra does not show significant differences. However,
in the case of the Re/5 computation, the inertial range is less pronounced than the Re/25
case and this could be attributed to a lack of dissipation. Because of the under-resolution
of the flow the energy is not sufficiently dissipated and, therefore, get accumulated in the
flow. However, such an estimation is only local and it is, therefore, difficult to draw a
overall conclusion.
Flow pattern at the spanwise mid-plane of the diffuser
In Fig. 6.20 the velocity fields are shown in diffuser passages at maximum and minimum flow rates. At the maximum flow rate, the main difference is in the size of the
recirculation zone. Both computations capture the recirculation, but in the case of the
original calculation the size is smaller. It is worth mentioning that the flow rate is not
identical, as the initial calculation has a 10 % higher flow rate. Furthermore, the distribution of the flow exiting the guide vane channel is different. From there, the flow can
exit the machine through the stay vane channel or through the gap between the trailing
edge of the guide vane and the leading edge of the stay vane. In standard operation of
the pump, the distribution is near 50%. But in the case of the initial computation the
flow is less likely to exit via the gap and most of the flow exiting the guide vane channel
goes to the stay vane channel. The flow rate in the gap can decrease until a flow rate of
EPFL - Laboratory for Hydraulic Machines

86

6.6. DISCUSSION

10-2

-1

E/U 1 l

-5/3

10

-4

-7

10-6
10-8
100

Re/5
Re/25

101

102

103

104

Figure 6.19: Comparison of the kinetic energy spectrum near the pressure side and the
leading edge of the guide vane at the mid-spanwise.
0.2Qo .
The situation where the channel is at its lowest flow rate shows that in both computations the minimum is reached when the stall cell occupies near the entire pitch of the
guide vane channel. However, the distance from the trailing edge to the separation point
is nearly the same. At that instant, it can be seen that the size of the recirculation zone
on the downstream channel differs. Indeed, the size for the reduced Reynolds number
computation is already bigger and, therefore, it modifies the flow in its vicinity. This
confirms that the rate of growth of a stall cell determines the propagation speed of the
stall cell.

6.6

Discussion

In this chapter, it was shown that the propagation mechanism of the rotating stall may
be regarded as the local growth and decay of a stall cell. The rate of fluctuation of the stall
cell governs the propagation speed. The growth of the stall cell reduces firstly the flow rate
passing through the guide vane channel and almost blocks the channel when the stall cell
reaches its maximum size, i.e. when it reaches the leading edge of the downstream guide
vane. The growth and decay are governed by the variation of the adverse pressure gradient
between the leading and the trailing edges of the guide vane and by the deviation of the
flow near the leading edge of a guide vane. Furthermore, as the recirculation is located
in a narrowed area, it induces a modification of the flow as soon as the recirculation zone
grows. Therefore, any changes in the size of the recirculation zone change the flow pattern
near the leading edge of the downstream guide vane. These changes are responsible for
the switch between the favorable and the unfavorable conditions for a recirculation zone
to grow. A schematic of the flow patterns in a guide vane channel is shown in Fig. 6.21.

EPFL - Laboratory for Hydraulic Machines

Chapter 6. Rotating Stall Analysis

87

Re/5
C/U1

1.00

Re/25
C/U1

1.00

0.75

0.75

0.50

0.50

0.25

0.25

0.00

0.00

HF

C/U1

1.00

C/U1

1.00

0.75

0.75

0.50

0.50

0.25

0.25

0.00

0.00

LF

Figure 6.20: Comparison of the computed flow fields in the guide vane channels at maximum (HF) and minimum (LF) flow rate for the initial computation (Re/5) and the
reduced Reynolds number computation (Re/25).

EPFL - Laboratory for Hydraulic Machines

88

6.6. DISCUSSION

a)

b)

FR

FR max

SP2

SP2

p
SP1

SP1

AOA

AOA0

c)

d)

FR

FR min

SP2

SP2

p max

p
SP1

SP1

AOA max

e)

AOA

f)

FR

FR

SP2

SP2

p
SP1

AOA

SP1

AOA0

Figure 6.21: Schematic of the variation of the flow within a guide vane channel.

