J. Wind Eng. Ind. Aerodyn.

136 (2015) 6272

Contents lists available at ScienceDirect

Journal of Wind Engineering

and Industrial Aerodynamics
journal homepage: www.elsevier.com/locate/jweia

A Pareto optimal multi-objective optimization for a horizontal axis

wind turbine blade airfoil sections utilizing exergy analysis and
neural networks
Seyed Mehdi Mortazavi a,n, Mohammad Reza Soltani b, Hamid Motieyan c
a

Department of Energy Engineering, College of Energy and Environment, Science and Research branch of Azad University, Tehran, Iran
Department of Aerospace Engineering, Sharif University of Technology, Tehran, Iran
c
Collage of Surveying Engineering, Department of GIS, Khaje Nasir University of Technology, Tehran, Iran
b

art ic l e i nf o

a b s t r a c t

Article history:
Received 18 May 2014
Received in revised form
1 September 2014
Accepted 13 October 2014
Available online 18 November 2014

In this study a multi-objective genetic algorithm is utilized to obtain a Pareto optimal set of solutions for
geometrical characteristics of airfoil sections for 10-meter blades of a horizontal axis wind turbine. The
performance of the airfoil sections during the process of energy conversion is evaluated deploying a 2D
incompressible unsteady CFD solver and the second law analysis. Articial neural networks are trained
employing CFD obtained data sets to represent objective functions in an algorithm which implements
exergetic performance and integrity characteristics as optimization objectives. The results show that utilizing
the second law approach along with Pareto optimality concept leads to a set of precise solutions which
represent minimum energy waste, maximum efciency, and topmost stability. Furthermore, enhanced rotor
performance coefcients are observed through a BEM study which compares conventional designs with the
second law obtained congurations. Exergy analysis is believed to be an efcient tool in the optimal design of
wind turbine blades with the capability of determining the amount of lost opportunities to do useful work.
& 2014 Elsevier Ltd. All rights reserved.

Keywords:
Wind turbine airfoils
Exergy analysis
CFD
Articial neural network
Pareto optimal set

1. Introduction
Depleting sources of fossil fuels and the environmental issues
associated with utilizing them have made development of alternative
and clean energy sources and their effective employment a considerably rapid process (IEA, 2013). Wind energy market is experiencing a
signicant growth which stands out among renewable energy technologies and is expected to extend even in a more substantial way.
Horizontal Axis Wind Turbines (HAWTs) have been demonstrated to
be the transcendent machine tools in extracting the energy from the
wind (Hau, 2006; Gipe, 2004); these machines possess a number of
advantages when compared with Vertical Axis Wind Turbines
(VAWTs), such as higher energy efciency, more mechanical stability,
less sensitiveness to off design conditions, etc. Thus, effective extraction of the wind stream energy with less energy waste, obtaining a
comprehensive understanding of the energy conversion phenomena,
and economic analysis of these systems have become an area of
concern for technology leaders as well as an area of interest for
researchers (Mehta et al., 2014; Arroyo et al., 2013; Castellani and
Vignaroli, 2013; Castellani and Garinei, 2013; McKenna et al., 2014;
Jung and Kwon, 2013; Jha, 2010; Burton et al., 2011; Wood, 2011). In

Corresponding author. Tel: 98 936 865 2980

E-mail address: mehdimortazavi1986@gmail.com (S.M. Mortazavi).

http://dx.doi.org/10.1016/j.jweia.2014.10.009
0167-6105/& 2014 Elsevier Ltd. All rights reserved.

this regard, optimization of geometrical properties of wind turbine

blades has been investigated using various approaches. The design of a
HAWT rotor blade is a complex task consisting of trade-off decisions.
The objective of this procedure can usually be addressed as nding the
optimum performance for a range of specied conditions (Burton et
al., 2011).
Benini and Toffolo (2002) optimized HAWT blades using the
BEM theory along with an evolutionary algorithm. In their study a
multi-objective optimization algorithm was employed for designing the blades of a stall-regulated HAWT. The BEM theory analyzed
the ow eld around the airfoil while the evolutionary algorithm
selected the decision variables.
Jureczko et al. (2005) used aerodynamic characteristics along
with modifying the composite material of the blade in order to
optimize the wind turbine blade. The complexity of the problem of
determining the optimum shape of the blade and the best composite material became the incentive to use multi-criteria optimum
design approach. The aerodynamic analysis was based on the BEM
theory and the effect of the blade material on the dynamic properties was evaluated using a Finite Element Model (FEM). A genetic
algorithm was employed to conduct the optimization.
Vitale and Rossi (2008) devised a computational method for the
design of wind turbine blades. They developed and introduced a
computer code for blade design of HAWTs with low capacities, which
provided the possibility of obtaining optimum wind turbine blade

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

Nomenclature
A
a
a0
c
Cp
_
Ex
,
F
H
I
_
m
k
P
Q
r
R
Re
S
t
T
U
,

V
y

Area [m2]
Axial induction factor
Tangential induction factor
Chord length [m]
Specic heat kJ=kgK
Exergy rate kW
External body force
Energy
Unit tensor
Mass ow rate kg=s
Thermal conductivity W=m2 K
Static pressure Pa
Heat [J]
Radius [m]
Gases constant J=Kmol
Reynolds number
Source term
Time [s]
Temperature [K]
Undisturbed velocity m=s
Velocity component m=s
Free stream velocity m=s
Turbulence wall Y Plus

