Caboni, Marco (2016) Probabilistic design optimization of horizontal axis

wind turbine rotors. PhD thesis

http://theses.gla.ac.uk/7338/

Copyright and moral rights for this thesis are retained by the author

A copy can be downloaded for personal non-commercial research or

study, without prior permission or charge

This thesis cannot be reproduced or quoted extensively from without first

obtaining permission in writing from the Author

The content must not be changed in any way or sold commercially in any
format or medium without the formal permission of the Author

When referring to this work, full bibliographic details including the

author, title, awarding institution and date of the thesis must be given.

Glasgow Theses Service

http://theses.gla.ac.uk/
[email protected]
Probabilistic Design Optimization of
Horizontal Axis Wind Turbine Rotors

Marco Caboni
School of Engineering
University of Glasgow

This thesis is submitted for the degree of

Doctor of Philosophy

May 2016
 Abstract

Considerable interest in renewable energy has increased in recent years due to the concerns
raised over the environmental impact of conventional energy sources and their price volatility.
In particular, wind power has enjoyed a dramatic global growth in installed capacity over
the past few decades. Nowadays, the advancement of wind turbine industry represents
a challenge for several engineering areas, including materials science, computer science,
aerodynamics, analytical design and analysis methods, testing and monitoring, and power
electronics. In particular, the technological improvement of wind turbines is currently
tied to the use of advanced design methodologies, allowing the designers to develop new
and more efficient design concepts. Integrating mathematical optimization techniques into
the multidisciplinary design of wind turbines constitutes a promising way to enhance the
profitability of these devices. In the literature, wind turbine design optimization is typically
performed deterministically. Deterministic optimizations do not consider any degree of
randomness affecting the inputs of the system under consideration, and result, therefore, in an
unique set of outputs. However, given the stochastic nature of the wind and the uncertainties
associated, for instance, with wind turbine operating conditions or geometric tolerances,
deterministically optimized designs may be inefficient. Therefore, one of the ways to further
improve the design of modern wind turbines is to take into account the aforementioned
sources of uncertainty in the optimization process, achieving robust configurations with
minimal performance sensitivity to factors causing variability.
The research work presented in this thesis deals with the development of a novel inte-
grated multidisciplinary design framework for the robust aeroservoelastic design optimization
of multi-megawatt horizontal axis wind turbine (HAWT) rotors, accounting for the stochastic
variability related to the input variables. The design system is based on a multidisciplinary
analysis module integrating several simulations tools needed to characterize the aeroservoe-
lastic behavior of wind turbines, and determine their economical performance by means of
the levelized cost of energy (LCOE). The reported design framework is portable and modular
in that any of its analysis modules can be replaced with counterparts of user-selected fidelity.
The presented technology is applied to the design of a 5-MW HAWT rotor to be used at
sites of wind power density class from 3 to 7, where the mean wind speed at 50 m above the
iv

ground ranges from 6.4 to 11.9 m/s. Assuming the mean wind speed to vary stochastically in
such range, the rotor design is optimized by minimizing the mean and standard deviation of
the LCOE. Airfoil shapes, spanwise distributions of blade chord and twist, internal structural
layup and rotor speed are optimized concurrently, subject to an extensive set of structural
and aeroelastic constraints. The effectiveness of the multidisciplinary and robust design
framework is demonstrated by showing that the probabilistically designed turbine achieves
more favorable probabilistic performance than those of the initial baseline turbine and a
turbine designed deterministically.

Keywords: horizontal axis wind turbine design, multidisciplinary design optimization, robust
design optimization, wind turbine aeroservoelasticity, uncertainty propagation, environmental
uncertainty.
Table of contents

List of tables ix

List of figures xi

Nomenclature xxv

1 Introduction 1
1.1 Multidisciplinary design optimization . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Wind turbine MDO literature review . . . . . . . . . . . . . . . . . 4
1.2 Robust design optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Structural constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Airfoil optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Objectives, novelty and overview of the thesis . . . . . . . . . . . . . . . . 12

2 Multidisciplinary analysis of horizontal axis wind turbine rotors 15

2.1 Wind turbine rotor aeroservoelasticity . . . . . . . . . . . . . . . . . . . . 15
2.1.1 Wind turbine rotor aerodynamics . . . . . . . . . . . . . . . . . . 16
2.1.2 Structural analysis of the rotor blades . . . . . . . . . . . . . . . . 25
2.2 Wind turbine airfoil aerodynamics . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Methods for the airfoil aerodynamic analysis . . . . . . . . . . . . 31
2.2.2 Rotational effects and 3D corrections . . . . . . . . . . . . . . . . 32
2.2.3 Global post-stall methods . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Cost of energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3 Robust design optimization 41

3.1 Mathematical formulation of robust design optimization problems . . . . . 42
3.2 Robust design optimization methods . . . . . . . . . . . . . . . . . . . . . 44
3.3 Robustness of constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Methods for uncertainty propagation . . . . . . . . . . . . . . . . . . . . . 45
vi Table of contents

3.4.1 Sampling methods . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.4.2 Deterministic methods . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Optimization algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.1 Derivative-based algorithms . . . . . . . . . . . . . . . . . . . . . 51
3.5.2 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.3 Pattern search algorithms . . . . . . . . . . . . . . . . . . . . . . . 55

4 Multidisciplinary and robust design optimization framework 59

4.1 Fully integrated multidisciplinary and robust design optimization framework 59
4.2 Integrated multidisciplinary analysis model . . . . . . . . . . . . . . . . . 61
4.2.1 Airfoil aerodynamics . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.2 Rotor aeroservoelasticity . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.3 Cost model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3 Uncertainty propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.4 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4.1 MOPED and IDEA . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4.2 MATLAB patternsearch function . . . . . . . . . . . . . . . . . . 72

5 Design space and reference turbine definition 75

5.1 Turbine parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1.1 External geometry parametrization . . . . . . . . . . . . . . . . . . 76
5.1.2 Parametrization of the rotor speed . . . . . . . . . . . . . . . . . . 78
5.1.3 Internal structural layup parametrization . . . . . . . . . . . . . . . 81
5.2 Design space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Reference turbine definition . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3.1 Definition of the reference turbine’s airfoils . . . . . . . . . . . . . 87
5.3.2 Definition of the reference turbine’s chord and twist distributions . . 88
5.3.3 Definition of the reference turbine’s rotor speed and internal struc-
tural layup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

6 HAWT design optimization under environmental uncertainty 95

6.1 Objective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.2.1 Design load cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2.2 Safety factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.3 Ultimate load analysis . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.4 Fatigue load analysis . . . . . . . . . . . . . . . . . . . . . . . . . 103
Table of contents vii

6.2.5 Constraint on maximum rotor speed . . . . . . . . . . . . . . . . . 105

6.2.6 Blade natural frequency analysis . . . . . . . . . . . . . . . . . . . 106
6.2.7 Geometric feasibility checks . . . . . . . . . . . . . . . . . . . . . 106
6.3 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.4 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

7 Conclusions and suggestions for further work 123

7.1 Summary and concluding remarks . . . . . . . . . . . . . . . . . . . . . . 123
7.1.1 Conclusions on the development of the design system . . . . . . . . 123
7.1.2 Conclusions on the application of the design system . . . . . . . . 124
7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

References 127
List of tables

5.1 Blade laminate constituent materials . . . . . . . . . . . . . . . . . . . . . 83

5.2 Mechanical properties of the materials used to define the composite laminates
of the blades . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Design variable bounds of the root airfoil . . . . . . . . . . . . . . . . . . 84
5.4 Design variable bounds of the midspan airfoil . . . . . . . . . . . . . . . . 84
5.5 Design variable bounds of the tip airfoil . . . . . . . . . . . . . . . . . . . 85
5.6 Design variable bounds of chord and twist distributions, torque control
parameter and thickness parameter . . . . . . . . . . . . . . . . . . . . . . 85
5.7 Torque control parameter, and thickness parameter of the reference turbine
and the NREL offshore 5-MW baseline one . . . . . . . . . . . . . . . . . 92
5.8 Blade mass of the reference turbine and the NREL offshore 5-MW baseline one 93

6.1 IEC DLCs considered in the structural verification of each wind turbine
generated during the optimization process . . . . . . . . . . . . . . . . . . 98
6.2 Total safety factor used in the ultimate and fatigue load analyses . . . . . . 101
6.3 Ultimate design blade root moment . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Comparison of the gross design specifications and overall performance of
the reference, deterministic and robust designs . . . . . . . . . . . . . . . . 109
6.5 Comparison of the structural characteristics of the reference, deterministic
and robust blade designs . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.6 Torque control parameter, and thickness parameter of the reference, deter-
ministic and robust designs . . . . . . . . . . . . . . . . . . . . . . . . . . 115
List of figures

1.1 Forecast of worldwide wind power installed capacity . . . . . . . . . . . . 1

1.2 Upscaling history of wind turbine size . . . . . . . . . . . . . . . . . . . . 2

2.1 Geometric and aerodynamic parameters of a generic blade strip . . . . . . . 19

2.2 Tip vortex pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Thrust coefficient versus axial induction factor . . . . . . . . . . . . . . . . 22
2.4 Tower shadow model at a given point . . . . . . . . . . . . . . . . . . . . . 23
2.5 Dynamic stall behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.6 Modeling of the turbine blade cross-section . . . . . . . . . . . . . . . . . 27
2.7 Flow field over a stalled section of a rotating blade . . . . . . . . . . . . . 33
2.8 Measured lift and drag coefficients of two airfoils at high AoAs . . . . . . . 36
2.9 Comparison of measured and calculated lift and drag coefficients for the
NREL S809 airfoil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.10 Cost estimate from the DOE/NREL scaling model . . . . . . . . . . . . . . 40

3.1 Objective function values at initial point and mesh points after the first polling 56

4.1 Overview of the fully integrated multidisciplinary and robust design opti-
mization framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2 Overview of the integrated multidisciplinary analysis model . . . . . . . . 61
4.3 Details of the module for the aerodynamic analysis of the blade airfoils . . 63
4.4 Comparison between the S809 aifoil experimental polars and those deter-
mined by means of XFOIL and AERODAS . . . . . . . . . . . . . . . . . 64
4.5 Details of the module for the aeroservoelastic analysis of the rotor . . . . . 65
4.6 Typical multi-megawatt HAWT power curve and indicative wind speed
distribution of a generic site . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.7 Details of the cost model . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.8 Uniform PDF of the uncertain variable ū . . . . . . . . . . . . . . . . . . . 69
xii List of figures

5.1 Airfoil parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5.2 Parametrization of chord distribution and twist distribution . . . . . . . . . 78
5.3 Rotor speed and tip-speed ratio against wind speed . . . . . . . . . . . . . 79
5.4 Generator torque versus generator speed response of the variable-speed
controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.5 Turbine electrical power against wind speed . . . . . . . . . . . . . . . . . 81
5.6 Representative composite blade cross-section . . . . . . . . . . . . . . . . 82
5.7 Planform view of the turbine blade configuration . . . . . . . . . . . . . . 82
5.8 Comparison between the NREL 5-MW baseline turbine’s root airfoil and the
parameterized reference one . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.9 Comparison between the NREL 5-MW baseline turbine’s midspan airfoil
and the parameterized reference one . . . . . . . . . . . . . . . . . . . . . 89
5.10 Comparison between the NREL 5-MW baseline turbine’s tip airfoil and the
parameterized reference one . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.11 Comparison between the reference turbine’s chord and twist distributions
and the NREL offshore 5-MW baseline wind turbine’s ones . . . . . . . . . 91
5.12 Rotor speed and tip-speed ratio as a function of wind speed . . . . . . . . . 93
5.13 Electrical power as a function of wind speed . . . . . . . . . . . . . . . . . 94

6.1 Total blade clearance in unloaded conditions . . . . . . . . . . . . . . . . . 103

6.2 FAST’s coned coordinate system . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Comparison of the base airfoil shapes of the reference, deterministic and
robust designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.4 Comparison of lift and drag coefficients of the reference, deterministic and
robust designs’ base airfoils . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.5 Comparison of chord and twist distributions of the reference, deterministic
and robust designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.6 Comparison of rotor speed and tip-speed ratio against wind speed of the
reference, deterministic and robust designs . . . . . . . . . . . . . . . . . . 114
6.7 Electrical power as a function of wind speed of the reference, deterministic
and robust designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.8 Comparison of the angle of attack and tangential force distributions along
the blade span of the reference, deterministic and robust designs . . . . . . 116
6.9 Displacement and applied forces of the reference blade . . . . . . . . . . . 118
6.10 Displacement and applied forces of the deterministic blade . . . . . . . . . 118
6.11 Displacement and applied forces of the robust blade . . . . . . . . . . . . . 119
6.12 Laminate normal stress of the reference blade . . . . . . . . . . . . . . . . 119
List of figures xiii

6.13 Laminate normal stress of the deterministic blade . . . . . . . . . . . . . . 120

6.14 Laminate normal stress of the robust blade . . . . . . . . . . . . . . . . . . 120
6.15 Buckling criteria of the reference blade . . . . . . . . . . . . . . . . . . . . 121
6.16 Buckling criteria of the deterministic blade . . . . . . . . . . . . . . . . . . 121
6.17 Buckling criteria of the robust blade . . . . . . . . . . . . . . . . . . . . . 122
 Acknowledgements

I would like to extend my gratitude to Dr. M. Sergio Campobasso for his valuable support and
supervision over the past few years. I would also like to extend my thanks to Dr. Edmondo
Minisci for his kind help and encouragement.
This project has been carried out with the support of the wind turbine manufacturer
Gaia-Wind and the Energy Technology Partnership, which are hereby acknowledged.
 Declaration

Part of the work presented in this thesis has been published in the following articles:

M. Caboni, M. S. Campobasso and E. Minisci. “Wind Turbine Design Optimization

Under Environmental Uncertainty”. Journal of Engineering for Gas Turbines and Power.
DOI: 10.1115/1.4032665.

