Using genetic algorithms to improve the thermodynamic eciency of

gas turbines designed by traditional methods

Jose M. Chaquet
a,
, Enrique J. Carmona
b
, Roque Corral
a
a
Technology and Methods Dep., Industria de TurboPropulsores S.A., Avda. de Castilla 2, 28830, San Fernando de
Henares, Madrid, Spain
b
Dpto. de Inteligencia Articial, Escuela Tecnica Superior de Ingeniera Informatica, Universidad Nacional de
Educacion a Distancia, Madrid, Spain
Abstract
A method for optimizing the thermodynamic eciency of aeronautical gas turbines designed by
classical methods is presented. This method is based in the transformation of the original con-
strained optimization problem into a new constrained free optimization problem which is solved by
a genetic algorithm. Basically, a set of geometric, aerodynamic and acoustic noise constraints must
be fullled during the optimization process. As a case study, the thermodynamic eciency of an
already optimized by traditional methods real aeronautical low pressure turbine design of 13 rows
has been successfully improved, increasing the turbine eciency by 0.047% and reducing the total
number of airfoils by 1.61%. In addition, experimental evidence of a strong correlation between
the total number of airfoils and the turbine eciency has been observed. This result would allow
us to use the total number of airfoils as a cheap substitute of the turbine eciency for a coarse
optimization at the rst design steps.
Keywords: Gas turbine, Thermodynamic eciency, Genetic algorithm, Throughow, Number of
airfoils
1. Introduction
The Low Pressure Turbine (LPT) module has a major contribution in gas turbine engines with
one third to the total weight and up to 20% to the total cost [12]. The LPT design process is a very
challenging task. A lot of dierent constraints must be taken into account and usually the nal
decision on the particular optimum conguration needs a trade-o among dierent requirements.
The extraction of work from the uid is done by means of several aerodynamic surfaces called
airfoils that are placed in an annular way to form rows. A turbine stage consists of two consecutive
rows, called stator and rotor. Stator airfoils are called vanes, whereas rotor airfoils are called
blades. The stator is attached to the casing and directs the ow towards the rotor, whereas the
rotor transmits the power to the turbine shaft. The number of airfoils in a row is called NumberO.
In the literature related to this area, few works have been focused on the importance of the
NumberOs. For instance, work [21] investigates the inuence of blade height and blade number on

