Beruflich Dokumente
Kultur Dokumente
the requirements for the Degree of Master in Science in the Program of Aeronautics
and Mechanical Engineering, Field of Aerodynamics, Propulsion and Energy.
Campo Montenegro
So Jos dos Campos, SP Brazil
2011
Cataloging-in-Publication Data
Documentation and Information Division
Ando, Victor Fujii
Genetic Algorithm for Preliminary Design Optimisation of High-Performance Axial-Flow
Compressors / Victor Fujii Ando.
So Jos dos Campos, 2011.
162f.
Thesis of master in science Program of Aeronautics and Mechanical Engineering. Field of
Aerodynamics, Propulsion and Energy Aeronautical Institute of Technology, 2011. Advisor: Prof. Dr.
Joo Roberto Barbosa.
1. Genetic Algorithm. 2. Axial-flow compressor. 3. Preliminary design. I. Aeronautics Institute of
Technology. II. Title
BIBLIOGRAPHIC REFERENCE
ANDO, Victor Fujii. Genetic Algorithm for Preliminary Design Optimisation of HighPerformance Axial-Flow Compressors. 2011. 162f. Thesis of master of sciences in
Aerodynamics, Propulsion and Energy Aeronautics Institute of Technology, So Jos dos
Campos.
CESSION OF RIGHTS
AUTOR NAME: Victor Fujii Ando
PUBLICATION TITLE: Genetic Algorithm
Axial-Flow Compressors
PUBLICATION KIND/YEAR:
Thesis / 2011
iii
Chairperson ITA
Advisor ITA
ITA
ITA
iv
Acknowledgements
This work was executed in the context of the Programa Integrado GraduaoMestrado PIGM. Under this programme, ITA Bachelor students from the last year
undertake disciplines from the post-graduate programme and are encouraged to develop the
Bachelor Thesis as an intermediate step towards the research to be conducted during the
Masters.
The author acknowledges the support of Fundao de Amparo Pesquisa do
Estado de So Paulo (FAPESP) to conduct this study.
The author would like to express his gratitude to his advisor, Prof. Barbosa, for the
guidance and the invaluable assistance, especially with regard to the axial-flow compressor
design program. The author is also indebted to Prof. Nelson Manzanares Filho, from UNIFEI,
who was very supportive with insightful discussions on Genetic Algorithms.
Thanks are also addressed to the colleagues from the Gas Turbine Group at ITA for
the amiable companionship.
Finally, the author conveys his thankfulness for the inestimable support of his family.
Resumo
Este trabalho apresenta uma abordagem para a otimizao de projeto preliminar de
compressores axiais de alto desempenho. No contexto do Grupo de Turbinas do ITA, o
projeto preliminar feito utilizando-se um programa computacional baseado no mtodo da
curvatura da linha de corrente, empregando-se correlaes da literatura para o cmputo das
perdas. A escolha de diversos parmetros do ciclo termodinmico e de geometrias depende da
longa experincia acumulada pelos membros do Grupo. Contudo, esse processo exige um
trabalho longo e exaustivo de tentativas e erros. Desse modo, a fim de auxiliar o projetista na
escolha de alguns parmetros, um programa de otimizao, chamado de REMOGA, foi
desenvolvido em linguagem FORTRAN, para fcil integrao com os programas
desenvolvidos pelo Grupo de Turbinas. O programa baseia-se em um algoritmo gentico
multi-objetivo, com codificao real e elitismo.
Em seguida, o REMOGA e o programa de projeto preliminar foram integrados para
o projeto de um compressor axial de cinco estgios. Para isso, foram variados os ngulos de
sada do escoamento dos estatores, a distribuio de temperatura nos estgios e a relao de
raios, visando a maiores eficincias e maiores razes de presso, mas controlando-se o
nmero de De Haller e o ngulo de arqueamento. Graas ao REMOGA, dezenas de milhares
de projetos puderam ser rapidamente avaliados. Finalmente, por meio de um critrio de
escolha, quatro solues foram tomadas para anlise, revelando que o programa desenvolvido
conseguiu encontrar solues mais eficientes e plausveis do que a originalmente proposta.
vi
Abstract
This work presents an approach to optimise the preliminary design of highperformance axial-flow compressors. The preliminary design within the Gas Turbine Group at
ITA is carried on with an in-house computational program based upon the streamline
curvature method, using correlations from the literature to assess the losses. The choice of
many parameters of the thermodynamic cycle and of geometries relies upon the expertise
from the members of the Group. Nevertheless, it is still a laborious and time-consuming task,
requiring successive trial and errors. Therefore, to support the compressor designer in the
choice of some parameters, an optimisation program, named REMOGA, was written in
FORTRAN language, allowing an easy integration with the programs developed by the Gas
Turbine Group. The program is based upon a multi-objective genetic algorithm, with real
codification and elitism.
Then the REMOGA and the preliminary design program were integrated to design a
5-stage axial-flow compressor. Therefore, the stator air outlet angles, the temperature
distribution and the hub-tip ratio were varied aiming at higher efficiencies and higher pressure
ratios, but controlling the de Haller number and the camber angle. Thanks to the REMOGA,
thousands of designs could be quickly evaluated. Finally, using a choice criterion, four
solutions were selected for further analysis, revealing that the developed program was
successful in finding more efficient and feasible compressor designs.
vii
LIST OF FIGURES
viii
Figure 20 Rotor row and stator row with velocity triangles in an axial-flow
compressor stage.................................................................................................... 50
Figure 21 Inlet and outlet relative velocity ratio is reduced with the increase of fluid
deflection. .............................................................................................................. 54
Figure 22 Streamline-blade leading edge coordinate system (s-m). [42] .............................. 58
Figure 23 Streamlines, stage rows and calculation nodes. Adapted from [42]. .................... 58
Figure 24 Overview of the SLC program algorithm. ............................................................ 59
Figure 25 Mapping between the decision space and the objective space. ............................. 61
Figure 26 Representation of dominance and indifference between solutions in a twoobjective minimisation problem. Solution a dominates b, but is indifferent
to c. ........................................................................................................................ 63
Figure 27 A convex function illustration............................................................................... 64
Figure 28 Illustrative region where a gradient-based algorithm can get stuck onto a
suboptimal solution................................................................................................ 66
Figure 29 Simple GA algorithm [48]. ................................................................................... 67
Figure 30 Chromosomal representation of decision variables. ............................................. 68
Figure 31 Tournament selection illustration.......................................................................... 69
Figure 32 Biological crossover illustration. .......................................................................... 70
Figure 33 Bit-wise crossover representation. ........................................................................ 70
Figure 34 Single-point crossover representation. .................................................................. 71
Figure 35 Two-point crossover representation. ..................................................................... 71
Figure 36 Mutation operator. ................................................................................................. 72
Figure 37 Algorithm of the REMOGA program. .................................................................. 72
ix
xi
xii
LIST OF TABLES
xiii
LIST OF SYMBOLS
LATIN SYMBOLS
C
absolute velocity
blade chord
dij
vector of objectives
objective space
enthalpy
htr
hub-to-tip ratio
mass flow
nc
niche count
npop
population size
pressure
radius
rank (.)
rank of a solution
rp
pressure ratio
pitch or spacing
Sh (.)
sharing function
temperature
tangential velocity
relative velocity
power
decision space
xiv
GREEK SYMBOLS
degree of reaction
isentropic efficiency
polytropic efficiency
camber angle
flow coefficient
angular velocity
SUBSCRIPTS
0
total property
rotor inlet
stator inlet
stator outlet
axial component
meridional component
xv
DOE
Design of Experiments
EA
Evolutionary Algorithm
GA
Genetic Algorithm
IGV
LHS
MOEA
MOGA
MOOP
N-S
Navier-Stokes
NSGA
OGV
RANS
Reynolds-Averaged Navier-Stokes
REMOGA
RSM
SBX
SLCM
SLCP
SOOP
xvi
CONTENTS
1
INTRODUCTION ......................................................................................................... 19
1.1
Motivation.............................................................................................................. 19
1.2
Objective ................................................................................................................ 20
1.3
Methodology .......................................................................................................... 20
1.4
Context................................................................................................................... 21
1.5
1.6
LITERATURE REVIEW............................................................................................. 25
2.1
Introduction............................................................................................................ 25
2.1.1
2.1.2
2.1.3
2.2
3
Introduction............................................................................................................ 36
3.1.1
3.1.2
3.1.3
3.1.4
3.1.5
3.2
3.3
History ..................................................................................................... 36
Classification ........................................................................................... 38
Gas turbine............................................................................................... 40
Basic operation ........................................................................................ 41
Nomenclature .......................................................................................... 43
Solvers ..................................................................................................... 25
Reference stage ........................................................................................ 26
Optimisation methods.............................................................................. 27
Introduction............................................................................................................ 56
4.2
4.3
xvii
Definitions ............................................................................................................. 60
5.1.1
5.1.2
5.1.3
5.1.4
5.2
5.3
5.4
5.5
METHODOLOGY........................................................................................................ 90
6.1
Test functions......................................................................................................... 85
5.5.1
5.5.2
5.5.3
5.6
6.2
6.3
6.4
7.2
7.3
7.4
xviii
7.5
9.2
APPENDIX B
19
1.1
INTRODUCTION
MOTIVATION
20
1.2
OBJECTIVE
1.3
METHODOLOGY
Aiming at the proposed objective, the work was divided in two parts:
1. Optimisation:
a. Literature review on the use of optimisation procedures in the design
of axial-flow compressors;
b. Study of multi-objective genetic algorithms;
21
1.4
CONTEXT
This present work was executed under the Programa Integrado GraduaoMestrado PIGM. This Program aims at the integration of the Undergraduate and the
Masters Programs by allowing the student from the last year of the undergraduate course at
Instituto Tecnolgico de Aeronutica (ITA) to undertake courses from the post-graduate
programs, shortening the necessary time to fulfil the requirements to the title of Master in
Science.
In this context, the Bachelor Thesis (Trabalho de Graduao TG) was
supervised to provide a well-developed start point to the Master Thesis. The TG of the author,
entitled Project Optimisation of High-Performance Axial-Flow Compressors was executed
under the supervision of Prof. Dr. Joo Roberto Barbosa (the same supervisor of this work). It
preliminarily validated the design optimisation procedure by coupling the Streamline
Curvature Method to a Multi-Objective Genetic Algorithm. In that work, the design variables
were the efficiency, hub-to-tip ratio and the stator air outlet angles via a multivariate
interpolation, which used four control points, namely hub and tip at the first row and hub and
tip at the last row. Diffusion factors and camber angles were controlled by means of a penalty
factor treated as objectives to be minimised.
22
The Master Thesis was developed under the scholarship from Fundao de Amparo
Pesquisa do Estado de So Paulo FAPESP (So Paulo State Research Foundation) at the
Centre for Reference on Gas Turbine at ITA.
1.5
Tomita [1] describes the research on gas turbine within DCTA. A summary of this
history is presented hereafter.
Plans to develop gas turbines in Brazil are found in the Plans of Foundation [2] of the
Centro Tcnico de Aeronutica CTA (Aeronautical Technical Centre), in 1947.
However, the research only flourished in the 1970s, with the establishment of a Research
Program at CTA. At the time a new turbine project was developed, thereby many
opportunities of partnerships with important manufacturers, like Rolls-Royce (UK), Garret
(USA), Pratt & Whitney (USA and Canada), Lucas Aerospace (UK), and Kongsberg
(Norway), succeeded and were valuable. Thereafter, the project was seriously hindered due to
lack of experienced professionals. A joint project with Rolls-Royce to design and manufacture
of a 300 kW turboprop to be mounted on aircrafts from Bandeirante class was halted as a
result of lack of personnel.
Thus, an ambitious program of training the CTA personnel commenced with
Cranfield Institute of Technology (currently Cranfield University). From this Institute,
engineers from CTA and ITA, working in research related to Gas Turbines, graduated,
including the supervisor of this work, who obtained his PhD degree in Cranfield in 1987.
Even with the present practice of importing gas turbines rather than designing and
manufacturing in Brazil, the necessity of specialists in those machines, mainly in performance
23
analysis and applications is evident. The process of choosing the correct turbine is vital, since
it undoubtedly allows a significant reduction in operation and maintenance costs.
Observing the current actions of the major players from the energy sector in Brazil,
or even big companies moving to the energy sector, one might again note a real requirement
for specialists in turbines and compressors. In this context, two companies should be
highlighted: Vale Solues em Energia (VSE), which is preparing to design and manufacture
its own gas turbines, to secure its high energy demand in mining operations; and General
Electric, which launched a massive investment program in Brazil in 2010.
CTA was renamed DCTA Comando-Geral de Tecnologia Aeroespacial
(Brazilian General Command for Aerospace Technology), but the efforts to implement a
modern Turbine Laboratory persist. According to Barbosa [3], it should include a compressor
test bed (1500 kW shaft power and up to 60,000 rpm); a turbine test bed (2000 kW brake
power and rotation speed up to 60,000 rpm) and a combustion chamber test bed (for hot gases
up to 1500 K; 1.0 MPa). The development of a small gas turbine for research should be
carried on, as well.
The research on gas turbine at ITA is conducted by the Centre for Reference in Gas
Turbine (CRTG Centro de Referncia em Turbinas a Gs). The Centre, which belongs to
the Mechanical Engineering Department of ITA, relies its research upon information of public
domain and upon many years of experience from its members.
The centrepiece of the research developed at CRTG is on numerical simulation.
Programs of design point performance, off-design performance, computational fluid
dynamics, transient performance, combustion chamber performance, noise prediction have
been written and are fully operational.
24
1.6
In chapter 1 the reader finds the introduction, where the motivation, objective and
methodology are presented. A brief history of the research on gas turbine within DCTA is
also presented.
Chapter 2 contains a review of studies published in axial-flow compressor
optimisation. A review of ASME Turbo Expo congresses since 2000 in this particular field is
also shortly conducted.
Chapters 3 and 4 provide the basic theory on axial-flow compressors and on the
streamline curvature method.
In chapter 5, the author starts with the basic ideas behind Genetic Algorithms and
then he details features, algorithms and models used in the REMOGA program, which was
developed as part of this work.
Chapter 6 describes how the integration of the SLC program and the REMOGA
program took place.
Chapter 7 shows the results obtained through the aforementioned integration and
analyse four solutions, which were selected among thousands of solutions proposed by the
REMOGA.
Chapters 8 and 9 conclude this work, suggesting future works as well.
Four appendixes are provided. The first contains a basic derivation of the SLC
method. The second contains the FORTRAN code of the developed optimisation program.
The third appendix offers the design parameters of the start-point axial-flow compressor. And
the last appendix provides further graphical information from the compressors analysed in this
work.
25
LITERATURE REVIEW
2.1
INTRODUCTION
Before proceeding with the comparative table, some preliminary concepts are
presented.
2.1.1
Solvers
Solvers can be defined as computational programs that solve a given mathematical
26
virtually impossible to solve, except for very simple cases, which are not of real-world
interest. Therefore numerical methods and algorithms are employed to obtain approximate
solutions. According to the problem, the user may choose a 2D or 3D solver, depending on
the desired accuracy and on the computational resources available, as well.
To describe turbulent flows, instantaneous quantities of the N-S equations are timeaveraged to provide an approximation, which is easier to calculate. The resulting equations
are called Reynolds-Averaged Navier-Stokes equations, or RANS.
A further simplification of the N-S equations can be carried out by ignoring viscosity
and heat conduction. The simplified equations are called Euler equations. If used per se it
provides very rough approximations in turbomachinery calculation, as viscosity plays an
important role. Nevertheless, Euler equations can be used accurately if losses are assessed by
correlations derived from experiments.
2.1.2
Reference stage
The most frequent reference stage used for academic purposes is the NASA Rotor
37, see Figure 1. As its flow field was used by the American Society of Mechanical Engineers
in 1994 in a CFD blind-test exercise, plenty of studies on the flow field in the aforementioned
rotor were derived [4].
27
NASA Rotor 37 was designed and tested at NASA Lewis Research Center (renamed
NASA Glenn Research Center) in the late 1970s. It is a low aspect ratio inlet with 36
multiple-circular-arc (MCA) blades. Rotor 37 has a pressure ratio of 2.106 at a mass flow of
20.19 kg/s.
2.1.3
Optimisation methods
A brief introduction to optimisation methods is provided in chapter 5.
2.2
A comparative table of works presented at ASME Turbo Expo from 2000 to 2011
regarding optimisation in axial-flow compressors is drawn to provide a panorama of the
theme, as well as its evolution. The works were primarily taken from the topic Design
Methods and CFD Modelling for Turbomachinery. Therefore, the following information was
taken, when applicable:
Solver: which method was used to obtain quantitative results from the design;
Reference stage: many optimisation studies are carried on long-timeestablished open-data stages, e.g., NASA rotor 37;
28
Table 1 Summary of recent works presented at ASME Turbo Expo on compressor optimisation.
Ref.
title
problem
opt. Method
design variables
objective
[5]
2000
C4 airfoil
gradient-based
SOOP
Navier-Stokes
min. total-to-total
pressure loss
coef.