M. S. Campobasso, E. Minisci and M. Caboni. “Aerodynamic Design Optimization of

Wind Turbine Rotors Under Geometric Uncertainty”. Wind Energy. DOI: 10.1002/we.1820.

M. Caboni, E. Minisci, and M. S. Campobasso, 2015. “Robust Aerodynamic Design

Optimization of Horizontal Axis Wind Turbine Rotors”. In Advances in Evolutionary and
Deterministic Methods for Design, Optimization and Control in Engineering and Sciences,
D. Greiner, B. Galván, J. Periaux, N. Gauger, K. Giannakoglou, and G. Winter, eds., Vol. 36
of Computational Methods in Applied Sciences. Springer. ISBN 978-3-319-11540-5.

M. Caboni, M. S. Campobasso and E. Minisci. “Wind Turbine Design Optimization

Under Environmental Uncertainty”. ASME paper GT2015-42674, presented at the ASME
Turbo Expo 2015: Turbine Technical Conference and Exposition, 15-19 June 2015, Montréal,
Canada.

M. S. Campobasso, M. Yan, J. Drofelnik, A. Piskopakis and M. Caboni. “Compressible

Reynolds-Averaged Navier-Stokes Analysis of Wind Turbine Turbulent Flows Using a Fully
Coupled Low-speed Preconditioned Multigrid Solver”. ASME paper GT2014-25562, pre-
sented at the ASME Turbo Expo 2014: Turbine Technical Conference and Exposition, 16-20
June 2014, Düsseldorf, Germany.

I declare that this thesis is the result of my own work and has not been submitted for any
other degree at the University of Glasgow or any other institution.
xviii List of figures

Marco Caboni
May 2016
Nomenclature

Greek Symbols

α Angle of attack

αp Power law profile coefficient

αw Weighted sum method weight

δ Out-of-plane tip deflection

δf Partial safety factors for loads

δm Partial safety factors for materials

δn Partial safety factor for consequence of failure and component classes

δT Total safety factor

κi Joint PDFs of the uncertainty variables

λ Tip-speed ratio

λr Local speed ratio

µLCOE Mean of LCOE

µX Mean of random variable X

νi Joint PDFs of the uncertainty parameters

ν12 Poisson ration

Ω Angular velocity of the rotor

ω Rotor speed
xx Nomenclature

ω1 Blade first natural frequency at a rotor speed of 12.1 rpm

ωg Generator speed

ϕ Angle of relative wind

ρ Density

σ1 Turbulence standard deviation

σC Compressive stress

σLCOE Standard deviation of LCOE

σr Local solidity

σT Tensile stress

σX2 Variance of random variable X

τc Generator torque

θp Section pitch angle

θ p,0 Blade pitch angle

θT Section twist angle

Other Symbols

a Axial induction factor

a′ Circumferential induction factor

askew Axial induction factor corrected for skewed wake

b Design variables

bL Design variable lower bound

bU Design variable upper bound

c Airfoil chord length

C0 Position continuity

C1 Tangent continuity
Nomenclature xxi

C2 Curvature continuity

CD Drag coefficient

CL Lift coefficient

CP Local power coefficient

CQ Local torque coefficient

CT Local thrust coefficient

cW Weibull distribution scale parameter

D Rotor diameter

dA Area of annular element

dFD Local drag force

dFL Local lift force

dP Local power

dQ Local torque

dr Radial width of annular element

D(RootMxc1) Lifetime damage caused by in-plane moment at the blade root

D(RootMyc1) Lifetime damage caused by out-of-plane moment at the blade root

dT Local thrust

E11 Principal Young’s modulus

E22 Lateral Young’s modulus

F Probabilistic objective function

f Objective function

fdet Deterministic objective function

Fhub Hub loss correction factor

FPr Prandtl’s correction factor

xxii Nomenclature

frob Robust objective function

Ftip Tip loss correction factor

G12 Shear modulus

Gj Probabilistic constraints

gj Constraints

I Turbulence intensity

Iref IEC standard’s hub height reference turbulence intensity

K Torque control parameter

Kskew Function of the skew angle

kW Weibull distribution shape parameter

Nb Number of blades

p Design parameters

Pe Turbine electrical power

ψ Azimuth angle

pX PDF of a generic variable X

R Tip radius

r Mean radius of annular element

Re Reynolds number

Rhub Hub radius

s Thickness parameter

U Freestream wind velocity

u Wind speed

ū Mean value of the wind speed distribution

ūL Lower bound of the mean value of the wind speed distribution
Nomenclature xxiii

Urel Relative wind velocity

ūU Upper bound of the mean value of the wind speed distribution

Ve50 Extreme wind speed with a recurrence period of 50 years

Vhub Hub height wind speed

Vref IEC standard’s hub height reference wind speed

wb Blade mass

z Generic uncertainties associated with the design variables and parameters

Z Vertical distance from the ground

zb Uncertainties associated with the design variables

zp Uncertainties associated with the design parameters

Acronyms / Abbreviations

1D One-dimensional

2D Two-dimensional

3D Three-dimensional

AEP Annual energy production

AoA Angle of attack

AR Blade aspect ratio

BEM Blade element momentum

BOS Balance of station

CFD Computational fluid dynamics

CLT Classical lamination theory

DE Differential evolution

DLC Design load case

EDA Estimation of distribution algorithm

xxiv Nomenclature

EW M Extreme wind speed model

FEM Finite element method

FIO Fully integrated optimization

FRC Fixed charge rate

GA Genetic algorithm

HAW T Horizontal axis wind turbine

HSS High-speed shaft

ICC Initial capital cost

IDEA Inflationary differential evolution algorithm

IEC International Electrotechnical Commission

LC Lease cost

LCOE Levelized cost of energy

LEP Leading edge panel

LRC Levelized replacement/overhaul cost

MBH Monotonic basin hopping

MDO Multidisciplinary design optimization

MOPED Multi-objective parzen-based estimation of distribution

NS Navier-Stokes

NSGA − II Non-dominated sorting genetic algorithm II

NT M Normal turbulence model

O&M Operations and maintenance

PDF Probability density function

PPI Producer price index

QoI Quantity of interest

Nomenclature xxv

RDO Robust design optimization

RMSE Root-mean-square error

TCC Turbine capital cost

T EP Trailing edge panel

URQ Univariate reduced quadrature

Chapter 1

Introduction

By the late 20th century, wind power had become one of the most promising alternative
sources of energy worldwide, achieving a rapid global growth in installed capacity. In
the foreseeable future, as shown in Fig. 1.1, the global wind power installed capacity is
expected to continue to grow at an average annual rate of about 8%, bringing the worldwide
installed capacity up to 600 GW by 2018 [1]. One of the main challenges associated with the

Fig. 1.1 Forecast of worldwide wind power installed capacity. Image reproduced from [1].

advancement of wind energy technology is making wind energy economically competitive

with the conventional sources of energy, such as oil, gas and coal. This is normally pursued
by lowering the levelized cost of energy (LCOE). This figure of merit can be roughly defined
as the ratio of the capital and the operations and maintenance (O&M) costs of an energy
2 Introduction

system to the energy capture of the energy system over its lifetime. Accordingly, the LCOE
is the minimum price at which energy must be sold for an energy project to break even. The
attribute “levelized” in the definition of the cost of energy refers to the fact that the LCOE
characterizes the cost of energy of a system by means of a single figure of merit, indeed
distributed (or levelized) over the life of the energy system. As depicted in Fig. 1.2, in order
to decrease the LCOE a general trend in wind turbine industry has been to increase the rotor
diameter, therefore increasing the installation height (or hub height) and turbine rated power.
Indeed, larger turbines can extract more energy form the wind, and although both capital and

Fig. 1.2 Upscaling history of wind turbine size. The picture shows the trend over the past
three decades in rotor diameter, installation height and turbine rated power. Image reproduced
from [2].

O&M costs increase, the overall LCOE decreases. During the past few years, offshore wind
turbines with rated power up to 8 MW have been developed and installed.
The further increase of wind turbine size represents one of the directions for the future
development of wind energy technology [3]. This however will inevitably lead to new
challenges, and new solutions are therefore required to tackle them. In fact, as demonstrated
by Ashuri [3], the upscaling of large offshore wind turbines, using the current dominant
concept (i.e., a three-blade, upwind, variable speed, pitch regulated wind turbine installed
on a monopile), presents significative technical and economical challenges for machines
approaching 10 MW. Indeed, for very large turbines, the extreme and fatigue loads, that
the structural components experience throughout their lifespan, become more severe. This
leads to a considerable weight growth associated with the blades and other components [4],
posing major challenges to the technical feasibility and economical profitability of wind
1.1 Multidisciplinary design optimization 3

turbines. In other words, we cannot keep designing profitable large wind turbines just by
upscaling the current concepts without introducing significant advancements in their design.
Therefore, in recent years there has been intensive research into the design of more effective
wind turbines, focusing on aspects beyond the mere upscaling of current wind turbines.
To this aim, advanced analysis and design methodologies have been devised, providing
better understanding of the operating behavior of wind turbines, and enabling designers to
develop new design strategies. In this research framework, the present thesis focuses on
multidisciplinary design optimization (MDO), which will be introduced and reviewed in
Sect. 1.1.

1.1 Multidisciplinary design optimization

To study the effect of changing the turbine size on the turbine’s components and charac-
teristics, two methods have been traditionally used. These methods, denoted by classical
upscaling methods, are indeed based on two approaches, namely on linear scaling laws [3, 5]
and existing data trends [3, 5]. The first approach is based on the assumption that all geomet-
rical parameters of the turbine scale linearly with the rotor diameter. This method, however,
fails to characterize accurately the technical characteristics and economical performances for
large scales. Therefore, it is commonly used only in conceptual design phases. In the second
approach, the correlations between the rotor diameter and other parameters are determined by
interpolating over real data, available for existing wind turbines. However, when designing
turbines larger than the existing ones, and therefore real data are not available, an extrap-
olation is needed, therefore introducing important uncertainties. Considering these facts,
classical upscaling methods are not well suited for the design of large scale wind turbines,
and therefore they cannot be used to achieve the aforementioned advancements needed to
make large wind turbines effective and profitable. Therefore, the design of large scale wind
turbines necessitates the use of alternative methods. These methods should be able to evaluate
the relevant design constraints, investigating the results of design modifications in terms of
cost of energy.
In general, the situation in which a designer need to improve the performance of a given
system is referred to as design problem. In order to solve design problems, several approaches
can be followed. Factors like the quality of the solution, time requirements and acceptance
of the solution represent different criteria to assess the design approaches. However, good
design approaches need to have in common the following characteristics:
4 Introduction

• A design approach should incorporate a mechanism for directing the search performed
by the designer (or a design solver). The search mechanism should reduce quickly the
design space, eliminating non feasible solutions, and leading to the solution.

• A design approach should lead to high quality solutions in a short time. This character-
istic is particularly important when dealing with problem involving a large number of
design variables and different disciplines.

• A design approach should incorporate an adequate amount of practical knowledge in

order to obtain sound solutions.

• A design approach should overcome the designers’ psychological inertia, preventing

them to came up with innovative solutions.

The design of wind turbines is a multidisciplinary process, integrating aerodynamic,

structural, environmental, manufacturing, transportability and cost considerations. The
traditional approach to wind turbine design is based on a manual process, relying on the
knowledge of the designer, plus intuition and trial-and-error. Such approaches are more
efficient for relatively simple design problems. For more complex design problems, involving
a high number of design variables, trial-and-error methods present significant drawbacks.
For example, a trial-and-error approach is not time effective, as the number of successful
trials (that concretely lead to design improvements) per unit of time is low. Other major
disadvantages of such methods are that they are not able to define a search direction in which
the optimum solution might be found, and also they do not incorporate any mechanism to
systematically explore the design space. The use of MDO [6] constitutes a promising way to
overcome the inherent limitations of traditional design approaches and improve the design of
wind turbines, therefore enhancing their profitability. MDO is a field of engineering that uses
mathematical optimization techniques to solve design problems involving a certain number
of disciplines. MDO techniques are able to directly and systematically search through the
design space for the optimum design, automatically fulfilling specified constraints. This
potentially allows designers to come up with configurations beyond their experience and
intuition, enabling them to develop new design concepts.