Corresponding author. Tel.: +34 91 207 9147

Email addresses: [email protected] (Jose M. Chaquet), [email protected] (Enrique J. Carmona),
[email protected] (Roque Corral)
Preprint submitted to Applied Soft Computing July 6, 2012
2
the performance of low head axial ow turbines for micro-hydro applications. The study concluded
that the inuence of blade number is higher than the blade height and that the choice of blade
number should be carefully made. Other work where the number of airfoils is optimized can be seen
in [13]. In that contribution the objectives were to simultaneously minimize the total pressure loss,
maximize the total aerodynamic loading and minimize the number of airfoils for a turbomachinery
cascade. Several constraints were taken into account, as xed mass ow rate, xed axial chord,
xed inlet and exit ow angles, etc.
Regarding the optimization process, the usual approach modies the airfoil shapes, but without
modifying the NumberOs. For instance in [8] we can see the aerodynamic optimization of highly
loaded turbine cascade blades for heavy duty gas turbine applications. The main target was the
reduction of the total pressure losses, which is equivalent to increase the thermodynamic eciency.
Other works have been reported where airfoil shape optimization is performed not only for gas
turbine engine airfoils [13, 17] but as well for Micro-Air-Vehicle airfoils [6], propellers [7], wind
turbine blades [11] and steam turbine airfoils [19].
In order to obtain performance variables such as thermodynamic eciency, pressure losses,
power, etc, the Navier-Stokes equations must be solved in the uid domains. This is done with
Computational Fluid Dynamics (CFD) solvers. One specic type of CFD solver for turbomachinery
applications is the so called throughow code [20, 22], which computes the ow variables along all
the rows. Navier-Stokes equations are circumferentially averaged solving the 2D ow eld over
a meridional plane. Using throughow models in the optimization process is appealing because
they can produce a fast design avoiding expensive full 3D CFD simulations [9, 23, 18]. Throughow
models have been used not only for airfoil optimization, but as well for other geometry optimization
as we can see in [14], where entropy minimization (or eciency maximization) optimizes hub and
shroud geometries and inlet and exit ow-eld for each blade row of a two-stage axial ow gas
turbine. Although throughow codes could give accurate solutions, in some cases throughow and
3D CFD solutions must be linked [5].
Several techniques have been used for Aerodynamic optimization, such as gradient-based schemes.
These methods require knowledge about the aerodynamic derivatives for each parameter, which are
normally expensive to compute. Furthermore, gradient-based methods cannot be applied to prob-
lems where there are discontinuities in the design space because the derivatives in these regions are
not dened [15]. In such cases Genetic Algorithms (GAs) have demonstrated good performance.
Multidisciplinary airfoil optimization (aerodynamic and acoustic) can be seen in [16] where a par-
allel GA was used to generate a family of aerodynamically ecient, low-noise rotor blade designs
which represent the Pareto optimal set. The multiple objectives of that work were to maximize
lift-to-drag of a rotor airfoil shape and to minimize an overall noise measure, including the eects
of loading and thickness noise of the airfoil. More recently [6], a GA for optimizing the shape of low
Reynolds number airfoils for generating maximum lift for Micro-Air-Vehicle (MAV) applications
has been developed where the GA computational eciency has been signicantly enhanced with
an articial neural network (ANN). The authors showed that the combined GA/ANN optimization
technique is capable of nding globally optimal airfoils accurately and eciently. Another example
of coupling a GA with an ANN can be seen in [11], where wind-turbine blades are optimized for
generating maximum lift to drag ratio.
Regarding thermal system engineering, several recent works have shown that GAs can be suc-
cessfully employed in the optimization of real power generation systems based on gas turbines. In
this way, [4] presents the simulation and multiobjective optimization of a gas turbine power plant
with preheater, [1] reports the optimization of a combined heat and power plant for cogeneration
3
purposes, [2] makes a thermodynamic and exergoeconomic modeling and optimization of a gas tur-
bine plant, and [3] gives an exergy, exergoeconomic and environmental analysis and optimization of
several combined cycle power plants. Among other important parameters, the gas turbine isentropic
eciency is investigated as design parameter. It can be appreciated that improving the design of
several key components (as the compressor, turbine or combustion chamber) important savings can
be obtained in the complete cycle, not only from the thermodynamic point of view, but as well from
other dimensions such as the environmental impact or the monetary costs. These complex systems
usually require a trade-o between dierent requirements, so multiobjective optimization must be
employed. Evolutionary algorithms such as GAs have proved to be very ecient for solving this
type of problems.
This paper presents a method for optimizing the thermodynamic eciency of a turbine while a
set of geometric, acoustic and aerodynamic restrictions are fullled. The optimization problem at
hand involves seeking a solution which maximizes the turbine eciency fullling at the same time
several constraints. In this context, the function to optimize is not guaranteed to be continuous.
Therefore, it is not recommendable to use gradient-based methods. Moreover, we are interested in
nding global optimum in a problem which naturally is multimodal due to the high dimensionality
and complex relationship among the control variables. The large size of the search space makes direct
methods such as exhaustive or random search impracticable [10]. Therefore the use of Evolutionary
Computation (EC) paradigms is more suitable to that type of problems: a smooth continuous and
derivable optimization tness function is not required, and the use of a population of candidate
solutions facilitates the global optimum nding. The ability of evolutionary algorithms to maintain
a population of potential solutions not only provides a means of escaping from one local optimum;
it also provides a means to manage large and discontinuous search spaces. As it will be shown
in the next section, our initial constraint satisfaction problem can be reduced to optimize a real
valued function of integer variables. This fact makes the use of a Genetic Algorithm (GA) more
appropriated than other EC paradigms such as Evolutionary Strategies, more suitable for functions
of real variables.
Unlike other approaches in the literature, the main control parameter used in this work for
optimizing the eciency is the number of airfoils for each row. This approach is applied to a
turbine that has already been designed using a classical methodology. Fluid variables are computed
using a throughow solver. The approach adopted to solve the optimization problem uses a Genetic
Algorithm (GA). The particular optimization performed considers the uid variables as constant, so
throughow models are not updated until the GA run is nished. This approach saves computational
time and makes the algorithm more robust. Since the GA geometry modication aects the ow
eld, there must be an iterative process between the throughow and the GA. Convergence is
achieved in only 2 or 3 global iterations in the given case studies.
In summary, the main contributions of the present work are the following:
Improving the thermodynamic eciency of an already designed by traditional methods aero-
nautical LPT using an evolutionary algorithm.
Demonstrating that, for a given number of airfoils, it is possible to fulll all the geometric
constraints with a correct election of other design parameters such as chords and gaps between
rows.
Showing that a GA can successfully deal this type of constrained optimization problems.
4
a)
b)
Ri
ci
gi
Si
g0
L
Figure 1: Real geometry (a) and simplied one (b)
Showing the importance of reducing the total number of airfoils to improve the turbine ef-
ciency, giving some possible recommendations for speed-up in the rst steps of the design
process.
In this article, rst a description of the problem to solve is presented (section 2). Then the GA ap-
proach (section 3) and how the algorithm interacts with other design tools (section 4) are described.
The results obtained in the optimization of a real 13 row aeronautical gas turbine are presented for
two dierent sets of constraints (section 5). Finally, some conclusions and future works are given
(section 6).
2. Problem description
The problem consists of optimizing the thermodynamic eciency of a LPT for a given ow-path
(Fig. 1a) and aerodynamic exit angles. The optimization process has to fulll a set of aerodynamic,
acoustic and geometric restrictions that may be reduced to a set of explicit analytical expressions.
As a consequence, the restrictions are extremely fast to evaluate.
In order to parametrize the problem, simplied geometry will be used to approximate each row
to a rectangle (Fig. 1b). For a turbine of M number of rows, each row is dened with only 5
parameters: NumberO (N
i
), gap (g
i
), chord (c
i
), span (S
i
) and mean radius (R
i
) where i goes
from 1 to M. The chord c
i
for a given row i is dened as the axial length between the leading and
the trailing edges, and the span S
i
is the length of the row in radial direction. The mean radius is
the distance between the middle point of the row to the turbine axis. There must also be one global
variable L, which is the total axial length of the turbine. The turbine inner and outer annuli are
supposed to be optimized in an outer loop and in this work are kept constant. Therefore the mean
2.1 Constraint description 5
radius and the span of all the rows are constant. NumberOs, gaps and chords will be modied in
order to nd optimum feasible congurations. The gap for row i is the distance between the trailing
edge of row i and the leading edge of the next row i + 1, or the exit station for the last row. The
initial gap g
0
, is dened as the distance between the inlet station and the leading edge of the rst
row (Fig. 1b) and can be computed using the following expression
g
0
= L
M

