You are on page 1of 8

43

Optimization method for the design of axial


hydraulic turbines

A Lipej
Turboins† titut, Rovs† nikova 7, 1210 Ljubljana, Slovenia

Abstract: Computational  uid dynamics (CFD) is becoming an increasingly reliable tool for the design of
water turbines. Using different CFD codes, it is possible to Ž nd out and compare criteria for classifying
runner blade geometry regarding the strengths of their characteristics. The Ž nal decision of runner geometry,
with demanding energetic and cavitation characteristics, always remains for the design engineer. To reach the
Ž nal result, the engineer has to compare the  ow analysis results of a great number of different geometries.
To replace a part of this work, an optimization algorithm has been developed. This optimization procedure
helps to check many more geometries with less human work. In this paper, a multiobjective genetic algorithm
for the design of axial runners is presented. For the  ow analysis, the CFX-TASC  ow code, with a standard
k–e turbulence model, has been used, because the code enables calculation within a rotating frame of
reference. For design of the initial geometry, within the optimization procedure, a special program has been
developed, which makes it possible to start the optimization procedure with a relatively high level of
efŽ ciency and transforms prescribed genetic parameters to the runner geometry.

Keywords: water turbines, efŽ ciency, cascade, numerical method, genetic algorithm

NOTATION a angle of absolute velocity (deg)


b chord angle of the runner blade (deg)
A area (m2) Z efŽ ciency
c absolute velocity (m/s) W angle of relative velocity (deg)
D diameter (m) r density (kg/m3)
Dz maximum diameter of the runner blade (m) c pressure number
E speciŽ c energy (J/kg) o angular velocity (s¡1)
f/l relative curvature of the proŽ le
F Ž tness function
gj constraints 1 INTRODUCTION
Kcm meridional velocity coefŽ cient
Ku2 outlet vortex coefŽ cient In the last few years, some researchers have started to work
Kc relative pressure number with an automated design of hydraulic turbines using
l/t chord–pitch ratio different optimization methods [1]. A lot of work has been
M torque (N/m) done in the Ž eld of automated design of a single airfoil, but
n rotational speed (s¡1) for the complete turbine optimization procedure there have
p pressure (N/m2) been only a few studies.
Q  owrate (m 3/s) This paper deals with numerical  ow analysis and design
S, T random numbers in the interval [0, 1] of optimal hydraulic shapes for axial water turbines. The main
u circumferential velocity (m/s) task is to Ž nd the shape and exact position of each proŽ le
va, vr, vu cylindrical velocity components (m/s) from the hub to the maximum radius of the runner (Fig. 1).
w relative velocity (m/s) Most real-life design procedures are complex tasks that
x, y vectors of variables (genotype) have to deal with multidisciplinary environments and not
always with clearly deŽ ned targets. The optimization process
must consider several different, usually con icting, objectives.
The design of axial runners is a very complex task, and
The MS was received on 8 April 2003 and was accepted after revision for one single criterion does not provide the desired result.
publication on 11 November 2003. In the present research, multiobjective genetic algorithm
A02803 # IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy
44 A LIPEJ