1.1.1 Wind turbine MDO literature review

The aim of wind turbine MDO is to find the best design of a wind turbine by optimizing a set of
parameters, denoted by design variables, without violating some constraints. Design variables
and constraints represent physical characteristics of wind energy systems, and may be directly
related with different disciplines, such as aerodynamics, solid mechanics and control theory.
1.1 Multidisciplinary design optimization 5

A review of the design constraints used in wind turbine MDO will be presented in Sect. 1.3.
Design variables commonly considered in wind turbine MDO applications include: twist
and chord distributions along the blade span [7–20], blade pitch angle [7, 20–22], rotor
speed [7, 8, 10, 16, 17, 19–25], rotor diameter [7, 16, 17, 21–25], rated power [17, 21–23, 25],
hub height [16, 21–23, 25], blade thickness to chord ratio [7, 10, 11] and the thickness of
the blade internal structural layup [7, 10, 14, 16, 17, 19, 25–27]. Lately, a few wind turbine
MDO studies have also focused on the optimization of the airfoil shapes [12, 15, 18, 19].
This particular application will be discussed in Sect. 1.4. The design optimization of wind
turbines is generally carried out through an iterative procedure, whereby a figure of merit,
denoted by objective function, is optimized by varying the design variables. Wind turbine
MDO problems have been formulated in terms of different objective functions, including:
rotor power [9, 12, 24], wind turbine annual energy production (AEP) [8, 17, 20], blade
mass [26], the ratio of turbine mass to AEP [14, 17, 19] and LCOE [7, 8, 10, 11, 14–18, 21–
23, 25]. Ning et al. [17] investigated the influence of the objective function choice on the
design optimization of a multi-megawatt horizontal axis wind turbine (HAWT). Considering
the LCOE minimization as the reference metric in wind turbine design optimization, these
authors concluded that maximizing AEP leads to suboptimal solutions, while minimizing
the ratio of turbine mass to AEP can be effective only for fixed rotor diameter designs.
The iterative design process, by which a high number of wind turbine configurations are
generated, is regulated by an optimization algorithm. Optimization algorithms encompass
derivative-based algorithms, as well as genetic and pattern search algorithms. A review of
these methods is included in Sect. 3.5. Both gradient-based algorithms [7, 10, 11, 14, 16–
18, 21, 22, 24] and evolutionary algorithms [8, 9, 12, 13, 15, 19, 20, 23, 25, 26] have been
used in wind turbine MDO. At each iteration, the objective function and constraints of the
optimized configurations are typically evaluated by means of a computational module, which
must be able to rapidly and accurately perform the analysis of a given wind turbine. A
complete wind turbine analysis involves determining the interactions between aerodynamics,
structure and control. The interaction between these disciplines is commonly denoted by
the term “aeroservoelasticity”. In the literature, several MDO methods (or architectures)
have been developed and presented [6]. Commonly, wind turbine MDO is performed by
using a simple sequential approach, in which the different disciplines are treated in cascade,
following successive optimization stages. For example, a simple sequential approach for the
MDO of wind turbines is to firstly optimize the aerodynamic shape of the rotor blades taking
into account no or only a limited number of structural considerations. Subsequently, once the
geometric shape of the blades is determined, and hence the aerodynamic loads on the rotor are
known, a structural optimization is performed to determine the internal structural layup of the
6 Introduction

blades. Clearly, this approach neglects the coupling between disciplines (i.e., aerodynamics
and structure), leading to suboptimal solutions. A more complex and effective MDO method
is represented by the fully integrated optimization (FIO) architecture [6]. In this approach, an
optimizer is directly coupled with a multidisciplinary analysis model, allowing the different
disciplines to interact with each other concurrently during the optimization process.
In recent years, several studies have been devoted to the MDO of wind turbines, encom-
passing a wide variety of approaches and techniques. Fuglsang and Madsen [7] presented a
MDO method for the aeroelastic design of HAWT rotors. Using a gradient-based approach,
these authors optimized the design of a 1.5-MW stall-regulated rotor, minimizing LCOE,
while enforcing multiple constraints on the ultimate1 and fatigue loads and the aerodynamic
noise emission. Design variables included the blade geometric shape, rotor regulation and the
aerodynamic characteristics of the airfoils (i.e., lift and drag coefficients as a function of the
incidence). LCOE was reduced by 3.5% and 7% respectively without and with optimization
of the aerodynamic characteristics of the airfoils. This paper showed the potential of the
integration of airfoil optimization within the rotor optimization. Fuglsang et al. [21, 22]
investigated the site-specific design optimization of wind turbines, coupling a gradient-based
optimization algorithm with an aeroelastic turbine model. These authors optimized the design
of a stall-regulated wind turbine for the minimum LCOE across different installation site
conditions, incorporating detailed wind climate information. Design variables included the
hub height, rated power and rotor speed and diameter. The optimizations perfomed by these
authors showed that there are significant differences in the optimum design obtained for
normal flat terrain and offshore, and that cost of energy for offshore wind turbines can be
significantly reduced compared with today’s wind turbines. Benini and Toffolo [8] used a
global multi-objective evolution-based search method to optimize HAWT conceptual designs,
investigating the choice of fundamental HAWT design parameters, such as its rotor diameter,
on the economy of whole wind farms. The core results showed that the minimization of
LCOE requires larger HAWTs having high AEP, but low blade loads and weights. As oppo-
site, the maximization of AEP density requires smaller HAWTs having low AEP but high
blade loads which result in expensive designs. The obtained Pareto front, however, allows us
to determine the limits of improvement for the AEP density and the corresponding increase
in LCOE. Kenway and Martins [10] used a gradient-based optimizer to design a 5 kW wind
turbine rotor. These authors considered seven groups of design variables, including: blade
chord and twist distributions, spar thickness, spar location, spar length, airfoil thickness, and
rotation speed. Subject to a set of structural constraints, the objective was to minimize LCOE
by keeping constant wind turbine total costs and maximizing AEP. The framework presented

1 Ultimate limit state generally corresponds to maximum load carrying capacity [28].
1.1 Multidisciplinary design optimization 7

by these authors showed the ability to simulate the average power expected from specific
localized wind distributions, allowing one for detailed site-specific optimization. Xudong et.
al. [11] developed a design tool based on an aeroelastic model coupled with a gradient-based
algorithm which was used to minimize LCOE of three wind turbine rotors of different sizes.
The design variables included the blade chord and twist distributions and the blade thickness
to chord ratio, while the constraints encompassed the rotor torque and thrust. This work
demonstrate that optimization tools in all cases represent a valuable tool for the design of
wind turbines to achieve the most efficient design. Maki et. al. [25] optimized the the rotor
diameter, the rotational speed, the maximum rated power, the hub height, the structural
characteristics of the blade, and the geometric characteristics of the blade (distribution of
thickness, twist angle, and chord) of a HAWT by using a multi-level system optimization
algorithm. This approach allowed these authors to coordinate and execute a network of two
single-disciplinary optimizations, namely the maximization of AEP, and the minimization
of the turbine blade moment at the root of the blade. Each single-disciplinary optimization
was performed by means of a genetic algorithm (GA), and LCOE was considered as the
top-level objective function. Bottasso et al. [14] described a procedure for the MDO of wind
turbines, where the blade twist and chord distributions and internal structure are optimized
by maximizing AEP and minimizing the rotor blade weight, subject to a set of structural
constraints. These authors devised a nested optimization approach solved by two consecutive
optimizations. The first optimization maximizes AEP assuming frozen structural parameters,
while the second one minimizes the blade weight by using the optimal aerodynamic ones
obtained through the first optimization. This procedure was demonstrated on the aeroelastic
optimization of two multi-megawatt HAWTs. Using a GA, Vesel and McNamara [15] opti-
mized a multi-megawatt turbine for minimum LCOE, integrating the rotor aerodynamics and
the structural bend-twist coupling behavior of the blades. Design variables included airfoil
shapes, chord and twist distributions, and the degree of the blade bend-twist coupling. LCOE
of the optimized turbine was decreased by over 6% than that of a baseline turbine. Ashuri et
al. [16] defined the minimum number of structural constraints that should be considered to
obtain a practical design. Using a gradient-based algorithm, these authors optimized the rotor
and tower of a wind turbine for minimum LCOE. Blade design variables encompassed chord
and twist distribution, blade length, rated rotor speed and structural thicknesses along the
span. The use of this methodology contributed to 2.3% reduction in the LCOE compared to
a baseline turbine. The results showed a significant improvement in the quality of the design
process by means of a realistic assessment of the LCOE and constraints, while keeping
the coupling of the disciplines, and by using numerical optimization. Bottasso et al. [18]
optimized concurrently the airfoil shapes and the chord and twist spanwise distributions
8 Introduction

of a 2-MW wind turbine blade subject to a set of structural constraints. A gradient-based

algorithm was used to solve the optimization problem, achieving a 6% reduction in LCOE
with respect to an initial reference turbine. This work highlights the improved optimization
capability allowed by the simultaneous design of the blade and of its airfoils, thereby achiev-
ing a true 3D optimization, in which aerodynamic and structure can interact simultaneously
in the course of the optimization.

1.2 Robust design optimization

In the literature, wind turbine design optimization is typically performed deterministically.
Deterministic optimizations do not consider any degree of randomness affecting the inputs of
the system under consideration, and result, therefore, in an unique set of outputs. However,
given the stochastic nature of the wind and the uncertainties associated, for instance, with
wind turbine operating conditions or geometric tolerances, deterministically optimized
designs may be inefficient. In other words, the performance of deterministically designed
wind turbines is likely to deteriorate when they operate in the presence of uncertainty, that
leads them to work under off-design conditions. Therefore, one of the ways to further improve
the design of modern wind turbines is to take into account the aforementioned sources
of uncertainty in the optimization process, achieving robust configurations with minimal
performance sensitivity to factors causing variability. Design optimization in the presence
of uncertainty is denoted by robust design optimization (RDO) [29]. The main feature of
RDO is that the objective function and constraints are treated probabilistically. Three basic
steps are required to perform the RDO of a given system. First of all, the uncertainties
affecting the physical inputs of the system need to be characterized probabilistically. This
is normally accomplished by defining each uncertain variable by means of a probability
density function (PDF). The second step consists of propagating the input uncertainties
throughout the computational analysis system in order to characterize probabilistically the
system outputs (i.e., objective function and constraints) by means of a PDF. The probabilistic
definition of the objective function and constraints is often expressed by their mean and the
variance2 . Uncertainty propagation is a computationally intensive step, involving a high
number of system analyses. A variety of numerical methods have been developed to carry
out uncertainty propagation, from sampling based approaches (e.g. Monte Carlo) to more
sophisticated stochastic spectral Galerkin approaches [31]. The third and last step consists
2 Variance is a measurement of the spread between numbers in a data set. The variance measures how far
each number in the set is from the mean. Variance is calculated by taking the differences between each number
in the set and the mean, squaring the differences (to make them positive) and dividing the sum of the squares by
the number of values in the set [30]. The square root of the variance is called the standard deviation.
1.2 Robust design optimization 9

of optimizing the probabilistic objective function subject to the probabilistic constraints.

When the PDFs of the objective function and constraints are expressed as their mean and
variance, the optimization is achieved by concurrently optimizing the mean of the objective
function and minimizing its variance, satisfying all constraints within a certain range of
variability. Despite the significant impact it may have on wind turbine performance, the RDO
of wind turbines has so far received little attention. Petrone et al. [12] optimized the airfoil
shapes and blade chord and twist distributions of a stall-regulated rotor for maximum mean
power coefficient and minimum acoustic emissions, considering the uncertainty on laminar-
to-turbulent transition caused by insect contamination. Uncertainty was propagated by means
of the stochastic simplex collocation method, and the optimization was carried out using a
multi-objective GA. The design obtained with this probabilistic procedure appeared to be
less sensitive to the presence of uncertainty than its deterministic counterpart. Campobasso
et al. [20] developed a robust optimization strategy for the aerodynamic design of HAWT
rotors accounting for the uncertainty of the blade geometry caused by manufacturing and
assembly errors. Their numerical studies aimed at maximizing the expectation of AEP
and minimizing its standard deviation by optimizing the chord and twist distributions, and
the rotor angular speed. Uncertainty was propagated with Monte Carlo sampling and
the univariate reduced quadrature (URQ) approach [32], and a two-stage multi-objective
evolution-based optimization strategy was used. The comparative analysis of a rotor design
obtained by considering the stochastic geometry errors and a rotor design obtained by
neglecting all uncertainties showed that AEP standard deviation of the former rotor was
less than 40% than that of the latter. The analyses indicated that a lower sensitivity of AEP
to rotor geometry errors can be achieved by lowering rotational speeds and compensating
for the reduction of power due to lower rotational velocities by shifting upwards the radial
profiles of CL . This increment of the aerodynamic loading is achieved by increasing the
angle of attack (to a large extent through lower values of the sectional pitch) and the chord of
the blade over most part of the blade length. Caboni et al. [19] generalized the probabilistic
design approach of [20] by including also the airfoil geometry in the probabilistic design
environment, and applied this technology to maximize the expectation and minimize the
standard deviation of AEP of a multi-megawatt rotor subject to geometric uncertainty. Both
robust optimization processes performed in this paper confirm that the search for the lower
sensitivity to geometry errors is pursued by adopting lower rotational speeds, and therefore
the robustness is actually obtained by moving to a range of higher values of the angle of
attack where the slope of the angle of attack/lift coefficient curve is lower than for lower
values of the angle of attack. Using a gradient-based optimizer, Ning et al. [17] designed a
wind turbine rotor to be used across diverse sites characterized by a different mean of the
10 Introduction

wind distribution. Subject to a set of structural constraints, blade chord, twist and spar cap
thickness distributions, rotor speed, diameter and rated power were optimized to minimize
the expected value of LCOE. The robust design achieved a 1.2% lower average LCOE than
that of a deterministically designed rotor. This paper showed the importance of combining
optimization with uncertainty quantification, along with a better understanding of the nature
of the uncertainties.

1.3 Structural constraints

To ensure structural integrity, wind turbines must satisfy an extensive set of structural require-
ments, relating to material ultimate and fatigue strength, deflections and structural stability.
Several national and international organizations, such as the International Electrotechnical
Commission (IEC), Germanischer Lloyd and Det Norske Veritas, prepare and publish stan-
dards covering the various aspects of the wind turbine design process. More specifically,
these standards prescribe a large number of combinations of environmental and operating
conditions under which wind turbine components need to be verified structurally. These
design combinations are normally referred to as design cases. In addition, the standards
define the procedures to be followed when calculating ultimate and fatigue loads, suggesting
the application of safety factors to cover uncertainties and variabilities of various natures
(such as in loads and materials) in the design procedure. Specifically, the structural analysis
of wind turbine components starts from predicting the loads that the wind turbine components
experience during their lifespan, considering a combination of different operating conditions
(e.g., normal, fault, and parked) and environment conditions (e.g., normal and extreme wind).
For each design case considered, load predictions are carried out by means of aeroelastic
codes, which are able to simulate the complex dynamic interaction between wind turbine aero-
dynamics and structure under various conditions. Subsequently, these loads are augmented
by means of safety factors in order to consider, for instance, possible errors in the aeroelastic
model, or the uncertainty related with the strength of constituent materials. A stress-strain
analysis is then required to verify the structural integrity of each component subject to the
calculated loads. Generally, this analysis step involves the calculation of ultimate stress,
fatigue damage, component deflection and structural stability, such as buckling. Structural
integrity is ensured when these criteria fulfill the structural strength requirements of materials
and components.
In the context of wind turbine optimization, all configurations generated during the
iterative design procedure should be assessed structurally, to make sure that they satisfy
the design requirements prescribed by the standards. In practice, during the optimization,
1.4 Airfoil optimization 11

the above-mentioned structural criteria, treated as design constraints, must be kept below
their critical values, for the full set of design cases. Due to the large number of aeroelastic
simulations required, this is a computationally expensive step, and in most cases impracticable.
Fortunately, a limited number of design cases have been proven to be the most likely design
driver for majority of turbine blades [16, 33, 34]. In particular, the design case in which the
wind turbine rotor is parked and subject to extreme wind conditions is one of the most critical
cases for ultimate stress, deflection and buckling. In the literature, only few HAWT design
optimization studies attempted to consider a sufficient number of design cases to obtain a
practical design [14, 16, 17], evaluating the turbine response with respect to ultimate strength,
fatigue failure, stability and critical deflections. In fact, wind turbine design optimization
is commonly subject only to a limited number of structural constraints, neglecting the
most relevant design cases. Typically, the structural verification in wind turbine design
optimization problems is carried out by using simple surrogates of more involved structural
analyses. However, this approach may neglect design cases or criteria that may predict
structural failure. In the majority of MDO studies reported in the literature [7–20], only
operational conditions are considered when evaluating structural criteria, neglecting extreme
conditions. Moreover, in these studies, a detailed stress-strain analysis is often replaced by
the analysis of simpler criteria, such as root bending moment [20], simplified function of
structural stresses [8] or rotor thrust [11].