[6]
2000
SOOP
3D NavierStokes
four-stage
ATKOM NPT
stacking lines
max. efficiency
[7]
2000
SOOP
Quasi-3D N-S
NASA rotor 37
gradient-based and
sensitivity analysis
max. adiabatic
efficiency
[8]
2001
SOOP
3D NavierStokes
NASA rotor 37
modified feasible
directions algorithm
max. adiabatic
efficiency
[9]
2002
MOOP
CFD code;
sliding mesh
and time
dependent
NACA 65-12-10
Multi-objective
Evolutionary
Algorithm
min. dynamic
loading and max.
time-avg.
efficiency
[10]
2002
SOOP
3D NavierStokes
NASA rotor 37
RSM
max. efficiency
[11]
2003
MOOP
Quasi-3D N-S
Evolution
Strategies and
MOGA
min. total
pressure loss and
min. deviation
angle
[12]
2003
Numerical optimization of
turbomachinery bladings
SOOP
Quasi-3D N-S
and 3D N-S
CONMIN
(gradient-based)
blade deformation
max. efficiency
[13]
2003
3D blade geometry
weigted sum:
aerodynamic
losses, maximum
Mach, etc
MOOP
solver
Mises (Euler
Q3D)
reference stage
Covariance Matrix
Adaption (CMA)
29
Ref.
title
problem
solver
reference stage
opt. Method
design variables
objective
[14]
2004
SOOP
3D NavierStokes
NASA rotor 37
ANN, GA,
Simulated
Annealing
blade parameters
efficiency and
weighted sum of
penalties
[15]
2005
MOOP
Reynoldsaveraged 2D NS
real-coded MOEA
max. static
pressure and min.
total pressure loss
[16]
2006
SOOP
3D NavierStokes
NASA rotor 37
DOE, RSM
(second-order
polynomial) and
GA
max. adiabatic
efficiency
[17]
2006
MOOP
3D-CFD
Rolls-Royce
datum design
DOE, Monte-Carlo
Simulation, NSGAII
[18]
2006
SOOP
3D NavierStokes
NASA rotor 37
DOE, RSM
(second-order
polynomial)
max. adiabatic
efficiency
[19]
2006
MOOP
asynchronous
MOEA, ANN
3D blade geometry
[20]
2006
SOOP
3D NavierStokes
blade geometry: 3D
NURBS, 65 control
points
min. constrained
augmented
functional
blade parameters:
CCGEOM
desirability
function, which
embraces
efficiency and
pressure ratio
IOSO
blade geometry
efficiency for
operation mode
[21]
2007
SOOP
CFD code
SWIFT
[22]
2007
SOOP
CFD NUMECA
NASA stage 35
30
Ref.
title
problem
solver
[23]
2008
reference stage
opt. Method
design variables
objective
MOOP
Blade-Gen,
Turbo-Grid,
CFX-Pre, CFXSolver
NASA rotor 37
Latin hypercube,
PRESS based
averaging, RSM
and gradient-based
6 design variables
defined by parametric
curves
efficiemcy, total
pressure and the
combination of
both
[24]
2008
SOOP
isentropic
efficiency at two
operating points
[25]
2008
MOOP
2D MISES
DOE (Latin
Hypercube or
SOBOL); NSGAII; Kriging RSM
2D blade profile
pressure loss at
DP, stall and
choke
[26]
2008
SOOP
blade-to-blade
MISES (Q3D)
GA (developed by
Carroll)
weigted sum:
losses and inlet
angle
[27]
2008
MOOP
3D-RANS and
throuflow
MAGELAN
IDAC3 of RWTH
Aachen
MOEA, ANN,
Kriging and
polynomial surfaces
chordwise s-Shift,
stagger variation, suction
side control points,
annulus. 23 parameters
efficiency
improvement and
diffusion factor in
stator 3
[28]
2009
SOOP
3D NavierStokes coupled
with BaldwinLomax
NASA rotor 37
Simple gradientbased
Multi-section blade
parameters
adiabatic
efficiency
[29]
2009
SLC method
NACA 10-stage
subsonic axial
compressor
total pressure at
surge-marginrelated operating
point
SOOP
Genetic Algorithm
31
Ref.
title
problem
[30]
2009
[31]
2010
SOOP
[32]
2010
SOOP
[33]
2011
[34]
2011
MOOP
solver
reference stage
opt. Method
design variables
objective
NASA rotor 37
Multiobjective
Differential
Evolution (MDE)
isentropic
efficiency and
min. maximum
stress
CFD NUMECA
NASA rotor 67
combination off
overall eficiency
and pressure ratio
3D-RANS
NASA rotor 37
circumferential grooves:
width, depth normalized
by tip chord
circumferential grooves:
width, depth. Angle
between axis of rotation
and camber tangent
MOGA and
gradient-based
improvements
efficiency, mass,
length, rotor
blade count and
stator blade count
3D-RANS
MOOP
ANSYS-CFX
MOOP
T-AXI:
axyisymmetric
solver
NASA rotor 37
32
From Table 1 one might notice that in the early 2000s, most of the optimisation
methods were based on gradient. Later, however, the use of EA was the rule. Similarly, a
tendency to MOOPs is observed, which is related to the spread of MOEA. Before, MOOPs
were mostly treated as SOOPs by means of encompassing many objectives in a single
objective function (weighted average).
Table 1 also shows that blade profile optimisation has been extensively studied in the
context of compressor optimisation. Evidently the techniques employed are closely related to
the computer capabilities. In 2000, Chung and Lee [7] used a quasi-3D Navier-Stokes solver
and a gradient-based method in a SOOP to optimise the NASA rotor 37 with eight design
variables. Nine years afterwards, Luo et al. [30] conducted a study on multi-disciplinary
optimisation of the same NASA rotor 37 using a 3D-RANS solver to the aero domain and
FEM to the mechanical domain using 19 design variables related to the blade suction surface
geometry. The optimisation aimed not only at higher isentropic efficiencies, but also at the
minimisation of the maximum mechanical stress. To achieve that, aero and mechanical mesh
were required and the aero solution had to be calculated to feed the FEM boundary
conditions, as may be clear in Figure 2.
Start
Preprocessing
Design
Parametrisation of 3D blade
variable
Generation
aero. mesh
Generation
mech. mesh
Surfaces
Parallel
MDE
CFD solution
Aero
efficiency
pressure
Aero
perf ormance
computation
FEM solution
Mechanics
perf ormance
computation
End
33
These two different approaches to the optimisation of the NASA rotor 37 highlight
the evolution of the optimisation capabilities in the 2000s decade. The major move from
simple SOOP gradient-based strategies to multidisciplinary optimisation involving several
design variables and objectives was certainly due to the advances in computer hardware, as
EAs require considerable amount of computational effort and are particularly suited to
parallel computing [35].
Gathering information from 44 Intel domestic processors, summarised in Figure 3, a
glimpse of the evolution of the processors in a decade can be put into perspective.
To plot Figure 3, the following processor families were taken into account: Pentium
III, Pentium 4, Pentium 4 HT, Celeron, Celeron D, Pentium D, Pentium Extreme Edition,
Core 2 Duo, Core 2 Quad, Core 2 Extreme, Core i3, Core i5, Core i7 and Core i7 Extreme
Edition.
Transistors and Cache memmory over time
1400
4.5
1200
4.0
1 thread
2 threads
1000
4 threads
800
12 threads
3.5
3.0
8 threads
Clock [GHz]
600
1
2
4
8
12
thread
threads
threads
threads
threads
2.5
2.0
1.5
400
1.0
200
0.5
bubble size: Lithography [32;250] nm
0
1998
2000
2002
2004
2006
year
2008
2010
2012
0.0
1998
2000
2002
2004
2006
year
2008
2010
2012
It is noticeable that a stabilisation in clock speed was reached close to 4 GHz, but the
increase of the number of transistors and of threads is still taking place. But the main benefit
in recent computation for MOGA is the parallelisation capabilities provided by multiple
threads.
34
Recalling Table 1, one may observe that among 30 selected papers, 24 were centred
on the blade profile. From blade section geometry through spline control points to blade
stacking line and from leading edge line to sweep, lean and skew, the theme has been
thoroughly explored. Similarly, the methods ranged from simple gradient-based ones to
various Evolutionary Algorithms, Response Surface Method, Latin Hypercube Sampling and
Artificial Neural Network. Predominantly, however, 3D or quasi-3D Navier-Stokes solvers
were employed.
Apart from blade geometry optimisation, Binini and Toffolo [9] studied the axial
distance between rows and circumferential clocking on dynamic loading and efficiency. The
optimisation was conducted via MOGA. Furthermore, Choi et al. [32] and Kim et al. [33]
carried an investigation on circumferential grooves targeting higher stall margin and peak
efficiency.
From 2000 to 2011, only Shadaram et al. [29] presented a work on compressor
optimisation using the Streamline Curvature Method at Turbo Expo. The study aimed at the
maximisation of the total pressure ratio at off-design condition of a 10-stage compressor by
means of changing the stagger angles of the inlet guide vane (IGV) and two rows of stator
vanes. To achieve that, a single-objective GA was employed.
Apart from researches published at Turbo Expo, Oyama and Liou [35] developed a
multiobjective design optimisation tool based on the SLC method and on a real-coded MOGA
aiming at higher efficiencies and pressure ratios of a 4-stage axial flow compressor. To
achieve that, they used design parameters at the rotor trailing edge and at the stator trailing
edge. At the former, total pressures and solidities are design variables, and at the latter, flow
angles and solidities. To avoid flow separation, the diffusion factor was constrained. The
study revealed hundreds of feasible Pareto-optimal solutions.
35
Keskin and Bestle [36] presented at the German Aerospace Congress 2005 a
procedure to automate a given Rolls-Royce preliminary design process to find Pareto-optimal
solutions for design conditions. A meanline prediction process was integrated to sampling
methods like Design of Experiments and Monte-Carlo Simulation and to a Multi-island
Genetic Algorithm (MIGA). Additionally, a gradient-based Lagrange-Newton type algorithm
is used. In order to reduce the number of design variables and keep the design freedom to save
computational costs, Bzier-spline parameterisation was employed to describe the annulus
lines and the stage pressure ratio distribution. In this manner, the control points of the Bziersplines were used as decision variables. The optimisation goal was overall polytropic
efficiency, overall pressure ratio and surge margin at design point. The constraints were: stage
loadings, relative rotor and absolute stator inlet Mach numbers, compressor exit Mach
number, Koch parameters, diffusion numbers and de Haller numbers. Keskin and Bestle
found that the efficiency could rise by 0.11% point keeping the surge margin constant or
improve the surge margin by 3.2% points without diminishing efficiency.
36
This chapter aims at providing the basic knowledge about axial-flow compressors. It
was written based primarily on the books of Saravanamuttoo [37], Aungier [38], Horlock
[39], Boyce [40] and Walsh [41] .
3.1
INTRODUCTION
The purpose of the compressor is to raise the total pressure of the working fluid to a
level required by the thermodynamic cycle. The pressure rise should consume the minimum
shaft power, as this component absorbs approximately one third of the turbine power.
3.1.1
History
Axial-flow compressors for aeronautical applications started their development in the
1930s and entered into service at the end of the WW2. The Germans took the lead with the
engine Junkers Jumo 004, which was mounted on many aircrafts, among them, the famous
Messerschmitt Me 262 Schwalbe, the world first operational jet-powered fighter aircraft.
Figure 4 Junkers Jumo 004 axial jet engine and Me 262. Source: <www.luftarchiv.de>
37
A British axial engine program was also carried (The Metropolitan-Vickers F.2 was
the first axial British design), but it was unsuccessful to deliver an engine to the war.
From 1940s to 2010s, there was a considerable technological leap in axial-flow
compressor design. Metallurgy technology, new materials, multi-spool configurations,
variable geometries, computational resources and test facilities contributed for the increase in
efficiency and achievement of higher pressure ratios with fewer stages. For the sake of
comparison, Table 2 provides some illustrative data about the Jumo 004 and the Rolls-Royce
Trent 1000 (Figure 5), certified in 2007 to show the evolution after a bit more than half of a
century.
Table 2 Comparison between Junkers Jumo 004 and Rolls-Royce Trent 1000.
Type
Entry
Pressure ratio
Spools
Number of stages
Average pressure ratio per stage
Thrust [kN]
38
3.1.2
Classification
Compressors are classified into two major groups: positive displacement and
Compressor
Positive displacement
Dynamic
Centrif ugal
Axial-f low
39
Therefore, it was recognised from the beginning of the gas turbine history that axialflow compressors would be capable of higher pressure ratio and higher efficiency than
centrifugal compressors[37].
P / Pdesign
Positive displacement
Head
1.10
Axial-Flow
Compressor
1.05
1.00
Centrifugal
0.95
Compressor
Centrifugal
Compressor
0.90
Axial-Flow
Compressor
0.85
0.80
Flow
0.90
0.95
1.00
1.05
1.10
Q / Qdesign
Another difference between axial-flow and centrifugal compressors is that the latter
has narrower operational range than the former. In an axial-flow compressor, a small variation
of flow rate around the design point results in great pressure ratio variation in comparison
with centrifugal compressors. Schematically, the comparison between centrifugal and axialflow compressors is shown in Figure 8.
High-performance axial-flow compressors seek high efficiencies and high pressure
ratios, but with few stages. This is almost contradictory, because then high air velocities are
required, but this normally incurs in higher friction and higher losses. Thus a tuned
40
temperature distribution along the stages is required, as well as a proper selection of the
airfoil.
3.1.3
Gas turbine
A simple and ideal gas turbine basically consists of three components: the
compressor, the combustion chamber and the turbine. The working fluid (e.g., air) enters the
compressor, which raises the pressure and the temperature of the fluid in an isentropic process
(ideally). The compressed fluid is then provided to the combustion chamber, wherein fuel is
added and burnt, leading to a dramatic increase in temperature and energy of the mixture in a
isobaric process. Finally, the working fluid expands isentropically in the turbine, transferring
energy to its blades. The turbine and the compressor are connected by a shaft, which transfers
mechanical energy from the turbine to the compressor. The turbine must extract energy in
excess to drive a load (e.g., propeller, generator, free turbine, etc.). Figure 9 shows a simple
gas turbine scheme and its related ideal temperature-entropy diagram.
T
P2
fuel
combustion chamber
P1
power output
air
exhaust gas
compressor
turbine
1
s
Figure 9 Schematic figure of the main components in a gas turbine and the Brayton cycle.
T
T
T3 1 4 T2 1 1
wcycle c p (T3 T4 ) c p (T2 T1 )
T2
T3
=
=
=
cycle
q23
c p (T3 T2 )
T3 T2
(1)
41
Ta
Tb
Pa
=
Pb
(2)
P4
P1
T3 1
T 1
P3 2 P2
1
cycle =
= 1
r
T3 T2
p
(3)
From Equation (3) one immediately notices the relevance of the compressor in the
overall engine efficiency.
3.1.4
Basic operation
An axial-flow compressor consists of a series of rotating blades and stationary
blades, as shown in Figure 10. The air first enters a row of rotating blades, where mechanical
energy from the shaft is transferred to the fluid to accelerate it. Then, the air with high
velocity is delivered to the stationary row, where it flows through a divergent nozzle and is
diffused, i.e., fluid kinetic energy is converted to static pressure rise.
rotor
rotation
Mechanical Energy Fluid kinetic energy
42
Figure 11 Visual aid to the common plane scheme of an axial-flow compressor stage.
Static pressure
Total pressure
Static temperature
Total temperature
Relative velocity
Absolute velocity
Enthalpy
Density
Rotor
Increase
Increase
Increase
Increase
Decrease
Increase
Increase
Increase
Stator
Increase
Small decrease
Increase
Constant
Decrease
Constant
Increase
43
3.1.5
Nomenclature
The literature presents many different nomenclatures for blade and cascade. In this
Point of
maximum camber
V2
V1
1
2
1
2
V1
V2
Letters C, V and U are used for absolute, relative to the rotor and tangential (or
peripheral) velocities, respectively. Subscripts 1, 2 and 3 denote respectively rotor inlet, rotor
outlet and stator outlet. Subscript 0 denotes total property. Subscripts w and a indicate the
whirl (tangential) and the axial components. The meridional direction m is given by the
composition of the radial and axial directions of the flow:
m =
Vw r + Va z
Vw2 + Va2
(4)
Greek letters and indicate absolute air and relative air angle; denotes blade
44
The distance from the leading edge to the trailing edge is the chord c. The distance
from one blade to another measured at constant axial coordinate is the space or pitch s.
The inlet velocities and the outlet velocities of a rotor row are usually drawn together
in a recurrent scheme named velocities triangles, as shown in Figure 14. If the row is purely
axial, then the meridional component is the axial component.
Cm1
Cm 2
U1
Vw1
Cw1
Vw 2
U2
Cw 2
3.2
3.2.1
DIMENSIONLESS PARAMETERS
Flow coefficient
The first dimensionless parameter commonly used in performance calculation is the
Ca1
U
(5)
45
3.2.2
c p T0 stage
h03 h01 c p constant
=
2
U
U2
(6)
For satisfactory operation Walsh and Fletcher [41] suggest [ 0.25, 0.5] .