Greek Letters

shape. The Blade Element Momentum (BEM) theory was used in the
algorithm of uid dynamics analysis. The software made developing
various designs for many different conditions affordable.
Leung et al., (2010) obtained the optimum design of the rotor of a
small wind turbine by using Computational Fluid Dynamics (CFD)
analysis. The objective of this study was performance evaluation of
small wind turbines which usually extract energy from low speed
wind streams. The study investigated the variation of the performance
of micro wind turbine with different design parameters. The results
showed that the performance of high-solidity wind rotors is more
desirable than that consisting of low-solidity rotors.
The overall efciency of a wind turbine might depend on
several contributing factors, such as generator performance, the
ratio of the gear box system and the height of the hub from the
ground, which are not investigated in this study. Generally, with
specied wind distribution and at a given range for rotational
speed of the rotor, the most important issue to address will be
selection of the airfoils and their geometrical properties along the
blade (Jha, 2010; Burton et al., 2011; Wood, 2011; Cengel and
Boles, 2006).
Typically, common design methodologies employ the rst law
of thermodynamics for wind power system analysis and design
(Benini and Toffolo, 2002; Jureczko et al., 2005; Vitale and Rossi,
2008; Leung et al., 2010). Through these methods a theoretical
maximum efciency can be predicted, but the loss of opportunities
to produce more useful work is not evaluated. According to Dincer
and Rossen (2012), to provide an efcient and effective use of the
energy sources, it is essential to consider both the quality and
quantity of the energy used to achieve a given objective. In this
regard, the second law of thermodynamics deals with the quality
of energy and it simply analyzes degradation of the energy during
a process, entropy generation and the lost opportunities to do
work (Dincer and Rossen, 2012; Dincer et al., 2014). Entropy-based
design and exergy analysis have been shown to identify the
maximum theoretical capability of energy system performance

63

Angle of attack deg

Twist angle deg
Exergy efciency
Dynamic viscosity kg=ms
Density [kg=m3 ]
Rotational speed rad=s
Shear stress N=m2 
Approach (ow) angle deg

Subscripts
0
at
D
dest
ef f
in
L
Loss
m
out
ph
product
rel
work
x
y

Reference condition
Atmosphere condition
Drag
Destruction
Effective
Inow
Lift
Loss
Mass
Outow
Physical
Products
Relative
Work
Streamwise direction
Vertical direction

in various applications, by employing the fact that higher levels of

entropy generation are associated with a lower level of useful
energy (Dincer and Rossen, 2012; Dincer et al., 2014; Koroneos
et al., 2003).
Therefore, exergy analysis together with stability characteristics can be considered as an advantageous tool for optimal design
of geometrical parameters of wind turbine rotors. In this study a
Mmulti-objective genetic algorithm (MOGA) is utilized to obtain a
Pareto optimal set of solutions for chord lengths and ow angles
for the airfoil sections of 10-meter blades of a HAWT rotor. The
objective functions in this algorithm are exergy efciency of the
section, loss of exergy around the airfoil, and the solidity of the
airfoil section. First, performances of three different NREL Dan
Somers' airfoil families are evaluated using an incompressible
unsteady CFD solver and second law approach which leads to
selection of an airfoil family for blade investigations. CFD simulations are conducted in order to evaluate exergy efciency and the
resultant dissipated exergy around the airfoil under various ow
conditions of the section's radius, chord length and ow angle.
Neural networks are trained in order to enable predicting the
values of exergy efciency and exergy loss under un-investigated
ow conditions and section radii. In the next step trained neural
networks along with the section's solidity are deployed as objective functions in MOGA to achieve Pareto optimal sets of solution
tailored for each of the blade sections. Moreover, power coefcients of the rotors designed by the second law approach are
compared with the power coefcients representing rotors designed utilizing conventional methods.

2. Second law analysis via CFD

CFD has emerged as an expeditious tool for apprehending
cognition of the ow eld in various engineering problems
including wind engineering, a few exhibits are ow analysis

64

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

(Jafari and Kosasih, 2014), wind farm evaluation (Choi et al., 2013),
wake behavior detection (Abdesalam et al., 2014), etc. (Blocken,
2014). In this study a 2D incompressible unsteady CFD solver, based
on the nite volume method is implemented to provide imperative
information for a second law analysis (i.e., exergy efciency and
dissipated exergy evaluation) around the airfoil in various conditions
of Reynolds number, section's radius, chord length and ow angle.
The data sets achieved in CFD simulations will further be deployed in
the learning process of the articial neural networks making it
possible to predict exergy efciency and exergy loss for those cases
which are not individually simulated. Note that the CFD approach is
homologous for all simulated airfoils and the same procedure is
preserved for each and every case.
2.1. Governing equations
2.1.1. NavireStokes equations
The mathematical model for the study of uid dynamics
problems is based on the fundamental mass, momentum and
energy conversion principles. The approach used in this study is
called Reynolds Averaged NavireStokes (RANS) modeling. These
equations govern transport of the averaged ow quantities, and
can be used for the entire ow eld.
 ,

: Sm
1
t

t


,

 ,,


, ,
:  :p : g F

,





H : H p : kef f T S
t

   
2 ,
,
,T
: I

3

2
3
4

For pertinent prediction of external ow properties, different

turbulence models have been developed. In this study the shear
stress transport k  model is employed, which is highly recommended for modeling ows with considerable adverse pressure.
The detailed descriptions of these turbulence models are given in
Wilcox (1993) and Menter (1994).
2.1.2. Exergy balance for ow around airfoil
Generally, the exergy efciency associated with energy conversion processes can be evaluated from the exergy balance relation.
In this case, the exergy balance equation can be expressed as
_ in Ex
_ work Ex
_ out Ex
_ dest Ex
_ loss
Ex