1.4 Airfoil optimization

Another important limitation affecting most reported HAWT design optimization studies is
that the airfoil design is not handled within the rotor optimization, although such inclusion
plays a fundamental role in both aerodynamic and structural turbine performance. In fact, a
commonly employed design procedure is to optimize the rotor blades by preselecting a set of
existing available airfoils. This choice clearly limits the exploration of the design space, and
inefficiently considers the cross-coupling interactions between the blade aerodynamics and
the structural properties of the rotor. The interaction between wind turbine aerodynamics
and structure is tied to the structural sizing of all wind turbine components, and therefore the
design of the airfoil shapes affects not only the aerodynamic performance of the rotor, but
directly determines the overall efficiency and LCOE of wind turbines. Designing the airfoils
together with the rest of the blade therefore results in a wider exploration of the design space,
and allows the airfoils to adapt to different local aerodynamic and structural conditions. In
the course of an automated airfoil optimization process, the airfoil shapes are allowed to
change, and therefore, at each iteration the aerodynamic characteristics of the airfoils need to
12 Introduction

be determined. Therefore, the key issues that need to be addressed in design optimization of
airfoils concern the definition of a suitable airfoil shape parametrization and the calculation of
the airfoil aerodynamic performance. In a majority of airfoil design optimization applications,
panel codes and Navier-Stokes (NS) computational fluid dynamics (CFD) solvers are used
to evaluate airfoil aerodynamic performance. Airfoil parametrization is needed to relate the
airfoil shape to numerical parameters, which are handled by the optimizer in order to find
the optimum. To reduce the computational burden associated with the optimization process,
airfoil parametrization should have a limited number of degrees of freedom, yet it should
be able to describe any shape with good accuracy, exploring as many design alternatives
as possible. Common parameterization techniques devised to numerically represent airfoil
geometry for design optimization studies include: basis vector, domain element, partial
differential equation, discrete, polynomial and spline, CAD-based, analytical, and freeform
deformation. An exhaustive review of these methods can be found in [35]. The issue of
optimizing the airfoil shapes with the rest of the blade was recently investigated by Petrone
et. al. [12], Vesel and McNamara [15], Bottasso et al. [18] and Caboni et. al. [19]. These
authors used XFOIL to determine the airfoil aerodynamic characteristics, and parametrized
the airfoil shapes by means of B-splines [12] and Bézier curves [15, 18, 19].

1.5 Objectives, novelty and overview of the thesis

The main drive of the research work reported in this thesis was two-fold. On the one hand, it
aimed at developing and assessing a novel integrated multidisciplinary design framework
for the probabilistic aeroservoelastic design optimization of HAWT rotors. On the other
hand, it aimed at demonstrating the effectiveness of the developed framework by performing
the design optimization of a multi-megawatt HAWT rotor under environmental uncertainty.
More specifically, the main objectives associated with the development of the design system
were to:
• integrate, in a common computational framework, several simulation tools needed to
assess the aeroservoelastic behavior of wind turbines, and determine their economical
performance, and

• couple the multidisciplinary analysis module with an uncertainty propagation algorithm

and an optimizer, and

• define the design space by developing a suitable wind turbine parametrization.

The main objective associated with the application of the design system was to optimize the
aeroservoelastic design of 5-MW HAWT rotor so as to minimize mean and standard deviation
1.5 Objectives, novelty and overview of the thesis 13

of its LCOE in the presence of uncertain wind speed. This probabilistic design scenario was
described by Ning et al. [17]. The presented probabilistically optimal design was compared
to a reference turbine based on the NREL 5-MW turbine [36], and a deterministic design,
obtained without considering uncertainties in the optimization process.
The main novelty of this thesis is the presentation and demonstration of a fully integrated
computational framework for the robust aeroservoelastic design optimization of HAWT rotors.
The key feature of such a framework is that it concurrently optimizes airfoil shapes, external
blade geometry and internal blade structure. As typical multidisciplinary HAWT rotor
design systems optimize only blade chord and twist distributions, and internal blade structure
making use of pre-selected airfoils, the design framework presented herein enables a wider
exploration of the feasible design space, possibly leading to radically new and more efficient
designs. In the developed framework, spanwise distributions of blade chord and twist, airfoil
shapes, rotor speed and blade internal structural layup are optimized concurrently, enforcing
structural constraints on ultimate, fatigue and buckling stresses, tip deflection and blade
natural frequency, considering an extensive set of design cases. The design system is also
probabilistic in that it can account for several uncertainty sources, and the study presented
below focuses on probabilistic HAWT rotor design accounting for uncertain site-dependent
mean wind speed. To the best of the writer’s knowledge, there are no published reports
on fully automated wind turbine multidisciplinary design optimization that simultaneously
include realistic structural analysis/constraints based on three-dimensional (3D) models, the
fully 3D blade shape parametrization of the blade, the airfoil geometry, and consider the
stochastic variability of the wind frequency distribution.
The thesis is organized as follows. Chapter 2 presents an overview of the main topics and
issues associated with the multidisciplinary analysis and design optimization of wind turbine
rotors, providing details on the theoretical aspects and computational models commonly
used to characterize the aeroservoelastic behavior of wind turbine rotors, as well as their
economical performance. Advantages and disadvantages of the use of these methods in
wind turbine design optimization are discussed. Chapter 3 presents the classical approach
to the design optimization under uncertainty. The classical formulation of the RDO is
explained along with the common approaches followed to solve such problems. This chapter
also present an overview of the most popular methods for uncertainty propagation and
optimization algorithms. A detailed description of the developed multidisciplinary and robust
design optimization framework is reported in Chapter 4. This includes an integrated analysis
module, based on aerodynamic, aeroelastic, structural and cost analysis models, coupled with
an uncertainty propagation and optimization strategy. Chapter 5 deals with the development
of the wind turbine parametrization, and it illustrates its use for the definition of the reference
14 Introduction

5-MW HAWT rotor. In particular, this chapter aims to present the definition of the design
space developed for the optimizations. Chapter 6 focuses on the HAWT design optimization
under environmental uncertainty, formulating the probabilistic and deterministic optimization
problems, and explaining the objective function and the constraints used. In this chapter
the probabilistic and deterministic turbines are compared with the reference one. The main
conclusions of the thesis and future work are summarized in Chapter 7.
Chapter 2

Multidisciplinary analysis of horizontal

axis wind turbine rotors

Wind turbines are complex devices, and their overall characteristics result from the interaction
of different disciplines, including aerodynamics, solid mechanics, control theory and eco-
nomics. As already stated in Chapter 1, the design optimization of wind turbines is pursued by
optimizing a figure of merit, or objective function, while fulfilling a number of requirements,
or constraints. The objective function typically characterizes the economical performance
of wind turbines, while the constraints ensure the structural integrity of their components
throughout the expected turbine life. Both the objective function and constraints depend on
the coupling among the aerodynamic and structural characteristics of the rotor, as well as the
control strategy. This highlights the multidisciplinary nature of the design optimization of
wind turbines. Therefore, the central part of a design system for the optimization of wind
turbines should be constituted by an integrated multidisciplinary analysis module used to
evaluate wind turbine aeroservoelastic performance. The aim of this chapter is to provide an
overview of the topics which are fundamental to understanding the behavior of HAWTs. In
particular, this chapter discusses the theoretical approaches and the computational tools that
are commonly used to model wind turbines. For this purpose, three analysis areas have been
identified and presented individually, namely: wind turbine rotor aeroservoelasticity, airfoil
aerodynamics and cost.

2.1 Wind turbine rotor aeroservoelasticity

Aeroelasticity is a discipline that studies the interactions between the inertial, elastic, and
aerodynamic forces that occur when an elastic body is exposed to a fluid flow. The com-
16 Multidisciplinary analysis of horizontal axis wind turbine rotors

bination of aeroelasticity with control theory is known as aeroservoelasticity. In order to

determine wind turbine power production and loads acting on the blades, the aeroservoelastic
response of wind turbine rotors needs to be accurately assessed. The aeroservoelastic analysis
of a wind turbine rotor is achieved by coupling the aerodynamic analysis with the structural
one, integrating the control strategy needed to regulate the turbine operating conditions. For
this purpose, several aeroelastic codes have been developed and used in the wind turbine
industry, including FLEX4 [37], Bladed [38], Phatas [39], FAST [40] and HAWC2 [41].
These codes model the dynamic behavior of wind turbine rotors by coupling wind turbine
aerodynamic and structural response to wind-inflow conditions, including a controller which,
for a given rotor, defines the rotational speed at which the rotor operates at each wind speed.
Aerodynamic analysis of the rotor, normally embedded in the aeroservoelastic analysis, is
performed by means of an aerodynamic module as reported in Sect. 2.1.1. Aeroservoelastic
simulations normally provide time-series data of the aerodynamic loads, as well as loads and
deflections of the structural components of the wind turbine. The aeroservoelastic analysis of
a wind turbine rotor, carried out with an aeroelastic code, generally requires the definition
of the blade shape in terms of chord and twist distributions along the blade span, the aero-
dynamic characteristics of the blade airfoils, the mode shapes and the distributed structural
properties of the blades, including cross-coupled stiffness properties and inertia properties.
Distributed structural properties of the blades are commonly calculated be means of structural
models as explained in Sect. 2.1.2. The evaluation of the aerodynamic characteristics of the
airfoil blades is described later on in Sect. 2.2.

2.1.1 Wind turbine rotor aerodynamics

The aerodynamic behavior of a wind turbine determines the flow field around its rotor, and
therefore it is closely related with the overall turbine performance. A correct evaluation of
the rotor aerodynamics is needed to efficiently predict the interaction of the aerodynamic
forces with the elastic structure of the rotor blades, and ensure their structural integrity.
Aerodynamic analysis of wind turbine rotors can be performed either my means of blade
element momentum (BEM) theory [42] or NS CFD [43]. BEM theory is based on simple
models and can perform the aerodynamic analysis of a given rotor very quickly. Some
of the most commonly used BEM codes include AeroDyn [44] and CCBlade [45], both
developed by NREL. AeroDyn is a plug-in type code for interfacing with various NREL
dynamics programs such as FAST, while CCBlade is a stand-alone code. CFD codes solve
the NS equations numerically, incorporating several models to account for turbulence and
laminar-turbulent transition. Both research [46] and commercial [47] CFD codes are currently
available for the aerodynamic analysis of wind turbines. These codes are able to determine a
2.1 Wind turbine rotor aeroservoelasticity 17

detailed description of the flow field around the rotor and several rotor diameters upstream
and downstream of the rotor plane. Despite the rapid increase in computer power, calculation
times associated with CFD are still excessive for their use in design optimizations requiring
hundreds or thousands of rotor analyses. Therefore these methods are generally used for
the evaluation of the rotor design, rather than being part of an iterative design optimization
process. For this reason, BEM theory is the most common method used for the prediction of
the aerodynamic rotor performance in wind turbine rotor design. The following two sections
present respectively an overview of the BEM theory and the corrections needed to overcome
its limitations.