EPFL - Laboratory for Hydraulic Machines

Part IV
Conclusions and Perspectives

EPFL - Laboratory for Hydraulic Machines

Chapter 7
Conclusions and Perspectives
7.1

Conclusions

A large scale computation of the rotating stall phenomenon was performed for the first
time using the Large Eddy Simulation technique. The case study selected is the reduced
scale model HYDRODYNA pump-turbine operating at part load condition (76% BEP)
in pumping mode. The computations were performed using the open source software
FrontFlow/Blue and the PRIMEHPC FX10 supercomputer of the University of Tokyo.
To ensure the viability of this investigation, a feasibility study was performed beforehand.
It was concluded that, with the computing power currently available, such an investigation
is feasible if the Reynolds number is reduced. If the reduction is by a factor 5 (Re/5)
regarding the experiment, the required mesh features 7 billions of elements. If it is further
reduced by a factor 5 (Re/25) the size of the required mesh reduced to 80 millions elements.
However, the flow initialization performed for the Re/5 case with a mesh composed of
85 million elements showed that the unsteady rotating stall phenomenon is already well
captured. Four distinct zones, i.e. four guide vane channels experiencing high outward
flows compared to the others, showed the presence of the instability. The estimation
of the propagation speed of the stall cells was done by monitoring the instantaneous
pressure in the clearance area between the impeller exit and the guide vane leading edge
and by monitoring the flow rate through the guide vane channels. The results showed
a propagation in the same direction as the impeller, where the propagation speed is
approximately 60% slower than the measured value on the reduced scale model. Such a
difference is associated to the grid resolution as a mesh of 7 billions elements is required
to compute such a Reynolds number.
The simulation for the second Reynolds number reduction (Re/25) produced an accurate and satisfactory outcome. The number of stall cells was correctly captured and the
propagation speed was only 3.4% slower than the measured value, compared to the 60% in
the initial computation. Furthermore, with the difference of 0.36% regarding the measurements, the computed specific energy coefficient showed a better agreement than in the
initial computation (3.35% difference). It is concluded that the LES approach is a powerful tool, where the traditional RANS approach reached its limit. Indeed, former studies
indicated that the rotating stall is not accurately captured by a RANS computation with
the k SST model. For this investigation, the flow rate had to be increased by 23% in
order to observe the presence of the stall cells. Furthermore, for this higher flow rate, the
EPFL - Laboratory for Hydraulic Machines

92

7.2. PERSPECTIVES

computed propagation speed was 38% faster compared to the experiment. Despite the
good agreement of the LES results with the measurement data at Re/25, the following
inconsistencies are observed. The flow patterns in the diffuser displayed substantial differences with respect to the experiment; the size and the recirculation area were not well
represented in the computation. On the other hand, the evolution of the flow rate within
the guide-vane channels agrees well with the measurements. Therefore, the rotating stall
phenomenon was globally well captured and the calculation could, therefore, be used for
further investigations.
This analysis showed that the stall mechanism is driven by the growth and the decay
of the stall cell. Furthermore, it was shown that both the rate of growth and the rate
of decay are almost constant, where the rate of growth is higher than the rate of decay.
The stall cell is located at the guide vane trailing edge on the suction side. As soon as it
starts growing, the flow within the guide vane channel is deviated toward the downstream
channel. The growth of the stall cell is such that at its maximum volume, the stall size
occupies the entire guide vane pitch. However, the orientation of the flow at the guide
vane inlet and the decrease of the adverse pressure gradient result in the impossibility
for such a cell to sustain. The cell size decreases until the favorable conditions for flow
recirculation are met again.