Fig. 1 Four-bladed axial runners

optimization with two criteria has been used, and the results been made. Accuracy of the numerical results is important
have been compared with measurements. The key points of in order to obtain a reliable objective function for the design
the genetic algorithms are the operators used for selection optimization procedure. The last part presents a multiobjec-
and reproduction which have a great in uence on the tive, genetic optimization algorithm, coupled with the
robustness and the efŽ ciency of the algorithm. design program and numerical  ow analysis.
The main purpose of the  ow analysis is to predict  ow
properties and losses and to obtain reliable criteria for
2 DESIGN OF AXIAL RUNNERS
classifying runners with good or bad characteristics. From
the pressure distribution on the runner blades, the torque on
the shaft was calculated. In the runner, most of the differ- Development of the new runner blade shapes can be done
ence in total pressure is converted to runner work, while a using different direct or inverse design methods [2]. The
small part presents  ow energy losses. The turbine efŽ - development of a new runner starts with the deŽ nition of
ciency can be calculated by the following parameters:
(a) nominal and maximal head,
X Mo
Mˆ pi bxi Ayi ¡ yi Axi c and Zˆ (b) nominal power,
i
rQ DE (c) predicted efŽ ciency,
(d) predicted cavitation coefŽ cient.
where M is the torque on the shaft, p is the static pressure, x
The simpliŽ cation of the  ow analysis is based on the
and y are components of distance from the axis, Ax and Ay
fundamental features of the  ow in axial turbines. Inside
are area projections, index i represents all elements on the
the channel, in front of the runner and downstream of
blade surface, o is the angular velocity and DE is obtained
the runner, the  ow is axisymmetric, especially for the
from the difference between total pressure, pt at the domain
optimal operating point. Inside the runner the stream
inlet and outlet
surfaces are cylindrical [3], which is the consequence of
Á ! the axisymmetric  ow. Taking into account the hypothesis
1 X X
DE ˆ ptj Qj ¡ ptj Qj of cylindrical stream surfaces, the three-dimensional  ow
rQ j(inlet) j(outlet) can be analysed as two-dimensional  ow through the axial
runners (Fig. 2), where c is absolute velocity, w is relative
where Q is the  owrate and index j represents all the velocity and u represents the rotational speed.
elements on the inlet and outlet cross-sections.
The design procedure is based on the use of theoretical
hydrodynamic cascade characteristics for potential  ow and
empirical  ow corrections on account of viscosity effects.
The resulting cascade and proŽ le camber means the optimal
solution for given initial parameters. The design program is
used Ž rst of all to start the optimization of the runner blade
geometry with relative high efŽ ciency, which means a
signiŽ cant reduction in the CPU time.
In this paper the theory of axial turbines and the design
procedure for axial runners are presented Ž rst. The second
part deals with three-dimensional turbulent  ow analysis of
different speciŽ c speed axial runners for various operating
regimes. A comparison of energetic and kinematic charac- Fig. 2 Velocity traingles at the inlet (index 1) and outlet
teristics between numerical and experimental results has (index 2) of the runner
Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy A02803 # IMechE 2004
OPTIMIZATION METHOD FOR THE DESIGN OF AXIAL HYDRAULIC TURBINES 45

Important coefŽ cients responsible for the inlet and outlet


conditions are the meridional velocity coefŽ cient
cm
Kcm ˆ
cmav

where cm is the meridional velocity and cmav is the average


meridional velocity, and the outlet vortex coefŽ cient
cu2
Ku2 ˆ
Dcu

where cu2 is the circumferential component of absolute velocity Fig. 4 Comparison of (a) calculation and (b) measurement of
at the outlet and Dcu is the difference between the inlet and relative efŽ ciency for three different runners (r1, r2, r3)
outlet circumferential components of absolute velocity. The in the same operating regime
value of the coefŽ cient Ku2 is zero, or varies linearly from ¡0.4
near the hub to 0.4 on the peripheral cylindrical section.
A computer program for the design of the optimal shape of carried out on the model of a four-bladed Kaplan turbine in
axial runner [4] also prepares the Ž les necessary for the grid
the Turboins† titut.
generationprocedure with the CFX-TASCgrid code [5]. Before
According to the experimental results, behind the hub
the solver can start the calculation,some data about the number
there is an area of small velocities, but in the case of
of elements and density of the elements have to be determined. numerical analysis this area is smaller. This is the conse-
quence of the horizontal pier in the draft tube, which was not
3 NUMERICAL FLOW ANALYSIS taken into account in the numerical analysis.
The numerically and experimentally obtained efŽ ciencies
The numerical grids for all the calculations through the for three different runners are also presented in Fig. 4. The com-
runner consist of 32 000 nodes. The computational domain parison shows that using the numerical results for classifying
is reduced to the  owŽ eld between two runner blades, a part the designed runners is successful. The efŽ ciency measure-
in front of the runner and the Ž rst part of the diffuser. The ment can be predicted with up to 0.3 per cent accuracy;
computational domain between the two runner blades can however, the accuracy of numerically obtained efŽ ciency is
be analysed because periodical boundary conditions can be about 2 per cent.
used. The  ow analysis in the axial runner can be started
when the boundary condition is prepared.
4 GENETIC ALGORITHM FOR
Comparing the numerical results with experimental ones
MULTIOBJECTIVE OPTIMIZATION
is a good test for the accuracy of the numerical method. On
the test rig installed in the Turboins† titut in Ljubljana, many
The basic idea of the optimization procedure is to Ž nd the
energetic and cavitation measurements were carried out,
shape of the axial runner (Figs. 1 and 10) that gives the
including some measurements of the  ow kinematics at
required energetic and cavitation characteristics. After deŽ n-
the outlet of the runner (Fig. 3). These measurements were
ing the design operating point, all the necessary parameters
have to be determined. Firstly, cylindrical sections are
selected. Then, the parameters are given as a function of
the relative diameter:
(a) the meridional velocity coefŽ cient (two variables),
(b) the outlet vortex coefŽ cient (two variables),
(c) the chord–pitch ratio (two variables),
(d) the relative proŽ le maximum thickness (two variables),
(e) the camber position (two variables).
All the parameters can be either linear or a higher-order
function of the relative diameter. The optimization has
therefore been limited to the treatment of ten variables.
For the genetic algorithm, the binary string (Table 1) for
the runner shape has to be determined. This string consists
Fig. 3 Comparison of experimentally and numerically of prescribed parameters that can be converted to the runner
obtained velocity components at the outlet of the geometry using the design program so that the runner is
runner: va , vr , vu —numerical results; va m , vr m , vu m — prepared for evaluation [6]. The shape of the runner alone is
experimental results not enough to provide all the characteristics required, and so
A02803 # IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy
46 A LIPEJ