Blade element momentum theory

The steady state aerodynamics of wind turbines is commonly analyzed by using momentum
and blade element theory. Momentum theory refers to a control volume analysis of the
forces acting on the blade based on the conservation of linear and angular momentum. Blade
element theory refers to an analysis of forces at a blade section, as a function of blade
geometry. According to the blade element theory, the forces on the blades of a wind turbine
are expressed as a function of lift and drag coefficients and the angle of attack (AoA). The
results of these approaches can be combined into what is known as strip theory or BEM theory.
The BEM theory is based on the subdivision of the rotor disk into concentric rings of radial
width dr and mean radius r. Each ring intersects the rotor blades forming blade elements
or strips. The flow data and the aerodynamic forces acting on each strip are determined by
solving two equations, obtained by combining linear and angular momentum conservation
and classic lift and drag theory. One equation results from equating the ring axial thrust
determined with the one-dimensional (1D) conservation of the linear momentum to the
axial thrust computed with the lift and drag forces acting on the blade strips intersected by
the ring. The other equation results from equating the ring torque determined with the 1D
conservation of the angular momentum to the torque produced by the lift and drag forces
acting on the intersected strips. The main geometric and aerodynamic parameters of a generic
strip are depicted in Fig. 2.1, in which the section lift and drag forces are denoted by dFL
and dFD respectively. Denoting by dT the thrust acting on a ring, the local thrust coefficient
is CT = dT /(0.5ρU 2 dA), where dA = 2πrdr is the area of the ring, and ρ and U are the
freestream density and velocity respectively. The local thrust coefficient computed using the
conservation of linear momentum is:

CT = 4a(1 − a) (2.1)
18 Multidisciplinary analysis of horizontal axis wind turbine rotors

where a is the axial induction factor. The local thrust coefficient computed using lift and drag
theory is:
σr (1 − a)2
CT = (CL cos ϕ +CD sin ϕ) (2.2)
sin2 ϕ
where σr = (Nb c)/(2πr) is the local solidity, Nb is the number of blades, c is the airfoil
chord length, and CL and CD are the lift and drag coefficients respectively. The symbol ϕ
denotes the angle of the relative wind velocity vector Urel on the rotor plane. Its expression
is ϕ = arctan [(1 − a)/((1 + a′ )λr )], where a′ is the circumferential induction factor and
λr = Ωr/U is the local speed ratio, where Ω is the angular speed of the rotor. Urel is
expressed as Urel = U(1 − a)/ sin ϕ. Equating Eqs. 2.1 and 2.2 yields one equation in the two
unknowns a and a′ , since CL and CD are ultimately also functions of the induction factors.
In fact, these force coefficients can be obtained with panel or CFD codes (see Sect. 2.2 for
details) or experimental data as functions of the Reynolds number (Re), which depends on
Urel and the relative AoA α, the angle between the airfoil chord and Urel . As shown in
Fig. 2.1, α = ϕ − θ p , where θ p is the section pitch angle. This parameter depends only on
geometric features, and its expression is θ p = θ p,0 + θT , where θ p,0 is the pitch angle of the
blade and θT is the section twist angle. Denoting by dQ the torque acting on a rotor ring, the
local torque coefficient is CQ = dQ/(0.5ρU 2 rdA). The local torque coefficient computed
using the conservation of angular momentum is:

CQ = 4a′ (1 − a)λr (2.3)

The local torque coefficient computed using lift and drag theory is:

σr (1 − a)2
CQ = (CL sin ϕ −CD cos ϕ) (2.4)
sin2 ϕ

Equating Eqs. 2.3 and 2.4 yields another equation in the two unknowns a and a′ . The
nonlinear system resulting by equating the two expressions of CT and CQ for each strip need
to be solved with an iterative routine based for instance on Newton’s method or the method
of successive substitution. The two-dimensional (2D) CL and CD data are stored in tables as
functions of Re and α, and such data are computed in a pre-processing step. Once the flow
state of each strip is known, the elemental power dP can be computed. The nondimensional
local power coefficient CP = dP/(0.5ρU 3 dA) can be expressed as follows:

σr (1 − a)2 λr
CP = (CL sin ϕ −CD cos ϕ) (2.5)
sin2 ϕ
2.1 Wind turbine rotor aeroservoelasticity 19

Fig. 2.1 Geometric and aerodynamic parameters of a generic blade strip. Image reproduced
from [42].

The mechanical power of a given rotor corresponding to a particular value of U and Ω is

determined by integrating dP from the blade root to its tip.

BEM theory corrections

Due to its simplicity, the BEM theory has several limitations. First of all, BEM calculations
are static. This assumes that the flow field around the airfoils is always in equilibrium, and
the flow accelerates instantaneously to adjust to new inflow or turbine operating conditions.
In practice, however, the time taken by the flow-field to reach a steady-state can be relatively
long, and, as explained below, the unsteady aerodynamic effects can play an important role
in defining wind turbine operating conditions. Alternative methods based on the generalized
dynamic wake model [48] have been used to overcome this limitation. Another limitation is
tied to the fact that the BEM theory assumes that momentum is balanced in a plane parallel
to the rotor. In the presence of large blade deflections this assumption will lead to inaccurate
aerodynamic predictions. Moreover, BEM theory assumes the forces acting on the blade are
essentially 2D, neglecting the complex 3D phenomena occurring over the rotating blades.
20 Multidisciplinary analysis of horizontal axis wind turbine rotors

As discussed in Sect. 2.2, this limitation is circumvented by means of corrections directly

applied on the static force coefficients of the airfoils. Other limitations, described below,
come from the inability of BEM theory to model tip and hub losses, flows characterized by
high induction factors, and skewed inflow.

Tip and hub loss correction. As shown in Fig. 2.2, helical vortices are shed from the
blade tips into the wake. Tip vortices play an important role in defining the induced velocity
field around the rotor. The most common approach to include tip loss in the BEM theory is

Fig. 2.2 Tip vortex pattern. Image reproduced from [44].

the one developed by Prandtl [42]. This method accounts for tip loss by means of a correction
factor Ftip defined as follows:
     
2 Nb (R − r)
Ftip = arccos exp − (2.6)
π 2r sin ϕ

where R is the tip radius. When the tip correction is used, Ftip increases as the radial position
approaches the blade tip, ranging from zero near the root to unity at the tip. To account
for the vortices being shed at the blade hub, a correction for hub loss was developed. This
correction is based on a correction factor Fhub expressed as:
     
2 Nb (r − Rhub )
Fhub = arccos exp − (2.7)
π 2Rhub sin ϕ
2.1 Wind turbine rotor aeroservoelasticity 21

where Rhub is the hub radius. The combined effect of tip and hub losses is taken into account
by means of the Prandtl’s correction factor FPr defined as:

FPr = Ftip Fhub (2.8)

FPr is used to modify the momentum part of the BEM theory, replacing Eqs. 2.1 and 2.3 with
the following ones:
CT = 4FPr a(1 − a) (2.9)

CQ = 4FPr a′ (1 − a)λr (2.10)

Glauert correction. During normal operation, wind turbines typically work in the windmill
state, in which the axial induction factor ranges from 0 to 0.5. When wind turbines operate
at higher tip-speed ratios1 (for example during startup or shutdown), the rotor enters in the
so-called turbulent wake state [42], in which the axial induction factor is greater than 0.5.
For axial induction factors greater than 0.5, BEM theory is not longer valid as, according
to momentum theory, this operating state results when some of the flow in the far wake
starts to propagate upstream. Flow reversal is not physically possible, and what actually
happens is that the flow patterns through the wind turbine become much more complex than
those predicted by momentum theory. Above an axial induction factor of 0.5, measured
data indicate that thrust coefficient increases up to about 2 at an axial induction factor of
1. To compensate for this limitation Glauert [49] developed a correction to the rotor thrust
coefficient based on experimental measurements. As shown in Fig. 2.3, in the windmill
state, for axial induction factors up to 0.4, the mathematical relation between CT and a is
expressed by the classical momentum equation. For axial induction factors greater than 0.4,
the Glauert correction takes over, intersecting tangentially the classical momentum curve. A
numerical problem arises when the Prandtl’s correction factor FPr is included in the classical
momentum theory, as shown in Eq. 2.9. In this case, the application of the classical Glauert
formulation leads to a gap between the classical momentum curve and the empirical one. In
practice, the classical momentum curve and the Glauert curve do not intersect each other
anymore. This gap creates a numerical discontinuity when a computer is used to iterate
for the induction factor. To deal with this problem, Buhl [50] derived a modification of the

1 Tip-speed ratio is the ratio between the tangential speed of the tip of a blade and the actual velocity of the
wind.
22 Multidisciplinary analysis of horizontal axis wind turbine rotors

Fig. 2.3 Thrust coefficient versus axial induction factor. Image reproduced from [42].

Glauert correction including the Prandtl’s correction factor as follows:

   
8 40 50
CT = + 4FPr − a+ − 4FPr a2 (2.11)
9 9 9

Taking into account the value of FPr explicitly, Eq. 2.11 always guarantees continuity between
the classical momentum curve and the Glauert one.

Skewed wake correction. Wind turbines often operate with a non-zero yaw angle relative
to the wind inflow. This determines a skewed wake behind the rotor. BEM theory needs to
be corrected to account for such operating condition. Most of the available skewed wake
corrections are based on an equation developed by Glauert [51]. This equation corrects the
axial induction factor as follows:
h r i
askew = a 1 + Kskew cos(ψ) (2.12)
R
where Kskew is a function of the skew angle and ψ is the azimuth angle. In this case, since
the axial induction factor depends on the value of ψ, the BEM equations outlined above
need to be solved for each azimuth position. This correction has a limitation primarily due
to the fact that it assumes a cylindrical wake, which is true only for lightly loaded rotors.
Better predictions of the aerodynamics of wind turbines operating in yaw conditions can be
achieved by using alternative methods based on the generalized dynamic wake model [48].
2.1 Wind turbine rotor aeroservoelasticity 23

Unsteady aerodynamic effects. The turbulence associated with the wind and unsteady
aerodynamic effects causes rapid fluctuations in the aerodynamic forces acting over the
rotor blades, generating vibrations and important material fatigue. In particular, unsteady
aerodynamic effects, such as those related to the tower shadow, dynamic stall and dynamic
inflow (shortly described below), play a fundamental role on wind turbine operation, and
their modeling can improve the accuracy of wind turbine rotor aerodynamic analysis.

Tower shadow. The wind speed experiences a deficit behind the tower. In downwind
turbines, this causes a rapid drop in the power extracted by the rotor blades, and structural
vibrations. A model accounting for the tower influence has been developed by Bak et al. [52].
This method models the influence of the tower on the local velocity field at all points around
the tower. Fig. 2.4 shows the tower shadow model at a given point. According to this model,
the tower wake decays in strength and grows in width as the distance from the tower increases.

Fig. 2.4 Tower shadow model at a given point. In this picture, U∞ is the freestream wind
velocity, and d represents a characteristic length of the model. Image reproduced from [44].

Dynamic stall. When rapid changes in the AoA occur, for example when the rotor
blades of a downwind turbine encounter the tower wake or due to the effects of rotor
yaw, wind shear and turbulence, turbine blades can experience lift forces that are different
(normally larger [53]) than those expected in static conditions. This effect is tied to the blade
stall behavior, and it is normally referred to as dynamic stall. Dynamic stall is an unsteady
mechanism than can occur when the AoA of an airfoil increases rapidly from below to above
the static stall AoA. In this case, the flow over the airfoil can remain attached at angles of
attack above the angle at which steady-state flow separation normally occurs. As shown in
Fig. 2.5, under these circumstances, the airfoil can generate a higher lift coefficient than that
24 Multidisciplinary analysis of horizontal axis wind turbine rotors

it would generate in the static cases. In extreme cases, dynamic stall can increase the lift

Fig. 2.5 Dynamic stall behavior. Cl and α represent the lift coefficient and the AoA, respec-
tively. Image reproduced from [54].

coefficient by a factor of three [53]. The flow over the airfoil can then separate suddenly
with the result that the lift coefficient drops and the drag coefficient increases. The loads
experienced by a blade subject to dynamic stall can be large, causing significant fatigue
damage. Dynamic stall also causes sudden variations of the pitching moment, resulting in
important loads on the rolling bearings for the blade pitch motion. Several methods have
been developed to model dynamic stall, such as those of Gormont [55] and Beddoes [56].
Details on the formulation of the Beddoes’s model and its interaction within a BEM code
can be found in the AeroDyn’s theory manual [44]. Dynamic stall methods are not included
in the design framework reported below.

Dynamic inflow. Dynamic inflow is related to the flow field response to turbulence
and changes in rotor operation conditions (for example due to changes in the blade pitch
angle or rotor speed). According to steady state aerodynamics, these changes should result
in instantaneous changes in the flow field upstream and downstream of the rotor. However,
during rapid changes the flow field cannot respond quickly enough to instantly establish
steady state conditions. This results in aerodynamic conditions which may be different from
the expected ones. The time scale of dynamic flow effects is on the order of D/U, the ratio
2.1 Wind turbine rotor aeroservoelasticity 25

of the rotor diameter to the mean ambient flow velocity [42]. Therefore, the time scale of
these effects is of the order of about 10 seconds [57]. Phenomena occurring slower than this
can be treated using a steady state analysis. More details on dynamic inflow and its modeling
can be found in [57–59]. Dynamic inflow methods are not included in the design framework
reported below.

2.1.2 Structural analysis of the rotor blades

The design of wind turbine blades is an involved process mainly due to the complexity of
their aerodynamic shape and internal structural characteristics. Indeed, wind turbine blades
are made of composite laminates, constituted by anisotropic2 layup distributed non-uniformly
along the blade span. This means that the internal structural layup of composite laminates
varies across the blade span from root to tip. As mentioned above, the structural analysis of
rotor blades is primarily required to compute mode shape and distributed stiffness and inertial
properties needed by wind turbine aeroelastic codes. Moreover, the structural analysis of
the blades is needed to verify their structural integrity against ultimate and fatigue limits.
Ultimate load analysis refers to the assessment of material strength (through a stress-strain
analysis), blade tip deflection and structural stability (i.e., buckling), while fatigue load
analysis concerns fatigue strength. The extraction of the structural properties of the blade and
the stress-strain, blade tip deflection and buckling analyses are normally carried out using
structural codes. Fatigue calculations involve the use of specific algorithms to process load
histories in order to determine the damage accumulation. The initial part of this section deals
with different methods commonly used to perform the structural analysis of wind turbine
blades and fatigue calculations.
The realistic ultimate and fatigue loads, that turbine components are subjected to, need to
be accurately assessed. Such loads are normally generated by running a series of aeroelastic
simulations under different operating and environmental conditions, covering most of the sit-
uations that wind turbines likely experience during their lifetime. As mentioned in Chapter 1,
these conditions are prescribed by standards, such as the IEC standard [28]. Wind conditions
considered by these standards are normally fed into aeroelastic codes by means of wind input
files. The concluding part of this section will describe the structure of such files, and explain
how they can be generated.