_ out represent the exergy of the incoming and

_ in and Ex
where Ex
_
outgoing ows respectively, Ex
work is the available energy within
_ dest Ex
_ loss stands for the
the work done in the process, and Ex
amount of exergy which is destructed and lost in the conversion
procedure (Dincer and Rossen, 2012).
The exergy of a ow with a temperature of T, a pressure equal
to P, and a velocity of V, deviating from a reference environment
with properties of T0,P0, and V0 includes two terms: physical and
kinetic exergy. The kinetic term of exergy is equivalent of kinetic
energy of the ow while the physical term is described as


 
 

T
P
Q
_ ph m
_ C p T  T 0 T at C p ln
Ex
 Rln
 loss
6
T0
P0
T at
In Eq. (6), P P at 7 =2V 2 , also T is determined through the
wind chill temperature, based on a model developed by Zecher
(1999). In this study the velocity of the outgoing ow is obtained
at a length equal to one chord of the airfoil after the airfoil itself,
providing a consistent manner for assessing the exergy associated
with the downstream ow. The exergy efciency can be calculated

from the following equation:

_
Ex
product
_ out
Exin  Ex

_
Denition of Ex
product can be deemed as a pivotal task of the
second law analysis since description of a convenient product
directly pertains to the quality and feasibility of the second law
method. Based on the aerodynamic ground, the instrumental
product of this process would be the resultant lift force acting
perpendicular to the surface of the airfoil; therefore, the exergy of
the lift force could be an expeditious demonstration of the exergy
of the products. The exergy of the lift force is given by
Exlif t F L  V y  V y0

Obviously, determining a legitimate dead state vertical vel_

ocity,V y0 , is vital for the effective evaluation of Ex
lif t . The reference
condition in the energy conversion process of an airfoil can be
referred to as a condition in which there are no vertical or
horizontal velocities neither are there any exergy destructions,
which will lead to all the available incoming energy converting to
useful work. In such a condition, the airfoil experiences a state of
oatation leading to equality of the forces of lift and gravity acting
on the airfoil (F L mg). Therefore, for every Reynolds number a
V y0 exists so that
_ in F L 0 V y0
Ex

Eq. (9) expresses the condition in which the exergy of the inow is
thoroughly converted to the exergy of the product with no
destruction. Note that if V y0 was set to be zero, the connotation
would be that the exergy required for achieving and maintaining
the lift is zero, which is not applicable for airfoil application.
V y is evaluated through CFD analysis, utilizing imaginary lines
which are literally generated according to airfoil's surface shape
and placed 10  7 meters above the upper surface and below the
lower surface of the airfoil. Utilizing this method will lead to
obtaining velocity values on the airfoils surface acceptably since
due to the no slip condition deployed on the surfaces, the solver
assumes that these values are zero on the actual airfoil surfaces.
Thus, V y is computed by averaging the node values on the
imaginary lines.
With the exergy efciency evaluated using the discussed
approach, the summation of exergy destruction and exergy loss
for each ow condition will be determined using Eq. (5), this
summation will be referred to as dissipated exergy in this
manuscript. Note that dissipated exergy would represent losses
from the viscous dissipation phenomenon and the accompanied
parasitic drag force acting on the airfoil's surface. The Viscous
Dissipation term is the destruction of uctuating velocity gradients by the action of viscous stresses. With this approach, an utter
cognition of the lost energy conversion opportunities will be
provided which can be perceived as the unique advantage of the
second law method.
2.2. Computational domain
Fig. 1(a) shows the 2D computational domain; the utilized grid is
comprised of quadratic elements. A high resolution grid is employed
near the airfoil surface due to the importance of this region in
boundary layer modeling, the resolution of the grid decreases when
moving far from the airfoil and near the far eld boundaries to make
calculations less time consuming. Fig. 1(b) shows the ne grid around
the airfoil. Note that this initial grid around the airfoil will undergo
adaption (renement) in the process of the simulation, in order to
match the convergence criteria for different ow conditions. Furthermore, the total grid nodes number might not be precisely the same
after adaption, due to preservation of the mentioned criteria for

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

65

Fig. 1. Computational domain around the airfoil: (a) complete grid and (b) ne grid around airfoil.

different cases. However, the number of total nodes for the simulation case consisting of a standard S814 airfoil encountering ow
regime with Reynolds number equal to 106 and an angle of attack of
151 emerges to be 450806 with exactly 614 nodes on the airfoil's
surface.
2.2.1. Solving methods
Despite the problem being naturally steady-state, the ow
around the airfoil inherently tends to instabilities at high angles
of attack; thus, a transient approach has been implemented
throughout all angles of attack in order to maintain a consistent
method of evaluation. The low Reynolds ow around the airfoil
sections is considered as incompressible; as a result, the pressurebased method is used. The pressurevelocity coupling algorithm
is PISO since the simulations are conducted in an unsteady
manner; the discretization of the convection terms in the transport equations is selected to be the second order upwind. A
detailed description of these methods can be found in Ferziger and
Peric (2013).
Three main criteria are majorly considered in order to guarantee the convergence of the solution.
1. Convergence of lift and drag coefcients on the airfoil.
2. Convergence of velocity magnitude on the effective area
adjacent to the airfoils surface.
3. Proper wall y .
Lift and drag coefcients are considered as substantial aerodynamic characteristics of the ow, thus the no-uctuation
behavior should be obtained during the solution for these parameters. Velocity magnitude of near airfoil region represents great
importance in evaluating V y ; therefore, this parameter should be
monitored as the solution converges. Turbulence wall y is a
factor of correct boundary layer modeling, thus during the solution
process, the grid is rened to obtain y between 1 and 5 to
guarantee the requirements for this parameter. This adaption
repeats until y lies within the proper condition and there are
no signicant changes in other solution parameters observed. In
other words, the independency of solution to the utilized grid is
obtained by changing the grid around the airfoil; with this
approach, for each and every simulation case a distinct grid
independency study is conducted, which renes the initial general
grid for different ow conditions (i.e., Re and AOA) and airfoil
shapes to fulll a coherent y requirement.
2.2.2. CFD results and airfoil selection
The National Renewable Energy Laboratory (NREL) has developed several families of special-purpose airfoils for HAWTs (NREL
Airfoil Families for HAWTs, 1995). Three airfoil families have been