i=1
(c
i
+g
i
) . (1)
2.1. Constraint description
The geometric constraints are dened with the following parameters for each row: maximum
aspect ratio (MA
i
), minimum pitch to chord ratio (mPC
i
), maximum pitch to chord ratio (MPC
i
),
minimum gap (mG
i
), minimum gap to chord ratio (mGC
i
), maximum gap to chord ratio (MGC
i
),
maximum NumberO (MN
i
) and the NumberO for each package (P
i
). The maximum aspect ratio
should be limited by structural and utter considerations. The pitch to chord ratio is limited in
order to maintain Zweiel coecients bounded. Gaps are bounded in order to prevent mechanical
interferences and by noise restrictions. The package parameter forces the NumberO to be a
multiple of P
i
. For the inlet gap g
0
, two constraints are given for bounding it between a minimum
and a maximum value: mG
0
and MG
0
.
It is well known that one way of reducing the generation of noise associated with pure tones
is to force the NumberOs ratio for two consecutive rows to lie within some specic intervals [12].
When the NumberO ratio fullls this conditions, the acoustic wave amplitudes decrease with the
axial distance, the stage is said to be cut-o and the perturbations do not propagate outside the
turbine. The cut-o condition also depends on the ow variables, but in our problem these are
assumed to remain constant. Noise constraints are given by four parameters:
i
,
i
,
i
and
i
.
These parameters dene two intervals [
i
,
i
] and [
i
,
i
] where the NumberO ratio of row i and
row i + 1 must be located. Always 0
i

i
1
i

i
. When both ranges are used, the
conguration is called Mixed cut-o. For the Direct cut-o mode, the [
i
,
i
] interval is chosen for
even rows and the [
i
,
i
] for odd rows, therefore there will be more vanes than blades. The opposite
is chosen for the Reverse cut-o mode.
Putting everything together, the mathematical problem consists of nding the M positive integer
numbers N
i
and the M positive real numbers c
i
and g
i
which fulll the following constraints for
i [1, M]
S
i
c
i
MA
i
, (2)
mPC
i

2 R
i
N
i
c
i
MPC
i
, (3)
mG
i
g
i
, (4)
mGC
i

g
i
c
i
MGC
i
, (5)
N
i
MN
i
, (6)
N
i
%P
i
= 0 , (7)
mG
0
g
0
MG
0
, (8)
2.2 Degrees of Freedom Reduction 6
if(i = M &Mixed)
N
i
N
i+1
[
i
,
i
] [
i
,
i
] , (9)
if
_
_
_
i = M & (Direct &i%2 = 0)
or
(Reverse &i%2 = 1)
_
_
_
N
i
N
i+1
[
i
,
i
] , (10)
if
_
_
_
i = M & (Direct &i%2 = 1)
or
(Reverse &i%2 = 0)
_
_
_
N
i
N
i+1
[
i
,
i
] . (11)
In equation (7) the symbol % means the remainder of the integer division. In equation (8) the rst
gap g
0
is obtained using the expression 1.
2.2. Degrees of Freedom Reduction
Taking into account the three parameters for each row (N
i
, c
i
and g
i
), there are 3M Degrees of
Freedom (DoF). In [10] it was shown that the problem may be reduced to that of nding the M
DoF associated with the number of airfoils for each individual row. This reduction is very desirable
because the computational cost is notably reduced. Nevertheless the DoF reduction must to fulll
the condition that any optimal solution is lost.
This section describes the outline of the DoF reduction. A detailed description can be obtained
from the paper by the same authors [10] where an easier problem was solved using the same set
of constraints and optimizing the total number of airfoils instead of the thermodynamic eciency.
The reduction consists of choosing the minimum feasible gap and chord for a given row. Smaller
gaps and chords give more room to other rows. On the one hand, the eect of the gap on the global
turbine eciency is negligible if the minimum gap constraint, equation (4), is fullled. On the other
hand, a minimum chord for the same NumberO gives as well more room to other rows.
For a given set of NumberOs N
i
, the optimum chords and gaps are chosen with the following
expressions:
c
i
c
i
(N
i
) = max
_
c
min,i
,
2 R
i
N
i
MPC
i
_
, (12)
g
i
g
i
(c
i
) = max (mG
i
, mGC
i
c
i
) . (13)
where c
min,i
is the minimum feasible chord:
c
min,i
= max
_
S
i
MA
i
,
mG
i
MGC
i
,
2 R
i
MN
i
MPC
i
_
. (14)
The range of N
i
to be explored is given by the following expressions, where floor
Pi
() function
is the largest integer value not greater than the argument and multiple of P
i
, and function ceil
Pi
()
is the smallest integer value not less than the argument and multiple of P
i
:
N
max,i
= floor
Pi
_
min
_
MN
i
,
2 R
i
mPC
i
c
min,i
__
, (15)
N
min,i
= ceil
Pi
_
min
_
N
max,i
,
2 R
i
MPC
i
c
min,i
__
. (16)
7
3. Genetic Algorithm
Once it has been demonstrated that each design conguration is determined by a set of Num-
berOs, an exhaustive search could be performed computing the thermodynamic eciency and
checking the constraints for all the possible solutions. Using expressions (15) and (16), the number
of congurations to be explored will be
M