Table 1 Example of genotype optimizations where the objective becomes a weighted


1 combination of the objectives, called the utility function U
4 CoefŽ cient of meridional velocity
8 n
X
7 CoefŽ cient of outlet vortex Uk ˆ Wik Fi (x)
3 iˆ1
4 Chord–pitch ratio
4 where x is the vector of variables and Wi represents the
9 Relative thickness
weights for the objectives Fi.
8
Positive of maximal curvature While traditional optimization algorithms require the use
2
of a utility function, the particular structure of genetic
algorithms (GAs) can face the multiobjective optimization
problem in a more direct way, developing populations in
numerical  ow analysis is necessary to obtain the criteria for which the diversities follow the con icting objectives.
evaluation. The efŽ ciency and relative pressure number can Pareto GA algorithms mainly differ from classical GA in
be used as criteria of runner successfulness. the selection process, even though other speciŽ c operators
The multiobjective optimization [7, 8] problem can be might be constructed. In the present case, multi-directional
expressed as follows crossover has been used. When a multiobjective optimiza-
tion task has to be performed, the Ž tness differences can be
max Fi (x) for i ˆ 1, n replaced with the differences in the respective scalar
gj < 0 for j ˆ 1, m products of the Ž tness vector in the direction given by the
Ž tness of the individual to be reproduced
and it is obvious that, in general, the solution is not unique if
x ˆ xi ‡ S ¢ sign(F ¡ F1 ) ¢ (xi ¡ xi1 )
the functions are not linearly dependent.
With the introduction of the Pareto dominance concept, it ‡ T ¢ sign(F ¡ F2 ) ¢ (xi ¡ xi2 )
is possible to divide any group of solutions into two
subgroups: the dominated and the non-dominated where
subgroups. Solutions belonging to the second group are
the ‘efŽ cient’ solutions, i.e. the ones for which it is not F ˆ Fi £ Fi, F1 ˆ F i1 £ F i F2 ˆ F i2 £ F i
possible to increase any objective value without adversely
The selection of the individual i1 and i1 can be done using
affecting the values of the remaining objectives.
any available selection scheme. In this way, global informa-
In more formal terms, and in the case of maximization
tion about the Ž tness landscape is retained in the whole
problems, it is possible to say that solution x dominates y (the
population, while each individual maintains and improves
Pareto dominance concept) if the following relationship is true
its specialization through the evolution.
x >p y () (8i Fi (x) > Fi ( y)) \ (9j Fj (x) > Fj (y))
5 RESULTS OF OPTIMIZATION
Under the strict continuity and derivability hypothesis,
classical optimization algorithms are capable of Ž nding the For each optimization procedure the limits of the initial
optimal value only in the single objective case, and therefore parameters are prescribed. The results in Fig. 5a show an
the problem of Ž nding the group of non-dominated solu- example of optimization where only small changes in the
tions (the Pareto set) is reduced to several single objective initial parameters are allowed. The highest  uctuation in

Fig. 5 EfŽ ciency distribution for runners in all generations: (a) small changes in initial parameters;
(b) non-linear parameter distribution
Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy A02803 # IMechE 2004
OPTIMIZATION METHOD FOR THE DESIGN OF AXIAL HYDRAULIC TURBINES 47