considered to be evaluated for employment in the design of the

blade; these families are all recommended for blade radius of 10
meters. The aforementioned airfoil families are introduced in
Table 1. The performances of these airfoils are evaluated through
low Reynolds numbers and various ow angles from 01 to 201.
Performance criteria are exergy efciency and exergy dissipation
as mentioned earlier.
In order to validate the transient simulation approach, the
numerically obtained lift and drag coefcients of S825 airfoil are
compared to the experimental data from NASALangley wind
tunnel test data in Fig. 2. The experimental data are available
from the ofcial website of the laboratory (NREL Airfoil data,
2013), the Reynolds number of the delineated simulations is106 .
As can be observed from Fig. 2(a), lift coefcients show great
agreement through all considered angles of attack. In addition, the
behavior of change in drag coefcient is believed to be captured in
an acceptable manner when considering the fact that all CFD
models overpredict the drag coefcient due to the implemented
integration methods.
When comparing the second law analysis results, it is observed
that the airfoil family entitled F1 transcends in both characteristics
approximately in majority of the investigated ow conditions.
However when consistency in performance is considered, the
other two airfoil families (i.e., F2 and F3) show slighter magnitudes
in changes of performance with probable variation in ow conditions. Furthermore, increase of inow velocity in constant ow
angle will lead to decline in exergy efciency for ow angles other
than 01 and increment in dimensionless dissipated exergy. Examples of these detected tendencies are delineated in the plots of
Fig. 3 while Table 2 presents more of the observed values
demonstrating the paradigms for different airfoils.
Note that dissipated exergy in Fig. 3 and Table 2 is non_
dimensionalized using of the term (mT)
in order to provide the
possibility of comparison between different Reynolds numbers; as
explained earlier, dissipated exergy demonstrates the amount of
lost energy production potential.
After the above discussion, airfoil family entitled F2 is considered for investigations including different chord lengths and
ow conditions occurring in a HAWT blade, due to its coherence in
performance characteristics and rather acceptable exergy efciency magnitudes. The schematic shapes of the selected airfoil
family are presented in Fig. 4.

2.2.3. Blade 2D investigation approach

As wind stream osculates the rotary blade of a wind turbine, it
is assumed that an incident ensuing velocity acts on each of the
cross-sectional blade elements. This velocity is a result of the wind
undistributed velocity and the rotational speed of the blade itself
(Jha, 2010; Burton et al., 2011; Wood, 2011). In this assumption

66

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

three-dimensional effects are ignored; Fig. 5 shows the blade

element velocities.
U 1 1  a
tan
r 1 a0

10

11

In Eqs. (10) and (11) stands for the rotational speed of the
rotor, r represents the radius of the section, and a and a0 are the
induction factors. Also is the ow angle and and represent
the angles of attack and twist of the section, respectively. If the
tangential and radial forces acting on an airfoil are known, a
reverse method can be utilized in which the induction factors can
be calculated using correction Eqs. (10) and (11). As a result, when
a CFD simulation is conducted for a known resultant velocity,
radius of the section, chord length and ow angle, it is possible to
evaluate in what combination of wind free stream velocity and
rotational speed, the assumed, initial resultant velocity and ow
angle occur. Hence, it is feasible to approximately investigate the
real conditions of a wind turbine blade. In this study, based on trial
and error and utilizing the discussed reverse method, resultant
free stream velocities and ow angles are employed so that they
represent different wind free stream velocities equal to 5, 10, 15,
and 20 m=s and the various rotational rates of the rotor including
30, 60, 100, and 140rpm. Further, for every considered airfoil
section and radius, various ow conditions are investigated
according to common restrictions on chord length and ow angle
for a 10-meters blade. It can be observed from these simulations
that at a constant inow velocity and ow angle, exergy efciency
and dissipated exergy both behave proportional to chord length of
the section, which is considered to be quite sensible since the lift
force, parasitic drag force, and viscous dissipation all tend to grow
as a result of that altercation.