i=1
(N
max,i
N
min,i
) /P
i
. (17)
As will be shown in section 5, huge numbers appear in real problems.
Owing to the multiple restrictions, it is dicult to dene a continuous and derivable opti-
mization function. Therefore methods based on the gradient of the optimization function are not
recommended. On the other hand a Genetic Algorithm (GA) could be used because of the char-
acteristics of the problem. First of all, the formalism of a GA easily transforms a Constrained
Optimization Problem (COP) into a Free Optimization Problem (FOP). Secondly, the optimization
function is not necessary to be continuous. Thirdly, the solution codication is easily done using a
numeric vector.
3.1. Individual representation
The rst step for dening a GA is to link the real world to the GA world. Objects forming
possible solutions within the original problem context are referred to as phenotypes, while their
encoding are called genotypes. In our problem, the phenotypes are vectors of natural numbers with
the NumberO for each row. Each NumberO can only change in the range given by (15) and (16).
The encoding of each genotype is a vector of natural numbers n
i
for each NumberO in a range
given by
n
i

_
0,
N
max,i
N
min,i
P
i
_
. (18)
The number of natural numbers n
i
will be M, one for each row. The way of decoding the genotype
into the phenotype consists of obtaining the NumberO N
i
associated with each gene n
i
:
N
i
= N
min,i
+n
i
P
i
. (19)
Knowing each N
i
, gaps and chords are obtained using the expressions (12) and (13). Equation (13)
does not take into account the constraint (8). A repairing process may be necessary if, on obtaining
the phenotype, g
0
does not meet that constraint. If mG
0
g
0
MG
0
the xing is not necessary.
On the other hand, if g
0
mG
0
the solution cannot be repaired and the individual receives a high
penalty in its tness. If g
0
> MG
0
a repairing process is needed. The repairing process is done in
the phenotypic space and this consists of distributing the amount g = g
0
MG
0
among the rest
of the gaps maintaining the constraints g
i
MGC
i
c
i
.
3.2. Fitness function
The role of the tness function F is to represent the requirements to be optimized. In this work
it is dened in such a way that our initial COP is transform into an FOP.
A penalty function F
C
is dened to deal with the constraints. Negative values are used for
unfeasible individuals and zero value for feasible ones. With the representation adopted, all the
constraints are satised except (8), (9), (10) and (11). The penalty function F
C
is computed as
3.3 Genetic Operators 8
F
C
=
_

_
1 exp
_

mG0g0
L
_
+

M
i=1
F
i
if g
0
< mG
0

M
i=1
F
i
if mG
0
g
0
MG
0
1 exp
_

g0MG0
L
_
+

M
i=1
F
i
if MG
0
< g
0
, (20)
where g
0
is computed using expression (1). The value of constant parameter is used to modulate
the exponential decreasing in the unfeasible regions. Its value is taken experimentally and does not
have a pronounced eect on the performance of the algorithm. F
i
deals with the noise restrictions
depending on the cut-o mode. For instance, for Mixed cut-o mode:
F
i
=
_

_
1 exp
_

i

Ni
Ni+1
__
if
Ni
Ni+1
<
i
0 if
i

Ni
Ni+1

i
1 exp
_

_
Ni
Ni+1

i
__
if
i
<
Ni
Ni+1

i+i
2
1 exp
_

i

Ni
Ni+1
__
if
i+i
2
<
Ni
Ni+1
<
i
0 if
i

Ni
Ni+1

i
1 exp
_

_
Ni
Ni+1

i
__
if
i
<
Ni
Ni+1
. (21)
The tness function solves the initial target problem: optimizing the thermodynamic turbine
eciency among the feasible individuals:
F

=
_
F
C
if F
C
< 0
if F
C
0
, (22)
where is the thermodynamic turbine eciency whose value is calculated as it will be described in
section 4.1.
3.3. Genetic Operators
The parent selection mechanism implemented here is the tournament method, i.e. k individuals
with replacement are chosen randomly from the population and the nal individual chosen will be
the best of these k in terms of their tness value.
Once the parents have been selected, there is a recombination probability p
r
that determines
whether the ospring of two parents are just a copy of the parents or a real recombination is
produced. Single point crossover is used, i.e. a single crossover point on both parent chromosomes
is selected. All genes beyond that point in either parent genotype are swapped between the two
parents.
Another parameter that controls the algorithm is the mutation probability p
m
. After performing
the crossover of the parents, the ospring is mutated. The mutation is done in each gene adding or
subtracting a random number with constant distribution between 0 and 10% of the total possibilities
for that gene given by expression (18). Modular arithmetic is used in order to maintain the genes
inside their boundaries.
A generational model is used, so for each generation all parents are replaced by their ospring.
Elitism was implemented swapping the worst individual for the best individual of the previous
generation after the mutation operator was applied.
The initialization was done by taking a random representation of possible solutions from the
design space and carrying out tness evaluations on all the individuals.
9
4. GA interaction with other design tools
Several tools are involved in the traditional design process of a gas turbine. The purpose of this
work is not to present these tools. However this section will describe briey two of them due to
their interactions with the GA: eciency estimator module and throughow code.
4.1. Eciency estimator
This module computes the thermodynamic eciency of a turbine design and, therefore, it can
calculate the tness function of each individual using Eq. (22). Turbine isentropic eciency is
used, i. e. other engine components as fan, compressor or combustion chamber are not considered.
Detailed geometry and the complete uid solution in the meridional plane are needed for the
eciency evaluation. All this information is not contained in the individual representation managed
by the GA. Therefore, extra data must be provided.
Since the mean radius and spans for each row are constant, it is assumed that the uid solution
does not change for all the possible individuals in the population. This hypothesis allows not to
update the uid solution with CFD simulations, which would need high computational eort. The
extra data provided to the GA for computing eciencies are associated with a specic individual
called reference individual, whose dening parameters are