Fig. 6 (a) Distribution of b and (b) distribution of f/l for all runners in the Ž rst generation

efŽ ciency is in the Ž rst generation; for later generations the design parameters. In Fig. 7b the chord–pitch ratio is
results are almost constant. The possibility of obtaining an presented. In the optimization procedure each parameter
optimal solution in such a case is very small, because the has a different in uence on the efŽ ciency, which is why
in uence of initial conditions is dominant. In further analy- some parameters remain equal through most generations.
sis the limits of the initial parameters are different, and also This can be seen in Fig. 7b, where the function l/t is very
a non-linear distribution of parameters can be used. similar up to the eighth generation and only in the last few
The distribution of efŽ ciency (Fig. 5b) and pressure generations do signiŽ cant differences occur. From the
coefŽ cient shows high  uctuations in the Ž rst few genera- difference in l/t in the last generations, some conclusions
tions. In this case (b) the limits of initial parameters are about the number of generations can be obtained. To obtain the
more distant than in case (a), and with a non-linear para- optimal solution, the number of generations has to be higher.
meter distribution. This can also be shown with some During the optimization procedure, the genotypes are
geometrical characteristics of the runner blade. different. In the Ž rst generation, when the parameters are
In Fig. 6 two main parameters are shown: the skewness of chosen randomly, the difference from average values is
the blade, b, and the relative curvature of the proŽ les, f /l, for quite large. In the following generations the situation
all runners in the Ž rst generation. A difference in  uctua- improves. In the last generation, the average value for all
tions between the Ž rst and last generation is obtained in all ten parameters is almost constant.
optimization cases. As a second criterion, the relative pressure number is
In all cases of the optimization procedure there are ten taken into account (Figs 8 and 9). The relative pressure
individuals in each generation. To keep the maximum CPU number is deŽ ned by the equation
time for one optimization procedure within about 2 days, the
number of generations can vary from 16 to 20. cnum
Kc ˆ 1 ¡ 1 ¡
The distribution of the b angle for all runners in the last co
generation is so uniform that the difference in Fig. 7a is
almost invisible. The design procedure of the runner can where cnum is a numerically obtained pressure number
also be presented by diagrams showing the change in the and co is the prescribed optimal value of the pressure

Fig. 7 (a) Distribution of b for all runners in the last generation and (b) distribution of the chord–pitch ratio
for six generations
A02803 # IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy
48 A LIPEJ

Fig. 8 (a) Distribution of efŽ ciency and (b) distribution of relative pressure number for all individuals

Fig. 9 (a) Distribution of average efŽ ciency and (b) distribution of relative pressure number for all generations

Fig. 10 (a) Comparison of shapes for all proŽ les near the hub and at maximum radius in the Ž rst (a), tenth (b)
and last generation (c), shown as an evolution process
Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy A02803 # IMechE 2004
OPTIMIZATION METHOD FOR THE DESIGN OF AXIAL HYDRAULIC TURBINES 49

Fig. 11 Pressure distribution for the designed runner from the sixth (a) and last generation (b)

number. The relative pressure number is expressed by the The design program, together with experience gained,
equation makes it possible to start the optimization procedure for the
runner with a high level of efŽ ciency. It is developed
E especially for use with numerical  ow analysis and an
co ˆ
(p2 =2)n2 D2 optimization algorithm. The design program for axial
runners is responsible for transformation from genotype to
As a result of the optimization procedure, the comparison geometry data [9].
of geometric parameters of runner blades through all To ascertain which geometry has good or bad char-
generations is presented. In Fig. 10 the shapes of two acteristics, numerical  ow analysis inside the runner is
proŽ les for each individual from different generations are used, and one of the possible presentations of the results
presented. Each runner is presented by the proŽ le near is the pressure distribution (Fig. 11). From a great number
the hub and the proŽ le at the maximum radius (Fig. 1). In of analysed runners with different speciŽ c speeds, data for
the Ž rst half of the generation, all of the parameters are the inlet meridional velocity and the outlet vortex have
highly variable. After the Ž rst half of the generation, the been obtained. The calculated values can be used instead
blade angles do not change any further, but the para- of the theoretical ones necessary for runner blade design.
meters responsible for length and maximum curvature still Using the calculated parameters, the initial results for the
vary. Figure 10 presents all geometries for each genera- energetic characteristics of the runners are better than
tion, but the difference between cases a to c is that in the when theoretical values are used. The Ž nal results for
Ž rst generation (a) the geometries are different, in the the efŽ ciency of the initial runner and the optimized
tenth generation some runners are already very similar (b) runner, obtained by model measurements, are presented
and in the last generation (c) the differences are so small in Table 2.
that proŽ les overlap each other. Also, the pressure distri-
bution for the geometry from different generations is
compared. 6 CONCLUSIONS