7.2

Perspectives

The recent development of computing power paves the way for the LES approach
towards industrial applications. However, in the study, the proper mesh size is only
reached through a reduction of the Reynolds number. However, on the way of the LES
approach to become an industrial standard, further development is required. First of all,
this approach depends upon a dense, high quality mesh in order to guarantee the stability
and the accuracy of the results. This makes the mesh generation very time consuming.
An existing alternative is the SPH method, which does not require a mesh generation but
suffers from its high demand in CPUs. Therefore, a promising development in the future
is the automatic grid generation. First step relies on a coarse Cartesian grid, which divides
each cell into 4 or 8 sub-cells, depending on the 2D or 3D nature of the problem, for a
better discretization of the geometry or a region with high gradient of the flow variables.
This mesh is known as voxel mesh and the success of this approach would be a decisive
breakthrough for the LES to be used as a standard tool in the industries.
Another issue combines both the computing resources and the storage capacity. To
perform large-scale LES computations, the traditional local clusters accommodated by
the industry to the present day are by far too small. These clusters consist of a few
hundreds cores at maximum, whereas one to three order of magnitude more is be required.
Therefore, specific HPC centers have to be built and dedicated to industrial and academic
research. Such centers already exist and opened recently to the industry. Two examples
are the K computer (RIKEN) in Japan and the Vulcan supercomputer (LLNL) in the USA,
both being accessible to the industry and possessing two petaflops scale computing power.
The industrial LES perspectives are thus encouraging. A further issue which remains to
be solved concerns the data storage and the transport. A small LES computation features
one million to 10 million grid points. Therefore, the data volume from the computation
reaches easily the size of tera Bytes (TB). The present network capabilities may be used
EPFL - Laboratory for Hydraulic Machines

Chapter 7. Conclusions and Perspectives

93

to manage data of the size of a few GB. If the volume reaches a few TB, a portable hard
disk drive may still be used. Above 4 TB, which is today the largest portable hard disk,
it becomes increasingly complicated to move the data. Therefore, the analysis has to be
done on-site of the HPC center, which is possible but for a limited amount of time.
Finally, further investigations should be performed to identify the reason for the two
different propagation speeds resulting from the two computations. The reason probably
lies within the computation of the boundary layer development. Finally, it would be
interesting to evaluate the behavior of boundary layer for a fluctuating flow rate in both
computations.

EPFL - Laboratory for Hydraulic Machines

Bibliography

EPFL - Laboratory for Hydraulic Machines

Bibliography
[1] Arndt, N., Acosta, A. J., Brennen, C., and Caughey, T. K. Experimental
investigation of rotor-stator interaction in a centrifugal pump with several vaned
diffusers. Journal of Turbomachinery 112, 1 (1990), 98108.
[2] Arndt, N., Acosta, A. J., Brennen, C. E., and Caughey, T. K. Rotor-stator
interaction in a diffuser pump. Journal of Turbomachinery 111 (1988), 213221.
[3] Braun, O. Part Load Flow in Radial Centrifugal Pumps. PhD thesis, EPFL, 2009.
[4] Byskov, R. K., Jacobsen, C. B., and Pedersen, N. Flow in a Centrifugal
Pump Impeller at Design and Off-Design ConditionsPart II: Large Eddy Simulations. Journal of Fluids Engineering 125, 1 (Jan. 2003), 7383.
[5] Chapman, D. K. Computational aerodynamics development and outlook. AIAA
journal 17, 12 (1979), 12931313.
[6] Choi, H., and Moin, P. Grid-point requirements for large eddy simulation: Chapmans estimates revisited. Physics of Fluids 24, 1 (2012), 011702.
[7] Clark, R. A., Ferziger, J. H., and Reynolds, W. C. Evaluation of subgridscale models using an accurately simulated turbulent flow. Journal of Fluid Mechanics 91, 01 (1979), 116.
tais, O. On the influence of coherent structures upon
[8] Da Silva, C., and Me
interscale interactions in turbulent plane jets. Journal of Fluid Mechanics 473, 1
(2002), 103145.
[9] de Jager, B. Rotating stall and surge control: A survey. In 34th IEEE Conference
on Decision and Control (1995), vol. 2, IEEE, pp. 18571862.
[10] Deardorff, J. A numerical study of three-dimensional turbulent channel flow at
large Reynolds numbers. Journal of Fluid Mechanics 41, 2 (1970), 453480.
[11] Dennard, R., and Gaensslen, F. Design of ion-implanted MOSFETs with very
small physical dimensions. IEEE Journal of Solid-State Circuits 9, 5 (1974), 256268.
[12] Desai, D. M., Bradicich, T. M., Champion, D., Holland, W. G., and
Kreuz, B. M. BladeCenter system overview. IBM Journal of Research and Development 49, 6 (2005), 809821.
[13] Dongarra, J. The LINPACK benchmark: past, present and future. Concurrency
and Computation: practice and experience 15, 9 (2003), 803820.
EPFL - Laboratory for Hydraulic Machines