N
i
, g
i
and c
i
for each row i. The input
data to the Eciency estimator can be classied into two sets. One set consists of pure uid
variables such as velocities, temperatures, pressures, Mach Numbers, etc. The other set consists
of variables associated with geometry such as axial and radial coordinates, pitch, lift coecients,
Reynolds Numbers, chords, etc. When the eciency of an individual is needed, the internal data
of the eciency estimator of the rst set of variables are maintained constant, but the second set
is updated using the geometry increments with the reference individual. For instance, the pitch for
a particular radial position j of row i is dened as
pitch
j
i
=
2r
j
i
N
i
. (23)
Therefore, the new pitch for a given individual can be computed with the pitch of the reference
individual as
pitch
j
i
=

pitch
j
i

N
i
N
i
. (24)
Similar expressions such as (24) can be obtained for the other variables of the second set of the
eciency estimator module.
4.2. Throughow
A ThroughFlow is a specic CFD code for turbomachinery design which computes the ow
variables along all the rows over the meridional plane. Navier-Stokes equations are circumferentially
averaged and the axisymmetric ow eld is obtained. Several formulations have been developed
[20]. The code used in this work corresponds to the classical formulation developed by Wu [22],
which consists of solving the equations on steady, inviscid ow in a relative reference frame.
Fig. 2 shows the new proposed method for the optimization of a throughow model. The
throughow uses the eciency estimator, so it can compute the turbine eciency using the uid
solution. The GA communicates with the throughow by two interfaces called export and import
which have been specically designed to carry out our implementation. In the export process,
10
Efficiency Estimator
ThroughFlow
Efficiency Estimator
Genetic Algorithm
> ?
GA TF
No
Yes
Export
Import

Finish
Start
Figure 2: GA relations with throughow and eciency estimator.
Property Value
Inlet Mass Flow 50.8 kg/s
Inlet Total Pressure 243 kPa
Inlet Total Temperature 943 K
Inlet Angle 18
o
Pressure ratio 5.78
Axial Length 1.225 m
Power 18.302 MW
Total NumberO 1486
Table 1: Turbine main properties
all the data needed by the GA is generated, particularly all the uid variables of the reference
individual for computing the eciency. After the run of the GA, the eciency of best individual in
the last generation
GA
is compared with the eciency of the throughow model
TF
. If
GA
>
TF
the throughow model is modied with the new geometry computed by the GA using the import
interface. NumberO, gaps and chords of the throughow model are modied accordingly. After
a run of the throughow model the loop is repeated until
GA

TF
. If this condition is fullled,
the GA cannot nd a better individual than the reference one, so the iterative loop is nished.
5. Case Studies
The methodology described in the previous sections was applied to the optimization of an
aeronautical LPT consisting of 6 stages and an Outlet Guide Vane (OGV), which gave a total of
13 rows (Fig. 1). The initial LPT was designed following a conventional methodology. In Table 1
the main properties of the turbine are shown.
Two sets of constraints were applied. The rst one, called Case A, are not realistic but try to
increase the number of possible feasible solutions in order to check the performance of the GA when
the solution space is big. On the other hand, in Case B the restrictions imposed were the same as
5.1 Case A 11
Property Value
Population 5 10
4
Generations 50
Tournament parameter 5
Recombination probability 0.8
Mutation probability 0.01
Table 2: GA parameters
Constrain Value
MA
i

S
i
/c
i
mPC
i
0.98 2

R
i
/

N
i
c
i
MPC
i
1.02 2

R
i
/

N
i
c
i
mG
i
0.95 g
i
mGC
i
0.4
MGC
i
max (0.75, g
i
/c
i
)
MN
i
200
P
i
_

_
1 if OGV
2 if rotor
5 if stator&

N
i
%5 = 0
6 if statot&

N
i
%6 = 0
7 default
[
i
,
i
] , [
i
,
i
] [0.6, 0.8] , [1.3, 1.6]
mG
0
0.8 g
0
MG
0
1.2 g
0
Cut-o mode Mixed
Table 3: Constraints for Case A
those used by the design team, so the case study must be considered as real. In both cases, the
control parameters of the GA were chosen by trial and error. The best results were obtained for the
parameters shown in table 2. All runs were performed using a 2.40 GHz Intel Core Duo machine
with a 4 GB of RAM memory and a Linux openSUSE 10.3 operative system.
5.1. Case A
As already highlighted, the constraints are articially chosen in order to increase the number of
possible feasible solutions. Using the hat symbol () to refer to the values in the initial throughow
model, which coincides with the rst reference individual, the constraints are given in Table 3.
Multiplying the possibilities for each row given by expression (17), 4.6 10
18
possible cong-
urations are obtained. If we use an exhaustive search and consider that each conguration was
evaluated in 10
6
seconds, the computing time would be 146235 years. So an exhaustive search
cannot be used in this case.
Owing the stochastic nature of the GA, several runs have been performed with dierent random
generator seeds in order to study the dispersion of the results. The same optimum solution is
obtained in all the runs, providing evidence about the robustness of the method. The average time
needed for each run was 12 minutes, and the number of dierent feasible individuals in the last
5.1 Case A 12
10 20 30
Generations
F
i
t
n
e
s
s