Table 2 Difference of efŽ ciency for This paper describes the complete optimization procedure
initial and Ž nal runner for all for the design of axial runners. The procedure consists of a
operating regimes design program, numerical  ow analysis and a multiobjec-
tive genetic algorithm. The inlet boundary condition can be
Q/Qopt (—) DZ (%)
obtained from a previous calculation of a tandem cascade.
0,39 1,7 On the basis of the numerical results, the genetic algorithm
0,58 1,3
0,81 0,8 makes new generations using operators for selection and
1 0,6 reproduction. In spite of the fact that three-dimensional
1,23 0,55 numerical  ow analysis is very time consuming, using
1,61 0,5
the program for runner blade design enables the whole
A02803 # IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy
50 A LIPEJ

optimization procedure to be implemented. The procedure is REFERENCES


completely automatic.
The main knowledge obtained using a multiobjective 1 Schilling, R., Thum, S., Muller, N., Kramer, S., Riedel, N. and
genetic optimization algorithm is the determination of initial Moser, W. Design optimization of hydraulic machinery bladings
parameters and the range of change in these parameters. by multi level CFD-technique. In Proceedings of 21st IAHR
During the optimization procedure the observation of para- Symposium on Hydraulic Machinery and Systems, Lausanne,
9–12 September, 2002, pp. 19–28.
meters in the early generations is very useful because, from
2 Rabbe, J. Hydro Power, 1985, pp. 407–437 (VDI Verlag,
 uctuations of functions, the correct tendency of optimi-
Düsseldorf).
zation can be stated. To obtain better and reliable results, 3 Barlit, V. V. Hydraulic Turbines (in Ukrainian), 1977,
more generations should be analysed in the optimization pp. 218–227 (Vyssha Shkola, Kiev).
procedure. 4 Jos†t, D., Lipej, A., Oberdank, K., Jamnik, M. and Velens† ek, B.
The important results of this research are the merging Numerical  ow analysis of a kaplan turbine. V. In Hydraulic
together of different methods, the appropriate selection of Machinery and Cavitation (Eds E. Cabrera, V. Espert and
optimization parameters and determination of criteria for F. Martinez), Proceedings of 18th IAHR Symposium, Valencia,
geometry selection with good characteristics. The software Spain, 16–19 September 1996, pp. 1123–1132 (Kluwer,
obtained is an efŽ cient and simple tool for the design of Dordrecht).
hydraulic machines. The method can also be used for the 5 TASC ow Users Manual, Version 2.2, 1993 (Waterloo, Canada).
6 Poloni, C. and Lipej, A. Progettazione di una turbina Kaplan
design of other types of turbine and pump.
con codici  uidodinamici viscosi tridimensionali e algoritmi
The beneŽ ts of this approach in terms of reduced design
di ottimizzione. Congresso Nazionale ATI, Cernobbio, Italy,
time are very important. Without the optimization method 22–26 September 1997, pp. 1033–1044 (SGEditoriali, Padova).
such a design procedure can take a few weeks, but using the 7 Goldberg, D. E. Genetic Algorithms in Search, Optimisation and
proposed method the time can be reduced to a few days. Machine Learning, 1988, pp. 197–211 (Addison-Wesley, Read-
The results of the optimization procedure help show how ing, Massachusetts).
each design parameter in uences the energetic and cavita- 8 Poloni, C., Ng, D. and Fearon, M. Parallelisation of genetic
tion characteristics of axial runners. algorithm for aerodynamic design optimisation. Adaptive
Using the genetic algorithm in combination with three- Computing in Engineering Design and Control, Plymouth
dimensional numerical  ow analysis was found to be useful University, UK, March 1996, pp. 59–64.
for the design of hydraulic machines, because the initial 9 Lipej, A. and Poloni, C. Design of kaplan runner using genetic
algorithm optimisation. XIX IAHR Symposium, Section on
geometry from the design program has a relatively high
Hydraulic Machinery and Cavitation, 9–11 September 1998,
level of efŽ ciency.
Singapore, Vol. I, pp. 138–147.

Proc. Instn Mech. Engrs Vol. 218 Part A: J. Power and Energy A02803 # IMechE 2004

You might also like