Professional Documents
Culture Documents
Approach to Two-Dimensional
Blade Profile Design for Steam
M. A. Trigg
Turbines
G. R. Tubby
In this paper a systematic approach to the optimization of two-dimensional blade
profiles is presented. A genetic optimizer has been developed that modifies the blade
A. G. Sheard profile and calculates its profile loss. This process is automatic, producing profile
designs significantly faster and with significantly lower loss than has previously
been possible. The optimizer developed uses a genetic algorithm to optimize a two-
Allen Steam Turbines,
Bedford, United Kingdom
dimensional profile, defined using 17 parameters, for minimum loss with a given flow
condition. The optimizer works with a "population" of two-dimensional profiles with
varied parameters. A CFD mesh is generated for each profile, and the result is
analyzed using a two-dimensional blade-to-blade solver, written for steady viscous
compressible flow, to determine profile loss. The loss is used as the measure of a
profile's ' 'fitness.'' The optimizer uses this information to select the members of the
next population, applying crossovers, mutations, and elitism in the process. Using
this method, the optimizer tends toward the best values for the parameters defining
the profile with minimum loss.
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
unsuitable for use, or not fitting into a smooth tapered and
twisted three-dimensional blade.
The automatic optimization technique described here avoids
these problems by using the viscous calculated profile loss as the
arbiter of aerodynamic excellence while allowing mechanically
necessary features to be incorporated by limiting the range of
the relevant parameters controlling the profile geometry.
The use of an automatic two-dimensional profile generation
technique does not change the blade design methodology in any
way. The throughflow code of Denton (1978) is used to produce
an annulus design, plus inlet and exit conditions from each blade
row. The two-dimensional CFD solver of Dawes (1983) is used
to produce two-dimensional profiles, which are then stacked and
analyzed with the three-dimensional solver of Dawes (1992). The
use of an optimizer simply takes advantage of the power of mod-
ern desk top computers to analyze many more two-dimensional
profiles than would be attempted manually.
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
bmorph
Completed
SLEQ TBRATB CFD Mesh BTOB CFD Post profiles for
Streamline parametric and Blade to results Processing input to 3D
Equilibrium profile boundary blade >• blade
calculation design conditions viscous design
program CFD process
1 Designer
4.0 Genetic Optimizer changes to the profile. The GA simply takes profile loss as a
Genetic Algorithms (GA's) were originally developed to single figure assessment of a profile's worth.
model computationally the theory of evolution, but have found The genetic optimizer process loop has no prescribed end
wider application in totally unrelated fields, described by, point, Fig. 3. The user must decide if the best profile so far is
among others, Goldberg (1989). The reason for the popularity good enough for the purpose or whether to let the GA continue.
of GA's is that they are powerful global optimizers that can
negotiate complex nonlinear search domains to provide optimal 4.2 Optimizer Implementation. GA's are generally ro-
design solutions. bust and relatively easy to apply once the requirements and the
Unlike conventional methods, which usually require the func- objectives have been identified clearly. They are ideal in a case
tion of interest to be well behaved, GA's are able to tolerate such as this where the GA requires no real understanding of
noisy and discontinuous function evaluations. Due to their sto- the complex flow analysis carried out by the viscous code, but
chastic nature, they are able to search the entire solution space receives just a single number for each result.
with more chance of finding the global optimum than conven- The practical implementation of a GA is often more complex
tional methods. They also do not suffer by getting stuck on a than it first appears. This is due to the effects of collecting
relative optimum and so failing to find an absolute optimum. together separate manual input programs to form an unsuper-
vised design/analysis program. The application of the GA to
4.1 Genetic Algorithm. The GA works with a population two-dimensional profile design proved to be a logical progres-
of individuals, in this case two-dimensional blade profiles. Pro- sion to the design process as all the core tools had already been
files are defined by the set of 17 parameters, described in Section developed for the blade designer, described in Section 2.0.
3.0. These parameters are the defining features or characteristics The GA implementation reported was named "BMORPH."
of the individual. They are coded into a binary string, which is The inputs to BMORPH consist of a range and number of
the "genetic code" of the profile. discrete levels for each of 17 parameters that define the profile.