(
E
f
f
i
c
i
e
n
c
y
)
Run 1
Run 2
Run 3
Run 4
Run 5
a)
0.0005
5 10
Row number
0
50
100
150
N
u
m
b
e
r
O
f
f
Optimum
b)
Figure 3: Case A results: best individual tness versus generation number for 5 runs (a) and NumberOs for each
row for all the feasible individuals in the last generation in one run (b).
Throughow GA
Iteration
0

N
i

0

N
i
1 0 1486 0.0039236 1334
2 0.003253 1334 0.0034802 1329
3 0.003408 1329 0.003408 1329
Table 4: Iterative process between throughow and GA for Case A. Eciencies are expressed relative to the initial
one
0
.
generation was around 30. Nevertheless, the optimal individual found in the last generation did not
change. In Fig. 3a the best individual tness versus generation number for 5 runs is plotted. The
evolving process involves two phases: an initial stage (F

< 0) for satisfying the constraints, and

a second stage (F

0) for maximizing the turbine ecienccy once the constraints are satised.
Fitness is only plotted if there is at least one feasible individual in the population. Only 7 to 8
generations are needed to have at least one individual that meets all the restrictions. In Fig. 3b
the number of airfoils for each row for all the feasible individuals in the last generation for one
run is shown. All the congurations are very similar due to the typical decrease of diversity in the
standard GA in last generations. We can observe that all the feasible congurations are pure Mixed
cut-o (N
1
< N
2
< N
3
< N
4
).
Following the iterative process shown in Fig. 2, only three iterations are needed to obtain the
optimal conguration as shown in Table 4. For each iteration, the thermodynamic eciency and
the total number of airfoils are given for the throughow model and for the best individual in the
last generation of the GA. We can check that the total number of airfoils in throughow models is
the same as in the GA at the previous iteration. The iterative process is stopped when the GA does
not change the input data from the throughow. Comparing the rst and last iteration we can see
that the GA has increased eciency a 0.36% and reduced the total number of airfoils by 10.56%.
In Fig. 4 the new geometry in meridional plane obtained by the GA is shown. Notice that the
5.2 Case B 13
Original geometry
New geometry
Figure 4: Optimized throughow geometry (continuous line) compared with the original one (dashed line) for Case
A
algorithm has changed the chord and the gap of some rows in order to fulll the constraints.
5.2. Case B
The constraints imposed for Case B were the same as those used by the design team, so the
case study must be considered as real (Table 5). Only the Reverse cut-o condition is considered
feasible. Owing to the mechanical restrictions, the rst and last row will not be modied.
Constrain Value
MA
i
_
_
_
2.128 if i = 1
7.142 if i = 2, 3, .., 12
1.244 if i = 13
mPC
i
0.98 2

R
i
/

N
i
c
i
MPC
i
1.01 2

R
i
/

N
i
c
i
mG
i
0.95 g
i
MN
i
_
200 if i = 13
20 if i = 13
mG
0
g
0
MG
0
1.0001 g
0
Table 5: Aerodynamic and Geometric constraints for Case B
The number of blades in each package for rotors are P
i
= 2, whereas P
1
= 66, P
3
= P
5
= P
7
= 6,
P
9
= P
11
= 5 and P
12
= 12. The feasible noise intervals are given in Table 6. Any noise constraints
are imposed for rows 1, 12 and 13.
As in Case A, exhaustive search is not feasible because the number of possible congurations
to be explored (2.3 10
18
) is too high. The average time needed for each run was 10 minutes, and
the number of dierent feasible individuals in the last generation was around 30. As in Case A,
the optimal individual found in the last generation did not change for the 5 runs performed. In
Fig. 5a the best individual tness versus generation number for 5 runs is shown. The constraints
5.3 Eciency-NumberO correlation 14
Row
i

i

i

i
2 0.685 0.86 1.37 1.73
3 0.6 0.79 1.21 1.58
4 0.7 0.85 1.4 1.69
5 0.61 0.77 1.23 1.54
6 0.695 0.83 1.39 1.66
7 0.62 0.76 1.24 1.52
8 0.69 0.82 1.38 1.63
9 0.62 0.75 1.25 1.5
10 0.675 0.8 1.35 1.61
11 0.63 0.75 1.25 1.5
Table 6: Noise Constraints for Case B
Throughow GA
Iteration
0

N
i

0

N
i
1 0 1486 0.000504 1462
2 0.000444 1462 0.000444 1462
Table 7: Iterative process between throughow and GA for Case B
used make the problem harder and more than 10 generations are needed to nd the rst feasible
individual versus the 7 to 8 generations in Case A.
Following the iterative process shown in Fig. 2, only two iterations are needed fto obtain the
optimal conguration as shown in Table 7. Turbine eciency increased by 0.047% and the total
number of airfoils were reduced to 1.61%. In Fig. 6 the new geometry in the meridional plane
obtained by the GA is shown. Notice that the algorithm has changed the chord and the gap of
some rows in order to fulll the constraints.
5.3. Eciency-NumberO correlation
A strong correlation was observed between the total number of airfoils