98

BIBLIOGRAPHY

[14] Emmons, H. W. A survey of stall propagation: experiment and theory. Journal of

Basic Engineering 81 (1959), 409416.
[15] Fujitsu. A new Generation 16-core Processor for Supercomputing SPARC64 IXfx.
2011.
[16] Germano, M., Piomelli, U., and Moin, P. A dynamic subgrid-scale eddy
viscosity model. Physics of Fluids 3, 7 (1991), 17601765.
[17] Gropp, W., Lusk, E., and Skjellum, A. Using MPI: portable parallel programming with the message-passing interface. the MIT Press, 1999.
[18] Gu, G., Sparks, A., and Banda, S. S. An overview of rotating stall and surge
control for axial flow compressors. IEEE Transactions on Control Systems Technology
7, 6 (1999), 639647.
[19] Guo, Y., Kato, C., Wang, H., and Yamade, Y. A Fractional-Step Large-Eddy
Simulation Method in Overset Finite-Element Mesh and Its Application to Turbomachinery. Nihon Kikai Gakkai Ryutai Kogaku Bumon Koenkai Koen Ronbunshu
(CD-ROM) 82 (2004), 120123.
[20] Hasmatuchi, V., Farhat, M., and Avellan, F. Hydrodynamics of a PumpTurbine Operating at Off-Design Conditions in Generating Mode. PhD thesis, EPFL,
2012.
[21] Hunt, J. C., and Morrison, J. F. Eddy structure in turbulent boundary layers.
European Journal of Mechanics - B/Fluids 19, 5 (2000), 673694.
[22] IBM Blue Gene Team. Overview of the IBM Blue Gene/P project. IBM Journal
of Research and Development 52, 1.2 (2008), 199220.
[23] Inoue, M., and Cumpsty, N. A. Experimental study of centrifugal impeller
discharge flow in vaneless and vaned diffusers. Journal of engineering for power 106,
2 (1984), 455467.
[24] Jones, W., and Launder, B. The prediction of laminarization with a twoequation model of turbulence. International Journal of Heat and Mass Transfer
15, 2 (Feb. 1972), 301314.
[25] Karypis, G., and Kumar, V. A parallel algorithm for multilevel graph partitioning and sparse matrix ordering. Journal of Parallel and Distributed Computing 48,
1 (1998), 7195.
[26] KATO, C., Kaiho, M., and Manabe, A. An Overset Finite-Element LargeEddy Simulation Method With Applications to Turbomachinery and Aeroacoustics.
Journal of Applied Mechanics 70, 1 (2003), 3243.
[27] Kato, C., Mukai, H., and Manabe, A. Large-eddy simulation of unsteady flow
in a mixed-flow pump. International Journal of Rotating Machinery 9, 5 (2003),
345351.
EPFL - Laboratory for Hydraulic Machines