The GA is linked into a profile design process loop, Fig. 2. Table 1 shows a typical example. This provides the means of
This loop converts the genetic code into an aerofoil definition, constraining parameters to the desired range of values, or of
produces a CFD mesh with aerodynamic boundary conditions fixing parameters that are not required to vary during a particular
for that aerofoil, and then calculates profile loss. study. Additional input parameters were a mutation rate, popula-
The GA can begin with a completely random set of parame- tion size, and flow parameters acting as boundary conditions
ters for the entire first population. Each profile is analyzed to for the BTOB analysis.
ascertain its loss, which is used as a measure of the profile's The value of the parameter is stored in the program as a
fitness in its environment: the specified flow conditions. The binary string, the number of bits depending on the number of
algorithm selects the individuals for the next population from discrete levels required by the user. All these binary strings are
the current population based on their fitness. For this selection, put together to form one long binary string for the individual
a ' 'roulette wheel'' model is used where the profiles with lower profile, Fig. 4.
loss have a higher likelihood of being selected. The optimization algorithm itself has a number of variables
The newly selected individuals are arranged in pairs and a and flags that allow control over its various functions. Once the
crossover site is selected at a random position along each string. initial implementation of the GA had been proven to work, a
The segment of the string that lies after the crossover site is great deal of work was carried out to optimize these variables
exchanged with that of the other individual in the pair. to provide a tool that was both consistent and easy to use.
The resulting binary strings are then randomly mutated at a The user is allowed a great degree of control over the starting
given mutation rate. This rate is several orders of magnitude point for an optimization run, enabling a start from a completely
higher than that observed in living organisms. Increased muta- random point using a random number seed or beginning with
tion rate was found to increase convergence rate; however, a population of given profiles. A completely random starting
above a critical mutation rate the optimization process broke value for each parameter for each profile in the first population
down, and became essentially random. avoids any preconceived notions of what constitutes a good
After all the manipulations have been carried out to create profile, but necessitates a high degree of robustness from the
the new population, each binary string is converted back into a program. Using a given start with specified values for any or
set of 17 parameters. These parameters are used to define the all parameters has not proved to be an important feature as
new aerofoil geometry for each individual in the population. initial convergence is rapid.
For each individual, a CFD mesh is generated and profile losses Although some work was carried out on parallelism, a much
calculated. The process is then ready to be repeated. This basic more effective method in terms of execution time on a single
loop continues with a general trend toward lower loss profiles. machine and complexity for this implementation was found.
The tools used by both designer and GA to produce a two- The method uses a variable population size for dealing with the
dimensional profile are identical, Fig. 2. The only difference is problem convergence rate of the GA, depending on the random
that a designer would view the CFD results and make intelligent start point. The initial population size is set at a value that is a
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
factor higher than the normal population size, typically 5 to 10 Table 1 BMORPH input parameters
times the size. This allows the GA to sample a large area of
Parameter Title Min Max Levels
the search space initially before carrying on for subsequent
Number of Blades 57 57 0
populations with a smaller population size. This allows a de-
Aerofoil Radius 280 280 0
pendable convergence of the GA within reasonable time scales. Axial Chord 15 100 512
A typical run would use a starting population of 100 to 200 Tangential Chord 3 25 512
with the following populations of around 20 to 40 individuals. Unguided Turning Angle 0.5 30 512
Blade Inlet Angle 5 80 1024
Blade Inlet Wedge Angle 5 80 256
LE Circle Radius 0.1 3 32
Blade Exit Angle -40 -80 256
C Input parameter
ranges and flow
conditions
I
Calculate number of
TE Circle Radius
Throat To Pitch Ratio
Tangent Prop. 32
Tangent Prop. 23
Tangent Prop. 21
0.9
0.32
0.1
0.1
0.0005
0.9
0.32
0.95
0.95
0.95
0
0
512
512
512
Tangent Prop. 12 0.17 1.2 512
discrete levels for Tangent Prop. 45 0.1 0.95 512
each parameter Tangent Prop. 54 0.1 0.95 512
"
Setup population array
of binary code strings & Another method used to reduce the overall run time is to
fill with random bits record all the individuals encountered, together with their re-
spective loss as the program proceeds. This enables the program
" to check whether a profile has already been analyzed, avoiding
Convert binary code the need to carry out a time-consuming viscous calculation
r+— strings to real parameters again. This allows the concept to be introduced of ' 'number of
unique BTOB runs" as a measure of the computational effort
used in a particular run of the optimizer.
t Elitism was implemented as an option in this GA, but not
For each profile in population found to be of significant benefit. A study of the effect of elitism
create mesh from geometry did not show it to produce an improvement in convergence rate.
run CFD for loss Both absolute fitness with scaling based on the inverse of the
fractional loss and relative fitness based on differences between
t the losses have been coded as options in the program, with the
latter being most commonly used. Future work may involve
Find minimum loss for
population multi-objective optimization using the mechanical properties of
the profile as an added fitness factor for the profile.