Efficiency improves as loading is reduced, but a decrease in stage loading implies
more stages. Thus, a compromise is in question for aero engines, as both high efficiency and
low weight (fewer stages) are mandatory.
3.2.3
Degree of reaction
The distribution of the flow diffusion taking place at the rotor and the stator rows is
indicated by the degree of reaction. The degree of reaction is the ratio between the static
enthalpy rises in the rotor and in the stage:
h2 h1 c p constant
T2 T1
=
h3 h1
T3 T1
(7)
Many preliminary compressor designs start with a 50% reaction, due to even
distribution of diffusion, leading to smaller losses.
3.2.4
htr =
rhub
.
rtip
(8)
46
High values of hub to tip ratio usually indicate short blades, hence, the tip clearance
becomes relatively higher. The tip clearance, as the name suggests, is the distance between the
blade tip and the compressor casing. High values of tip clearance lead to lower efficiencies,
due to leakage flow through the spacing. Figure 15 shows the hub to tip ratio and the tip
clearance in an actual compressor.
Tip clearance
rtip
rhub
Figure 15 Hub to tip ratio and tip clearance.
Low values of hub to tip ratio yield long blades, hence more pronounced secondary
losses, as well as, more difficult mechanical mounting in the rotor disc.
3.2.5
kinetic energy has been adiabatically converted to internal energy. A subscript 0 is used to
denote total properties. In a given point, the total enthalpy and total temperature are,
respectively:
h0= h +
C2
,
2
(9)
T0= T +
C2
,
2c p
(10)
where h is the static enthalpy, C is the absolute velocity of the fluid and cp is the specific heat
at constant pressure.
47
c =
h01
h02
.
h02 h01
(11)
=
c
(12)
1
T01 T01 p01
=
.
c =
T02 T01
T02
1
T01 T01
T01
(13)
For later compressor stages, as the pressure is already high, it is much more difficult
to increase the pressure. This can be explained by the fact that the isobaric lines in a T-s
diagram are divergent (to the right), as shown in Figure 16. Noticeably, the compressor
requires more energy to compress the fluid in the first stage than in the last stage, even for the
same pressure ratio.
p4 p2
=
p3 p1
Last stage
p4
p3
p2
p1
First stage
s
Figure 16 Divergent isobaric lines and the increased compression difficulty in the last stages.
Hence the efficiency of the latter stages tends to be smaller than the initial stages,
even when the technological level is the same.
48
This fact revealed the necessity of another definition of efficiency for multi-stage
compressor: the polytropic efficiency or small-stage efficiency, which is defined as a constant
isentropic efficiency of an elemental stage throughout the whole compression stage:
=
,c
dT
= constant
dT
(14)
The idea of the polytropic efficiency can be visualised in Figure 17. The polytropic
efficiency represents the particular technological level for a particular design. Thus, it is
reasonable in preliminary design to consider constant polytropic efficiency for all stages.
dT
dT
Elemental stage
s
3.3
In section 3.1.3, a very simple gas turbine thermodynamic cycle was presented.
Nevertheless, the compression is not isentropic, the combustion is not isobaric and the
expansion is not isentropic. Thus the overall efficiency is smaller. From now on, the
discussion here will focus on the compressor side.
49
T
P2
P3
3
4
P4
P1
1
s
Firstly, to support the derivation, consider the compressor stage in a temperatureentropy diagram in Figure 19. Assuming adiabatic process, one immediately finds that the
power input to the compressor rotor is given by:
p (T02 T01 )
=
W mc
(15)
T02 = T03
(16)
The power is solely transferred to the rotor, which delivers air at high speed to the
stator. The stator, through diffusion, transforms kinetic energy to static pressure rise.
To proceed with the blade preliminary geometry, its angles are written together with
aerodynamic and thermodynamic equations.
A compressor stage with its velocity triangles is shown in Figure 20. Assuming that
C
=
C=
Ca 2 , simple trigonometry yields:
a
a1
U
= tan 1 + tan 1 ,
Ca
(17)
50
U
= tan 2 + tan 2 .
Ca
p02
02
T02 =T03
T03
C32
2cp
03
3
C22
2cp
p2
p01
T01
p03
03
p3
T2
(18)
2
p1
01
T1
C12
2cp
s
Figure 19 Axial-flow compressor stage in a T-s diagram.
1
C1
V1
rotor
C1
Ca1
Cw1
C2
U
V2 2
stator
C2
Ca2
C3
Cw2
Figure 20 Rotor row and stator row with velocity triangles in an axial-flow compressor stage.
51
(19)
(20)
In the compressor, the flow enters with tangential velocity Cw1 at radius r1 and
leaves with tangential velocity Cw 2 at radius r2 . Thus, the required torque for a mass flow
rate m is
=
T m ( r2Cw 2 r1Cw1 ) ,
(21)
W m ( r2Cw 2 r1Cw1 ) .
=
(22)
( Cw 2 Cw1 ) .
=
W m r1 ( Cw 2 Cw=
mU
1)
(23)
a ( tan 2 tan 1 ) ,
=
W mUC
(24)
a ( tan 1 tan 2 ) .
=
W mUC
(25)
Equation (25) shows that more power is used by the compressor if the blade has
higher camber angle, thus more power is transferred to the fluid in this condition. Later,
however, it will be shown that there is a limit for this camber, otherwise, 1 90 and
52
Continuing with the derivation, if the compressor power input is transferred to the
fluid to raise its pressure, the whole power input contributes to the total pressure rise:
T03 =T02
UCa
( tan 1 tan 2 ) .
cp
(26)
=
c
rp 1
T02
1
T01
c
T01
(T02 T01=)
rp
UCa
1 c
( tan 1 tan 2=) rp 1
c pT01
(27)
UCa
1
=
rp c
( tan 1 tan 2 ) + 1
c pT01
Equation (27) provides wise advices on how to obtain higher pressure ratios per
stage. High values of compressor efficiency, rotational speed, axial velocity and camber angle
and low values of inlet total temperature, cp and provide elevated pressure ratios per stage.
Usually, the designer has no control on the working fluid and ambient conditions,
thus, changes in c p , and T01 are not case of study here. The rotational velocity is limited by
material technology and the compressor efficiency is given by the technological level at
disposal. The axial velocity does play an important role, but is limited due to high losses.
Advanced aero engines can handle axial velocities up to 200 m/s. Thus, major analysis is
focused on angles, which are related to the temperature rise.
3.3.1
Tip speed
The rotational velocity of a gas turbine is limited by material technology. This
happens due to elevated levels of centrifugal tensile stress under which the blades are
submitted, and its maximum value, occurring at the blade root is given by:
53
( ct )max =
b 2
Sr
S ( r ) .r dr ,
rt
rr
(28)
where:
angular velocity;
Sr
r
rr
radius;
radius at blade root;
rt
Present technology imposes a 400 m/s limit at the blade tip. In fans, however, this
figure reaches 450 m/s. To evaluate the angular velocity, at which material limitation is
critical, let the tip speed limit be 350 m/s, then for a 5 cm radius microturbine and a 1.5 m
radius high bypass-ratio turbofan:
U = .r =
U 350
rad
=
= 7000
N = 66845 rpm
r 0.05
s
(29)
U 350
rad
=
= 233
N = 2228 rpm
r
1.5
s
(30)
U = .r =
3.3.2
per stage. Consider the case, which the relative inlet angle 1 is kept constant and the angle
54
2
V2
2 2
V2
C2
V1
C2
C1
Ca
U
Figure 21 Inlet and outlet relative velocity ratio is reduced with the increase of fluid deflection.
It is clear that high fluid deflection results in lower outlet relative velocity, this
means that more kinetic energy is converted to static pressure. In other words, high fluid
deflections, hence camber angle, entails a high rate of diffusion. Due to excessive losses, a
limit of diffusion exists and in preliminary design it is quantified by the de Haller number,
defined as:
deHaller
=
V2
> 0.69 .
V1
(31)
Originally, the limit was 0.72, but accumulated experience pushed this figure to 0.69.
3.3.3
Compressor surge
The surge is an unstable operation of the compressor, characterised by a sudden drop
of delivery pressure and by intense aerodynamic pulsation, which propagates from its origin
to the whole engine. The phenomenon yet very harmful to the engine is still not fully
understood. Usually it is related to excessive vibrations and a particular noise. The surge is
seen as the lower limit of stable operation, beyond which reversal of the flow is expected.
55
3.3.4
Compressor choke
From the gas dynamics, it is known that the maximum mass flow rate through a
nozzle is reached when the throat is at Mach 1. No matter what is done to increase the
pressure ratio, no extra flow pass through this nozzle. As the space between the blades forms
a nozzle, the compressor choke happens when the blade throats choke.
Thus, the operational range of a compressor, for each rotational speed, is bounded by
surge and choke.
56
THE
STREAMLINE
CURVATURE
COMPUTATIONAL
PROGRAM
4.1
INTRODUCTION
57
Due to the complex flow field observed in axial-flow compressors, many early
computational models failed to accurately predict performance characteristics. Nevertheless,
the SLCP demonstrated to be a fast and reliable performance prediction tool, as shown in [42]
by comparing its results with a real three-stage transonic compressor and with commercial
codes [47].
4.2
58
casing
blade trailing edge
m
streamline
IGV
rotor
stator
nodes
streamlines
Figure 23 Streamlines, stage rows and calculation nodes. Adapted from [42].
In the SLCM, the flow is divided into concentric streamtubes, wherein the flow is
axisymmetric. The flow is calculated according to inviscid equations and the losses due to
viscosity are incorporated as entropy increase, pressure decrease, etc. at the trailing edge, by
means of empirical correlations.
The basic derivation of the method can be found in Appendix A.
59
4.3
COMPUTATIONAL PROGRAM
The design mode of the SLC program is based on reference [42]. It is written in
FORTRAN language and it has been continuously updated. The SLC program is interactive,
fully modular and its high flexibility allows new features to be easily integrated.
At the present time, a simplification of the main structure of the program is
illustrated in Figure 24. Many convergence loops and subroutines were omitted for the sake of
clarity.
start
Channel inlet
Channel outlet
Channel intake
Vortex
Inlet
Radial equilibrium
Outlet
Incidence
Velocity triangles
Blading
Boundary layer
New grid
Efficiency
calculation
N
Efficiency
converged?
Y
Print tables
Write complete
geometry
Write optimisation
input file
Calculated efficiency
De Haller number
Camber
Pressure ratio
end
Losses
DCA
60
This chapter aims at providing the reader, who is not familiar with Genetic
Algorithms, with the basis to proceed without loss of understanding. To start with, important
definitions about Multiple Objective Optimisation Problem (MOOP) are presented. Next, the
fundamentals of Genetic Algorithms (GAs) are explained. Finally, a real-coded elitist multiobjective genetic algorithm program (REMOGA) developed in FORTRAN is detailed.
Optimisation may be defined as the search for solution(s) which correspond to
minimum (or maximum) of one or more objectives, satisfying given constraints. A singleobjective optimisation problem (SOOP) usually has one single optimal solution. A MOOP
accounts for multiple objectives, which may be conflicting. In this case, normally one obtains
not one, but a set of optimal solutions named Pareto optimal solutions. To compare two
solutions in a MOOP, the concept of dominance is introduced to encompass the idea that if a
certain solution a dominates solution b, then solution a is at least better in one objective and is
better or at least equal in all the other objectives.
5.1
DEFINITIONS
To accurately describe a MOOP, the specific vocabulary and definitions are made
clear. The definitions presented hereinafter are extracted or adapted from Deb [48] and Bche
61
[49]. To start with, a MOOP can be described by a vector of decision variables x and the
corresponding vector of objectives, f = f ( x ) .
5.1.1
Definition 1 The multi-objective optimisation problem is defined as the search for the set of
solutions x , which minimises/maximises:
=
min/max
f ( x)
( f ( x ) , f ( x ) , , f ( x ) ) F
1
( x1 , x2 , xn ) X
=
subject to
g ( x ) ( g1 ( x ) , g 2 ( x ) , , g p ( x ) ) 0
=
h ( x ) (=
h1 ( x ) , h2 ( x ) , , hq ( x ) ) 0
with =
x
xi( L ) xi xi(U ) ,
(32)
i=
1, 2, , n,
(
where X n is the n-dimensional decision space bounded by xi
L)
( )
and xi , and F m is
U
the m-dimensional objective space. The functions g ( x ) and h ( x ) are the constraint
functions. Any solution which satisfies every constraint, totalling p+q constraints and 2n
variable bounds is called a feasible solution, otherwise, infeasible solution.
An illustrative mapping from a 3-dimensional decision space to a 2-dimensional
objective space is shown in Figure 25.
Decision space
Objective space
x3
f2
z
x2
x1
f1
Figure 25 Mapping between the decision space and the objective space.
62
5.1.2
Domination
To compare different solutions from (32), an ordering among different solutions is
ab
i {1, 2, , m} :
fi ( a ) / fi ( b )
j {1, 2, , m} :
f j (a) f j (b )
(33)
b is dominated by a, or;
a is non-dominated by b.
If Definition 2 does not apply, then it is said that a does not dominate b, or, a / b .
63
f2
b
ab
a / c
f1
Figure 26 Representation of dominance and indifference between solutions in a two-objective
minimisation problem. Solution a dominates b, but is indifferent to c.
5.1.3
Pareto-optimal set
If in a given set of solutions, all possible pairwise comparisons are performed, one
eventually finds which solution dominates which and which solutions are not dominated with
respect to each other. This leads to an important set, named Non-dominated set:
64
P P is a non-dominated set a P, b P : a
/ b
(34)
Definition 5 (Globally Pareto-Optimal set). When the set P is the entire search space, i.e.,
P = X , then the non-dominated set P is called globally Pareto-optimal set.
For the sake of concision, the globally Pareto-optimal set is often referred to as
Pareto set.
5.1.4
Convexity
f ( a + (1 ) b ) f ( a ) + (1 ) f ( b ) ,
[ 0,1]
(35)
A convex f : function is illustrated in Figure 27. Note that the line segment
joining a and b is always greater or equal the function evaluated between those values.
f (x)
f ( a ) + (1 ) f ( b )
f (b )
f (a)
f ( a + (1 ) b )
a + (1 ) b
65
m
Definition 7 Let F . F is said to be a convex set if, given any two points members of
5.2
According to Goldberg [50], optimisation and search techniques fall onto three main
methods: calculus-based, enumerative and random. A brief description of each one is
provided to elucidate the reader the reason of the success of GAs in the turbomachinery
context.
Calculus-based methods are divided into two categories: indirect and direct. Indirect
methods rely on solving the set of equations provided by equalling the gradient of the
objective functions to zero. Direct methods are based on the iterative hill-climbing concept,
i.e., starting from a given point the gradient is calculated to provide the climb-direction of the
next point. Successively, a local optimum is found eventually. The main disadvantages of the
method are:
Almost implies that the objective function surface has to be known a priori.
66
Figure 28 Illustrative region where a gradient-based algorithm can get stuck onto a suboptimal solution.
To treat MOOPs, calculus-based methods most commonly convert the MOOP into a
SOOP summing all objectives with certain weights. The problem that arises is with the choice
of the weights and the fact that the optimisation will result in a single solution, rather than a
compromise Pareto set.
min/max
f (x)
=
w j f j ( x ), wi 1
=
=j 1 =i 1
( x1 , x2 , xn ) X
subject to
=
g ( x ) ( g1 ( x ) , g 2 ( x ) , , g p ( x ) ) 0
h1 ( x ) , h2 ( x ) , , hq ( x ) ) 0
=
h ( x ) (=
with =
x
xi( L ) xi xi(U ) ,
(36)
1, 2, , n,
i=
The enumerative method is the least intelligent algorithm and applicable to few a
simple cases. It relies on evaluating the objective function at every point, given a finite search
space; unquestionably, the maximum is found, as the whole search space is covered.
Nevertheless, it only fits problems, which full enumeration is practicable; however, it is
computationally inefficient and expensive, as there is no reason to explore unnecessary
regions. Enumerative methods are definitely infeasible for real-world multi-dimensional
engineering problems.
Random search algorithms are gaining popularity in Turbomachinery design as
shown in chapter 2. Goldberg [50] distinguishes random walks and randomized techniques.
67
The former implies a random scheme to search and save the best, while the latter, where GAs
are included, uses random choices as guiding tool.
5.3
start
Initialise
population
gen=0
Evaluation
gen = gen+1
Assign fitness
Condition
satisfied
?
No
Reproduction
Crossover
Mutation
Yes
Stop
68
x1 = 9 cm
8-bit string (1001 1100). If a refined step is desired, than more bits should be used.
Chromosome
10011100
A = 54 cm
x1 = 9
x2 = 12
x2 = 12 cm
Figure 30 Chromosomal representation of decision variables.
With the binary string representation in mind, the understanding of the selection,
crossover and mutation operations will be more transparent in the following sections.
GAs strategies usually evaluate not a single solution per time, but a pool of solutions.
Thus, recalling Figure 29, the first step after the algorithm is started is Initialise population.
This means that a certain amount of solutions in the design space is chosen by a given criteria
(which can follow a certain rule or be purely random).