2 Anisotropy is the property of being directionally dependent, as opposed to isotropy, which implies identical
properties in all directions.
26 Multidisciplinary analysis of horizontal axis wind turbine rotors

Methods for the structural analysis of wind turbine blades

Finite element method (FEM) codes, such as ANSYS [60], Abaqus [61], SolidWorks [62]
and NuMAD [63], can be used to perform the structural analysis of wind turbine blades,
accurately accounting for complex geometric shapes and composite structural layup. FEMs
rely on numerical approximation techniques that divide a component or structure into discrete
regions (the finite elements) and the response is described by a set of functions that represent
the displacements or stresses in those regions [64]. These models are able to accurately
provide the span-variant properties of the blades, and describe the strain-stress fields in detail.
The use of FEM techniques is however computationally expensive, requiring the generation
of a computational mesh, and complex post-processing. In the preliminary design stages,
where a huge number of different configurations may be evaluated, FEM approaches may
become impractical. Therefore the structural analysis in the aeroelastic design optimization
of wind turbine rotors generally relies on simpler and faster models. The rest of this section
will review some of these simplified structural models tailored towards composite rotor
blades, and featuring various solution techniques.
PreComp [65] is a popular NREL code developed to provide span-variant structural
properties for composite blades. These structural properties encompass: flap, lag (edgewise),
axial (with respect to the blade pitch axis), and torsion stiffnesses, as well as orientation of
principal axes, density, and moments of inertia. This code is based on the classical lamination
theory (CLT) with a shear-flow approach. Details on the CLT and the shear-flow theory can
be found respectively in [66, 67] and [68]. It should be noted that PreComp cannot be used
to perform stress-strain analysis. Therefore, in the framework of this research, alternative
structural codes, not based on FEMs, yet featuring stress-strain analysis capabilities, have
been reviewed.
Co-Blade [69] is a computationally efficient open source structural analysis and design
code developed by Sale. This code includes all of the same capabilities of PreComp,
adding analysis of load induced strain, stress, deflection, buckling, optimization capabilities,
and graphical post-processing capabilities. Making use of a blend of CLT, Euler-Bernoulli
theory [68, 70, 71] and shear flow theory applied to composite beams, Co-Blade predicts both
the distributed structural properties of composite wind turbine blades, and their deformation
and material stress fields. The Co-Blade’s technical approach models the turbine blade as
a cantilever beam subject to aerodynamic loads, self-weight, buoyancy3 , and centrifugal
forces. As a consequence, the beam undergoes bending, axial deflection (i.e., tension
and compression along the longitudinal axis of the beam), and twist (i.e. torsion about the
longitudinal axis of the beam). The linear differential equations of equilibrium for a cantilever
3 Buoyancy is an upward force exerted by a fluid that opposes the weight of an immersed object.
2.1 Wind turbine rotor aeroservoelasticity 27

beam are then used to determine the shear force and bending moment distributions along
the beam length. Co-Blade considers the beam cross sections are assumed to be thin-walled,
closed, and single- or multi-cellular. The cross-section of the cantilever beam is discretized as
a connection of flat composite laminates, as illustrated in Fig. 2.6. Although each composite

Fig. 2.6 Modeling of the turbine blade cross-section. Each section is discretized as a
connection of composite laminate plates, each made up of a number of different laminas
(θ represent the orientation of each lamina’s principal material direction with respect to the
blade axis). Image reproduced from [69].

laminate is a stack of a number of different laminas (each characterized by its own material
and constitutive properties), the CLT is used to evaluate the effective mechanical properties
(i.e., Young’s modulus, shear modulus, Poisson’s ratio, thickness, and density) of each
laminate, treating it as a single structural element. Therefore, the beam cross-section is made
up of a number of discrete areas, or panels, of homogeneous material (represented through
different colors in Fig. 2.6). The panels, made up of consistent flat composite laminates, are
characterized by the effective mechanical properties computed via CLT. Each panel then
contributes to the global cross-sectional properties, which are computed by the method of
Young’s modulus weighted properties [70, 71]. Once that global cross-sectional properties
are known, the deflections and the beam effective axial stress and effective beam shear stress
are computed under the assumption of an Euler-Bernoulli beam. The calculation of the beam
effective shear stress is based on a shear flow approach. The distributions of the effective
beam stresses from the Euler-Bernoulli theory are eventually converted into the equivalent
in-plane distributed loads on the flat composite laminates, so that the strains and stresses of
each lamina can be recovered by means of CLT. Co-Blade follows the approach described
in [72, 73] to perform the linear buckling analysis of the blade. In this approach the top
and the bottom surfaces of the blade are modeled as curved plates subject to a combination
of compression and shear loads, while the shear webs (connecting the top surface to the
bottom one) are modeled as flat plates subject to combined bending and shear. Practically,
panel buckling is treated by means of a buckling criteria R expressed by a dimensionless
28 Multidisciplinary analysis of horizontal axis wind turbine rotors

number which is lower than 1 if the effective stresses in a panel have not exceeded the critical
buckling stresses.
Another simple and fast structural code has been developed and presented by Ashuri
et al. [74]. This code, based on the Euler-Bernoulli theory, enables both the extraction of
structural properties of a composite wind turbine blade and the calculation of its bending
stress field. In fact, unlike Precomp, this code includes the calculation of the cross-sectional
area moments of inertia, allowing one to determine the bending stresses σ . For this purpose,
the classic formula for determining σ , in a beam under simple bending, can be used as
follows:
My
σ= (2.13)
Ix
where M is the moment about the neutral axis4 , y is the perpendicular distance to the neutral
axis, and Ix is the cross-sectional area moments of inertia about the neutral axis.
As mentioned above, blade mode shapes are required to perform rotor aeroelastic simula-
tions by means of aeroelastic codes. Moreover, to avoid blade resonance issues, rotor blades
should be designed considering their natural frequencies. As explained in Chapter 6, the first
natural frequency of the blade should be above the maximum rotor blade passing frequency.
All the structural codes described in this section cannot cope with the calculation of either
the blade mode shapes or the blade natural frequencies. The preprocessor BModes [75] is
a NREL tool able to generate coupled modes and natural frequencies for a turbine blade
or a tower. BModes uses distributed inertial and stiffness properties of the blade and tower
along their longitudinal axis. For blade mode shapes calculation, BModes also requires the
rotational speed of the blade to calculate the rotational stiffening and its pitch angle as input.
From the solution of the associated eigenvectors, polynomial expressions of the mode shapes
are calculated. The calculation of the the natural frequencies are carried out in the same way,
including the stiffening effects of rotation.

Fatigue

Fatigue is defined as the progressive and localized structural damage that occurs when a
material is subject to cyclic loading. Due to the spatiotemporal variability of the wind and
the rotation of the rotor, wind turbine components, such as the rotor blades, tower and drive
train, indeed experience cycling loads throughout their lifespan, and therefore they need to
be designed against fatigue. The most common approach to design wind turbine mechanical

4 The
neutral axis is an axis in the cross section of a beam subject to bending along which there are no
longitudinal stresses or strains.
2.1 Wind turbine rotor aeroservoelasticity 29

parts against fatigue is to keep structural stress below threshold of fatigue limit, normally
represented by a parameter indicating the fatigue damage accumulation.
Fatigue analysis for wind turbines is typically carried out by running a large number
of aeroelastic (time-domain) simulations under different wind conditions, determining the
fatigue loads expected over the lifetime of the turbine. Fatigue loads on each primary
component of the turbine are then post-processed. MLife [76] is a NREL code, created to
compute fatigue estimates resulting from time-series load data files. This code accumulates
fatigue damage due to fluctuating loads. These fluctuating loads are broken down into
individual hysteresis cycles each characterized by a load mean and range, using the rainflow
counting method according to the ASTM standard [77]. In practice, the rainflow counting
algorithm reduces a complex spectrum of varying loads into a simple set (or series) of cycles
defined by a given mean and amplitude. MLife assumes that the fatigue damage accumulates
linearly and independently for each of these cycles according to the Miner’s rule [78]. Thus,
the total damage resulting from all cycles is given by:
ni
D=∑ (2.14)
i Ni (LiRF )

where LiRF denotes a given cycle’s load range about a fixed load-mean value and Ni (LiRF )
represents the number of cycles, characterized by LiRF , that would lead to failure. ni represents
the actual count of cycles characterized by LiRF . When ni is equal to Ni (LiRF ) failure occurs,
and this corresponds to D equal to 1. Load ranges are related to cycles to failure by means of
the S-N curve (or Wöler curve) which can be modeled as follows:
!m
Lult − |LMF |
Ni = 1 RF

(2.15)
2 Li

where Lult is the ultimate design load of the component, LMF is the fixed lead mean and m is
the Wöler exponent, depending on the considered component.

Wind input files for structural verification

Standards, such as the IEC standard, prescribe wind conditions at which aeroelastic simula-
tions should be carried out in order to verify structurally wind turbine components. These
conditions encompass normal, extreme, steady and turbulent wind speeds. Several codes
have been devised to generate wind files for HAWT structural verification suitable for aeroe-
lastic simulations. In particular, NREL developed wind simulators primarily for use with
AeroDyn-based programs, such as IECWind [79] and TurbSim [80]. In AeroDyn, wind
30 Multidisciplinary analysis of horizontal axis wind turbine rotors

conditions are prescribed by two types of wind input files, namely the hub height wind files
and the full field turbulence files. The former typology represents wind conditions, varying
in time, by means of the values of horizontal wind speed at the hub, wind direction, vertical
wind speed, horizontal wind shear, power law vertical wind shear, linear vertical wind shear
and gust velocity. For the exact definition of these parameters one can refer to [44]. The
latter wind file typology represents all three components of the wind vector varying in space
and time. This permits a detailed simulation of a wind field with the appropriate scales and
correlation of atmospheric turbulence. IECWind is a program developed to create AeroDyn
hub height wind files, modeling the steady conditions outlined in the IEC standard. TurbSim
is a tool able to generate numerical simulations of a full field flow, containing turbulence
structures that reflect realistic spatiotemporal turbulent velocity fields. TurbSim can be used
to create AeroDyn full field turbulence files, according to the turbulent conditions prescribed
by the the IEC standard.

2.2 Wind turbine airfoil aerodynamics

Because of its simplicity, short execution time and robustness, BEM theory is the most widely
used approach for the aerodynamic analysis and design of wind turbine rotors. Using BEM
codes, the aerodynamic characteristics of the airfoils making up the rotor blades must be
provided in terms of lift and drag coefficients as a function of Re and AoA. Aerodynamic
characteristics used in BEM calculations are typically based on 2D airfoil lift and drag
polars determined either by means of wind tunnel experiments, or computational models
such as panel codes, and NS CFD. It should be noted that it is unlikely that experimental
airfoil data can be used in design optimization systems which optimize concurrently the
design of the blade planform and its airfoils, because many diverse and new airfoils are
examined by the optimizer, and, in general, the optimal design will feature new airfoils
for which experimental data do not exist. This highlights the need of using sufficiently
fast computational tools for airfoil aerodynamics. Sect. 2.2.1 will review the most popular
approaches for the aerodynamic analysis of wind turbine airfoils.
Using only 2D airfoil polars, however, BEM codes are not able to model the complex 3D
effects that occur over the rotor blades. These effects are dominated by those associated with
blade rotation, applying centrifugal and Coriolis forces to the air flow. Several empirical and
semi-empirical corrections, called 3D corrections, have been devised to account for these
effects. 3D corrections, described in Sect. 2.2.2, are normally applied directly to the 2D
airfoil polars.
2.2 Wind turbine airfoil aerodynamics 31

Under normal operating conditions, wind turbine blades can experience very high AoAs,
causing aerodynamic stall. Therefore, to model these situations using BEM theory, 2D airfoil
polars must be provided in both pre- and post-stall regions. However, wind tunnel airfoil tests
are normally available until stall, and due to model limitations, airfoil analysis codes often
cannot predict accurately airfoil performance beyond stall. To deal with this issue global
post-stall methods have been developed. These empirical methods, often accounting for all
3D effects (including the rotational effects), are able to determine pre- and post-stall airfoil
characteristics starting from the pre-stall 2D polars. Sect. 2.2.3 will illustrate the most used
global post-stall methods.

2.2.1 Methods for the airfoil aerodynamic analysis

Over the last two decades, several panel codes have been developed and used to perform
the aerodynamic analysis and design of new airfoils. Tangler et al. [81] used the Eppler
Airfoil Design and Analysis Code to develop the NREL S8xx airfoil series. However, the
most popular panel code is XFOIL [81], developed by Drela at the Massachusetts Institute of
Technology (MIT). XFOIL is a panel code with a strong viscid–inviscid interaction scheme,
giving realistic boundary layer properties. The code uses the eN method to predict transition.
The use of XFOIL enables a rapid and efficient calculation of the airfoil performance
in the linear region of the lift curve below stall; the code, however, is known to usually
overestimate the maximum lift coefficient [82], and not to provide reliable predictions of the
force coefficients beyond the stall inception point, after which significant flow separations
occur. The near stall predictions of XFOIL appear to be particularly inaccurate for thicker
airfoils like the NREL S809 airfoil [83]. Improved near-stall force predictions could be
obtained with RFOIL [82], the variant of XFOIL developed at Delft University of Technology
(TU Delft), the National Aerospace Laboratory of the Netherlands (NLR) and the Energy
research Centre of the Netherlands (ECN). RFOIL features a better convergence around the
maximum lift due to the use of different velocity profiles for the turbulent boundary layer
and due to modifications in the calculation of the turbulent boundary layer shape factor [84].
This improves the prediction of the maximum lift coefficient. An alternative to panel codes
for the 2D aerodynamic analysis of airfoils is represented by the use of CFD codes. With
respect to panel codes, CFD codes give better agreement with experimental measures, and
provide a more detailed description of the flow field around the airfoil. The use of CFD is
in fact reaching a level of maturity enabling it to accurately predict airfoil aerodynamics
well beyond the AoA of maximum lift [85]. However, the computational and preparation
time associated with the use of CFD codes is much more onerous than that required by panel
32 Multidisciplinary analysis of horizontal axis wind turbine rotors

codes. For this reason, the use of panel codes represents the preferred approach in most of
the wind turbine design optimization applications.