" 4.3 Optimizer Results. In application, BMORPH has
been found to reach a practical optimum within 1000 unique
Calculate genetic weighting for BTOB runs, with only very small reductions in loss after this
each profile in population point. Figure 5 shows a typical run of BMORPH, with the
based on minimum loss lowest loss profile from key populations illustrating the progres-
n " sion of the optimizer. Initially the profiles are clearly absurd;
however, after no more than 250 unique runs, the loss is roughly
/ Output results for / comparable to that typically accepted as satisfactory by a de-
/ °uTOn,p°puia,ion r
"
signer producing the profile without the aid of an optimizer.
Once BMORPH has produced a profile that contains no major
flaws, reductions of profile loss become less frequent as the
profile is "fine tuned."
Select profiles for next
population from current When the minimum loss for each population analyzed is
using roulette model viewed in conjunction with the minimum overall loss, Fig. 6,
the advantage of the genetic algorithm within BMORPH is
w apparent. Having essentially converged on a local optimum after
Select randomly for each pair of 500 unique runs other profiles are evaluated, with considerably
strings a cut position and cross inferior performance, for about 250 unique runs, before an im-
over the pairs provement in profile loss is produced. A further 1000 unique
runs are required before the next reduction in profile loss, during
t which time the lowest loss profile for each population evaluated
often has a profile loss no lower than that achieved after the
Randomly mutate strings first 250 unique runs.
—<- based on given mutation
rate An extensive study with different starting points did not sig-
nificantly reduce or increase the time taken to reach an effective
optimum. Allowing BMORPH to run significantly longer than
i—-^ 1000 unique runs did produce reductions in loss; however, they
were so small as to be of little significance. Theoretically further
Geometry output of
populations may generate an improved profile no matter how
optimised profile
many have gone before. In practice 1000 unique runs was picked
Fig. 3 The genetic optimizer BMORPH process loop as a practical end point.
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
P1 P2 i P3 P4 P16 P17
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
Standard 1960s Design
Manually Improved Profile
Bmorph Optimised Profile
The manually improved profile comes to the point of separa- It must be remembered that optimization is only a form of
tion, Fig. 8, but just does not separate. In this respect the manual design exploration and so is a tool for the designer and not a
improvement has been successful, as the gross loss-producing replacement for the designer.
flow feature has been eliminated. A study of the static pressure
distribution about the 1960s profile and manually improved 6.0 Conclusions
profile, Fig. 7, reveals that the manually improved profile still 1 A genetic optimizer has been developed for minimizing two-
retains an undesirable area of diffusion on the suction surface. dimensional blade profile loss.
The manual improvements to the profile, therefore, constitute 2 The genetic optimizer is unsupervised, only requiring setup
a minimization of the bad features of the 1960s profile. and occasional monitoring. The time required for an experi-
The profile produced by BMORPH is fundamentally different enced designer to produce a profile is correspondingly re-
from both the original and manually improved profile, Fig. 7. duced by approximately an order of magnitude compared
The undesirable diffusion on the suction surface is totally elimi- to that taken previously.
nated, and the pressure is maintained on the pressure surface 3 The genetic optimizer has been shown to reduce two-dimen-
for a significantly greater fraction of surface length. The skin sional blade profile loss by typically 10-20 percent com-
friction, Fig. 8, does not exhibit the peaks and troughs of the pared to unoptimized blade designs.
previous profiles. 4 The genetic optimizer has proven to be a useful development
The BMORPH profile is more than an incremental improve- tool for design exploration, showing trends and general be-
ment over its predecessors, it is a completely new profile design havior. This has facilitated a better understanding by the
without the undesirable flow features of the original. The elimi- designer of the effect on profile performance associated with
nation of these features has effected a 19 percent reduction in a change in blade profile geometric parameters.
profile loss over the 1960s profile and a 5 percent reduction 5 The genetic optimizer has been implemented without chang-
over the manually improved profile. An optimum profile might ing blade design methodology. The laborious task of produc-
be judged mechanically unsuitable, but the performance sacri- ing two-dimensional profiles has been automated, which has
ficed in using a modified or constrained profile can be resulted in the designer focusing on other aspects of blade
determined. design.