Next, in the evaluation block, the objective functions are calculated and the fitness
is assigned to each solution. Fitness is the biological correspondent to objective and in
SOOP is the same as objective, but the concept goes beyond when dealing with MOOP as
will be treated afterwards.
Then the GA operators are performed to search optimal solutions.
69
5.3.1
population and eliminates bad solutions. It can be performed via tournament selection [50], in
which each solution plays against another solution and the fittest wins and the other is
eliminated. If each individual from the initial population plays the tournament twice (against
different individuals), the population size is kept. Moreover, the average fitness is improved.
Recalling the triangle area minimisation example, let the initial population be of
eight individuals and given as shown in the leftmost table from Figure 31. Each individual #1,
#2,...#8 has initially assigned values of x1 and x2. The two tables in the centre of Figure 31
illustrate the matches of each tournament. In the first tournament, solutions number 1, 4, 5
and 8 are winners. Prior to the second tournament, the matches are shuffled. Then, in the
second tournament solutions 6, 4, 8 and 7 are winners. Therefore, after the selection operator,
solutions 1, 4, 5, 6, 7 and 8 survive and solutions 2 and 3 are eliminated. In other words, the
best solutions are copied and the worst solutions are eliminated, keeping the population size
constant.
prior selec.
#
x1 x2 fit.
Tournament 1
#
x1 x2 fit.
Tournament 2
#
x1 x2 fit.
after selec.
#
x1 x2 fit.
21
21
16
21
15
30
15
30
21
16
21
21
21
15
30
16
16
av. fit.
14
The selection aims at passing to the next generation only the fittest. Were the
selection the unique operator, the best solution among the initial pool of candidates would be
70
found after some iterations. But the optimum may not be any of the initial candidates.
Therefore, the crossover and mutation operators come to provide further exploration of the
design space.
5.3.2
Crossover operator
The biological crossover, illustrated in Figure 32, is an exchange of genetic material
between chromosomes, which occurs during the meiosis. The crossover results in a new
arrangement and more diversity.
The computational parallel is performed thanks to the binary string concept. The
crossover can be a simple exchange of one or some bits in a pair of chromosomes. A bit-wise
crossover is illustrated in Figure 33. Differently from the selection operator, the crossover
generates new solutions, contributing to the exploration of the design space.
h = 9
b = 12
1001
1100
1001
0100
54
h = 10
b = 6
30
1010
0110
1010
1110
h = 8
b = 12
48
h = 10
b = 14
70
71
h = 9
b = 12
1001
1100
1001
0110
54
h = 10
b = 6
30
1010
0110
1010
1100
h = 9
b = 6
27
h = 10
b = 12
60
Lastly, the two-point crossover is operated by exchanging the bits between two
selected bits, as depicted in Figure 35. Bits in position 4, 5 and 6 are exchanged, resulting in
triangles with areas 16 and 77.
h = 9
b = 12
1001
1100
1000
0100
54
h = 10
b = 6
30
1010
0110
1011
1110
h = 8
b = 4
16
h = 11
b = 14
77
5.3.3
Mutation operator
To secure diversity, additionally to the crossover, the mutation operator is used. Its
working principle is simple: with a certain probability pmut , the operator changes a bit from 1
to 0, or vice versa. This probability affects the convergence and the exploration of the design
space.
72
54
h = 9
b = 12
Chromosome
1 0 0 1 1 1 00
Chromosome
1 0 1 1 1 1 00
h = 11
b = 12
66
Now that the basic concepts of GAs were presented, it will be easier to the reader,
who is not familiar with GAs to follow the mechanisms used in the real-coded multi-objective
genetic algorithm, which was developed as part of this work.
5.4
start
gen=1
gen = gen+1
Read initial
population
Sort according to
fitness
Evaluate
objectives
Write parent
population
Calculate rank
Crowded Tournament
Selection
Sort according to
rank
Real Polynomial
Mutation
Assess shared
fitness
parent offspring
Stop
Yes
Gen >
gen_max?
No
73
The basic algorithm is inevitably similar to the one presented in Figure 29, but its
features are expanded due to the multi-objective and real-coded approaches. The obtained
algorithm, which will be detailed hereafter, is depicted in Figure 37.
5.4.1
Multi-objective formulation
As aforementioned, the fitness is the equivalent to objective when dealing with a
SOOP, but its concept is changed when the problem is a MOOP. As the selection operator
uses the fitness to identify the best solutions and to eliminate the bad ones, the fitness must
encompass the concept of domination for MOOP.
5.4.1.1
rank ( i ) = 1 + ni .
(37)
Thus the non-dominated solutions are assigned a rank equal to 1 and worse solutions
are assigned values greater than 1. In this way an initial MOOP fitness of an individual, or its
raw fitness, fit R , is defined as the population size minus its rank, see Equation (38).
Proceeding in this way, better solutions are assigned with greater fitness values.
fit R=
( i ) npop rank ( i )
(38)
74
f m( j ) be the m-th objective of the j-th individual of a certain population of size npop, then the
rank assignment algorithm is depicted in Figure 38.
i=1
n i
rank=1
n=1
start
Y
flag1=0
flag2=0
j=1
Y
flag1=1
N
Y
flag2=1
N
j=j+1
N
j > nobj
Y
flag1=0
AND
flag2=1
rank=rank+1
N
i=i+1
N
Y
i>npop
rank(fn)=rank
N
n>npop
n=n+1
Y
end
The next step in the algorithm is the sorting of the population according to the
calculated rank.
5.4.1.2
Sorting
A simple Bubble sort algorithm was implemented for the sorting of the population.
The choice of this simple algorithm, rather than more efficient ones, e.g. Quicksort, was due
to the computationally small population per generation, which hardly exceeds 500.
The pseudocode of the Bubble sort algorithm is shown in Figure 39.
75
5.4.1.3
guaranteed. To avoid concentration of solutions in the objective space and encourage a broad
exploration of the codomain, solutions that are too close to other solutions should have its
fitness penalised. This penalty must encompass both the distance to other solutions and the
number of solutions that are nearby. Fonseca and Fleming [53] have also come up with clever
ideas about niche of solutions. The niche control can be visually understood in Figure 40.
f2
Without niche
penalty
f2
f1
With niche
penalty
f1
Figure 40 Optimisation (a) without niche penalty and (b) with niche penalty
76
The first step towards niche avoidance is the calculation of the normalised distance
between any two solutions i and j in a rank r:
1
m f (i ) f ( j ) 2 2
k
k
, rank ( i ) rank ( j ) .
dij =
=
max
min
k =1 f k f k
(39)
dij
Sh=
( dij ) max 1 , 0 ,
share
(40)
nci =
( ri )
Sh ( d ) ,
j =1
ij
(41)
where ( ri ) is the number of solutions in rank i. The niche count indicates the crowding
around a solution; high values of nci should be avoided in order to maintain diversity.
77
Sharing function
1.0
Sharing function
alpha
0.8
0.5
1.0
1.5
2.0
0.6
0.4
0.2
0.0
0
0.2
1.2
The next step is to use the niche count in the fitness. This can be done by dividing the
raw fitness by the niche count. This new fitness should be referred to as shared fitness. An
isolated solution has a niche count of 1, thus its shared fitness is not penalised, but another
solution in a crowded area has a niche count greater than 1, thus, the shared fitness is lower
than the raw fitness.
It was not revealed hitherto which value of share to use. A simple approach is to
define a fixed value based on a rough knowledge of the Pareto optimal region. Nevertheless,
rarely it is known a priori. Fonseca and Fleming [53] suggest a value based on the smallest
m-hypercube (m is the number of objectives) which contains the objective space and in its
minimum subdivisions in smaller m-hypercubes of edge share . Eventually it is suggested the
solution of the following polynomial equation:
m
( i i + share ) ( i i )
m 1
=
i 1 =i 1
npop=
. share
= 0, share > 0 ,
share
(42)
78
min f1 , , min f m ) (1 , , m ) ,
(=
( max f1 , , max f m ) = ( 1 , , m ) .
(43)
In the developed program, a simplification was used and the share was calculated as:
share
=
npop
1
m
1 m
max ( fi ) min ( fi ) .
m i =1
(44)
In Equation (44), the share was taken as if a m-hypercube with edge equal to the
average of the length of each objective was divided in such a way that npop m-hypercubes of
edge share would fit in this larger m-hypercube.
This idea may be easily understood with a 2-dimensional objective space, as shown
in Figure 42. Let 16 solutions be scattered in such a way that the difference of the maximum
value and minimum value for objectives 1 and 2 are, respectively, 6 and 4. The average edge
length is 5. Now, consider that all 16 individuals shown in Figure 42 are to be distributed in a
2-dimensional square set of edge 5. It can be achieved if the square can be divided into 16
smaller squares, what means a small square of edge 1.25, whose value will be used as share .
Naturally, this square may not be so orderly divided if the number of solutions is not a square
of a natural number, thus the mathematical formulation given by (44) expands the explained
idea not only for this case, but also for hyper-dimensional sets.
f2
6
16 indv.
f1
share
79
5.4.2
Tournament Selection (CTS). Differently from the GA Tournament Selection, the CTS uses
both the rank and the shared fitness to perform the selection of the fittest and elimination of
the bad solutions. The CTS operator works as follows: a solution i wins a tournament against
solution j, if:
if rank ( i ) =
rank ( j )
(45)
f2
4
1
56
2
1 vs. 4 : 1
4 vs. 5 : 4
3 vs. 1 : 3
f1
Figure 43 Crowded tournament selection operator.
DO i=1,nsel
CALL selection
END DO
Figure 44 Multiple selections.
80
5.4.3
approach. If a small step is desired, the string has to be very long, requiring more
computational effort. Moreover, even with a considerable refinement, the real space is not
fully explored. Hence, a real-coded genetic algorithm is necessary and a different crossover
and mutation mechanisms are required.
To assess the real-coded crossover as similar to the binary string, the simulated
binary crossover (SBX) from Deb and Agrawal [54] is used. First, the parameter i is
calculated as follows:
1
c +1 ,
u
2
(
)
i
1
i =
1
c 1
,
2 (1 ui )
if ui 0.5;
(46)
otherwise,
where ui ( 0,1) is a random number and c is a control parameter, which will be explained
afterwards.
Then, if xi(1,t ) represents the i-th parameter of solution 1 of generation t, the offspring
is calculated as follows:
xi(1, 1=)
1
(1 + i ) xi(1,t ) + (1 i ) xi( 2,t ) ,
(47)
xi(
1
(1 i ) xi(1,t ) + (1 + i ) xi( 2,t ) .
(48)
t+
2,t +1)
Considering the parent solutions xi(1,t ) = 1 and xi( 2,t ) = 1 , equations (47) and (48)
result in an offspring probability density distributions as shown in the histogram of Figure 45,
81
with step of 0.1. From it, one notes that the parameter c is responsible for the concentration
of the probability near the parent solutions.
Offspring histogram
20%
3.0
eta_c = 1
2.0
eta_c = 1
15%
1.0
0.0
10%
-1.0
5%
-2.0
0%
-3.0 -2.0 -1.0 0.0
20%
eta_c = 2
15%
1.0
2.0
3.0
-3.0
0.2
0.0
0.4
3.0
eta_c = 2
2.0
0.6
0.8
1.0
0.6
0.8
1.0
0.2
0.4
0.6
0.8
random number value
1.0
1.0
0.0
10%
-1.0
5%
-2.0
0%
-3.0 -2.0 -1.0 0.0
20%
eta_c = 4
15%
1.0
2.0
3.0
-3.0
0.0
0.2
0.4
3.0
eta_c = 4
2.0
1.0
0.0
10%
-1.0
5%
-2.0
0%
-3.0 -2.0 -1.0 0.0 1.0 2.0
offspring solution
3.0
-3.0
0.0
High values of c showed slower convergence but more stability, small values of c
granted faster convergence for simple tests, but incurred in divergence or even loss of
optimum regions.
The implemented crossover received an elite-preserving operator. To avoid the loss
of good solutions due to disadvantageous crossover, good solutions should bypass the
crossover. This was carried as follows: if solution i has rank ( i ) relit , where relit is given by
the user, then with a certain user-defined probability (preferably high), solution i undergoes
the crossover with another solution, but one of the offspring will receive the value from i
without change and the other one will receive a value given by the SBX.
82
Differently from what is commonly found, the elite-preserving operator used is not
on-off, but it allows for a probability of occurrence of elitism. No study was conducted by the
author to analyse if it is any better, but it is an additional feature.
5.4.4
operator. Deb and Goyal [55] proposed the following polynomial function:
)
yi(=
xi(
1,t +1)
1,t +1
+ max i ,
(49)
where yi(1,t +1) is the solution xi(1,t +1) after the mutation, and:
=
max max xi( n ,t +1) min xi( n ,t +1) .
n
(50)
m +1 1,
u
2
(
)
i
i =
1
1 2 (1 u ) m +1 ,
i
if ui < 0.5,
(51)
if ui 0.5.
83
8.0%
step:
0.02
eta_m = 1
6.0%
eta_m = 1
0.8
0.4
0.0
4.0%
-0.4
2.0%
-0.8
0.0%
-1.0
-0.5
8.0%
eta_m = 2
0.0
0.5
1.0
step:
0.02
6.0%
-1.2
0.0
1.2
x1
= 0
delta = 1
0.2
0.4
0.6
0.8
1.0
eta_m = 2
0.8
0.4
0.0
4.0%
-0.4
2.0%
-0.8
0.0%
-1.0
-0.5
8.0%
eta_m = 4
6.0%
0.0
0.5
1.0
step:
0.02
-1.2
0.0
1.2
x1
= 0
delta = 1
0.2
0.4
0.6
0.8
1.0
eta_m = 4
0.8
0.4
0.0
4.0%
-0.4
2.0%
-0.8
0.0%
-1.0
-0.5
0.0
0.5
mutated solutions
1.0
-1.2
0.0
x1
= 0
delta = 1
0.2
0.4
0.6
0.8
random number
1.0
To visualise the results from the program, the code generates dat files which are
imported by an Excel sheet, which was prepared to provide a graphical interface. The Excel
sheet created for visualisation is also capable of displaying an animation of the evolution of
the population along the generations and the result after each operator for each generation.
To exemplify the selection, crossover and mutation operators using the graphical
interface, a simple 2-minimisation objective test function being operated is show in Figure
47.
84
selection
Objective space
crossover
Objective space
mutation
Objective space
Objective space
10
Gen
8
4
f2
6
4
2
0
10
Gen
5
f2
6
4
2
0
10
Gen
6
f2
6
4
2
0
0
6
f1
10 0
6
f1
10 0
10 0
f1
10
f1
The first row shows the objective space in the fourth generation, the second row
represents the fifth generation and the third row shows the sixth generation. From the first
column to the second column, the selection operator effect is observed. One should note that
this operator does not create any new solution. Visually it does only remove some nonoptimal solutions, as a result of the tournament; but actually a duplication of good solutions
takes place. From the second to the third column, the crossover operator effect is observed. In
this case, new solutions are created as a result of pair-wise combination. Lastly, from the third
to the fourth column, the mutation takes place. Its effect is similar to the effect of the
crossover in the sense that it further explores the objective space. Being an easy and known
a priori optimisation problem, the REMOGA parameters were chosen to provide a quick
convergence, as one notices the rapid migration of the solutions towards the Pareto optimal
set.
85
5.5
TEST FUNCTIONS
The REMOGA program was verified using test functions found in the literature.
First, a very simple test of a convex problem was carried only to check the functionality of the
selection, crossover and mutation operators. Then a 2-objective optimisation problem with a
non-convex Pareto, as well as a discontinuous set, was successfully tested.
5.5.1
minimise
f1 ( x1 , x2=
) x12 + x22 ,
f 2 ( x1 , x2 ) = ( x1 + 2 ) + x22 ,
2
(52)
5 xi 5
c 30;=
m 0.5;
1.0 (from Eq. 8) and=
=
pmut 0.05;
=
=
pelit 0.8,
(53)
is relatively fast and its evolution per generation is illustrated in Figure 48, which shows the
decision and objective spaces at the first, 10th and 100th generations.
This test function was the same used to illustrate the steps of the algorithm in Figure
47. Nevertheless, in Figure 48 one can also follow the behaviour of the decision space at the
left. At the right, two dials labelled f1 and f2 show the average of each objective function.
86
Decision space
Objective space
10
Gen.: 1
Gen.: 1
4
f2
x2
6
0
10
10
-2
-4
-6
0
-6
-4
-2
0
x1
f2
10
f1
Decision space
Objective space
10
Gen.: 10
Gen.: 10
4
f2
x2
0
f1
10
10
-2
-4
-6
0
-6
-4
-2
0
x1
f2
10
f1
Decision space
Objective space
10
Gen.: 100
Gen.: 100
4
f2
x2
0
f1
10
10
-2
2
-4
0
0
-6
-6
-4
-2
0
x1
0
f1
f2
10
f1
Figure 48 Testing a simple MOOP. Population in the 1st, 10th and 100th generations.
Figure 49 shows this MOOP after 100 generations displaying every individual of
each generation. As 120 generations were used, a total of 12,000 individuals were evaluated.
87
f2
4
2
0
0
10
f1
Figure 49 Simple convex test function after 100 generations.