Table 1
NREL airfoil families under study.
Airfoil family

Root

Medial

Tip

F1
F2
F3

S814
S815
S811

S825
S812
S809

S826
S813
S810

3. Articial neural networks (ANNs)

An articial neural network (ANN) is a mathematical model
inspired by the structure and function of biological neural networks
in the brain. ANNs consist of a number of non-linear processing units
(i.e., neurons) which are connected to each other by means of weights.
ANNs can learn a specied task by adjusting the mentioned weights
and are demonstrated to be able to represent complex non-linear
behaviors if they are trained expeditiously (Krose and Smagt, 1996).
Through the training process, a network learns general properties of
the relationship between inputs and output of a system by comparing
generated output to target value within a back propagation rule
(Kumar, 2004; Schalkoff, 1997). In this study neural networks are
utilized in order to predict values of exergy efciency and dissipated
exergy around the airfoil surface for section radii, chord lengths, and
ow conditions which are not numerically evaluated.
Multi-layer perceptron (MLP) neural networks are feed-forward networks extremely capable of recognizing any type of relationship between input and output data (Fausett, 1993). Two MLP
networks are trained for prediction of the relationship between
geometrical characteristics of the airfoil sections and ow conditions as inputs, and exergy efciency and dissipated exergy around
airfoil surface as output. Data sets for training these MLPs are
provided by numerical simulations of ow osculating the airfoils,
which are explained in the previous sections.
According to Krose and Smagt (1996), the situations in which
the performance of the MLP improves with a second (or third, etc.)
hidden layer are quite rare, thus one hidden layer would be
sufcient for the large majority of problems. In this study in order
to detect the efcient number of neurons in the hidden layer, a
trial and error method has been utilized. Based on this approach,
the number of neurons changes in an ascending order until the
performance of the MLP tends to degrade; this method is implemented based on the idea that additional nodes provide additional
weights to store or release signals to the network during iteration.
Furthermore, since the MLP is meant to detect a nonlinear
behavior, continuous LogSigmoid function is used for transfer
purpose between the layers. The Sigmoid function introduces nonlinearity in the network, it is differentiable and its derivatives are
fast to compute (Schalkoff, 1997; Fausett, 1993), therefore this
function is deemed expeditious for employment in this study.
Five different section radii (i.e., 2, 4, 6, 8, and 10 m) are considered
to train the required MLPs, for each section of the blade there are 480
data sets with different wind stream velocities, blade rotational rate,
chord length, and ow conditions, thus with employing a total of 2400
data sets the MLPs are trained. The networks are sensitive to changes

Fig. 2. Comparison of CFD results for S825 with wind tunnel test data, Re 106 . (a) Lift coefcient and (b) drag coefcient.

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

67

Fig. 3. Plots of second law performance characteristics of the understudy airfoil. (Re 106 ).
Table 2
Low Reynolds number second law performance evaluation for under study airfoils.
Reynolds number

3  105

Performance characteristic -

Exergy efciency

Angle of attack -

S811
S814
S815
S809
S825
S812
S810
S826
S813

0.0753
0.1489
0.0662
0.0784
0.5713
0.1468
0.1132
0.7056
0.2448

0.5395
0.6687
0.5806
0.5590
0.9284
0.6473
0.5517
0.8835
0.7025

Reynolds Number

6.5  105

Performance characteristic -

Exergy efciency

Angle of attack -

S811
S814
S815
S809
S825
S812
S810
S826
S813

0.0296
0.1039
0.0581
0.0359
0.2507
0.0667
0.0439
0.2899
0.0991

0.2699
0.2923
0.3069
0.2443
0.3885
0.2300
0.2395
0.4410
0.3182

Dimensionless exergy dissipation

10
0.5210
0.6070
0.5841
0.4157
0.7213
0.4574
0.3921
0.7026
0.4865

15
0.4558
0.4709
0.4850
0.3312
0.4070
0.3901
0.2740
0.4140
0.3759

20
0.4639
0.3712
0.5049
0.3060
0.3810
0.3154
0.2354
0.3722
0.2799

0.0205
0.0133
0.0190
0.0106
0.0044
0.0102
0.0092
0.0027
0.0079

0.0071
0.0053
0.0072
0.0059
0.0011
0.0049
0.0062
0.0021
0.0039

10
0.0138
0.0111
0.0120
0.0168
0.0068
0.0161
0.0160
0.0074
0.0135

15
0.0198
0.0208
0.0200
0.0250
0.0248
0.0209
0.0263
0.0246
0.0210

20
0.0166
0.0273
0.0159
0.0204
0.0188
0.0202
0.0205
0.0187
0.0202

Dimensionless exergy dissipation

10
0.2667
0.3596
0.3353
0.2529
0.3672
0.2603
0.1954
0.3509
0.2511

15
0.2466
0.2602
0.2632
0.1928
0.2087
0.1953
0.1381
0.1968
0.1812

in wind free stream velocity, blade rotational rate, radius, airfoil type,
ow angle, and chord length. Note that 70% of the data sets are used
for training the network, 15% of them are employed to validate the
training process, and the remaining 15% is applied to test the network

20
0.2417
0.2599
0.2660
0.1567
0.2015
0.1808
0.0685
0.1637
0.1285

0.0647
0.0558
0.0671
0.0486
0.0345
0.0498
0.0469
0.0298
0.0414

0.0485
0.0505
0.0519
0.0441
0.0379
0.0472
0.0446
0.0322
0.0373

10
0.0793
0.0749
0.0799
0.0885
0.0681
0.0892
0.0938
0.0699
0.0853

15
0.1382
0.1449
0.1425
0.1377
0.1650
0.1393
0.1388
0.1682
0.1380

20
0.1195
0.1220
0.1206
0.1280
0.1289
0.1265
0.1294
0.1297
0.1267

performance against the unseen data. The performance of the trained

MLP for predicting dissipated exergy is shown in Fig. 6.
Analogous network topology, training strategy and number of
data sets are used to train an MLP for predicting exergy efciency for

68

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

Fig. 4. Geometrical depiction of the selected airfoil family.

Fig. 5. Blade element velocities and including angles.

different possible un-investigated cases. Table 3 summarizes the

network characteristics for each MLP. In this table MSE refers to
Mean Square Error, closer to zero values of which indicate higher
network performance in behavior modeling; regression values are
also included, closer to one values of which demonstrate the
network's more precise output predicting capability. Further and
complementary information about ANNs, different network topologies, various training strategies, and the methods of performance
evaluation can be found in Krose and Smagt (1996), Kumar (2004),
Schalkoff (1997) and Fausett (1993).