2.2.2 Rotational effects and 3D corrections

When used with accurate 2D airfoil lift and drag polars, BEM theory provides reasonable
performance predictions when the airfoils are at AoAs below stall. However, when a
significant part of the blade is subject to AoAs above stall, which lead to separated flow
conditions, the direct use of 2D airfoil polars results in a poor agreement between measured
and predicted power production and loads [86]. In fact, in the presence of local blade stall,
the BEM theory lacks the ability to model the complex 3D effects occurring over rotating
blades. These effects are primarily associated with the centrifugal and Coriolis effects,
which are important mainly over the inboard part of the blade, experiencing a high degree
of flow separation. In particular, rotational effects lead to the augmentation of rotating
blade aerodynamic properties, including stall delay and lift enhancement. In fact, blade
rotation delays turbine blade stall and significantly amplifies blade aerodynamic loads. This
phenomenon depends on viscous flow effects in the presence of centrifugal and Coriolis
influences, and occurs routinely during turbine operation.
At present, the physics behind rotational augmentation of HAWT blade aerodynamics
remains only partially characterized and understood. Recently, the combined use of experi-
mental measurements and CFD models have provided new insights into turbine blade flow
fields and aerodynamic forces produced in the presence of rotational augmentation [87, 88].
In particular, these studies highlighted that the mechanisms underlying augmented aerody-
namic force production during rotating conditions can be related to the flow field resulting
over the stalled portion of the blade, involving boundary layer separation and shear layer
impingement. In a stalled blade section, boundary layer separation normally occurs at some
point over the suction surface. As the AoA increases, boundary layer separation moves
forward towards the leading edge, starting from the trailing edge vicinity. As shown in
Fig. 2.7, when the boundary layer separation reaches the leading edge, the resulting shear
layer arches over the blade suction surface, and finally impinges on the aft portion of the blade
chord. Between separation and impingement, the shear layer surrounds a region of intense
recirculation strongly reminiscent of vortical structure, within which the flow is directed
outboard and forward towards the leading edge. The degree to which rotating forces are
augmented is correlated with the size of this vortical structure. Separation reaches the leading
edge first at inboard locations and arrives later at the leading edge at outboard locations.
For this reason, at inboard blade regions the vortical structure is bigger and stronger, and
2.2 Wind turbine airfoil aerodynamics 33

Fig. 2.7 Flow field over a stalled section of a rotating blade, showing boundary layer separa-
tion and shear layer impingement. A region of intense recirculation above the blade’s suction
surface is also depicted. Image reproduced from [87].

therefore these regions undergo the greatest degree of rotational augmentation. At outboard
blade regions, where the flow is attached, there is little or no rotational augmentation.
Several 3D correction models have been developed to deal with centrifugal and Coriolis
forces. In general, these models correct the 2D lift and drag coefficients at each radial
station of the blade, accounting for different flow conditions along the blade span. The
most commonly used 3D corrections include those developed by: Snel et. al. [89], Du and
Selig [90], Chaviaropoulos and Hansen [91], Lindenburg [92], Bak et al. [93] and Eggers et
al. [94]. The first four of these models determine the 3D lift and drag coefficients, denoted
respectively by CL,3D and CD,3D , by means of the following equations:
c 
CL,3D = CL,2D + fCL , ... ∆CL (2.16)
r

c 
CD,3D = CD,2D + fCD , ... ∆CD (2.17)
r
in which CL,2D and CD,2D are respectively the 2D lift and drag coefficients, and c/r is the
local blade solidity, where c is the local chord and r is the local radius. ∆CL and ∆CD denote
respectively the difference between the CL and CD that would exist if the flow did not separate
(taken respectively as CL = 2π(α − αlift=0 ) and CD = CD,α=0 , where α is the AoA, αlift=0 is
34 Multidisciplinary analysis of horizontal axis wind turbine rotors

the zero-lift AoA and CD,α=0 is the drag coefficient at an AoA of zero) and the CL and CD
measured in a 2D configuration including separation (i.e., CL,2D and CD,2D ). Hence:

∆CL = 2π(α − αlift=0 ) −CL,2D (2.18)

∆CD = CD,α=0 −CD,2D (2.19)

Accounting for the local blade solidity, these corrections vary along the blade span. In
particular, the closer the considered radial section is with respect to the blade root (i.e., the
lower r is), the stronger the 3D correction on the aerodynamic polars is. In fact, as already
mentioned above, blade sections near the blade root are subjected to stronger 3D effects due
to rotation.

The functions fCL and fCD depend on the selected model. The model developed by Snel et.
al. provides a correction only for CL . According to this method, the function fCL is expressed
as follows:  c 2
fCL = 3 (2.20)
r
The model of Du and Selig expresses fCL and fCD as follows:
" d R
#
1 1.6(c/r) a − (c/r) Λ r
fCL = −1 (2.21)
2π 0.1267 b + (c/r) Λd Rr

" d R
#
1 1.6(c/r) a − (c/r) 2Λ r
fCD = −1 (2.22)
2π 0.1267 b + (c/r) 2Λd Rr

ΩR
Λ=p (2.23)
U 2 + (Ωr)2
where Ω is the rotor speed, R is the tip radius, U is the freestream wind velocity and a = b = d
= 1. According to the Chaviaropoulos and Hansen model, fCL and fCD are expressed as:
 c h
fCL = fCD = a cosn (θ p,0 ) (2.24)
r
where θ p,0 is the blade pitch angle, a = 2.2, h = 1 and n = 4. The Lindenburg’s model
accounts for a correction just for CL as follows:
 2  
Ωr c 2
fCL = 3.1 (2.25)
Urel r
2.2 Wind turbine airfoil aerodynamics 35

where Urel is the relative wind velocity. An example of the application of 3D correction
models to a stall-regulated wind turbine can be found in [83].

2.2.3 Global post-stall methods

The accurate prediction of airfoil aerodynamic characteristics in both pre- and post-stall
regions is a key step in the BEM-based rotor analysis and design, and this holds for both
stall- and pitch-regulated wind turbines. Indeed, in the former case, flow stall is induced to
control the mechanical power extracted by the rotor, while in the latter case blade airfoils
experience very high AoAs during the rotor start and stop transients, and in parked conditions,
especially over the inboard part of the blade. For this reason, a number of global post-stall
methods have been developed and presented. These codes, generally derived empirically,
provide a method for calculating lift and drag polars of wind turbine airfoils, in both the
pre-stall and post-stall regions, including all aerodynamic 3D effects (tip and hub losses,
rotational effects, etc.). In fact, the attribute “ global” of such models refers to their ability to
capture the net aerodynamic effects related to the blade geometry rather than detailed flow
physics characteristics (as is the case of 3D corrections seen in Sect. 2.2.2). Typically, global
post-stall methods correct and extend the pre-stall airfoil lift and drag coefficient curves.
Figure 2.8 shows an example of measured force coefficients at high AoAs for two typical
wind turbine airfoils measured in a wind tunnel.
Viterna and Corrigan [96] proposed an empirical method to correct and extend 2D airfoil
aerodynamic characteristics. This method evaluates the lift and drag coefficients at high
AoA starting from the stall point of the airfoil lift curve. The equations of this method
were developed by matching the experimental power curves of different rotors using a BEM
code. This means that all aerodynamic effects are included, and this specifically covers the
rotational effects that are dominant at the inboard part of the blade. The Viterna and Corrigan
method start from correcting the 2D pre-stall polars in order to account for the blade aspect
ratio effects. In fact, 2D polars do not consider the effect due to a finite blade aspect ratio,
such as tip and hub losses. The size of the aspect-ratio effects on airfoil coefficients in the
attached regime can be estimated using the classical equations for converting infinite-length
airfoil data to finite-length data, from the work of Munk, Glauert, and Prandtl [97]. According
to these equations, the finite aspect ratio pre-stall lift and drag coefficients and the AoA can
be calculated as follows:
CL = CL0 (2.26)

CL2
CD = CD0 + (2.27)
πAR
36 Multidisciplinary analysis of horizontal axis wind turbine rotors

Fig. 2.8 Measured lift and drag coefficients of two airfoils at high AoAs. Image reproduced
from [95].

57.3CL
α = α0 + (2.28)
πAR
where 0 is a subscript denoting infinite aspect ratio (or 2D) data, and AR is the blade aspect
ratio. In this model, AR is defined as the blade radius divided by the chord length at 75% of the
span. The equations for the calculation of the post-stall aerodynamic coefficients constituting
the Viterna and Corrigan method are partly based on the variation of the maximum flat plate
drag coefficient, CD,max , with AR, as given in Eqs. 2.29 and 2.30:

CD,max = 1.111 + 0.018AR for AR < 50 (2.29)

CD,max = 2.01 for AR ≥ 50 (2.30)

Post-stall drag coefficient is then determined as follows:

CD = CD,max sin2 α + K1 cos α (2.31)

2.2 Wind turbine airfoil aerodynamics 37

where the parameter K1 is expressed as:

CD,s −CD,max sin2 αs

K1 = (2.32)
cos αs

Post-stall lift coefficient can be calculated as:

1 cos2 α
CL = CD,max sin 2α + K2 (2.33)
2 sin α
where:
sin αs
K2 = (CL,s −CD,max sin αs cos2 αs ) (2.34)
cos2 α
The index s denotes stall conditions defining the starting point for the Viterna and Corrigan
curves. In the stall point, the pre-stall characteristics match with the calculated post-stall
performance of the airfoil. Recently, Tangler and Kocurek [86] used the Viterna and Cor-
rigan method to calculate the power output of a stall regulated wind turbine, using various
combinations of input parameters of the correction model.

Recently, an alternative set of empirical equations for modeling lift and drag coefficients
in the pre- and post-stall regimes, called AERODAS, has been developed by Spera [98],
extending the Viterna and Corrigan model. Based on empirically derived equations, AERO-
DAS provides a method for calculating stall and post-stall lift and drag characteristics of
rotating airfoils, using as input a limited amount of pre-stall 2D aerodynamic data of the blade
airfoils. More precisely, the equations of the AERODAS model were determined through an
empirical approach based on the trends of experimental data available for a wide variety of
airfoils. These trends were then modeled by a set of algebraic equations providing the best
fit of the available experimental data. An important feature of the AERODAS equations is
that they correct the infinite-length airfoil (or 2D) data for the effect of blade aspect ratio,
explicitly taking into account the tip and the hub losses. Figure 2.9 presents the comparison
between experimental 2D lift and drag coefficients for the NREL S809 airfoil [99], and those
calculated through the AERODAS equations for an infinite blade aspect ratio and a blade
aspect ratio of 15.3.

An alternative method, called StC, has been developed and presented by Lindenburg [100].
This approach is based on empirical relations that define the pre- and post-stall lift and drag
curve, including the effects associated with blade rotation and finite aspect ratio.

AirfoilPrep [101] is a NREL spreadsheet devised to generate the airfoil data needed by
BEM codes. This code applies rotational augmentation corrections for 3D delayed stall
to pre-stall 2D airfoil polars, by using Du’s method [90] to augment the lift and Eggers’
38 Multidisciplinary analysis of horizontal axis wind turbine rotors

Fig. 2.9 Comparison of measured and calculated lift and drag coefficients for the NREL S809
airfoil. Infinite blade aspect ratio test data [99] are compared with the coefficients calculated
by means of AERODAS, using an infinite blade aspect ratio and a blade aspect ratio of 15.3.
Image reproduced from [98].

method [94] to modify the drag. The corrected polars are then extended to the full 360 deg
range of AoAs by means the Viterna and Corrigan method [96].

2.3 Cost of energy

As already mentioned in Chapter 1, the reference metric in wind turbine design optimization
is the LCOE, defining the economical performance of wind turbine systems distributed (or
levelized) over their lifetime. Therefore, a reliable model is needed to accurately estimate the
cost of electricity generated by wind turbines. For three-bladed, upwind, pitch-controlled,
variable-speed wind turbines, the DOE/NREL scaling model [102] is a tool developed by
NREL to accomplish this goal. The purpose of this model is to evaluate the impact of any
change of the design of a wind turbine on the system cost. To do so, the model considers
several elements of the system, such as the initial capital cost (ICC), the O&M cost, the
levelized replacement/overhaul cost (LRC) and the AEP. To determine the cost of wind turbine
2.3 Cost of energy 39

components for different sizes and configuration, the DOE/NREL scaling model uses scaling
relationships. Component costs, all referred to 2002 United States dollars, in most cases
are function of rotor diameter, machine rating and tower height. In the DOE/NREL scaling
model, cost scaling functions have been developed for major components and subsystems.
Much of the data used to develop scaling functions for machines of greater than 1 to 2 MW is
based on conceptual designs. Many components are scaled using functions that are close to a
cubic relationship. This is what would normally be expected for technologies that did not
undergo design innovations as they grew in size [102]. According to the DOE/NREL scaling
model, the LCOE, expressed in $/kWh, can be calculated by using the following equation:

(FCR · ICC) + LC + O&M + LRC

LCOE = (2.35)
AEP
where FCR is the fixed charge rate and LC is the lease cost. The ICC, expressed in $, is the
sum of the wind turbine capital cost (TCC) and the balance of station (BOS) cost. The primary
wind turbine components considered within the TCC include: rotor, drive train, nacelle,
control, safety system, and condition monitoring and tower. BOS cost includes the costs
associated with foundation/support structure, roads, civil work, assembly and installation,
electrical interface/connections and engineering permits. When offshore wind turbines are
assessed, additional components are included in the BOS cost, including: marinization,
port and staging equipment, personal access equipment, scour protection, surety bond, and
offshore warranty premium. The FCR represents the annual amount per dollar of ICC needed
to cover the capital cost, a return on debt and equity, and various other fixed charges, such
as depreciation, income tax, and property tax and insurance. According to the DOE/NREL
scaling model, FCR is set equal to 11.85% per year. The product of FCR and ICC appearing in
the LCOE formula is therefore expressed in $/year. LC represents the annual cost charged for
the rental or lease of the land or ocean bottom. LC is expressed in units of $/year. O&M cost
is the yearly cost associated with operations and maintenance, and it is expressed in $/year.
O&M cost normally includes labour, parts, and supplies for scheduled and unscheduled
turbine maintenance, parts and supplies for equipment and facilities maintenance, and
labour for administration and support. LRC is the cost associated with major replacements
and overhauls, distributed over the wind turbine’s lifetime, and it is expressed in $/year.
Measured in kWh/year, AEP is the annual energy output of the turbine based on a given
annual average wind speed distribution. The AEP calculation is adjusted for mechanical and
electrical conversion losses. As an example, the cost estimate from the DOE/NREL scaling
model citecost for a 3 MW offshore wind turbine is reported in Fig. 2.10.
40 Multidisciplinary analysis of horizontal axis wind turbine rotors