5.5.2
minimise
minimise
2
2
f1 ( x1 , x2 ) =
1 exp ( x1 1) ( x2 + 1) ,
2
2
f 2 ( x1 , x2 ) =
1 exp ( x1 + 1) ( x2 1) ,
4 xi 4
(54)
=
c 2,=
pmut 0.05,
=
m 0.5,
=
1.0,=
pelit 0.75,
(55)
and using a population of npop = 120 and 200 generations, a non-convex Pareto is
successfully obtained as shown in Figure 50.
88
f2
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
f1
Figure 50 Non-convex test function from Fonseca and Fleming [56].
5.5.3
[48]. It has a non-convex and disconnected Pareto-optimal set and it is defined as follows (the
actual problem is a maximisation one, but here it was multiplied by -1 to convert it into a
minimisation problem):
2
2
minimise
f1 ( x1 , x2 ) =1 + ( A1 B1 ) + ( A2 B2 ) ,
2
2
f 2 ( x1 , x2 ) = ( x1 + 3) + ( x2 + 1) ,
minimise
xi , i =
1, 2,
where
0.5 1.0
2.0 1.5
=
a =
,
b
1.5 2.0
1.0 0.5
(56)
Using the following GA settings for a population of npop = 120 and 500 generations
89
=
c 200;=
pmut 0.10;
=
m 0.5;
=
1.0;=
pelit 0.8,
(57)
results in two disconnected Pareto sets, which can be observed in Figure 51.
objective 2
25
20
Pareto
optimal
regions
15
10
5
0
0
4
6
objective 1
10
5.6
This chapter provided the reader with the basic concepts of GAs and details of the
program, which was written to perform real-coded multi-objective optimisation. Starting from
definitions and proceeding with the essential ideas behind GA operators, namely: selection,
crossover and mutation, the reader was prepared to a more complex real-coded multiobjective program. The basic structure is maintained, but the fitness is associated with
dominance. The crossover and mutation operators are substituted for equations instead of
bitwise operations.
90
METHODOLOGY
The objective of this chapter is to provide the reader with the methods through which
an initial compressor was optimised by coupling the SLC program and the real-coded elitist
multi-objective genetic algorithm program (REMOGA) described in the previous chapter.
Illustratively, the integration depicted in Figure 52, where the REMOGA acts reading the
SLCP output and proposing inputs.
Modified SLCP
File 1
SLCP input
REMOGA output
File 2
SLCP output
REMOGA input
REMOGA
6.1
Initially, the preliminary design was chosen to be carried on with 5 streamlines and
15 rows, where 3 rows were dummy to simulate the inlet channel, rotors and stators rows after
one another to shape 5 stages and 2 final dummies to simulate the compressor outlet.
When recalling the SLCP, a generic grid node will be defined by the tangent to the
blade edge (i) and the streamline (j). So, i varies from 1 to 16 and j from 1 to 5. The innermost
streamline is j=1 and the outermost streamline is j=5. The tangent to the blade edges i defines
the rows given by Table 4.
91
i
12
23
34
45
56
67
78
89
Row
Dummy inlet channel
Dummy inlet channel
Dummy inlet channel
Rotor 1
Stator 1
Rotor 2
Stator 2
Rotor 3
i
9 10
10 11
11 12
12 13
13 14
14 15
15 16
Row
Stator 3
Rotor 4
Stator 4
Rotor 5
Stator 5
Dummy compressor outlet
Dummy compressor outlet
To conduct the integration of the SLC program with the REMOGA program, some
specific modifications were required. The SLCP was adapted to work as a blackbox, which
receives a couple of input information and returns another group of output information, for
any input.
Modified SLCP
Camber angle penalty
Neither message boxes, nor infinite loops should exist in the program. As the
REMOGA calls the SLCP many times, the integration should run without full human
supervision. Therefore, several routines were reviewed to receive stop criteria and argument
tests to avoid crashes, e.g., calculation of square root of a negative number. In case of
unfeasible solution, i.e., if a stop criteria like the square root one is reached, the program
actually returns a feasible, but bad solution with regard to the objectives. This was
conducted so that the bad solution is killed in the optimisation process, but never crashes
or stops without an output file.
This was a relatively time-consuming task, requiring long hours of problem-seeking
simulations followed by laborious debugging.
92
Processor
Processor release date
Clock
Cores / Threads
Memory
Operating System
Av. time original SLCP
Av. time modified SLCP
6.1.1
Notebook
Intel Core 2 Duo P7450
2009 1st Quarter
2.13 GHz
2/2
4.00 GB
Windows 7 Home Premium
35.34 s
3.05 s
REMOGA program is a file named optimisation.dat and is as simple as shown in Figure 54.
93
This file contains the objectives that the REMOGA has to minimise. To operate
maximisation objectives they were converted to minimisation of the opposite value.
In this work, the four objectives were:
6.1.1.1
j =meanline.
(58)
i =ibegin
limitation is carried on for every streamline. Thus, the parameter to be minimised, the camber
angle penalty, was defined as:
94
pen
=
iend
jend
=i ibegin
=j jbegin
min ( 0, ij 40 )
(59)
6.1.2
which contains more than 100 parameters to be chosen by the designer. Information such as
ambient conditions, number of stages, mass flow rate, tip speed limit, stator air outlet angles,
hub-to-tip ratio of the first row, space to chord ratios, aspect ratios, blade profiles, tip
clearance, etc. have to be properly set. One can find the complete input file in Appendix C.
The second input file is a selection of parameters from the original input file to be
used in the REMOGA. The SLCP actually first reads the original input file and then
immediately reads the second one. So, modifications can be done only in the second file, what
eases the handling of parameters.
The preliminary design optimisation was carried in two steps: search and refinement.
The first step was a preliminary search and was executed with large ranges for the following
12 variables:
Stator air outlet angle for each stator, assuming no radial variation of this
angle, and
Hub-to-tip ratio.
95
The refinement step used the history information from the search step to focus on
regions, where promising solutions are located. Moreover, the stator air outlet angle was
allowed to vary linearly from hub to tip, totalling 17 design variables.
6.2
The MOOP formulation for the axial-flow compressor search is, then:
maximise
maximise
minimise
minimise
subject to
6.3
calculated
pr
pendH
pen
0.85 input 0.90
(60)
( 3 )1 , ( 3 )2 , ( 3 )3 , ( 3 )4 [10, 40]
( 3 )5 [ 0, 20]
htr [ 0.40, 0.60]
REMOGA SETTINGS
The REMOGA program was set with conservative parameters, as the objective space
is unknown, but is clearly a not simple one. Preliminary tests revealed that the SLC is very
sensitive with regard to some parameters and incur in large number of unfeasible solutions.
The concept for multiple selections and rank-based elitism in the REMOGA was idealised
because of tests in which more than 60% of solutions were unfeasible. Hence, multiple
selections help eliminating unfeasible solutions and rank-based elitism avoids the loss of
converged solutions.
96
6.4
radius [m]
0.15
0.10
0.05
0.00
-0.10
-0.05
0.00
0.10
0.15
0.05
axial coordinate [m]
0.20
0.25
0.30
In summary, the target isentropic efficiency was set as 85%; all stator air outlet
angles were set as 25, hub to tip ratio of 0.55 and stagewise temperature rise weights
distribution as shown Figure 56.
97
1.10
30.00
1.05
25.00
1.00
20.00
0.95
15.00
0.90
10.00
0.85
5.00
0.00
0.80
s1
s2
s3
s4
s5
S1
S2
S3
S4
S5
The pressure and temperature distribution of the human design is show in Figure 57,
from where a pressure ratio of 4.901 is observed.
Pressure [MPa]
0.50
500
tt1(i)
0.40
400
0.30
300
0.20
200
0.10
100
0.00
Temperature [K]
0.60
0
4
8
10
compressor row
12
Although the resulting preliminary design does not satisfy all conditions, it is
potentially feasible. Thus, it could proceed to a thorough performance analysis and further
detailing, thereby meeting all remaining restrictions.
The camber angles and the de Haller number distribution along the nodes are shown
in Figure 58 and Figure 59. The white region indicates satisfactory regions, while the grey
region indicates unsatisfactory regions. As the de Haller number is considered only at the
meanline, the light grey region indicates an indifferent region.
On one hand, the overall camber angle distribution is reasonable, but the 4th stator
presents some high values, but not dramatic. On the other hand, the de Haller number
98
distribution is quite problematic. Not only are there five rows in the meanline with de Haller
numbers below 0.69, but also the first stator has an abnormally high values of de Haller
number. Moreover, the 4th stator has a steep spanwise de Haller number distribution.
The stage loading distribution shown in Figure 60 reveals that the majority of the
nodes are within the recommended interval, but the outermost streamline concentrates low
values of and the innermost streamline of the first rotor has an abnormal high value.
original solution - camber angle - rotor
5
streamline
R1
R2
R3
R4
R5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
streamline
S1
S2
S3
S4
S5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
99
streamline
R1
R2
R3
R4
R5
3
2
1
0.40
0.60
0.50
0.80
0.70
de Haller number [ - ]
0.90
1.00
1.10
streamline
S1
S2
S3
S4
S5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
1.10
streamline
R1
R2
R3
R4
R5
3
2
1
0.10
0.20
0.30
0.50
0.40
loading coefficient [ - ]
0.60
0.70
Finally, the number of blades in each row is shown in Figure 61. The total number of
blades is 529.
100
4 5 6 7 8 9 10 11 12 13
# blades 20 29 40 47 54 59 64 67 66 83
The blade chord of each row is show in Figure 62. The values are all reasonable.
3.25
2.5
2.69
2.0
2.18
1.97
1.5
1.77 1.65
1.53 1.56 1.60
1.0
1.35
0.5
0.0
4
10
11
12
13
101
7.1
The last REMOGA run aiming at optima compressors counted with 20,000 designs.
From this set, a subset with 6,162 (30.8%) solutions contains the feasible solutions and
another intersecting subset with 18,437 (92.2%) solutions contains the unique solutions. The
intersection of those sets, i.e., unique and feasible solutions delivers 4,625 solutions (23.1%).
Figure 63 illustrates the distribution of solutions in an Euler diagram.
Feasible
6,162
(30.8%)
Unique
18,437
(92.2%)
Intersection
4,625
(23.1%)
Figure 63 Euler diagram representing the sets of feasible and unique solutions.
(61)
leads to a final subset that contains feasible, unique and limited solutions with regard to the
penalties. This subset contains 3,064 solutions. The history of those solutions can be seen in
Figure 64 to Figure 65.
102
20
40
60
generation
80
100
20
40
60
generation
80
100
103
1.500
1.500
1.400
1.400
1.300
1.300
1.200
1.200
1.100
1.100
1.000
1.000
0.900
0.900
0.800
0.800
0.700
0.700
0
20
40
60
generation
80
100
20
40
60
generation
80
100
1.500
1.500
1.400
1.400
1.300
1.300
1.200
1.200
1.100
1.100
1.000
1.000
0.900
0.900
0.800
0.800
0.700
0.700
0
20
40
60
generation
80
100
20
40
60
generation
20
40
60
generation
80
100
80
100
104
History: angle S1
History: angle S2
30.0
30.0
25.0
25.0
20.0
20.0
15.0
15.0
10.0
10.0
5.0
5.0
0
20
40
60
generation
80
100
History: angle S3
20
40
60
generation
80
100
80
100
History: angle S4
30.0
30.0
25.0
25.0
20.0
20.0
15.0
15.0
10.0
10.0
5.0
5.0
0
20
40
60
generation
80
100
20
40
60
generation
History: angle S5
30.0
25.0
20.0
15.0
10.0
5.0
0
20
40
60
generation
80
100
105
7.2
To look for better solutions, an initial approach is to plot the objectives and check
where the REMOGA results are located in comparison to the initial design. Observing Figure
68 may lead to precipitated conclusions that the human design is undoubtedly better than the
REMOGA solutions. However, attention shall be conveyed to the fact that the air outlet angle
of the last stator row is limited in 20 in the REMOGA designs, differently from the initial
design, with a high angle of 25.
REMOGA designs
REMOGA designs
5.5
5.5
Human design
5.3
5.1
Pressure ratio
5.1
Pressure ratio
Human design
5.3
4.9
4.7
4.5
4.3
4.1
4.9
4.7
4.5
4.3
4.1
3.9
3.9
3.7
3.7
3.5
3.5
0
10
20
30
40
Camber angle penalty
50
10
20
5th stage 3 [deg]
30
Figure 68 Pressure ratio vs. camber penalty and last stage stator outlet angle for the limited subset of
solutions.
A high stator air outlet angle hinders the proper functioning of the combustion
chamber, due to elevated swirl velocities, hence the choice to limit it in 20.
Figure 69 shows that REMOGA designs present less potential problems with excess
of diffusion, indicated by the de Haller number penalty. Moreover, Figure 69 indicates that
the optimisation is capable of finding better solutions.
106
REMOGA designs
REMOGA designs
5.5
50
Human design
5.1
Pressure ratio
Human design
5.3
4.9
4.7
4.5
4.3
4.1
3.9
3.7
3.5
0.000
0.100
0.300
0.200
de Haller number penalty
40
30
20
10
0
0.000
0.100
0.200
de Haller penalty
0.300
A solution, whose analysis is of interest, is the one that is at the right of the human
design solution (Figure 70). It has a similar pressure ratio and camber angle penalty and has a
lower stator air outlet angle.
REMOGA designs
5.5
Human design
5.3
Solution 1
Pressure ratio
5.1
4.9
4.7
4.5
4.3
4.1
3.9
3.7
3.5
0
10
20
30
40
Camber angle penalty
50
Figure 70 Solution 1.
Let this solution be referred to as solution 1. Another solution of interest it the one
with the highest pressure ratio, but with no de Haller number penalty, nor camber angle
penalty (hereinafter, solution 2). Solution 2 has isentropic efficiency of 90% and pressure
ratio of 4.0.
107
7.3
Solution 1 and solution 2 were obtained with the following temperature weights and
stator outlet air angles:
Solution 1
Stage temperature weights
1.40
1.20
20.00
1.00
0.80
15.00
0.60
10.00
0.40
5.00
0.20
0.00
0.00
d1
d2
d3
d4
d5
S1
S2
S3
S4
S5
Solution 2
Stage temperature weights
1.60
1.40
1.20
1.00
0.80
0.60
0.40
0.20
0.00
d1
d2
d3
d4
d5
S1
S2
S3
S4
S5
Both solutions present low temperature weights in the last two stages. Similarly,
stator outlet air angles of initial stages are lower and increase in later stages.
7.3.1
Overview
To provide a first visual idea of the optima compressors geometries, the streamlines
and nodes are displayed in red together with the original compressor, shown in grey. Figure
72 shows the streamlines and nodes of solution 1. It is noticeable that the kink in the later
stages does not exist anymore, but a kink in the initial stages is now present.
108
radius [m]
0.15
0.10
0.05
0.00
-0.10
-0.05
0.00
0.05
0.10
0.15
axial coordinate [m]
0.20
0.25
0.30
Solution 2, whose streamlines are shown in Figure 73, has no kinks, neither in the
inlet, nor in the outlet.
radius [m]
0.15
0.10
0.05
0.00
-0.10
-0.05
0.00
0.05
0.10
0.15
axial coordinate [m]
0.20
0.25
0.30
109
The pressure and temperature rise along the stages obtained by the solutions in study
are shown in Figure 74.
Pressure [MPa]
0.50
500.0
tt1(i)
0.40
400.0
0.30
300.0
0.20
200.0
0.10
100.0
0.00
Temperature [K]
0.60
0.0
4
8
10
compressor row
12
Pressure [MPa]
0.50
500.0
tt1(i)
0.40
400.0
0.30
300.0
0.20
200.0
0.10
100.0
0.00
Temperature [K]
0.60
0.0
4
10
8
compressor row
12
Solution 1 has a total of 507 blades and solution 2, a total of 515 blades, both have
less than 529 from the original design. This positive result came without setting the number of
blades as an objective. The number of blades per row and the blade chord per row are
depicted in Figure 75 and Figure 76.
110
4 5 6 7 8 9 10 11 12 13
# blades 18 29 32 41 46 53 60 71 78 79
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
3.47
2.93
2.46
2.18
1.91
1.72
1.53 1.40
1.28 1.30
10
11
12
13
Figure 75 Number of blades and blade chord of each row for solution 1.
4 5 6 7 8 9 10 11 12 13
# blades 18 29 32 41 46 53 62 71 80 83
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
3.50
2.94
2.45
2.16
1.88
1.68
1.49 1.35
1.22 1.24
10
11
12
13
Figure 76 Number of blades and blade chord of each row for solution 2.
7.3.2
Camber angle
While the original design presents high camber angles in the 4th stator, with angles of
ca. 45, solutions 1 and 2 perform better in this aspect, as shown in Figure 77 and Figure 78.
Solution 1 works slightly beyond the established limit of 40 in outermost streamlines of
stators 2 and 5. Solution 2 not only does have all camber angles inferior to 40, but also has a
smoother spanwise distribution.