4. Multi-objective optimization methodology

4.1. Pareto optimality concept
As mentioned earlier, the objective of this study is to determine
the geometrical properties of airfoil sections of wind turbine blade
in a manner that exergy dissipation around the airfoil is minimized while exergy efciency along with solidity of the airfoil is
maximized. CFD simulation revealed that exergy efciency and
exergy dissipation around the airfoil do not mandatorily show
satisfying behaviors simultaneously. For instance, in some ow
conditions if the dissipated exergy is to be at its minimum possible
value, exergy efciency will suffer degradation. To address this
issue of optimization, it is usually convenient to introduce a set of

solutions as possible optimum solutions rather than sticking to

merely one solution as the most valuable (Zhou et al., 2011). As a
result, the scope of the method is to achieve the best trade-off
performance between two main objectives: lower dissipation of
exergy and higher exergy efciency plus higher solidity.
In a problem with two objective functions off 1 x andf 2 x,
solution x is a Pareto optimal solution if, and only if, there is no
other solution x0 such that f 1 x0 Z f 1 x and f 2 x0 g f 2 x or such
that f 1 x0 g f 1 x and f 2 x0 Z f 2 x. The solutions which are in this
set are called Pareto optimal set and the corresponding objective
function values are referred to as Pareto front (Doyle, 1985).
4.1.1. Genetic algorithm
Evolutionary multi-objective algorithms have formerly been
employed in the design of wind turbine blades; studies of Benini
and Toffolo (2002) and Hu et al. (2013) can be referred to as
notable exhibits. In this study in order to obtain Pareto optimal
solutions for each section, the multi-objective algorithm is iterated
separately for each section radius. The step-by-step process of this
algorithm is described in this section.
4.1.2. Generating random population
Pertinent to the radius of the section, the rst random population dening chord length and ow angle should be generated.
There are some constraints imposed on chord length and ow
angle in each section so that initial solutions are appropriate for
the associated section. The initial random solutions are presented
in the form of chromosomes each including 9 gens. The rst gen
from left to right of a chromosome denes the round part of the
chord length, the next three gens dene the decimal part of the
chord length. Gen numbers 6 and 7 describe the round part of the
ow angle, and nally the next three digits represent the decimal
part of the ow angle. For each section a random population with
100 solutions is generated.
4.1.3. Objective functions and parents selection
As mentioned earlier two trained MLP networks will be serving
as objective functions in the genetic algorithm, and another

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

69

Fig. 6. The performance of the trained MLP for predicting dissipated exergy.

Table 3
Characteristics of the trained multi-layer perceptron networks.
Articial Neural Network for exergy
efciency

Network type
Training algorithm
Number of neurons in hidden
layer
MSE of training
MSE of validation
MSE of testing
Regression of training
Regression of validation
Regression of testing

Articial Neural Network for dissipated

exergy

Network type
Training algorithm
Number of neurons in hidden
layer
MSE of training
MSE of validation
MSE of testing
Regression of training
Regression of validation
Regression of testing

objective function will be section solidity.

Bcr
2 r

12

In Eq. (12) r denes the radius of the section, B is the number of

blades on the rotor and c expresses the chord length. Since this
objective function is inuential only when the chord length
changes and also since it shows the same behavior as exergy
efciency when chord length variation occurs, these two functions
are combined in order to make Pareto optimal selection a less
complex task. Thus, there are two main objective functions:

 Minimum exergy dissipation

 Maximum exergy efciency maximum section solidity
For generating children from each population, parents with the
best characteristics need to be considered. In this study nondominated sorting concept is employed in order to classify the
solutions of each population (Srinivas and Dep, 1994). In the rst
step all the solutions of the population are compared considering
Pareto optimality concept, the best individuals which are not
dominated by any other individuals are selected and put into the

MLP with 1 hidden

layer
Back propagation L-M
15
0.0654125
0.0584269
0.0975148
0.99954
0.99847
0.99896
MLP with 1 hidden
layer
Back propagation L-M
15
0.0864601
0.0657832
0.0672287
0.99972
0.99979
0.99986

rst layer. The next step is similar to the rst one, except the
solutions of the rst layer are not considered as candidates for the
second layer. With this procedure all the individuals are sorted in
layers with a descending order for objective function merits. Thus,
considering both tness functions, solutions in the rst layer are
more creditable than solutions in the second layer, and so on.
The number of parents in each population is set to be 10% of the
number of the whole population, which leads to selecting 10
parents from the initial random population. Since the best individuals are meant to be selected as parents, solutions sorted in the
rst layer are considered. If there are not enough individuals in the
rst layer, solutions sorted in the second layer are considered,
and so on.

4.1.4. Cross-over and mutation

Since both chord length and ow angle are described with
three decimal digits and there are certain physical constraints on
the range of these numbers, the cross-over in this study will be
performed as shown in Fig. 7. As it is observed in this depiction,
the child takes its chord length from one parent and its ow angle
from the other. From each pair of parents two children are

70

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

Fig. 7. The pattern of cross-over for an example of a possible solution in the form of a chromosome.

Fig. 8. The Pareto optimal set suggested by MOGA for the section positioned at r 1
and r 2.

Fig. 9. The Pareto optimal set suggested by MOGA for the section positioned at r 3
and r 4.

generated, this approach will generate children with new properties which may not be generated in the initial random population.
In the process of mutation some consideration must be made,
rst, the mutation in the digit of gen must be in the range of
previously mentioned physical constraints, moreover, the mutation should not take place on gens dening the second and third
decimal digits, due to the fact that mutation on the second and
third decimal digits does not change the tness function values
signicantly.

tailored for a 10 m blade. Note that these plots delineate the Pareto
optimal set (not the Pareto front) after 100 iterations for each
section, and according to the Pareto optimality concept all of the
suggested combinations of ow angle and chord length are equally
meritorious. Foremost, it is observed that although all initial
populations are generated within the specied ranges, the MOGA
enforces more conned ranges for both chord length and ow
angle. The next discernible point in the plots deals with the overall
paradigm of the positioning of the solution sets for different
sections, since the ow properties of the radii 1 and 2, 3 and 4,
5 and 6, 7 and 8, and 9 and 10 m are more likely similar, an
analogous pattern of positioning of the optimal solution points is
detectable for each mentioned pair of sections.
Moreover, in order to evaluate the performance of a rotor
consisting of blades designed through second law analysis, an open
source BEM code Qblade is employed (http://www.q-blade.org,
2013). The software is capable of conducting performance analysis
for various rotor congurations which provides the opportunity of
comparison between different blade designs. Qblade utilizes polar
data proles to investigate a rotor performance through an iterative
process.
Three different blade designs with capricious selection of chord
length and twist angle among the second law optimization results
are considered to be compared with blade congurations obtained
from conventional design methods targeting the power coefcient.
The customary methods utilized for dening the geometrical
characteristics of the blades include linear distribution, Betz
method, and Schmitz method for chord length determination
and linear distribution, optimal lift to drag ratio, and simultaneous

4.1.5. Selecting new population and number of iterations

The last step of every evolutionary cycle will be selecting
individuals for the next generation, to achieve this objective
several approaches can be utilized. In this study a combination
of parents and children with the highest merits are selected in
order to prevent losing individuals with high values of tness
function. The number of iterations for each section of blade is set
to 100, which implies that after 100 iterations solutions sorted in
the rst layer are presented as Pareto optimal set for each section.

5. Results and discussion

Since the design of a 10 m blade usually requires more than
5 blade sections, and due to the fact that ANNs and specically
MLPs possess the possibility of rather accurate interpolation,
characteristics of 10 blade sections are obtained through the
MOGA. The results of the algorithm for each section of the blade
are shown in Figs. 812, which represent the design characteristics

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

71

Table 4
Characteristics for rotor blade congurations under study.
Design name

Chord distribution

Twist distribution

Second Law Cong.1, 2, 3

Cong.1
Cong.2
Cong.3

Pareto optimal sets

Linear distribution
Betz equation
Schmitz equation

Pareto optimal sets

Linear distribution
Optimal lift to drag
Simultaneous stall

Fig. 10. The Pareto optimal set suggested by MOGA for the section positioned at
r 5 and r 6.

Fig. 13. Performance coefcients comparison between different second law

designed rotors and rotor designs obtained from conventional methods.

Fig. 11. The Pareto optimal set suggested by MOGA for the section positioned at
r 7 and r 8.

coefcients versus tip speed ratio for the aforementioned designed

blades demonstrating the domination of the second law designs in
majority of the tip speed ratios. It is perceived that second lawbased blade designs possess average advantages approximately
equal to 8.5%, 15.4%, and 12.1% over congurations 1, 2, and 3,
respectively. Furthermore, it is observed that despite the random
selection of the characteristics for second law designs, homologous values of C p are achieved via second law optimized blades,
which posits the idea that employing Pareto optimal concept has
led to rather equally creditable solutions.

6. Conclusion

Fig. 12. The Pareto optimal set suggested by MOGA for the section positioned at
r 9 and r 10.

stall for twist angle selection. Table 4 tabulates the information

about the utilized methods in each design. Note that all the blade
designs are conducted in a manner to achieve the best performance at a tip speed ratio equal to 6.5. Fig. 13 contains power

In this study the second law analysis of energy conversion in

wind turbine airfoils is conducted deploying a 2D incompressible
unsteady CFD solver. Several NREL airfoil sections are evaluated by
exergetic performance characteristics under low Reynolds number
ows, leading to selection of an airfoil family for blade investigations. Each simulation case implements a distinct grid since the
independency of the solution to the grid is obtained by adapting
the mesh around the airfoil maintaining the same y requirements. Neural networks are trained to represent objective functions of exergy efciency and dissipated exergy in the multiobjective genetic algorithm the results of which illustrate a more
precise, conned expression of the combinations of geometrical
characteristics for different sections along the blade. The Pareto
optimal solution sets represent a beneciary trade-off selection
between more efciency and solidity and less waste in energy
production opportunities. Employing a commercial open source
BEM code further reveals the worthiness of the blades designed by
the second law analysis method in comparison with the blade
congurations resulting from the usual design methods. An exergy
analysis approach is believed to be a useful tool in assessing the