N
i
and turbine ef-
ciency . In order to study this correlation, the GA was run for both cases A and B using the
penalty function F
C
as the tness function. In a post-processing step, the eciency of all the
feasible individuals in the last generation was computed.
In Fig. 7 all (

N
i
, ) pairs are plotted. The original conguration and the optimum solution
found for both cases are also shown. As can be observed, all the pairs are distributed along a
straight line with negative slope. We can also observed that the solution space for Case A is bigger
than for Case B. 66248 solutions were found for Case A, compared with 4059 for Case B. The
dierence between both cases can be seen too in the range of eciency and total NumberO where
pairs are distributed.
5.4. Comparative study
It is not straightforward to make a comparative study between our results and other previous
approaches because the number of airfoils normally is maintained constant. On the other hand,
tness function and number and type of constraints do not match. Furthermore, in this study an
5.4 Comparative study 15
10 20 30 40
Generations
F
i
t
n
e
s
s

(
E
f
f
i
c
i
e
n
c
y
)
Run 1
Run 2
Run 3
Run 4
Run 5
a)
0.0001
5 10
Row number
0
50
100
150
N
u
m
b
e
r
O
f
f
Optimum
b)
Figure 5: Case B results: best individual tness versus generation number for 5 runs (a) and NumberOs for each
row for all the feasible individuals in the last generation in one run (b).
Original geometry
New geometry
Figure 6: Optimized throughow geometry (continuous line) compared with the original one (dashed line) for Case
B
16
Figure 7: Eciency-Total NumberO pairs
already optimized design by traditional methods is used as initial point, meanwhile some other
works have more room for improvements.
The work presented in [8] optimizes a unique airfoil shape by means of 15 parameters. In our
study, 13 rows are optimized simultaneously with 39 parameters. The number of constraints are
also very dierent, going from two relative complex constraints (smooth acceleration on the suction
and pressure surface and an ecient cooling of the blade) in [8] to around one hundred simple
constraints in the present contribution. A maximal reduction of the total pressure losses coecient
(which is directly related with the eciency) by about 20% was achieved in [8], meanwhile in the
present work the maximum increment in the global eciency is around 1.61%. This discrepancy
can be explained by the fact that in the present contribution the initial design corresponds to an
already optimized turbine with traditional methods.
The work presented in [13] tries to reduce the number of airfoils in a turbomachinery cascade
among other targets. A maximum reduction respect to the initial design of 2% of the number of
blades was obtained. In the present work, the reduction in the total NumberO is 10.56% for Case
A and 1.61% for Case B.
In [17] an axial turbine rotor cascade shape optimization with unsteady passing wakes was
performed to obtain improved aerodynamic performance. The objective function was dened either
as minimization of total pressure loss or as maximization of lift, while the mass ow rate was xed
during the optimization. The design variables were geometric parameters characterizing airfoil
leading edge, camber, stagger angle, and inter-row axial spacing. The optimization results indicated
that only minor improvements were possible in the unsteady rotor/stator aerodynamics by varying
these geometric parameters. These results are similar to what has been obtained for our realistic
Case B.
6. Conclusions and Future Work
A GA has been applied to improve the thermodynamic eciency optimization of an LPT gas
turbine fullling a set of restrictions. The turbine model used as input to the GA corresponds to
17
the nal design of a turbine based on a standard design methodology. Two sets of restrictions,
called Case A and Case B, were used to measure the algorithm performances. The rst one is less
realistic and checks the performance of the GA when the solution space is large. The second one
is the same as the one used by the design team to obtain the turbine model used as input. In
Case A, the algorithm increased eciency by 0.36% and reduced the total number of airfoils by
10.56%. In Case B, these values dropped to 0.047% for eciency and 1.61% for the total number
of airfoils. The hypothesis of maintaining the uid properties constant during the GA runs requires
less computational eort and only two or three global loops with the throughow code are needed to
achieve the convergence. Experimental evidence of a strong correlation between turbine eciency
and the total NumberO was observed.
As a future work new tness functions could be implemented to perform the optimization of
other parameters such as the total turbine weight. Multi-objective optimization could be performed
using Pareto Front techniques. Other constraints could be considered as well. The DoF reduction
would have to be modied to adjust to the new constraints. For instance, a new noise constraint
could be considered imposing not only a cut-o mode with the appropriate N
i
/N
i+1
intervals, but
also a minimum cut-o decay, i.e. the acoustic tones produced by the turbine decay exponentially
with the distance at least with a specied rate. Fluid properties must be known for cut-o decay
computations, so a similar technique to that implemented for eciency could be used.