111
streamline
R1
R2
R3
R4
R5
3
2
1
0
10
15
30
25
20
camber angle [ deg ]
35
40
45
50
streamline
S1
S2
S3
S4
S5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
streamline
R1
R2
R3
R4
R5
3
2
1
0
10
15
30
25
20
camber angle [ deg ]
35
40
45
50
streamline
S1
S2
S3
S4
S5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
112
7.3.3
De Haller number
The de Haller number was an issue in the original design, as there are values as low
as 0.64 (meanline). Imposing the de Haller penalty as objective, the de Haller number at the
meanline was successfully controlled, as shown in Figure 79 and Figure 80. The yellow
rectangle highlights the region of interest for the de Haller number, i.e., meanline and values
higher than 0.69.
Both solutions 1 and 2 have acceptable figures. Solution 2 has a very well-behaved
de Haller number distribution. Solution 1 has an abnormal value for the last stator row at
streamline 5.
7.3.4
Stage loading
The stage loading distribution is better in solution 1, as more nodes are located
within the recommended range (white region), as shown in Figure 81. However, both
solutions face low values at streamline 5 and rotor 1 has a strange behaviour at streamline 1.
Further investigation should be carried on to check whether this high loading is real or is the
result of some numerical problem.
113
streamline
R1
R2
R3
R4
R5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
streamline
S1
S2
S3
S4
S5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
streamline
R1
R2
R3
R4
R5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
streamline
S1
S2
S3
S4
S5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
114
streamline
R1
R2
R3
R4
R5
3
2
1
0.10
0.20
0.30
0.50
0.40
loading coefficient [ - ]
0.60
0.70
streamline
R1
R2
R3
R4
R5
3
2
1
0.10
0.20
0.30
0.40
0.50
loading coefficient [ - ]
0.60
0.70
7.4
As detailed in the previous chapter, the first run was a search procedure. This step
revealed two interesting solutions, but more importantly, through the history of feasible
solutions, the ranges of promising design variable for a second optimisation run were
determined.
Hence, the refinement formulation is given by (62):
115
maximise
maximise
minimise
minimise
subject to
calculated
pr
pendH
pen
0.87 input 0.90
Tw1 , Tw2 , Tw3 [1.10,1.30]
(62)
In the refinement step, 31200 solutions were tested in 156 generations, resulting in
4372 (14%) feasible and unique solutions. Filtering according to the same criteria from (61)
2857 (9.2%) solutions were analysed. Again, the more relevant conflicting objectives are the
pressure ratio and the camber angle penalty, as shown in Figure 82.
The de Haller number is not a major issue in the solutions obtained. There is a good
concentration of solutions with very little de Haller number penalty, as shown in Figure 83.
REMOGA designs
5.3
Pressure ratio
5.1
4.9
4.7
4.5
4.3
4.1
3.9
3.7
Human design
3.5
0
10
20
30
40
Camber angle penalty
50
Figure 82 Pressure ratio vs. camber angle penalty from the refinement run.
116
REMOGA designs
REMOGA designs
5.5
50
Human design
Camber angle penalty
5.3
Pressure ratio
5.1
4.9
4.7
4.5
4.3
4.1
3.9
3.7
40
30
20
10
Human design
3.5
0
0
0.1
0.2
de Haller number penalty
0.3
0.1
0.2
de Haller penalty
0.3
Proceeding similarly to the search step, two solutions from the refinement step are
going to be detailed. Solution 3 was chosen due to its high pressure ratio and proximity to the
human design. It has a pressure ratio of 4.836 and isentropic efficiency of 87.2%. Solution 4
was chosen as the solution with the highest pressure ratio and no de Haller penalty nor camber
angle penalty. This filter yields to a solution with 4.399:1 pressure ratio and 87.1% isentropic
efficiency.
REMOGA designs
5.3
Pressure ratio
5.1
Solution 3
4.9
4.7
4.5
4.3
4.1
3.9
3.7
Human design
3.5
0
10
20
30
40
Camber angle penalty
50
117
7.5
1.20
hub
15.00
1.00
tip
0.80
10.00
0.60
0.40
5.00
0.20
0.00
0.00
d1
d2
d3
d4
d5
Solution 4
Stage temperature weights
1.40
1.20
15.00
1.00
hub
tip
0.80
10.00
0.60
0.40
5.00
0.20
0.00
0.00
d1
d2
d3
d4
d5
7.5.1
Overview
The streamlines distribution is shown in Figure 86. Little visual difference is noticed
between them. Similarly to solutions 1 and 2, the outer diameters are also smaller. A small
kink is observed at approximately 0.02 m of axial distance, but no kink at the outlet.
118
radius [m]
0.15
0.10
0.05
0.00
-0.10
-0.05
0.00
0.05
0.10
0.15
axial coordinate [m]
0.20
0.25
0.30
0.25
0.30
radius [m]
0.15
0.10
0.05
0.00
-0.10
-0.05
0.00
0.05
0.10
0.15
axial coordinate [m]
0.20
Pressure and temperature distributions along the rows are shown in Figure 87. Notice
that they have a similar pattern, but solution 3 is slightly superior in every rotor row; this has
a considerable impact in the overall pressure ratio of 4.836 vs. 4.399.
119
500.0
tt1(i)
0.40
400.0
0.30
300.0
0.20
200.0
0.10
100.0
0.00
0.0
Temperature [K]
1.50
pt1(i)
0.50
Pressure [MPa]
600.0
1.40
1.30
1.20
1.10
1.00
0.90
8
10
compressor row
9 10 11 12 13
12
500.0
tt1(i)
0.40
400.0
0.30
300.0
0.20
200.0
0.10
100.0
0.00
0.0
Temperature [K]
1.50
pt1(i)
0.50
Pressure [MPa]
600.0
0.60
1.40
1.30
1.20
1.10
1.00
0.90
4
8
10
compressor row
9 10 11 12 13
12
As Figure 88 and Figure 89 show, solution 3 totals 493 blades and solution 4 requires
511 blades. The blade chords are also within an acceptable range. The smallest one solution
4, rotor of the 5th stage has a blade chord of 1.24 cm.
R2
S2
R3
S3
1.31
S1
1.28
R1
1.40
R1 S1 R2 S2 R3 S3 R4 S4 R5 S5
# blades 18 29 32 37 44 51 60 67 76 79
1.52
1.71
20
1.90
40
2.18
60
2.47
80
2.98
100
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
3.56
R4
S4
R5
S5
Figure 88 Number of blades and blade chord of each row for solution 3.
120
R2
S2
R3
S3
1.27
S1
1.24
R1
1.36
R1 S1 R2 S2 R3 S3 R4 S4 R5 S5
# blades 18 29 32 41 46 53 62 71 80 79
1.48
1.67
20
1.86
40
2.13
60
2.43
80
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
0.00
2.94
100
3.52
R4
S4
R5
S5
Figure 89 Number of blades and blade chord of each row for solution 4.
7.5.2
Camber angle
Figure 90 and Figure 91 show the distribution of camber angles of solutions 3 and 4,
respectively. They are below the upper limit of 40 (except some nodes in solution 3).
Solution 4 presents a smoother almost linear variation from hub to tip (except stator 2 and 5).
7.5.3
De Haller number
De Haller numbers at the meanline are also above 0.69 (Figure 92 and Figure 93)
Solution 3 has an abnormal value for the streamline 5 of stator 5, which certainly is not real.
121
streamline
R1
R2
R3
R4
R5
3
2
1
0
10
15
30
25
20
camber angle [ deg ]
35
40
45
50
streamline
S1
S2
S3
S4
S5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
streamline
R1
R2
R3
R4
R5
3
2
1
0
10
15
30
25
20
camber angle [ deg ]
35
40
45
50
streamline
S1
S2
S3
S4
S5
3
2
1
0
10
15
20
25
30
camber angle [ deg ]
35
40
45
50
122
streamline
R1
R2
R3
R4
R5
3
2
1
0.40
0.50
0.60
0.80
0.70
de Haller number [ - ]
0.90
1.00
streamline
S1
S2
S3
S4
S5
3
2
1
0.40
0.50
0.60
0.70
0.80
de Haller number [ - ]
0.90
1.00
streamline
R1
R2
R3
R4
R5
3
2
1
0.40
0.50
0.60
0.80
0.70
de Haller number [ - ]
0.90
1.00
streamline
S1
S2
S3
S4
S5
3
2
1
0.40
0.50
0.60
0.80
0.70
de Haller number [ - ]
0.90
1.00
123
7.5.4
Stage loading
Finally, the stage loading are plotted in Figure 94. No major improvement from
solutions 1 and 2 to solutions 3 and 4 is seen. But again, the outermost nodes are not very
loaded.
solution 3 - loading coefficient - rotor
5
streamline
R1
R2
R3
R4
R5
3
2
1
0.10
0.20
0.30
0.40
0.50
loading coefficient [ - ]
0.60
0.70
streamline
R1
R2
R3
R4
R5
3
2
1
0.10
0.20
0.30
0.50
0.40
loading coefficient [ - ]
0.60
0.70
Eventually, no noticeable improvement was noticed between the solutions from the
search step to the solutions from the refinement step. However, solutions 3 and 4 meet similar
results from solutions 1 and 2, but the former is more strict with the stator air outlet angle,
whose upper bound is 18 instead of 20. The compressor revealed to be very sensitive to this
parameter, so that the human design has fewer restrictions to camber angle using the
aforementioned angle in 25. This angle is limited due to restrictions in the combustion
chamber.
Appendix D provides further graphical information from the analysed solutions.
124
CONCLUSIONS
125
Eventually, four solutions (two from the search and two from the refinement) were
analysed. Although they revealed to be consistent in terms of de geometry, number of blades,
de Haller number, camber angle, etc., none reached the pressure ratio of 5.
The procedure developed in this work revealed that the preliminary design of an
axial-flow compressor can be optimised thanks to the full automation and inherent
intelligence of the developed multi-objective genetic algorithm. More than 100,000 designs
were evaluated during this work and this took just some weeks, while a non-automated
procedure would require some minutes per design and analysis. If the manual process of
deciding upon design variables, running the SLCP and analysing the results takes around 30
minutes for each design, then the 100,000 evaluations would require 50,000 hours, or 2083
days of non-stop work. In this sense this work contributed to the preliminary design of axialflow compressors.
126
FURTHER WORK
Throughout the development of this work, ideas for further development naturally
come. Some are related to improvements and others to works that can be derived from the
obtained results.
9.1
IMPROVEMENTS
The REMOGA, as presented, was not conceived at once. It suffered many changes
throughout the integration with the SLCP. The main changes were related to the high error
level of the SLCP. Nevertheless, time is limited and some ideas or models to improve the
algorithm were not tested.
The first one is the normalisation of the fitness, so that the average fitness is kept
constant though the generations. This might improve the efficacy of the other selection
operators.
The second REMOGA proposed improvement is the treatment of unfeasible
solutions. At the present moment, unfeasible solutions receive penalised objective values, in a
way that they would be killed in the selection operation. However, when working with high
level of errors, the tournament selection can create disputes between two identical penalised
solutions, passing a penalised solution to the next generation. A suggestion for this aspect is
to substitute penalised solutions by existing feasible solutions with low rank. It may have a
127
similar effect of the selection operator, but grants elimination of unfeasible solutions and the
crossover is, then more effective for the generation of new solutions.
A limiting issue of the REMOGA is that it runs one design per time and does not
benefit of multiple threads existing in modern processors. A relatively simple approach would
be the execution of a batch program, which manages the executable files in different folders,
so that there is no mixing of input and output files.
9.2
9.2.1
SUGGESTION OF WORKS
Detailed project
After some optimum preliminary design solutions were found, the next natural step is
the deep investigation of them. Existing in-house design-point and off-design point
performance programs can be used to assess the feasibility of the obtained compressors.
Investigations can also be carried on with commercial programs, like Concepts NREC.
9.2.2
Robust optimisation
Existing cooperation in turbomachinery optimisation research between ITA and
Universidade Federal de Itajub (UNIFEI) can use the procedure developed in this work to
conduct further optimisation studies in robust optimisation, which are themes of research at
UNIFEI. This would be an enriching further study to be conducted in the axial-flow
compressor design, in order to understand the impact of input variables that are given by
probability distributions.
128
REFERENCES
129
11. SONODA, T. et al. Advanced high turning compressor airfoils for low Reynolds number
condition. Part 1: design and optimization. In: ASME TURBO EXPO, 48., 2003,
Atlanta. Proceedings... New York: ASME, 2003. CD-ROM.
12. BURGUBURU, S. et al. Numerical optimization of turbomachinery bladings. In: ASME
TURBO EXPO, 48., 2003, Atlanta. Proceedings... New York: ASME, 2003. CD-ROM.
13. BCHE, D.; GUIDATI, G.; STOLL, P. Automated design optimization of compressor
blades for stationary, large-scale turbomachinery. In: ASME TURBO EXPO, 48., 2003,
Atlanta. Proceedings... New York: ASME, 2003. CD-ROM.
14. DEMEULENAERE, A.; LIGOUT, A.; HIRSCH, C. Application of multipoint
optimization to the design of turbomachinery blades. In: ASME TURBO EXPO, 49.,
2004, Vienna. Proceedings... New York: ASME, 2004. CD-ROM.
15. JUN, L. et al. Multiobjective optimization approach to turbomachinery blades design. In:
ASME TURBO EXPO, 50., 2005, Reno-Tahoe. Proceedings... New York: ASME,
2005. CD-ROM.
16. YI, W.; HUANG, H.; HAN, W. Design optimization of transonic compressor rotor using
CFD and Genetic Algorithm. In: ASME TURBO EXPO, 51., 2006, Barcelona.
Proceedings... New York: ASME, 2006. CD-ROM.
17. KESKIN, A.; DUTTA, A. K.; BESTLE, D. Modern compressor aerodynamic blading
process using multi-objective optimization. In: ASME TURBO EXPO, 51., 2006,
Barcelona. Proceedings... New York: ASME, 2006. CD-ROM.
18. JANG, C.-M.; SAMAD, A.; KIM, K.-Y. Optimal design of swept, leaned and skewed
blades in a transonic axial compressor. In: ASME TURBO EXPO, 51., 2006, Barcelona.
Proceedings... New York: ASME, 2006. CD-ROM.
19. VO, C. et al. Automated Multiobjective optimisation in axial compressor blade design.
In: ASME TURBO EXPO, 51., 2006, Barcelona. Proceedings... New York: ASME,
2006. CD-ROM.
20. PAPADIMITRIOU, D. I.; GIANNAKOGLOU, K. C. Compressor blade optimization
using a continuous adjoint formulation. In: ASME TURBO EXPO, 51., 2006, Barcelona.
Proceedings... New York: ASME, 2006. CD-ROM.
21. IYENGAR, V.; SANKAR, L. N.; DENNEY, R. A first-principles based methodology
for design of axial compressor configurations. In: ASME TURBO EXPO, 52., 2007,
Montreal. Proceedings... New York: ASME, 2007. CD-ROM.
22. KUZMENKO, M. L. et al. Optimization of the gas turbine engine parts using methods of
numerical simulation. In: ASME TURBO EXPO, 52., 2007, Montreal. Proceedings...
New York: ASME, 2007. CD-ROM.
23. SAMAD, A.; KIM, K.-Y. Stacking and thickness optimization of a compressor blade
using weighted average surrogate model. In: ASME TURBO EXPO, 53., 2008, Berlin.
130
131
132
133
APPENDIX A
SLC SUMMARY
The detailed derivation can be found in the PhD thesis of Barbosa [42] and in the
Report of Frost [58]. This section aims at providing basic guidance to the reader who is not
familiar with the method.
From the velocity triangles:
C= V + U .
(63)
Cm1
Cm 2
U1
Vw1
Vw 2
Cw1
U2
Cw 2
C = Cr r + C + Cz z ,
(64)
V =Vr r + V + Vz z ,
(65)
U = r z ,
(66)
where r , and z indicate the radial, tangential and axial components, respectively.
From Equations (63) to (66),
134
C r = Vr ,
(67)
C=
V + r ,
(68)
Cz = Vz .
(69)
For a non-inertial coordinate frame rotating with the rotor at a angular velocity ,
the inviscid and steady-flow equation of motion is
DV
P=
+ ( r ) + 2 V + F ,
Dt
where
(70)
D ( )
is the material derivative, whence:
Dt
DV
V V V
V
= Cr
+
+ Vz
.
Dt
r
r
z
(71)
Combining (70) and (71), a mathematical representation of a steady-state nonviscous flow in cylindrical coordinates rotating about a fixed axis is obtained:
V W V
V
P Cr
=
+
+ Cz
+ ( r ) + 2 V + F .
r
r
z
(72)
( r ) =
2 r r :
Cr V Cr
Cr V2
1 P
= Cr
+
+ Vz
2 r 2V ,
r
r
z
r
r
(73)
V V V
V C V
1 P
= Cr + + Cz + r + 2Cr ,
r
r
r
z
r
(74)
Cz V Cz
Cz
1 P
.
= Cr
+
+ Cz
r
z
z
r
(75)
135
C=
C
m
r + Cz ,
=
Cm
Cr2 + Cz2 .
(76)
(77)
As the streamline is tangent to the velocity and with aid of Figure 22:
Cr = Cm sin ,
(78)
Cz = Cm cos .