The manual improvement to the original profile was under-
taken by an experienced turbine designer, Fig. 2. Over one Acknowledgments
week, a systematic study was conducted with the aim of identi- The work reported in this paper was undertaken within the
fying which features affected profile loss. The resulting im- Product Technology department of Allen Steam Turbines; the
proved profile was not manually optimized; however, it is con- authors offer thanks to other members of the department and
sidered typical of what would be produced without an optimizer. company whose contribution is acknowledged.
The BMORPH optimized profile was set up in an hour, and The CFD flow solver used by the genetic optimizer was
ran unsupervised overnight. While it is not typical to devote an written by Professor W. N. Dawes, Whittle Laboratory, Cam-
entire week to improve one two-dimensional blade profile, it is bridge University, England. The assistance of Professor Dawes
illustrative of the effort that can be spent minimizing profiles with the CFD solver, and during the production of this paper
loss without the aid of an optimizer. Within this context, the is acknowledged.
19 percent reduction in profile loss produced by BMORPH The work reported was funded by Allen Steam Turbines, a
overnight was considered extraordinarily good. Rolls-Royce Industrial Power Group company. The authors of-
The genetic optimizer has been compared with pre-CFD fer thanks to David Beighton, General Manager—Allen Steam
blade technology and that typically achieved by an experienced Turbines, for permission to publish the work reported in this
turbine designer working without optimization tools. The opti- paper.
mizer produced a profile with 19 percent lower profile loss than
the 1960s design in approximately 10 percent of the time that References
would typically be spent on a profile design without the aid of Cravero, C , and Dawes, W. N., 1997, "Through Flow Design Using an Auto-
an optimizer. matic Optimization Strategy," ASME Paper No. 97-GT-294.
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
Cofer, J. I., 1996, "Advances in Steam Path Technology," ASME Journal of Faux, I. D., and Pratt, M. J., 1979, Computational Geometry for Design &
Engineering for Gas Turbines and Power, Vol. 118, pp. 337-352. Manufacture, Ellis Horwood Ltd.
Dawes, W. N., 1983,' 'Computation of viscous compressible flow in blade cascades Goel, S., Cofer, J. I., and Singh, H., 1996, "Turbine Aerofoil Design Optimiza-
using an implicit iterative replacement algorithm," TPRD/M/1377/N83. tion," ASME Paper No. 96-GT-158.
Dawes, W. N., 1986, "Application of Full Navier-Stokes Solvers to Turboma- Goldberg, D. E., 1989, Genetic Algorithms in Search, Optimization, and Ma-
chinery Flow Problems," VKI Lecture Series 2: Numerical Techniques for Vis- chine Learning, Addison-Wesley Publishing Company Inc.
cous Flow Calculations in Turbomachinery Blading, Jan. 20-24. Pritchard, L. J., 1985, "An Eleven Parameter Axial Turbine Aerofoil Geometry
Dawes, W. N., 1992, "Toward improved through flow capability: the use of Model," ASME Paper No. 85-GT-219.
three dimensional viscous flow solvers in a multistage environment," ASME Shelton, M. L„ Gregory, B. A., Lamson, S. H., Moses, H. L., Doughty,
JOURNAL OF TURBOMACHINERY, Vol. 114, pp. 8-17. R. L., and Kiss, T., 1993, "Optimization of a Transonic Turbine Airfoil
Denton, J. D., 1978, "Throughflow calculations for transonic axial flow tur- Using Artificial Intelligence, CFD and Cascade Testing," ASME Paper No.
bines," ASME Journal of Engineering for Power, Vol. 100, No. 2, pp. 212-218. 93-GT-161.
Downloaded 08 Mar 2009 to 194.225.236.227. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm