Optimal Design of Horizontal-Axis Wind Turbines

Using Blade-Element Theory and Evolutionary Computation

Ernesto Benini – Andrea Toffolo

Dipartimento di Ingegneria Meccanica – Università di Padova

Via Venezia 1, 35131 Padova, Italy

Phone +39-049-8276799 Fax +39-049-8276785

e-mail: benini@[Link], [Link]@[Link]

Abstract

This paper describes a multi-objective optimization method for the design of stall-regulated

horizontal-axis wind turbines. Two modules are used for this purpose: an aerodynamic model

implementing the blade-element theory and a multi-objective evolutionary algorithm. The former

provides a sufficiently accurate solution of the flow field around the rotor disc; the latter handles the

decision variables of the optimization problem, i.e. the main geometrical parameters of the rotor

configuration, and promotes function optimization. The scope of the method is to achieve the best

trade-off performance between two objectives: annual energy production per square meter of wind

park (to be maximized) and cost of energy (to be minimized). Examples of the best solutions found

by the method are described and their performance compared with those of commercial wind

turbines.

Keywords: HAWT, Stall-regulated turbines, Aerodynamic Design, Evolutionary Computation,

Genetic Algorithms, Multi-Objective Design

REVISED 1
Nomenclature
a = Axial induction factor
A = Rotor area, m2
a’ = Tangential induction factor
B = Number of blades
c = Chord length, m
CD = Drag coefficient
CL = Lift coefficient
Cp.r = Rotor power coefficient
Ca = Axial thrust coefficient
D = Drag, N
F = Tip-loss factor
Fa = Axial thrust, N
Ft = Tangential thrust, N
Fc = Centrifugal force, N
fi = i-th fitness function
F = Vector of fitness functions
I = Moment of inertia of hub profile, m4
L = Lift, N
Mb = Bending moment, Nm
P = Power, W
R = Turbine radius, m
r = Local radius, m
s = Shell thickness, m
S = Contour area of hub profile, m2
U = Undisturbed wind speed, m/s
V = Wind speed at the rotor, m/s
W = Relative velocity, m/s
xi = i-th decision variable
x = Vector of decision variables

Greek symbols
 = Incidence angle, deg
 = Safety factor in stress calculation
 = Tilt angle, deg
 = Angle between the mean relative

velocity and tangential direction,

deg
c = Blade twist angle, deg
 = RU,tip-speed ratio
 = Population size
 = Hub-tip ratio
 = Coning angle, deg
 = Density, kg/m3
max = Maximum stress in hub profile,

N/m2
 = Rotational speed, s-1

REVISED 2
INTRODUCTION

The aerodynamic design of a Horizontal-Axis Wind Turbine (HAWT) rotor is a

complex procedure characterized by several trade-off decisions aimed at finding the

optimum overall performance and economy. The decision-making process is very

difficult and the design trends are not uniquely established [1]: a number of different

commercial turbine types in use today have been derived from both theoretical and

empirical methods, but there is no clear evidence on which of these have to be regarded

as optimal.

In fact, when a fixed-blade-pitch-angle wind turbine (also named stall-regulated

turbine) is considered, the most controversial issues regard, for a given Weibull wind

distribution, the choices of turbine geometrical parameters, tip-speed ratio and type of

airfoils [1]. It is well understood that such parameters play a decisive role in

determining the overall aerodynamic performance of a HAWT; however it is not yet

ascertained which combination of them is to be preferred to achieve the best design

objectives, especially when a multi-objective problem is tackled. The two fundamental

objectives of the design of a HAWT turbine are to maximize its Annual Energy

Production (AEP) and to minimize the Cost of Energy (COE) produced [2].

A number of papers have recently described how to deal with such a problem using

numerical models of rotor aerodynamics and optimization techniques [3-6]. In these

papers, optimization methods are described where a single objective function is taken

into account at each time with the presence of constraints; the problem involving

multiple objectives, namely maximum AEP and minimum COE, is addressed using a

single objective function where AEP and COE are combined by means of appropriate

weights [4,5]. These methods, therefore, do not have a multi-objective optimization

capability since they were not conceived to obtain the real set of trade-off solutions

among multiple (and often conflicting) objectives. In many circumstances, however, the

designer is interested in knowing the complete set of optimal turbine configurations

which correspond to the desired objectives. As an example, for a given (or target) COE

and a given wind distribution, the designer wants to know which rotor gives the

maximum AEP. On the other hand, the designer might want to know how to achieve the

minimum COE for a desired AEP. In general, the knowledge of how AEP varies as a

function of COE for a given set of HAWTs is fundamental for exploiting a windy site.

Using the well-known terminology of optimization, this is equivalent to search for the

set of Pareto-optimal design solutions with respect to AEP and COE.

Also, the choice of decision variables to reach desired objectives deserves to be

discussed. While COE is a general figure of merit and leads to homogeneous

comparisons among different HAWTs, AEP strongly depends on turbine size and rated

power. When maximum AEP is to be pursued, in fact, Selig and Coverstone-Carroll [4],

as well as Giguère et al. [5] show which is the best blade shape for a given turbine rated

power and rotor radius. However, to achieve more general design indications, the

influence of rotor radius and rated power should be investigated during the optimization

process. When this is the case, maximization of AEP would indeed drive the search

toward very large rotors with very thin blades, thus making the results of limited

engineering interest because of their infeasibility. Instead of using AEP tout court, a

new figure, namely AEP density (AEP/R2), should be employed in order to obtain more

general and effective design guidelines. The concept of AEP density is introduced and

described later on in the paper.

 This paper illustrates a two-objective optimization method that uses a Multi-

Objective Evolutionary Algorithm (MOEA) and a mathematical model for HAWT

aerodynamics. The method is used to obtain the best trade-off solutions between turbine

AEP density and COE. The MOEA handles the design parameters chosen for

optimization and searches for the group of optimal solutions following the basic

principles of Genetic Programming and Pareto concepts [7,8]. The mathematical model

utilizes blade element theory [9] to determine the performance of a HAWT, as well as to

measure the fitness functions of the optimization. The result is an integrated procedure

that helps the designer in the choice of fundamental design parameters of HAWTs.

In the following sections, the mathematical model of a wind turbine and the design

problem are first outlined. Next, the MOEA is described in detail. Finally, the results of

the design optimization problem are presented and discussed critically.

THE FLOW MODEL

The flow model used to predict the aerodynamic behavior of HAWTs is based on the

modified blade element theory (BET) proposed by Wilson and Walker [9]. A Prandtl

tip-loss model is included and a thrust correction form is employed to improve the

accuracy of the model for highly loaded rotors.

Despite being developed many years ago, the strip theory has demonstrated a good

reliability for analyzing the performance of HAWTs [10]. Because of the wake

expansion through the rotor disc, in fact, shed vortices in the wake do not interact

strongly with the main flow. Therefore, the tip vortices move outboard of the rotor,

making the estimation of induced velocities that affect the flow field more accurate.
Moreover, the strip theory is fully analytical in its formulation, and this fact makes the

model easy to implement and almost inexpensive to run.

The following system of non-linear equations is solved at each radial strip of radius

r, for a given wind speed, using a Newton-Raphson scheme (Figures 1-2):


 
2
 r 
a ' 1  a '    a 1  a  cos 2


  R 

 tan

 1
1 

a
a '
R c os
r 
   B r  R   
 
2
F
  a rc cos   exp  
 
   2 r si n  

C a  4 aF  1  a  if a  0.2
 
  4 F  0.04  0.6 a  if a  0.2

  
 
C B c cos 2
 a  1  a  2  C L cos  C D sin 
si n 2
  
2 r


 c  

(1)

The annual wind velocity variation is modeled as a Weibull distribution with mean

wind speed of 8.0 m/s, and a shape parameter of 2 (Rayleigh distribution). The cut-in

speed and the cut-out speed are fixed at 3 m/s and 25 m/s respectively.

THE PROBLEM OF HAWT OPTIMAL DESIGN

The design of a stall-regulated HAWT rotor is formulated as a constrained multi-

objective optimization problem with respect to geometric and operational

characteristics.

For a given wind farm area suitable for wind energy exploitation, and for a given

wind distribution, it is of primary interest to obtain the maximum global AEP at the

minimum COE. The AEP of the wind farm is related to the number and type of the

turbines to be installed. The COE is by definition the ratio of the total costs from

manufacture and erection of the wind turbines to the global AEP [5]. Since the number

of turbines that can be installed in the area is inversely proportional to the square of

turbine radius, the maximum AEP of the wind farm is to be expected when the ratio of
AEP of a single turbine to its rotor area reaches the maximum value. Therefore, AEP

density, hereafter defined as

f1(x) = AEP/R2 [kWh/y/m2] (2)

should be maximized.

The COE, which is to be minimized to obtain optimum economic convenience, is

calculated using the following expression given by Giguère et al. [5]:

TC  BOS
f2(x) = FCR  O & M [$/kWh] (3)
AEP

where TC is the turbine cost, proportional to blade weight (20$/kg, assuming that the

blades represent 20% of the total turbine cost), BOS is the balance of station,

proportional to turbine rated power (200$/kW), FCR is the fixed charge rate (11% per

year) and O & M is the cost for operation and maintenance (0.01$/kWh). This cost

model is based on the assumption that total turbine cost can be reconstructed on the

basis of the cost of turbine blade alone. This requires that for a given rated power, the

proportion among turbine component costs does not change. However, it does not take

into account the possible variations of such proportions that might occur for very

different turbine designs, e.g. in the case of very different tip speeds. In fact, the tip

speed influences the maximum torque needed to achieve a prescribed rated power,

which in turns affects the design and cost of the drive train. In this case, the cost model

does not capture such effect.

 The problem consists of finding the set of decision variables x=(x1,…,xn) that

simultaneously optimizes two objective functions F(x)=(f1(x),f2(x)): since these criteria

do not feature the same optimal solutions, the best possible trade-offs are to be found

between the two criteria. The tool used in multi-objective optimization to determine

whether or not a solution is really one of the best possible trade-off is Pareto optimality

(Figure 3): a solution x is said to be Pareto optimal if and only if there is no other

solution x' such that f1(x')  f1(x) and f2(x') > f2(x) or such that f1(x') > f1(x) and f2(x') 

f2(x). The set of Pareto optimal solutions is also called the Pareto Optimal Set and the

set of the corresponding objective function values is called the Pareto Front.

The decision variables are:

 Tip speed R. A range 40 <R <80 m/s is assumed here, according to the

values usually adopted in the practice. Both  and R are then obtained by

imposing a fixed turbine rated power (see below).

 Hub/tip ratio rmin/R. There is a lack of knowledge in the open literature

regarding this variable, although it is widely recognized that it has a preeminent

role for determining the operating range of the turbine, since it is linked to the

appearance of the stall phenomenon in the airfoils at the root. A range

0.05<<0.2 is assumed.

 Chord distribution along the blade c/R=f(r/R). The chord length is defined as a

function of blade radius using Bezier curves (Figure 4). Four control points are

used (one at hub, one at tip and two at intermediate stations), so six parameters

are needed to define their coordinates (the four ordinates plus the two abscissas

of the intermediate stations).

  Twist distribution along the blade c =f(r/R). The twist is defined as a function

of blade radius in the same way as the chord distribution (Figure 4).

The design parameters (fixed before any optimization run) are:

 Airfoil properties at any radius (CL and CD as functions of the incidence angle

). Airfoil family made up of three airfoils (root, primary and tip airfoils) is

considered. In particular, the distribution of these airfoils along the blade and

their thickness are fixed. NACA 63-2-21, NACA 63-2-18 and NACA 63-2-15

are chosen as root, primary and tip airfoils respectively [11]. The experimental

airfoil lift and drag coefficients [11] are considered over a wide range of angles

of attack (from 0 to 20 degrees), and are corrected according to Corrigan’s stall-

delay model [12] to account for three-dimensional and rotational effects. A

post-stall model is also applied to extrapolate airfoil data up to 90 degrees, as

described in [13]. The Reynolds number of the experiments is Re = 3106 and

airfoil properties are not modified to account for Reynolds variation effects.

 Shell (skin) thickness distribution along the blade. The distribution is supposed

to be parabolic with radius (Figure 5), the maximum thickness being located at

blade root. The actual value of hub thickness is determined by a blade structural

model, part of which is implemented as suggested in [5], considering bending

stress due to blade loading and tensile stress due to centrifugal forces in the hub

profile. In the structural model, airfoil section at the hub is modeled as an I-

beam without a shear web [5] (Figure 6). The skin thickness of the hub profile

is determined according to a loading condition corresponding to the prescribed

rated power. This loading condition is compared to the maximum allowable

 stress condition of the material using a safety factor. In detail, first the torque

corresponding to the prescribed rated power and rotational speed is derived.

Then the bending moment and the centrifugal force due to blade rotation are

computed for the hub profile. The required moment of inertia I of the I-beam

representing the hub profile, being a function of the skin thickness when the

other geometrical parameters have been fixed, is finally derived by computing

the maximum overall tensile and bending stress:

 M 0.21c hub Fc 
 max    b   (4)
 I  s 2 S 

The maximum stress level for the hub profile is derived from an allowable stress

for E-glass using a safety margin  equal to 10. An allowable material strength

max equal to 94 MPa is used.

 Coning angle . It is used for structural reasons and for the accomplishment of

an acceptable accommodation for the tower rather than for aerodynamic

purposes. A value of =0° is assumed.

 Tilt angle , i.e. the angle between wind direction and rotor axis of rotation. It is

often used for tower accommodation purposes. A value  =0° is assumed.

 Number of blades B. In the evolution of modern HAWTs, the conventional and

proven approach is to use three blades, thus in this paper it is assumed B=3.

The quantities calculated by the flow model are:

 Axial induction factor a(r/R);

 Tangential induction factor a'(r/R);

 Tip-loss factor F(r/R);

  Coefficient of axial thrust Ca(r/R);

 Angle of the relative velocity (r/R);

 Incidence angle (r/R).

The constraints imposed are the following:

 Rated power of the turbine, which is supposed to be fixed. Since the model

evaluates the rotor power coefficient Cp.r for the values of  corresponding to

3<U<25 m/s, the power density curve P/A as a function of wind speed can be

calculated. Assuming that turbine rated power corresponds to the maximum

value of this curve, R, and then , are obtained.

 The axial induction factor a must not exceed 0.5 in any strip, otherwise the axial

velocity beyond the rotor would become negative (feasibility condition).

 Shell thickness must not exceed half of blade profile thickness at any radius

(feasibility condition).

EVOLUTIONARY ALGORITHMS FOR SOLVING MULTI-OBJECTIVE

DESIGN PROBLEMS

Evolutionary algorithms are a hybrid deterministic/stochastic optimization tool

implementing an imitation of the natural evolution process, in which biological

organisms, generation by generation, become “optimized” with respect to the

environmental conditions. Multi-objective evolutionary algorithms (MOEAs) have

largely proven to be superior than classical search algorithms in Pareto-based multi-

objective optimization problems, since they already deal with a “population” of

solutions and they are able to find a set of multiple optimal solutions in a single run,

whereas classical techniques have to be applied repeatedly with different parameters in

order to find different Pareto-optimal solutions.

The two main goals of a MOEA are 1) convergence to the true Pareto optimal set

and 2) the spread of the solutions over the entire Pareto optimal set, introducing no bias

towards any particular zone of it; the latter requires the use of a diversity-preserving

mechanism during the search, otherwise the algorithm would prematurely converge

over the few Pareto-optimal solutions found in the early stages of the search. The

mechanism included in the MOEA used for the solution of the optimization problem

proposed in this paper was developed by Toffolo and Benini [14]. This technique, called

Genetic Diversity Evaluation Method (GeDEM), completely differs from traditional

mechanisms which attempt to form stable sub-populations, called “niches”, along the

Pareto optimal set through some modifications to the selection and/or replacement

process. The basic idea that inspired the method is that the evaluation process cannot

emphasize only fitness, but should also encourage the exploration of the search space by

promoting the most genetically different individuals. Thus, the GeDEM performs an

additional evaluation step considering a measure of genetic diversity as an objective of a

two-criteria optimization problem, the other objective being the Pareto rank scored in

the objectives of the original optimization problem. The GeDEM ranks the solutions

according to Pareto optimality, with slight modifications that are introduced to consider

the special case of solutions having the same rank in the objectives of the original

optimization problem.
 The structure of the MOEA used in the present work (Figure 7) follows the main

steps of a Evolution Strategy [8]. Decision variables are not coded, but they are

considered with their actual real values.

First of all, parents' selection is performed, each individual having the same

probability of being chosen because the actual struggle for survival occurs after the

evaluation step, when only individuals survive among the available. Then parents

enter the reproduction step, generating  offspring through a crossover strategy in which

the decision variable values of the offspring fall in a range defined by the decision

variable values of the parents; some decision variable values of the offspring are also

randomly mutated with a very low probability. Now the whole population of

individuals is checked for possible clones; clones are then destroyed and replaced

with new randomly generated individuals, to encourage the exploration of the search

space and to give the GeDEM a clearer picture of genetic diversity within the

population. The objective function values of the  offspring are then evaluated (a Pareto

ranking of the individuals according to Goldberg's scheme [7] is performed in

multi-objective optimization problems) and the whole population of individuals is

processed to determine the genetic diversity each individual. In this algorithm, the

chosen measure of genetic diversity for an individual is the minimum normalized

Euclidean distance in the decision variable space from another individual. The next step

of the algorithm is the application of the GeDEM, performed as described in the

previous section; the  best ranked solutions are then selected to survive, while the

remaining  are eliminated.

 Finally, the number of generations elapsed is compared to the established maximum

number of generations: if this termination condition is met, the process stops, otherwise

the surviving solutions become the starting population for the next generation.

APPLICATION OF THE METHOD

In this section, the results concerning the application of the method outlined above are

presented.

First of all, three values of turbine rated power were selected as being representative

of stall-regulated HAWT operation: 600 kW, 800 kW and 1000 kW. In this range,

several commercial turbine rotors have been developed and distributed, and their

technical specifications and performance in terms of AEP are public knowledge. The

COE of these turbines can be also calculated with good approximation using Eq.(3) in

cases where blade weight is given by the manufacturer. The performance of these

turbines were adopted as reference and used for comparison with results of the

optimization.

In all the simulations, a population of 100 individuals was used. Optimizations were

carried out on a Pentium III PC. Calculations were stopped after 150 generations. One

optimization run was performed to maximize AEP density and minimize COE, since the

flow models evaluates the Cp.r -  curve of a non-dimensional blade. In fact, once the tip

speed, , the non-dimensional chord distribution and the twist distribution are given,

AEP density is not affected by the size of the HAWT. By imposing the three selected

rated powers, as mentioned above, the corresponding radii, blade weights, AEPs, etc.

are obtained and then COE is evaluated.

 Figure 8 shows the three Pareto fronts obtained, where reference commercial

turbines are shown as filled markers. These values of the objective functions are the best

trade-offs between maximum AEP density and minimum COE. The points on the left

hand side of the fronts identify design solutions having low costs of produced energy

yet low AEP density, whereas the points on the right hand side of the front represent

configurations having high AEP density but high costs of energy.

It is worth noting that, according to the cost and structural models implemented,

COE always increases with turbine rated power. Therefore, according to these models,

the recently developed large HAWTs would be less suitable for the two-dimensional

layout of a wind farm than for stand-alone installations, in which AEP represents the

energetic objective (note the increasing trend of AEP as turbine rated power increases,

as shown in Figure 9).

The position of commercial turbines is very close to the minimum COE. The slope of

the Pareto front near the economic minimum is quite steep, and an improvement in AEP

density causes about the same proportional increase in COE (see the detail in Figure 8).

Configurations having high AEP density (>6000 kWh/y/m2) are too costly (from two to

six times the minimum COE) to be suggested as design solutions.

Figure 9 illustrates some features of the optimized HAWTs as functions of AEP

density. Both radius and AEP tend to decrease for higher AEP density, so the highest

wind farm annual energy production is achieved using a large number of small (yet

expensive) turbines having relatively low AEPs. On the other hand, bigger radii and

higher AEPs are preferred for commercial HAWTs, but the increase in the size is not

counterbalanced by the increase in energy production, at least in a wind farm

installation. From the economic point of view, however, the cost of these larger turbines
decreases as AEP increases, since the optimized blade weight is approximately constant

in the range of low AEP densities (see Eq.(3) and Fig. 9). On the contrary, for high AEP

densities, turbine cost is higher (due to higher blade weights) and AEP is lower; thus

COE grows rapidly. Blade tip speed, which is the same for the three rated powers

considered, grows with higher AEP densities until the increase in blade weight makes

R slower due to excessive centrifugal stresses depending on .

Figure 10 reports blade chord distribution and blade twist distribution of four

optimized configurations extracted at different positions on the Pareto front (see Figure

8). The solution named T1 is the closest to the minimum COE, and has a big radius and

a high AEP. Its non-dimensional chord distribution is the thinnest, but its radius is

nearly two times the radius of T4. The blade twist distribution indicates that T1 has the

lowest twist angles, that is, for a given wind speed and blade tip speed, the highest

incidence angles; the tangential component of the lift force, however, is lower in this

case. Considering the modifications of blade geometry from T1 to T4, the non-

dimensional blade chord progressively enlarges and the twist angle distributions shift to

somewhat higher angles.

CONCLUSIONS

This paper presents a multi-objective optimization method for the design of Horizontal-

Axis Wind Turbines (HAWTs), based on the coupling of an aerodynamic model

implementing the blade-element theory and an evolutionary algorithm. The method

attempts to address general indications for the choice of the fundamental parameters of
HAWTs that allow the realization of the highest values of the AEP density at the

minimum COE. This multi-objective problem is handled using Pareto concepts.

The main results obtained indicate that the minimization of COE requires large-

sized HAWTs having high AEPs but low blade loads and low blade weights. On the

other hand, the maximization of AEP density requires small sized HAWTs having

relatively low AEPs but high blade loads which result in overly expensive design

solutions. The determination of the Pareto front, however, makes it possible to establish

the margins of improvement for the AEP density and the corresponding increase in

COE.

References

[1] Fuglsang, P. L., Madsen, H. Å., 1995, A design study of a 1 MW stall regulated

rotor, Risø-R-799(EN), Risø National Laboratory, Roskilde, Denmark.

[2] Morgan, C. A., Garrad A. D., 1988, “The Design of Optimum Rotors for Horizontal

Axis Wind Turbines,” in Milborrow D. J. (Editor), Wind Energy Conversion,

Proceedings of the 1988 Tenth BWEA Wind Energy Conference, pp. 143-147,

Mechanical Engineering Publications Ltd., London.

[3] Belessis, M. A., Stamos, D. G., Voutsinas, S. G., 1996, “Investigation of the

Capabilities of a Genetic Optimization Algorithm in Designing Wind Turbine

Rotors,” Proceedings, European Union Wind Energy Conference and Exhibition,

Goteborg, Sweden, May 20-24.

[4] Selig, M. S., Coverstone-Carroll, V. L., 1996, “Application of a Genetic Algorithm

to Wind Turbine Design,” ASME Journal Energy Resources Technology, Vol.

118.

[5] Giguère, P., Selig, M. S., Tangler, J. L., 1999, Blade Design Trade-Offs Using Low-

Lift Airfoils for Stall-Regulated HAWTs, NREL/CP-500-26091, Golden, Colorado:

National Renewable Energy Laboratory.

[6] Fuglsang, P., Madsen H. Å., 1999, “Optimization method for wind turbine rotors,”

Journal of Wind Engineering and Industrial Aerodynamics, Vol. 80, pp. 191–206.

[7] Goldberg, D. E., 1989, Genetic Algorithms in Search, Optimization and Machine

Learning, Addison-Wesley, New York.

[8] Schwefel, H. P., 1995, Evolution and Optimum Seeking, John Wiley & Sons.

[9] Wilson, R. E., Walker, S. N., 1984, Performance Analysis of Horizontal Axis Wind

Turbines, Corvallis, Oregon: Oregon State University.

[10] Spera, D. A. (Editor), 1995, Wind Turbine Technology, American Society of

Mechanical Engineers, New York.

[11] Bertagnolio, F., Sørensen, N., Johansen, J., Fuglsang, P., 2001, Wind Turbine

Airfoil Catalogue, Risø-R-1280(EN), Risø National Laboratory, Roskilde,

Denmark.

[12] Corrigan, J. J., Schillings, J. J., 1994, “Empirical Model for Stall Delay Due to

Rotation,” American Helicopter Society Aeromechanics Specialists Conference,

San Francisco, CA, January.

[13] Raj, N., Selig, M. S., 1998, “Capabilities of WTPREP for Blade Airfoil Data for

HAWTs,” American Wind Energy Association WINDPOWER 1998 Conference,

Bakersfield, CA, April.

[14] Toffolo, A., Benini, E., 2000, Genetic Diversity as an Objective in Evolutionary

Algorithms. Submitted for publication in Evolutionary Computation.