(79)
r z
=
+
.
m m r m z
(80)
Cr
r
= sin
=
,
m
Cm
(81)
Cz
z
= cos
=
,
m
Cm
(82)
Cr
r
= tan
=
.
z
Cz
(83)
= sin + cos , or
m
r
z
Cm= Cr + Cz .
m
r
z
(84)
(85)
136
Moreover:
Cr Cr Cr Cz Cr
Cr
C
1 Cr
=
+
=
+ Cz r
Cr
m Cm r Cm z
m Cm
r
z
(86)
Cz Cr Cz Cz Cz
Cr
C
1 Cz
=
+
=
+ Cz z ,
Cr
m Cm r Cm z
m Cm
r
z
(87)
( rC )
C
r
.
= r + C
m
m
m
(88)
( rC )
C C C
r C
=
+ CZ + r .
Cr
Cm
r
m
r
z
(89)
Having in mind that C= U + V and from (86) - (89) substituted into (73) - (75), one
gets:
C
C C 2
C C 2
1 P
1 P
= Cr r + C z r = Cm r ,
r
z
m
r
r
r
r
C
C C V
Cm ( rC )
,
0
+ Cz + r
=
r
z
r
r
m
(91)
C z
C z
C z
1 P
1 P
= Cr
+ Cz
= Cm
r
z
m
z
z
(92)
0 Cr
=
(90)
The right-hand side of Equations (90)-(92) are written in terms of the streamline mcoordinate. Now, the left-hand side should be written in terms of the s-coordinate, which is
along the blade edge. Therefore, an analogous procedure is conducted by noticing that
s = s ( r , z ) , so:
r z
=
+
s s r s z
(93)
137
Recalling Figure 22 and that the angle is defined as the angle between the r-axis
and s-axis:
r
= cos
s
(94)
z
= sin
s
(95)
= cos + sin
s
r
z
(96)
cos ( + ) tan ( )
cos 2 ( ) s
R
c
2U
cos ( )
tan ( )
Cm
S
r
I
2
2 Cm
+ Cm
Cm= cos ( ) T
s
s
s
tan 2 ( ) cos ( ) +
+ sin ( + ) Cm
m
Cm
(97)
where,
tan ( + )
sec ( + ) +
+
C
Cm
Rc
s
Cm
=
.
2
m (1 M m ) 1 S
1
2
+ R m (1 + M m ) r sin ( )
2
m
(98)
Equations (97) and (98) form a system of partial differential equations. The system can
be solved if it is previously know that flow properties vary smoothly at the blade edges.
138
APPENDIX B
B.1
OPTIMISATION PROGRAM
MAIN PROGRAM
PROGRAM main
INCLUDE "zcommon.f90"
REAL*8
REAL*8
REAL*8
INTEGER
INTEGER(4)
::
::
::
::
::
var1
mem1(20)
objective
i,j,imax,jmax
seed_time
!population comumns
INTEGER :: nvarb,nvare,nobjb,nobje,nrank,nfit
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
nrank=nobje+1
nfit=nrank+1
!Initialise generation
gen=1
!'Randomise' the random operator
seed_time=TIME()-1318600000
CALL SEED(seed_time)
139
DO gen=1,ngen
!Evaluate Objective Functions
CALL eval_objective
!Assign fitness by rank
CALL fitness
!Write parent population prior selection
CALL write_parent
!Selection
DO i=1,nsel
CALL selection
ENDDO
!Write parent population after selection
CALL write_parent
!Cross-over
CALL cross_over
!Write offspring population prior mutation
CALL write_offspring
!Mutation
CALL mutation
!Write offspring population after mutation
CALL write_offspring
!Parent receives offspring
parent=offspring
ENDDO
ENDPROGRAM main
B.2
GLOBAL VARIABLES
!zcommon.f90
USE IFPORT
INTEGER,PARAMETER :: npop=200,nvar=12,nobj=4,totcol=nvar+nobj+3
INTEGER :: obj_selec,gen,ngen,rankel,nsel
140
REAL :: parent(npop,totcol),offspring(npop,totcol)
REAL :: upper_var(nvar),lower_var(nvar)
B.3
&
&
!-------------------------------------------------------!
!
Subroutine to read the initial user-defined
!
population, the GA parameters and the
!
variable limits
!
!-------------------------------------------------------!developed by VICTOR FUJII ANDO, MSc Aer-Mech. Eng.
!18th of October 2011
!--------------------------------------------------------
SUBROUTINE read_ini
INCLUDE "zcommon.f90"
INTEGER :: i
!population comumns
INTEGER :: nvarb,nvare,nobjb,nobje,nrank,nfit
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
nrank=nobje+1
nfit=nrank+1
OPEN(3,file="ini.dat")
OPEN(4,file="GA_parameters.dat")
!Read initial population from an external file
DO i=1,npop
parent(i,1)=REAL(i)
READ(3,*) parent(i,nvarb+ 0),parent(i,nvarb+ 1),parent(i,nvarb+ 2),&
&parent(i,nvarb+ 3),parent(i,nvarb+ 4),parent(i,nvarb+ 5),&
&parent(i,nvarb+ 6),parent(i,nvarb+ 7),parent(i,nvarb+ 8),&
&parent(i,nvarb+ 9),parent(i,nvarb+10),parent(i,nvarb+11)
parent(i,nvarb)=-parent(i,nvarb)
ENDDO
!Define variable limits
!effisen_given
lower_var(1)=-0.90
141
upper_var(1)=-0.85
!temp dist 1
lower_var(2)=0.80
upper_var(2)=1.35
!temp dist 2
lower_var(3)=0.80
upper_var(3)=1.35
!temp dist 3
lower_var(4)=0.80
upper_var(4)=1.35
!temp dist 4
lower_var(5)=0.80
upper_var(5)=1.35
!temp dist 5
lower_var(6)=0.80
upper_var(6)=1.35
!alfa2 - S1
lower_var(7)=10
upper_var(7)=40
!alfa2 - S2
lower_var(8)=10
upper_var(8)=40
!alfa2 - S3
lower_var(9)=10
upper_var(9)=40
!alfa2 - S4
lower_var(10)=10
upper_var(10)=40
!alfa2 - S5
lower_var(11)=0
upper_var(11)=20
!hub-tip ratio
lower_var(12)=0.40
upper_var(12)=0.60
!Read GA parameters
READ(4,*)obj_selec
READ(4,*)ngen
READ(4,*)nsel
READ(4,*)rankel
READ(4,*)eta_c
READ(4,*)p_mut
READ(4,*)eta_m
READ(4,*)alpha_sh
READ(4,*)p_elit
!To simplify calculations:
!redefine Crossover parameter "eta_c"
eta_c=1/(1+eta_c)
!To simplify calculations:
!redefine Mutation paramenter "eta_m"
eta_m=1/(eta_m+1)
ENDSUBROUTINE read_ini
142
B.4
EVALUATING OBJECTIVES
!-------------------------------------------------------!
!
Subroutine to evaluate the objective functions
!
!-------------------------------------------------------!developed by VICTOR FUJII ANDO, MSc Aer-Mech. Eng.
!18th of October 2011
!Last update: 21th Nov 2011
!-------------------------------------------------------!Features
! 1. Simple 2-objective convex front test
! 2. A 2-objective non-convex front test
! 3. Poloni et al. discontinuous front test function
! 4. User-defined, from an external program
!-------------------------------------------------------!Revision 1: 08th Nov. 2011
!
Use result from axial design program
!
!Revision 2: 21st Nov. 2011
!
Selection of objectives using obj_selec
!--------------------------------------------------------
SUBROUTINE eval_objective
INCLUDE "zcommon.f90"
INTEGER
INTEGER
INTEGER(4)
CHARACTER*8
REAL
REAL
::
::
::
::
::
::
i,j,count
nvarb,nvare,nobjb,nobje
int_time
char_time
x1,x2
A1,A2,B1,B2
LOGICAL(4) result
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
!Pressure ratio
!Isentropic
!itype dist 1=temp
!number of weights
!dist1
!dist2
!dist3
!dist4
!dist5
143
tip ratio
D1
D2
D3
R1
S1
R2
S2
R3
S3
R4
S4
R5
S5
D1
D2
D3
D1
D2
D3
R1
S1
R2
S2
R3
S3
R4
S4
R5
S5
D1
D2
D3
int_time=TIME()
CALL TIME(char_time)
WRITE(*,10)gen,i,char_time
10 FORMAT ('
gen = ',I3.3,'
n = ',I3.3,'
time: ',A8)
WRITE(*,11)parent(i,nvarb) ,parent(i,nvarb+1),parent(i,nvarb+
2),parent(i,nvarb+ 3),&
&parent(i,nvarb+4),parent(i,nvarb+5),parent(i,nvarb+
6),parent(i,nvarb+ 7),&
&parent(i,nvarb+8),parent(i,nvarb+9),parent(i,nvarb+10),parent(i,nvarb+11)
11 FORMAT (' ',<nvar>F10.5)
parent(i,nvarb)=-parent(i,nvarb)
result=SYSTEMQQ('axial_dp')
OPEN(6,file='optimisation.dat')
144
count=0
DO WHILE ((.NOT. EOF(6)).AND.(count.LE.nobj))
READ(6,*)parent(i,nobjb+count)
count=count+1
ENDDO
parent(i,nobjb)=-parent(i,nobjb)
parent(i,nobjb+3)=-parent(i,nobjb+3)
IF (count .LT.2) THEN
parent(i,nobjb)=0
parent(i,nobjb+1)=500
parent(i,nobjb+2)=600
parent(i,nobjb+3)=0
WRITE(*,*)'
out: --------- fail ---------'
ELSE
WRITE(*,20),parent(i,nobjb),parent(i,nobjb+1),parent(i,nobjb+2),parent(i,no
bjb+3)
20 FORMAT(' ',<nobj>F10.5)
ENDIF
CLOSE(6,status='delete')
ENDDO
ENDIF
!------------------------------------------------------------!Test functions
!------------------------------------------------------------! Simple
! Fonseca and Fleming
! Poloni
!------------------------------------------------------------IF(obj_selec.NE.1)THEN
DO i=1,npop
DO j=nobjb,nobje
x1=parent(i,2)
x2=parent(i,3)
IF(obj_selec.EQ.4)THEN
A1=0.5*SIN(1.)-2.0*COS(1.)+1.0*SIN(2.)-1.5*COS(2.)
A2=1.5*SIN(1.)-1.0*COS(1.)+2.0*SIN(2.)-0.5*COS(2.)
B1=0.5*SIN(x1)-2.0*COS(x1)+1.0*SIN(x2)-1.5*COS(x2)
B2=1.5*SIN(x1)-1.0*COS(x1)+2.0*SIN(x2)-0.5*COS(x2)
ENDIF
IF (j.EQ. 4) THEN
IF(obj_selec.EQ.2)parent(i,j)=x1**2+x2**2
IF(obj_selec.EQ.3)parent(i,j)=1-EXP(-(x1-1)**2-(x2+1)**2)
IF(obj_selec.EQ.4)parent(i,j)=(1+(A1-B1)**2+(A2-B2)**2)
ENDIF
IF (j.EQ.5) THEN
IF(obj_selec.EQ.2)parent(i,j)=(2+x1)**2+x2**2
IF(obj_selec.EQ.3)parent(i,j)=1-EXP(-(x1+1)**2-(x2-1)**2)
IF(obj_selec.EQ.4)parent(i,j)=((x1+3)**2+(x2+1)**2)
ENDIF
ENDDO
145
ENDDO
ENDIF
!-------------------------------------------------------------
ENDSUBROUTINE eval_objective
B.5
FITNESS SUBROUTINE
!-------------------------------------------------------!
!
Subroutine to perform the calculation of
!
FITNESS, SHARED FITNESS and RANK
!
!-------------------------------------------------------!developed by VICTOR FUJII ANDO, MSc Aer-Mech. Eng.
!18th of October 2011
!-------------------------------------------------------!Structure
! 1. Assignement of rank of each solution
! 2. Sorting of the population using simple BubbleSort
! 3. Definition of array mu(n), which gives the number
!
of solutions with rank n
! 4. Calculation of Fitness by the fitness-averaging
!
method
!--------------------------------------------------------
SUBROUTINE fitness
INCLUDE "zcommon.f90"
INTEGER
INTEGER
REAL
REAL
::
::
::
::
i,j,n,swap,memo_int
flag1,flag2,rank,mu(npop)
memo(totcol),memo_real,sum
niche(npop)
INTEGER :: nvarb,nvare,nobjb,nobje,nrank,nfit
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
nrank=nobje+1
nfit=nrank+1
! Assign rank
DO n=1,npop
rank=1
DO i=1,npop
IF (n .NE. i) THEN
flag1=0
flag2=0
DO j=nobjb,nobje
IF (parent(i,j).GT.parent(n,j)) flag1=1
IF (parent(i,j).LT.parent(n,j)) flag2=1
ENDDO
IF ((flag1.EQ.0).AND.(flag2.EQ.1)) rank=rank+1
146
ENDIF
ENDDO
parent(n,nrank)=rank
ENDDO
!\Assign rank
! Assign fitness
DO n=1,npop
sum=0
!Memo_real represents the number of solutions with rank
!equal to the rank of parent n, inclusive
memo_real=REAL(mu(INT(parent(n,nrank))))
IF (parent(n,nrank).EQ.1) THEN
parent(n,nfit)=REAL(npop)-0.5*(memo_real-1)
ELSE
memo_int=INT(parent(n,nrank)-1)
DO j=1,memo_int
sum=sum+(REAL(mu(j))-0.5*(memo_real-1))
ENDDO
parent(n,nfit)=REAL(npop)-sum
ENDIF
ENDDO
!\Assign fitness
147
ENDSUBROUTINE fitness
B.5.1
!-------------------------------------------------------!
!
Subroutine to perform the calculation of
!
NICHE COUNT
!
!-------------------------------------------------------!developed by VICTOR FUJII ANDO, MSc Aer-Mech. Eng.
!18th of October 2011
!-------------------------------------------------------!Structure
! 1. Calculation of nomalised distance
! 2. Dynamic update of Sigma share
! 3. Calculation of Sharing function
! 4. Calculation of Niche count
!-------------------------------------------------------SUBROUTINE niche_count(mu,niche)
INCLUDE "zcommon.f90"
INTEGER :: n,i,j,auxi1,auxi2
INTEGER :: mu(npop)
!Normalized Euclidian distance between solutions
REAL
:: distance(npop,npop)
!Sharing function
REAL
:: sharing(npop,npop)
!Niche count
REAL
:: niche(npop)
!Sharing parameter - dynamic update
REAL
:: sigma_sh
148
!Auxiliary variables
REAL
:: aux1,aux2,aux3,sum
REAL
:: delta_min,delta_max
INTEGER :: nvarb,nvare,nobjb,nobje,nrank,nfit
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
nrank=nobje+1
nfit=nrank+1
149
B.6
DO n=1,npop/2
IF (parent(2*n-1,nrank).LT.parent(2*n,nrank)) THEN
memory(2*n-1,:)=parent(2*n-1,:)
ELSEIF ((parent(2*n-1,nrank).EQ.parent(2*n,nrank)).AND.(parent(2*n1,nfit).GT.parent(2*n,nfit))) THEN
memory(2*n-1,:)=parent(2*n-1,:)
ELSE
memory(2*n-1,:)=parent(2*n,:)
ENDIF
ENDDO
CALL shuffle
150
DO n=1,npop/2
IF (parent(2*n-1,nrank).LT.parent(2*n,nrank)) THEN
memory(2*n,:)=parent(2*n-1,:)
ELSEIF ((parent(2*n-1,nrank).EQ.parent(2*n,nrank)).AND.(parent(2*n1,nfit).GT.parent(2*n,nfit))) THEN
memory(2*n,:)=parent(2*n-1,:)
ELSE
memory(2*n,:)=parent(2*n,:)
ENDIF
ENDDO
parent=memory
CALL shuffle
ENDSUBROUTINE selection
B.7
DO n=1,npop-1,2
DO i=nvarb,nvare
j=i-nvarb+1
! Random number
seed_time=TIME()-1318600000
CALL SEED(seed_time)
rand1=RAND()
!\Random number
151
! Offspring calculation
offspring( n,1)=parent( n,1)
offspring(n+1,1)=parent(n+1,1)
IF(((parent(n,nrank)-1).LT.0.01).AND.(RAND().LT.p_elit))THEN
offspring( n,i)=parent(n,i)
offspring(n+1,i)=0.5*((1betaq)*parent(n,i)+(1+betaq)*parent(n+1,i))
ELSEIF(((parent(n+1,nrank)-1).LT.0.01).AND.(RAND().LT.p_elit))THEN
offspring(n+1,i)=parent(n+1,i)
offspring( n,i)=0.5*((1+betaq)*parent(n,i)+(1betaq)*parent(n+1,i))
ELSE
offspring( n,i)=0.5*((1+betaq)*parent(n,i)+(1betaq)*parent(n+1,i))
offspring(n+1,i)=0.5*((1betaq)*parent(n,i)+(1+betaq)*parent(n+1,i))
ENDIF
!\Offspring calculation
! Check limits
IF (offspring(n,i).LT.lower_var(j))offspring(n,i)=lower_var(j)
IF (offspring(n,i).GT.upper_var(j))offspring(n,i)=upper_var(j)
IF (offspring(n+1,i).LT.lower_var(j))offspring(n+1,i)=lower_var(j)
IF (offspring(n+1,i).GT.upper_var(j))offspring(n+1,i)=upper_var(j)
!\Check limits
ENDDO
ENDDO
ENDSUBROUTINE cross_over
B.8
!-------------------------------------------------------!