7. Acknowledgements
The authors wish to thank Industria de Turbopropulsores S. A., for allowing the publication of
this paper and for its support during the project.
References
References
[1] P. Ahmadi and I. Dincer. Exergoenvironmental analysis and optimization of a cogeneration
plant system using multimodal genetic algorithm (mga). Energy, 35:51615172, 2010.
[2] P. Ahmadi and I. Dincer. Thermodynamic and exergoenvironmental analyses, and multi-
objective optimization of a gas turbine power plant. Applied Thermal Engineering, 31:2529
2540, 2011.
[3] P. Ahmadi, I. Dincer, and M. A. Rosen. Exergy, exergoeconomic and environmental analyses
and evolutionary algorithm based multi-objective optimization of combined cycle power plants.
Energy, 10:58865898, 2011.
[4] H. Barzegar Avval, P. Ahmadi, A. R. Ghaarizadeh, and M. H. Saidi. Thermo-economic-
environmental multiobjective optimization of a gas turbine power plant with preheater using
evolutionary algorithm. International Journal of Energy Research, 35:389403, 2011.
[5] K. Becker, M. Lawerenz, C. Voss, and R. Moenig. Multi-objective optimization in axial com-
pressor design using a linked cfd-solver. In Proceddings of ASME Turbo Expo 2008: Power for
Land, Sea, and Air GT2008-51131, pages 25332542, 2008.
REFERENCES 18
[6] M. Bellman, J. Straccia, B. Morgan, K. Maschmeyer, and R. Agarwal. Improving genetic
algorithm eciency with an articial neural network for optimization of low reynolds number
airfoils. In 47th AIAA Aerospace Sciences Meeting Including The New Horizons Forum and
Aerospace Exposition, 5-8 January, Orlando, Florida, 2009.
[7] C. Burger, R. Harteld, and J. Burkhalter. Performance and noise optimization
of a propeller using the vortex lattice method and a genetic algorithm. In 48th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference,
23-26 April, Honolulu, Hawaii, 2007.
[8] P. Cardamone. Aerodynamic optimisation of highly loaded turbine cascade blades for heavy
duty gas turbine applications. Masters thesis, Bundeswehr Munchen, Munich, 2006.
[9] M. Casey, F. Gersbach, and C. Robinson. An optimization technique for radial compressor
impellers. In Proceddings of ASME Turbo Expo 2008: Power for Land, Sea, and Air GT2008-
50561, pages 24012411, 2008.
[10] J. M. Chaquet, E. J. Carmona, and R. Corral. Optimizing the number of airfoils in turbine
design using genetic algorithms. In Proceddings of 23rd IEA/AIE, Cordoba, Spain, 2010.
[11] X. Chen and R. Agarwal. Optimization of atback airfoils for wind turbine blades using a
genetic algorithm with an articial neural network. In 48th AIAA Aerospace Sciences Meeting
Including the New Horizons Forum and Aerospace Exposition, 4-7 January, Orlando, Florida,
2010.
[12] P. de la Calzada. Aerothermodynamic design of low pressure turbines. In Aeroengine design:
from state of the art turbofans towards innovative architecture. Von Karman Institute Lecture
Series LS-2008-04, Brussels, Belgium, 2008.
[13] B. H. Dennis, G. S. Dulikravich, and Z.-X. Han. Constrained optimization of turbomachinery
airfoil cascade shapes using a navier-stokes solver and a genetic/sqp algorithm. AIAA Journal
of Propulsion and Power, 17(5):201210, 2001.
[14] G. S. Dulikravich, T. J. Martin, B. H. Dennis, and N. F. Foster. Multidisciplinaryhybrid
constrained ga optimization. In Evolutionary Algorithms in Engineering and Computer Science:
Recent Advances and Industrial Applications, EUROGEN99, Finland, May 30 - June 3, pages
231260, 1999.
[15] P. Gage and I. Kroo. A role of genetic algorithms in a preliminary design envioroment. In
AIAA paper No. 93-3933. Aircraft Design, Systems and Operations Meeting, Monterey, CA,
USA, 1993.
[16] B. R. Jones, W. A. Crossley, and A. S. Lyrintzis. Aerodynamic and aeroacoustic optimization
of airfoils via a parallel genetic algorithm. In 7th AIAI/USAF/NASA/ISSMO Symposium on
Multidisciplinary Analysis and Optimization, 1998.
[17] E.-S. Lee, G. S. Dulikravich, and B. H. Dennis. Rotor cascade shape optimization with unsteady
passing wakes using implicit dual time stepping and genetic algorithm. In Proceedings of the
9th International Symposium on Transport Phenomena and Dynamics of Rotating Machinery
(ISROMAC-9), Honolulu, HI, February 10-14, 2002.
REFERENCES 19
[18] M. V. Petrovic, G. S. Dulikravich, and T. J. Martin. Maximizing multistage turbine e-
ciency by optimizing hub and shroud shapes and inlet and exit conditions of each blade row.
International Journal of Turbo & Jet-Engines, 17:267278, 2000.
[19] X. Qin, L. Chen, F. Sun, and C. Wu. Optimization for a steam turbine stage eciency using
a genetic algorithm. Applied Thermal Engineering, 23:23072316, 2003.
[20] J.-F. Simon. Contribution to throughow modelling for axial ow turbomachines. Masters
thesis, University of Liege, Liege, Belgium, 2007.
[21] P. Singh and F. Nestmann. Experimental investigation of the inuence of blade height and
blade number on the performance of low head axial ow turbines. Renewable Energy, 36:272
281, 2011.
[22] C.-H. Wu. A general theory of three-dimensional ow in subsonic or supersonic turbomachines
of axial, radial and mixed ow types. In National Advisory Committee for Aeronautics Technical
Note 2604, 1952.
[23] H. Wu, Q. Li, and S. Zhou. Optimization of highly loaded fan rotor based on throughow
model. In Proceddings of ASME Turbo Expo 2007: Power for Land, Sea, and Air GT2007-
27603, pages 13211331, 2007.