BIBLIOGRAPHY

99

[28] Kerr, R., Domaradzki, J., and Barbier, G. Small-scale properties of nonlinear
interactions and subgrid-scale energy transfer in isotropic turbulence. Physics of
Fluids 8, 1 (1996), 197208.
[29] Kline, S. On the nature of stall. Journal of Basic Engineering 81, 3 (1959), 305320.
[30] Kline, S. J., Reynolds, W. C., Schraub, F. A., and Runstadler, P. W.
The structure of turbulent boundary layers. Journal of Fluid Mechanics 30, 04 (1967),
741773.
[31] Kolmogorov, A. N. The local structure of turbulence in incompressible viscous
fluid for very large Reynolds number. Dokl. Akad. Nauk SSSR 30, 4 (1941), 301305.
[32] Koseff, J. R., and Street, R. L. The Lid-Driven Cavity Flow: A Synthesis
of Qualitative and Quantitative Observations. Journal of Fluids Engineering 106, 4
(1984), 390398.
[33] Krain, H. A study on centrifugal impeller and diffuser flow. Journal of Engineering
Power 103 (1981), 688697.
hringer, K., and Pap, E. Time-resolved particle imaging ve[34] Krause, N., Za
locimetry for the investigation of rotating stall in a radial pump. Experiments in
Fluids 39, 2 (2005), 192201.
[35] Launder, B., and Sharma, B. Application of the energy-dissipation model of
turbulence to the calculation of flow near a spinning disc. Letters in Heat and Mass
Transfer 1, 2 (1974), 131137.
[36] Lesieur, M., and Metais, O. New trends in large-eddy simulations of turbulence.
Annual Review of Fluid Mechanics 28 (1996), 4582.
[37] Lilly, D. A proposed modification of the Germano subgrid-scale closure method.
Physics of Fluids 4, 3 (1992), 633635.
[38] Lucius, A., and Brenner, G. Unsteady cfd simulations of a pump in part load
conditions using scale-adaptive simulation. International Journal of Heat and Fluid
Flow 31, 6 (2010), 11131118.
[39] Meneveau, C. Statistics of turbulence subgrid-scale stresses: Necessary conditions
and experimental tests. Physics of Fluids 6, 2 (1994), 815833.
[40] Menter, F. R. Zonal Two Equation k- Turbulence Models for Aerodynamic
Flows. AIAA Paper 1993-2906 (1993).
[41] Menter, F. R. Two-Equation Eddy-Viscosity Turbulence Models for Engineering
Applications. AIAA Journal 32, 8 (1994), 15981605.
[42] Moore, G. Cramming more components onto integrated circuits.
[43] Moore, G. Progress in Digital Integrated Electronics. IEEE Int. Electron Devices
Meeting Tech. Dig. 21 (1975), 1113.
EPFL - Laboratory for Hydraulic Machines

100

BIBLIOGRAPHY

[44] ONeil, J., and Meneveau, C. Subgrid-scale stresses and their modelling in a
turbulent plane wake. Journal of Fluid Mechanics 349 (1997), 253293.
[45] Pedersen, N., Larsen, P. S., and Jacobsen, C. B. Flow in a Centrifugal
Pump Impeller at Design and Off-Design ConditionsPart I: Particle Image Velocimetry (PIV) and Laser Doppler Velocimetry (LDV) Measurements. Journal of
Fluids Engineering 125, 1 (Jan. 2003), 6172.
[46] Piomelli, U., and Balaras, E. Wall-layer models for large-eddy simulations.
Annual review of fluid mechanics 34, 1 (2002), 349374.
[47] Pope, S. B. Turbulent Flows. Cambridge university press, 2000.
[48] Robinson, S. Coherent motions in the turbulent boundary layer. Annual Review
of Fluid Mechanics 23, 1 (1991), 601639.
[49] Roth, S., Farhat, M., and Avellan, F. Fluid-Structure Coupling Effects on
the Dynamic Response of Pump-Turbine Guide Vanes. PhD thesis, EPFL, 2012.
[50] Sagaut, P. Large Eddy Simulation for Incompressible Flows: An Introduction.
Springer, 2006.
[51] Sano, T., and Nakamura, Y. Alternate blade stall and rotating stall in a vaned
diffuser. JSME International Journal Series B 45, 4 (2002), 810819.
[52] Sano, T., Yoshida, Y., Tsujimoto, Y., Nakamura, Y., and Matsushima,
T. Numerical Study of Rotating Stall in a Pump Vaned Diffuser. Journal of Fluids
Engineering 124, 2 (June 2002), 363370.
[53] Schlichting, H., and Kestin, J. Boundary-layer theory. McGraw-Hill New York,
1968.
[54] Sinha, M., and Katz, J. The onset and development of rotating stall within a centrifugal pump with a vaned diffuser. Paper FEDSM99-7314, The 3rd ASME/JSME
Joint Fluid Engineering Conference (1999), 17.
[55] Sinha, M., and Katz, J. Quantitative Visualization of the Flow in a Centrifugal
Pump With Diffuser VanesI: On Flow Structures and Turbulence. Journal of Fluids
Engineering 122, 1 (Mar. 2000), 97107.
[56] Sinha, M., Pinarbasi, A., and Katz, J. The Flow Structure During Onset and
Developed States of Rotating Stall Within a Vaned Diffuser of a Centrifugal Pump.
Journal of Fluids Engineering 123, 3 (Sept. 2001), 490499.
[57] Smagorinsky, J. General circulation experiments with the primitive equations.
Monthly weather review 91, 3 (1963), 99164.
[58] Tamura, A., Kikuchi, K., and Takahashi, T. Residual cutting method for
elliptic boundary value problems. Journal of Computational Physics 137, 2 (1997),
247264.
EPFL - Laboratory for Hydraulic Machines

Biography

101

[59] Torbergsen, E. A. Impeller/Diffuser Interaction forces in Centrifugal Pumps.