!Subroutine to perform the polynomial mutation
!
!-------------------------------------------------------!developed by VICTOR FUJII ANDO, MSc Aer-Mech. Eng.
!18th of October 2011
!--------------------------------------------------------
SUBROUTINE mutation
152
INCLUDE "zcommon.f90"
INTEGER :: n,i,j,seed_time
REAL
:: delta,rand1,rand2
REAL
:: min_var(nvar),max_var(nvar),delta_var(nvar)
INTEGER :: nvarb,nvare,nobjb,nobje,nrank,nfit
nvarb=2
nvare=1+nvar
nobjb=nvare+1
nobje=nobjb+nobj-1
nrank=nobje+1
nfit=nrank+1
DO i=nvarb,nvare
min_var(i-nvarb+1)=MINVAL(offspring(:,i))
max_var(i-nvarb+1)=MAXVAL(offspring(:,i))
delta_var(i-nvarb+1)=max_var(i-nvarb+1)-min_var(i-nvarb+1)
ENDDO
DO n=1,npop
rand2=RAND()
IF (rand2.LT.p_mut)THEN
DO i=nvarb,nvare
! Random number
rand1=RAND()
!\Random number
! Parameter delta
IF (rand1.LT.0.5) THEN
delta=(2*rand1)**eta_m-1
ELSE
delta=1-(2*(1-rand1))**eta_m
ENDIF
offspring(n,i)=offspring(n,i)+delta_var(i-nvarb+1)*delta
! Check limits
IF (offspring(n,i).LT.lower_var(i-nvarb+1))offspring(n,i)=lower_var(invarb+1)
IF (offspring(n,i).GT.upper_var(i-nvarb+1))offspring(n,i)=upper_var(invarb+1)
!\Check limits
ENDDO
ENDIF
ENDDO
ENDSUBROUTINE mutation
153
APPENDIX C
The original SLCP input file as given by the human design, which is the starting
point of this work follows:
0
=0 => no IGV
=1 => with IGV
=99 sample file
101325. 288.0 0.0
Pa
Ta Flight Mach No
3
ndummybefore
6.000 5
5
5
pr
nstage
jmax njmax
0.850
effisen_given 0.85828 0.84669
0.01
ac_eff=var_effic (delta effic para acerto)
0.001
var_effic (precisao)
0.0001
check_effic
0.250
f15=lim wtotal
8.20
mass flow kg/s
1 400.0 400.00
ispd (1=U; 2=RPM)
1
flag 1=temp
2=press
0
iogv
3
ndummyafter
1
itype_dist
=1 temperature
=2 pressure
5
num_weigths
0.950
temp or press distribution - stage 1
1.100
temp or press distribution - stage 2
1.250
temp or press distribution - stage 3
0.950
temp or press distribution - stage 4
0.800
temp or press distribution - stage 5
2
iloss (1=Swan
2=Msarratt)
2
iaxial_chanel (=1 linear rotor-estator; =2 linear rotor-rotor)
0.00 distrib of etac stage 1
0.00 distrib of etac stage 2
0.00 distrib of etac stage 3
0.00 distrib of etac stage 4
0.00 distrib of etac stage 5
0
idistalfa2 =0 linear
=1 for each streamline
25.0 25.0 25.0 25.0 25.0
alfa2 S 1 jmax streamlines
25.0 25.0 25.0 25.0 25.0
alfa2 S 2 jmax streamlines
25.0 25.0 25.0 25.0 25.0
alfa2 S 3 jmax streamlines
25.0 25.0 25.0 25.0 25.0
alfa2 S 4 jmax streamlines
25.0 25.0 25.0 25.0 25.0
alfa2 S 5 jmax streamlines
0.55
hub-tip ratio
0.000
dr1p
10.0
dvinlet
10.0
dvoutlet
1.00
xkoxamb
0.60
dstall
0.50
0.30
axial Mach N. at inlet and outlet of the compressor
1.00
xkb1 - Dummy1 before
1.00
xkb1 - Dummy2 before
1.00
xkb1 - Dummy3 before
1.00 - xkb1 - R 1
0.98 - xkb1 - S 1
0.97 - xkb1 - R 2
0.95 - xkb1 - S 2
0.94 - xkb1 - R 3
0.92 - xkb1 - S 3
154
0.91 - xkb1 - R 4
0.89 - xkb1 - S 4
0.88 - xkb1 - R 5
0.86 - xkb1 - S 5
0.86
xkb1 - Dummy1 after
0.86
xkb1 - Dummy2 after
0.86
xkb1 - Dummy3 after
0.86
xkb1 - Dummy4 after
cod
cod=constant outer diam cid=constant inner diameter cmd=constant mean
diameter
v1u
distrib of Vu
1 2
iradeq1 (1=old 2=new)
1 2
iradeq2 (1=old 2=new)
1 0
ikt
=0 single precision
=1 double precision
1 0
iio10
=0 single precision
=1 double precision
1 0
iin
=0 single precision
=1 double precision
1 0
ikdelta
=0 single precision
=1 double precision
1 0
ido10
=0 single precision
=1 double precision
1 0
inb
=0 single precision
=1 double precision
1 0
id0di2d
=0 single precision
=1 double precision
1 0
ii0ci2d
=0 single precision
=1 double precision
1 0
id0cd2d
=0 single precision
=1 double precision
1
secondary loss
1=Griepentropg_s method 2=Howell
1.00 0.30 0.30
lslr srlr ssls
0.90 0.90
relax relax1 - for convergence purposes only
1
ivortx
dca - R1
blade type - dca/b65s/c4/myblade
dca - S1
dca - R2
dca - S2
dca - R3
dca - S3
dca - R4
dca - S4
dca - R5
dca - S5
1.00 - s/c Dummy 1 before
1.00 - s/c Dummy 2 before
1.00 - s/c Dummy 3 before
1.10 - s/c - R 1
0.95 - s/c - S 1
0.93 - s/c - R 2
0.91 - s/c - S 2
0.89 - s/c - R 3
0.87 - s/c - S 3
0.85 - s/c - R 4
0.83 - s/c - S 4
0.81 - s/c - R 5
0.79 - s/c - S 5
0.59 - s/c Dummy 1 after
0.59 - s/c Dummy 2 after
0.59 - s/c Dummy 3 after
3.00 - h/c Dummy 1 before
3.00 - h/c Dummy 2 before
3.00 - h/c Dummy 3 before
2.00 - h/c - R 1
2.00 - h/c - S 1
2.00 - h/c - R 2
2.00 - h/c - S 2
2.00 - h/c - R 3
2.00 - h/c - S 3
2.00 - h/c - R 4
2.00 - h/c - S 4
2.00 - h/c - R 5
2.00 - h/c - S 5
1.88 - h/c Dummy 1 after
1.88 - h/c Dummy 2 after
1.88 - h/c Dummy 3 after
155
0.01
tip clearance R 1
0.01
tip clearance S 1
0.01
tip clearance R 2
0.01
tip clearance S 2
0.01
tip clearance R 3
0.01
tip clearance S 3
0.01
tip clearance R 4
0.01
tip clearance S 4
0.01
tip clearance R 5
0.01
tip clearance S 5
0.00001
wmcn1 - tolerance for mass flow !era 0.001
0.010 min space between rows - Dummy 1 before
0.010 min space between rows - Dummy 2 before
0.010 min space between rows - Dummy 3 before
0.010 min space between rows - R 1
0.010 min space between rows - S 1
0.010 min space between rows - R 2
0.010 min space between rows - S 2
0.010 min space between rows - R 3
0.010 min space between rows - S 3
0.010 min space between rows - R 4
0.010 min space between rows - S 4
0.010 min space between rows - R 5
0.010 min space between rows - S 5
0.010 min space between rows - Dummy 1 after
0.010 min space between rows - Dummy 2 after
0.010 min space between rows - Dummy 3 after
10.0 8.0
t/c - R 1 - hub - tip thickness-chord ratio %
8.0 8.0
t/c - S 1 - hub - tip
10.0 8.0
t/c - R 2 - hub - tip thickness-chord ratio %
8.0 8.0
t/c - S 2 - hub - tip
10.0 8.0
t/c - R 3 - hub - tip thickness-chord ratio %
8.0 8.0
t/c - S 3 - hub - tip
10.0 8.0
t/c - R 4 - hub - tip thickness-chord ratio %
8.0 8.0
t/c - S 4 - hub - tip
10.0 8.0
t/c - R 5 - hub - tip thickness-chord ratio %
8.0 8.0
t/c - S 5 - hub - tip
6.0 6.0
r/t - R 1 - hub - tip leading and trailig edges radius %
6.0 6.0
r/t - S 1 - hub - tip
6.0 6.0
r/t - R 2 - hub - tip leading and trailig edges radius %
6.0 6.0
r/t - S 2 - hub - tip
6.0 6.0
r/t - R 3 - hub - tip leading and trailig edges radius %
6.0 6.0
r/t - S 3 - hub - tip
6.0 6.0
r/t - R 4 - hub - tip leading and trailig edges radius %
6.0 6.0
r/t - S 4 - hub - tip
6.0 6.0
r/t - R 5 - hub - tip leading and trailig edges radius %
6.0 6.0
r/t - S 5 - hub - tip
170.0
V1a - R 1
168.0
V1a - R 2
166.0
V1a - R 3
164.0
V1a - R 4
160.0
V1a - R 5
0
idevmod (0=carter deviation correlation)
45.
angbullet - spinner
0.001
rmin
0.5
alfakb - relaxation factor for boundary layer
0.50
percentage of stage pressure loss attributed to the
0.50
percentage of stage pressure loss attributed to the
0.50
percentage of stage pressure loss attributed to the
0.50
percentage of stage pressure loss attributed to the
0.50
percentage of stage pressure loss attributed to the
10.0
flare
0.5
xkzdummy - % dummy no bullet
0.001
wrdes (streamlines reposition tolerance)
1.0d0
xacochamb
6.50 6.50 6.50 6.50 6.50 additional incidence R 1
3.50 3.50 3.50 3.50 3.50 additional incidence S 1
6.50 6.50 6.50 6.50 6.50 additional incidence R 2
rotor
rotor
rotor
rotor
rotor
R
R
R
R
R
1
2
3
4
5
156
3.50
6.50
3.50
6.50
3.50
6.50
3.50
0.72
-0.005
0.000
0.007
0.045
0.070
0.070
0.070
0.070
0.070
50
0.40d0
0.40d0
0.25d0
0.25d0
3.50
6.50
3.50
6.50
3.50
6.50
3.50
3.50
6.50
3.50
6.50
3.50
6.50
3.50
no
no
no
no
radeq2
radeq2
radeq2
radeq2
157
APPENDIX D
ADDITIONAL
INFORMATION
OBTAINED SOLUTIONS
FROM
THE
158
streamline
R1
R2
R3
R4
R5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
R1
R2
R3
R4
R5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
R1
R2
R3
R4
R5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
R1
R2
R3
R4
R5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
D.1
R1
R2
R3
R4
R5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
159
streamline
S1
S2
S3
S4
S5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
S1
S2
S3
S4
S5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
S1
S2
S3
S4
S5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
streamline
S1
S2
S3
S4
S5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
D.2
S1
S2
S3
S4
S5
3
2
1
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Inlet Mach [ - ]
1.00
1.10
1.20
1.30
160
streamline
R1
R2
R3
R4
R5
3
2
1
0.00
0.05
0.10
0.15
total loss [ - ]
0.20
0.25
streamline
R1
R2
R3
R4
R5
3
2
1
0.00
0.05
0.10
0.15
total loss [ - ]
0.20
0.25
streamline
R1
R2
R3
R4
R5
3
2
1
0.00
0.05
0.10
0.15
total loss [ - ]
0.20
0.25
streamline
R1
R2
R3
R4
R5
3
2
1
0.00
0.05
0.10
0.15
total loss [ - ]
0.20
0.25
D.3
R1
R2
R3
R4
R5
3
2
1
0.00
0.05
0.10
0.15
total loss [ - ]
0.20
0.25
161
streamline
S1
S2
S3
S4
S5
3
2
1
0.00
0.01
0.02
0.03
0.04
0.05
total loss [ - ]
0.06
0.07
0.08
streamline
S1
S2
S3
S4
S5
3
2
1
0.00
0.01
0.02
0.03
0.04
0.05
total loss [ - ]
0.06
0.07
0.08
streamline
S1
S2
S3
S4
S5
3
2
1
0.00
0.01
0.02
0.03
0.04
0.05
total loss [ - ]
0.06
0.07
0.08
streamline
S1
S2
S3
S4
S5
3
2
1
0.00
0.01
0.02
0.03
0.04
0.05
total loss [ - ]
0.06
0.07
0.08
D.4
S1
S2
S3
S4
S5
3
2
1
0.00
0.01
0.02
0.03
0.04
0.05
total loss [ - ]
0.06
0.07
0.08
162
streamline
R1
R2
R3
R4
R5
3
2
1
-8.00
-7.80
-7.60
-7.40
-7.20
-7.00
-6.80
incid. angle [ deg ]
-6.60
-6.40
-6.20
-6.00
streamline
R1
R2
R3
R4
R5
3
2
1
-8.00
-7.80
-7.60
-7.40
-7.20
-7.00
-6.80
incid. angle [ deg ]
-6.60
-6.40
-6.20
-6.00
streamline
R1
R2
R3
R4
R5
3
2
1
-8.00
-7.80
-7.60
-7.40
-7.20
-7.00
-6.80
incid. angle [ deg ]
-6.60
-6.40
-6.20
-6.00
streamline
R1
R2
R3
R4
R5
3
2
1
-8.00
-7.80
-7.60
-7.40
-7.20
-7.00
-6.80
incid. angle [ deg ]
-6.60
-6.40
-6.20
-6.00
streamline
D.5
R1
R2
R3
R4
R5
3
2
1
-8.00
-7.80
-7.60
-7.40
-7.20
-7.00
-6.80
incid. angle [ deg ]
-6.60
-6.40
-6.20
-6.00
163
streamline
S1
S2
S3
S4
S5
3
2
1
-5.40
-5.20
-5.00
-4.80
-4.60
-4.40
incid. angle [ deg ]
-4.20
-4.00
-3.80
-3.60
streamline
S1
S2
S3
S4
S5
3
2
1
-5.40
-5.20
-5.00
-4.80
-4.60
-4.40
incid. angle [ deg ]
-4.20
-4.00
-3.80
-3.60
streamline
S1
S2
S3
S4
S5
3
2
1
-5.40
-5.20
-5.00
-4.80
-4.60
-4.40
incid. angle [ deg ]
-4.20
-4.00
-3.80
-3.60
streamline
S1
S2
S3
S4
S5
3
2
1
-5.40
-5.20
-5.00
-4.80
-4.40
-4.60
incid. angle [ deg ]
-4.20
-4.00
-3.80
-3.60
streamline
D.6
S1
S2
S3
S4
S5
3
2
1
-5.40
-5.20
-5.00
-4.80
-4.60
-4.40
incid. angle [ deg ]
-4.20
-4.00
-3.80
-3.60
CLASSIFICAO/TIPO
DM
5.
2.
DATA
3.
REGISTRO N
4.
N DE PGINAS
162
TTULO E SUBTTULO:
AUTOR(ES):
INSTITUIO(ES)/RGO(S) INTERNO(S)/DIVISO(ES):
RESUMO:
This work presents an approach to optimise the preliminary design of high-performance axial-flow
compressors. The preliminary design within the Gas Turbine Group at ITA, is carried on with an in-house
computational program based upon the streamline curvature method, using correlations from the literature
to assess the losses. The choice of many parameters of the thermodynamic cycle and of geometries relies
upon the expertise from the members of the Group. Nevertheless, it is still a laborious and timeconsuming task, requiring successive trial and errors. Therefore, to support the compressor designer in the
choice of some parameters, an optimisation program, named REMOGA, was written in FORTRAN
language, allowing an easy integration with the programs developed by the Gas Turbine Group. The
program is based upon a multi-objective genetic algorithm, with real codification and elitism.
Then the REMOGA and the preliminary design program were integrated to design a 5-stage axialflow compressor. Therefore, the stator air outlet angles, the temperature distribution and the hub-tip ratio
were varied aiming at higher efficiencies and higher pressure ratios, but controlling the de Haller number
and the camber angle. Thanks to the REMOGA, thousands of designs could be quickly evaluated. Finally,
using a choice criterion, four solutions were selected for further analysis, revealing that the developed
program was successful in finding more efficient and feasible compressor designs.
12.
GRAU DE SIGILO:
(X ) OSTENSIVO
( ) RESERVADO
( ) CONFIDENCIAL
( ) SECRETO