72

S.M. Mortazavi et al. / J. Wind Eng. Ind. Aerodyn. 136 (2015) 6272

performance of wind turbine blades since it provides substantial

insights over the quality of energy conversion.
References
Abdesalam, A.M., Boopathi, K., Gomathinayagam, S., Kumar, S.S.H.K., Ramalinga, V.,
2014. Experimental and numerical studies on the wake behavior of a horizontal
axis wind turbine. J. Wind Eng. Ind. Aerodyn. 2014 (128), 5465.
Arroyo, A., Manana, M., Gomez, C., Fernandez, I., Delgado, F., Zobaa, A.F., 2013. A
methodology for the low-cost optimization of small wind turbine performance.
Appl. Energy 104, 19.
Available at: http//www.q-blade.org. (accessed 19.04.13).
Benini, E., Toffolo, A. 2002. Optimal design of horizontal-axis wind turbines using
blade-element theory and evolutionary computation. Dipartimento di Ingegneria Meccanica, Universita di Padova,Via Venezia 1, 35131 Padova, Italy.
Blocken, B. (2014). 50 years of computational wind engineering: past, present and
future. J. Wind Eng. Ind. Aerodyn. 2014; 129:69-102.
Burton, T., Jenkins, N., Sharpe, D., Bossanyi, E., 2011. Handbook of Wind Energy.
Wiley Sons, Chichester, West Sussex, PO19 8SQ, United Kingdom.
Castellani, F., Garinei, A., 2013. On the way to harness high-altitude wind power:
dening the operational asset for an airship wind generator. Appl. Energy 112,
592600.
Castellani, F., Vignaroli, A., 2013. An application of the actuator disc model for wind
turbine wakes calculations. Appl. Energy 101, 400432.
Cengel, Y.A., Boles, M.A., 2006. Thermodynamics: An Engineering Approach, 5th
edition McGraw Hill, Nevada, Reno.
Choi, N.J., Nam, S.H., Jeong, J.H., Kim, K.C., 2013. Numerical study on the horizontal
axis wind turbine arrangement in a wind farm: effect of separation distance on
the turbine aerodynamic power output. J. Wind Eng. Ind. Aerodyn. 2013 (117),
1117.
Dincer, I., Rossen, M.A., 2012. Exergy, Energy, Environment and Sustainable
Development, Second Edition Elsevier Science, Burlington, Massachusetts.
Dincer, I., Midili, A., Kucuk, H., 2014. Progress in Exergy, Energy, and the Environment. Springer, Cham Heidelberg New York Dordrecht London.
Doyle, J. 1985. Reasoned assumptions and pareto optimality. Proceedings of the 9th
IJCAI, pp. 8790.
Fausett, L.V., 1993. Fundamentals of Neural Networks: architecture (Algorithms and
Applications). Pearson.
Ferziger, J.H., Peric, M., 2013. Computational Methods for Fluid Dynamics. Springer.
Gipe, P., 2004. Wind Power, Revised Edition: Renewable Energy for Home, Farm,
and Business. Chelsea Green Publishing.
Hau, E., 2006. Wind Turbines: Fundamentals, Technologies, Application, Economics.
Springer.

Hu, W., Park, D., Choi, D., 2013. Structural optimization procedure of a composite
wind turbine blade for reducing both material cost and blade weight. Eng.
Optim. 45 (12), 14691487.
IEA World Energy Outlook1- www.iea.org (Accessed 26.04.13).
Jafari, S.A.H., Kosasih, B., 2014. Flow analysis of shrouded small wind turbines with
a simple frustum diffuser with computational uid dynamic simulations. J.
Wind Eng. Ind. Aerodyn. 2014 (125), 102110.
Jha, A.R., 2010. Wind Turbine Technology. CRC press.
Jung, S., Kwon, S.D., 2013. Weighted error functions in articial neural networks for
improved wind energy potential estimation. Appl. Energy 111, 778790.
Jureczko, M., Pawlak, M., Mezyk, A., 2005. Optimization of wind turbine blades. J.
Mater. Process. Technol. 167, 463471.
Koroneos, C., Spachos, N., Moussiopoulos, N., 2003. Exergy analysis of renewable
energy sources. Renew. Energy 2003 (28), 295310.
Krose, B., Smagt, P., 1996. An Introduction to Neural Networks. The University of
Amsterdam, Amsterdam.
Kumar, S., 2004. Neural Networks. Tata McGraw-Hill Education.
Leung, D., Deng, Y., Leung, M. 2010. Design Optimization of a Cost-Effective Micro
Wind Turbine. In: Proceedings of the World Congress on Engineering 2010, vol
II WCE 2010, June 30July 2, 2010, London, U.K.
McKenna, R., Hollnaicher, S., Fichtner, W., 2014. Cost-potential curves for onshore
wind energy: a high-resolution analysis for Germany. Appl. Energy 115,
103115.
Mehta, D., Van Zuijlen, A.H., Koren, B., Holierhoek, J.G., Bijl, H., 2014. Large Eddy
simulation of wind farm aerodynamics: a review. J. Wind Eng. Ind. Aerodyn.
2014 (133), 117.
Menter, F.R., 1994. Two-equation eddy-viscosity turbulence models for engineering. AIAA J. 32 (8), 15981605.
NREL Airfoil Families for HAWTs, 1995. (accessed 29.01.13), http://www.awea.org.
NREL Airfoil data. Available at: www.wind.nrel.gov/airfoils/AirfoilData. (accessed
15.03.13).
Schalkoff, R., 1997. Articial Neural Networks. Toronto, the McGraw-Hill Companies, Inc..
Srinivas, N., Dep, K., 1994. Multi-objective optimization using non-dominated
sorting in genetic algorithms. Evol. Comput. 2 (3), 221248.
Vitale, A.J., Rossi, A.P., 2008. Computational method for the design of wind turbine
blades. Int. J. Hydrog. Energy 33, 34663470.
Wilcox, D., 1993. Turbulence modeling for CFD. DCW Industries, Inc., La Canada, CA.
Wood, D., 2011. Small Wind Turbines: Analysis (Design, and Application). Springer.
Zecher, J.B., 1999. A new approach to an accurate wind chill factor. Bull. Am.
Meteorol. Soc. 80 (9), 1893-9.
Zhou, A., Qu, B., Li, H., Zhao, S., Suganthan, P, Zhang, Q., 2011. Multi-objective
evolutionary algorithms: a survey of the state of the art. Swarm Evol. Comput, 1
(1), 3249.