PhD thesis, Norwegian University of Science and Technology, 1998.
[60] Uzun, A., Blaisdell, G., and Lyrintzis, A. Sensitivity to the Smagorinsky
constant in turbulent jet simulations. AIAA journal 41, 10 (2003), 20772079.
[61] Van der Vorst, H. A. Bi-CGSTAB: A fast and smoothly converging variant of
Bi-CG for solution of non-symmetric linear systems. SIAM J. Sci. Stat. Comput. 13,
2 (1992), 631644.
[62] Wilcox, D. C. Reassessment of the scale-determining equation for advanced turbulence models. AIAA Journal 26, 11 (1988), 12991310.
[63] Wilcox, D. C. Turbulence modeling for CFD, vol. 3. DCW industries La Canada,
2006.
[64] Wuibaut, G., Dupont, P., Caignaert, G., and Stanislas, M. PIV measurements in the impeller and the vaneless diffuser of a radial flow pump in design
and off-design operating conditions. Journal of Fluids Engineering 124, 3 (2002),
791797.
[65] Zang, Y., and Street, R. L. A composite multigrid method for calculating unsteady incompressible flows in geometrically complex domains. International Journal
for Numerical Methods in Fluids 20, 5 (1995), 341361.
[66] Zobeiri, A., and Avellan, F. Investigations of time dependent flow phenomena in
a turbine and a pump-turbine of Francis type: rotor-stator interactions and precessing
vortex rope. PhD thesis, EPFL, 2009.

EPFL - Laboratory for Hydraulic Machines

Biography

EPFL - Laboratory for Hydraulic Machines

Olivier PACOT
LMH - EPFL
Av. de Cour, 33 bis
CH-1007 Lausanne (VD)
Tel. prof.:
Portable:
E-mail prof.:
E-mail prive:

+41 21 693 2563

+41 78 681 3526
[email protected]
[email protected]

Ne le 22 janvier 1983
Nationalite franco-suisse
Celibataire

FORMATION
Doctorat `
es sciences techniques
2010-2014:

Ecole
polytechnique federale de Lausanne (EPFL), Suisse
Dipl.-Ing. M
ecanicien
2003-2009:

Ecole polytechnique federale de Lausanne (EPFL), Suisse

Coll`
ege Madame de Sta
el, Suisse
1998-2003:
Maturite federale

EXPERIENCES

Ecole
polytechnique f
ed
erale de Lausanne (EPFL), Suisse.
2009-2014:
Assistant doctorant au Laboratoire de Machines Hydrauliques
The University of Tokyo, Japon.
2012-2014:
Visiting Researcher Associate
Bobst SA, Roseland NJ, USA
2007:
Stage de 3 mois
Creation dune base de donnees repertoriant tout essai sur machine
Elaboration dun protocol pour test de performance
Test dun materiau demagnetisant les feuilles plastiques

COMPETENCES
INFORMATIQUES
Syst`emes:
Windows, Linux.
Program.:
C, C++, Fortran, Python, Shell, MATLAB, Labview
FFB, CFX, Fluent, ICEM, OpenFoam, VisIt, Paraview.
CFD:
Bureautique: MS-Office, LaTeX.
Adobe Illustrator
Multimedia:
BREVETS
Instructeur suisse
Ski:
Aviation (PPL):En cours
Parapente:
En cours
SPORTS & LOISIRS
Sports:
Kitesurf, Parapente, Telemark, Volley, aviron.
Lecture, Cuisine, Aviation.
Loisirs:
LANGUES
Francais
Allemand
Anglais

Langue maternelle.
Intermediaire.
Avance.