Fig. 2.10 Cost estimate from the DOE/NREL scaling model [102] for a 3 MW offshore wind
turbine (rotor diameter equal to 90 m and hub height equal to 80 m).
Chapter 3

Robust design optimization

In the last decades, advanced computer-based simulation approaches and tools have been
successfully integrated in the design optimization of engineering systems, shortening the
overall design cycle, providing better understanding of the operating behavior of the systems
under consideration, and lowering design costs. However, it is still difficult to determine a
confidence level in the information provided by numerical methods and models. This com-
plexity is primarily related to the uncertainties associated with the inputs of any computation
attempting to represent a physical system, or the errors affecting numerical algorithms and
computer codes. Therefore, in order to ensure high design quality and reliability, uncertainty
quantification should be considered and integrated in the optimization process. RDO is a field
of optimization theory that deals with optimization problems in which a certain measure of
robustness is sought against uncertainty. In other words, RDO aims at finding a design with
the best possible performance expectation and minimal performance sensitivity to factors
causing uncertainty. RDO problems are generally formulated and solved in terms of multiple
objective functions, typically involving the mean and variance of one or more deterministic
objective functions. RDO is performed by following a sequence of three main stages. The
first stage consists of identifying, qualifying and quantifying the sources of uncertainty
associated with the system to be optimized. In this stage, the uncertainty of the random
input variables of the system is characterized probabilistically, by means of a given PDF. The
second stage consists of propagating the input uncertainty through the analysis system to
obtain a probabilistic description of the objective functions and constraints. Robust forms of
objectives and constraints are generally expressed as a function of the mean and variance
of their deterministic counterparts. Finally, the third stage consists of optimizing the robust
objectives subject to the robust constraints. The robust optimal design is such that the mean
of the objectives is optimized and their variance are minimized. The aim of this chapter is
to describe the mathematical formulation of a RDO problem, providing an overview of the
42 Robust design optimization

most common approaches used for uncertainty propagation and numerical optimization. In
the following sections, the methods included in the robust design optimization framework
developed in this work will be highlighted.

3.1 Mathematical formulation of robust design optimiza-

tion problems
An optimization problem aims at finding the best configuration of a system from all feasible
configurations. This is normally accomplished by optimizing a set of parameters character-
izing the system under consideration, denoted by design variables b ∈ ℜn , where n is the
number of design variables, without violating some constraints g j , j = 1, 2, ..., r, where r
is the number of constraints. The optimization process is driven by an objective function
f , which is typically minimized by varying the aforementioned design variables. A deter-
ministic design optimization problem (i.e., an optimization problem that does not consider
uncertainty) can be formulated as follows:

Find: b ∈ ℜn
to minimize: f (bb)
(3.1)
subject to: g j (bb) ≤ 0, j = 1, 2, ..., r
and: b L ≤ b ≤ bU

where b L and bU are respectively the lower and upper bounds of the design variables’
variability ranges. The evaluation of the objective function and constraints is performed
numerically, through a numerical model of the system under investigation. Considering
the input uncertainty associated either with the design variables or other parameters of the
system, the objective function and constraints can be modified as follows:

f (bb) −→ f (bb + z b , p + z p ) (3.2)

g j (bb) −→ g j (bb + z b , p + z p ) (3.3)

where p ∈ ℜm is a constant vector defining the design parameters, and m is the number of
design parameters. Like the design variables, the design parameters refer to some charac-
teristics of the system under consideration. Design parameters, however, do not drive the
optimization process (i.e., they are not optimized) as the design variables do. In a proba-
3.1 Mathematical formulation of robust design optimization problems 43

bilistic optimization framework, design parameters can be affected by uncertainty (as well
as the design variables), and therefore they need to be treated explicitly in the optimization
problem. z b ∈ ℜn and z p ∈ ℜm are respectively the uncertainties associated with the design
variables and the design parameters. Therefore, in a probabilistic framework, f and g j are
random quantities induced by the uncertainties zbi and zip , respectively. The RDO problem
is practically treated by introducing deterministic operators applied to f and g j , denoted
respectively by F and G j , in order to eliminate the random character of f and g j , and thus
to reduce their dependencies on z. Hence, the resulting design optimization problem under
uncertainty can be expressed as:

Find: b ∈ ℜn
to minimize: F(bb, p , z )
(3.4)
subject to: G j (bb, p , z ) ≤ 0, j = 1, 2, ..., r
and: b L ≤ b ≤ bU

Generally, the probabilistic objective function and constraints F and G j are defined in
terms of the statistical moments (i.e., mean, variance, skewness, kurtosis, etc.) of f and g j ,
respectively. Considering the objective function f in the form given by Eq. 3.2, the mean µ f
and the variance σ 2f can be expressed as follows [29]:
Z Z Z
b p
µ f (bb, p ) = E[ f (bb + z , p + z )] = ... f (bb + z b , p + z p )κ1 (zb1 )...κn (zbn )
(3.5)
×ν1 (z1p )...ν0 (z0p )dzb1 ...dzbn dz1p ...dz0p

Z Z Z
σ 2f (bb, p ) = E[( f (bb + z b , p + z p ) − µ f (bb, p ))2 ] = ... [ f (bb + z b , p + z p ) − µ f ]2
(3.6)
×κ1 (zb1 )...κn (zbn )ν1 (z1p )...ν0 (z0p )dzb1 ...dzbn dz1p ...dz0p

where κi (zbi ) and νi (zip ) are the joint PDFs of the uncertainty factors zbi and zip , respectively.

The RDO problem formulated in Eq. 3.4 highlights that the robustness can be sought
against the variability of both the objective function and constraints. The combined prob-
abilistic considerations of the objective function and constraints is known as reliability
optimization. In practice, reliability optimization extends the concept of RDO, the main
aim of which is to maximize the performance (expressed by the objective function) and
simultaneously to minimize the sensitivity of the performance with respect to random param-
eters. Indeed, reliability optimization aims at optimizing the objective function, limiting its
44 Robust design optimization

variations with respect to the variations of the uncertain variables (as explained in Sect. 3.2),
while satisfying all constraints within a range of tolerance (as described in Sect. 3.3).

3.2 Robust design optimization methods

For both constrained and non-constrained RDO problems, the robustness of the objective
function is pursued by minimizing the probabilistic objective function F. A number of
approaches have been used to solve RDO problems, formulating the minimization of F in
different ways. The simplest RDO approach, known as Bayes Principle [12], is to define the
probabilistic objective function F as the mean of f , denoted by µ f . Hence:

F(bb, p , z ) = µ f (bb, p ) (3.7)

Another simple approach is the weighted sum method [29], in which F is defined as the
weighted averaged of the mean and variance of f , respectively denoted by µ f and σ 2f . Thus:

F(bb, p , z ) = αw [µ f (bb, p )/µ ∗f ] + (1 − αw )[σ 2f (bb, p )/σ 2∗

f ] (3.8)

where αw and (1 − αw ) are the weights associated with µ f and σ 2f , respectively. µ ∗f and
σ 2∗ 2
f are used to normalize µ f and σ f , respectively. The approach known as constrained
optimization [12], aims at minimize µ f , keeping σ 2f lower or equal to a certain value σ 2∗
f .
Thus, the constrained optimization can be formulated as follows:

minimize: µ (bb, p )
f
(3.9)
subject to: σ 2 (bb, p ) ≤ σ 2∗
f f

Finally, the last approach reviewed in this section is to formulate the RDO problem as
a multi-objective approach [12], whereby the mean and the variance of f are minimized
concurrently as follows: 
minimize: µ (bb, p )
f
(3.10)
2
minimize: σ (bb, p)
f

3.3 Robustness of constraints

In a deterministic constrained optimization, the optimizer ensures that a given constraint does
not exceed a certain limit value. In the presence of uncertainty, constraints are characterized
by a certain variability, defined by a given PDF. A very simple approach to treat uncertain
3.4 Methods for uncertainty propagation 45

constraints in the course of an optimization process is to consider their mean value. This
means that the mean value of each constraint must be kept below a certain threshold. In this
case, the probabilistic form of constraints G j becomes:

G j (bb, p , z ) = µg j (bb, p ) (3.11)

where µg j is the mean of g j . Considering just the mean value of each constraint however
does not guarantee that all values that the constraints assume are below their limit values.
Therefore, if a high reliability related to the constraints is sought, each constraint needs be
verified probabilistically, within a confidence interval. This means that a given constraint
does not exceed a specific limit for a given number of cases (expressed by a confidence
level). Confidence intervals of the constraints depend on their PDFs. Expressing confidence
intervals in terms of mean and standard deviation, the probabilistic form of constraints G j
can be written as:
G j (bb, p, z) = µg j (bb, p) + kg j σg j (bb, p) (3.12)

where σg j is the standard deviation of g j , and kg j is a suitably chosen coefficient, determining

the length of the confidence interval.

3.4 Methods for uncertainty propagation

Within a probabilistic optimization framework, the problem of uncertainty propagation
consists of characterizing probabilistically the objective function and constraints given the
PDFs of all the uncertain input parameters. The output quantities of interest (QoIs), i.e., the
objective function and constraints, are characterized probabilistically by defining a PDF. As
already stated, the QoIs are typically defined probabilistically by their statistical moments,
such as mean and variance. Let’s simplify Eq. 3.5 and Eq. 3.6 by considering a generic
objective function f depending on the uncertain variables z . In this case, the mean value and
the variance of f , denoted respectively by µ f and σ 2f , can be written as:
Z +∞
µ f = E[ f (zz)] = f (zz)pz (zz)dzz (3.13)
−∞

Z +∞
σ 2f 2
= E[( f (zz) − µ f ) ] = [ f (zz) − µ f ]2 pz (zz)dzz (3.14)
−∞
where pz is the joint PDF of the random variable z . For engineering problems of practical
interest, the solution of Eq. 3.13 and Eq. 3.14 is normally achieved numerically, by means of
46 Robust design optimization

an uncertainty propagation method. Several methods have been developed to deal with this
problem. The numerical approximation of the statistical moments needs a trade-off between
computational time and accuracy. In this section, two approaches are described, namely
sampling and deterministic methods. In the robust design optimization framework developed
in this work the Latin hypercube sampling and the deterministic URQ technique (reviewed
below) are included.

3.4.1 Sampling methods

Techniques based on sampling are simple uncertainty propagation approaches, relying on

a number of simulations (or realizations) performed by selecting given input values. Once
all realizations have been performed, the outputs are collected and analyzed in order to
characterize statistically the outcome. Common approaches based on sampling are the Monte
Carlo method [103] and the Latin hypercube sampling strategy [104].

Monte Carlo method

The Monte Carlo method is the oldest and the most used sampling technique. It is based on a
random sampling within the variability ranges of the stochastic inputs. More precisely, it
generates a random sample of m points, taken for each uncertain input variable of a model.
Each of the m points corresponds to the sampling point z i . The selection of these sample
points is independent of the PDFs associated with the input variables. The outcome is
generally organized as a histogram and the estimations for the mean and the variance of f
are determined as follows:
1 m
µ f = ∑ f (zzi ) (3.15)
m i=1

1 m
σ 2f = ∑ [ f (zzi ) − µ f ]2 (3.16)
m − 1 i=1
The advantages of this method are that it is simple and does not require any modification of
the analysis system, as the method requires a sequence of deterministic analyses. However,
in order to obtain an accurate estimation of the statistical moments, a large number of
realizations are required, as the method converges to the exact solution only for a number
of samples going to infinity. The convergence however does not depend on the number of
uncertain variables. The use of Monte Carlo methods, although they always provide the right
answer, might be prohibitive in some cases due to time requirements.
3.4 Methods for uncertainty propagation 47

Variance reduction techniques [105], such as control variates, antithetic variables, strati-
fied, Latin hypercube (described below), and descriptive sampling [106], have been developed
to achieve faster convergence for Monte Carlo methods.

Latin hypercube sampling

Latin hypercube sampling is one of the methods developed to accelerate the Monte Carlo
approach, increasing its efficiency. Using the Monte Carlo method it might happen that a
large number of sample points are clustered closely, maybe leaving unexplored other regions.
Latin hypercube sampling aims at spreading the sample points more evenly across all possible
values. In this method, the range of variability of each input stochastic variable is divided in
M intervals with equal probability. M samples are then taken, one from each interval. The
samples for each input are then shuffled so that there is no correlation between the inputs.
The estimation of the mean and variance of f is then obtained by means of Eq. 3.15 and
Eq. 3.16.
Both Monte Carlo method and Latin hypercube sampling are unbiased1 estimation
techniques. Using these methods the estimated statistical moments approach their theoretical
values as the sample size increases. Given a number of samples, the Latin hypercube sampling
achieves a better estimation of the statistical moments with respect to the Monte Carlo method.
In other words, Latin hypercube sampling requires a lower number of samples than the Monte
Carlo method to achieve the same accuracy. In particular, it has been demonstrated [107] that
the same accuracy can be achieved by means of Monte Carlo method and Latin hypercube
sampling by using m2 and m samples, respectively.

3.4.2 Deterministic methods

Deterministic methods rely on the assumption that the probabilistic forms (such as means
and variances) of the outputs of interests can be expressed by approximated functions. This
approximation simplifies the original stochastic problem, reducing the computational require-
ments associated with uncertainty propagation. Therefore, this approximation requires the
formulation of new mathematical problems. The most popular deterministic methods include
the Taylor-based method of moments [103] and the stochastic expansion method [103]. The
former method is based on the representation of the outputs by means of a lower order
Taylor expansion, while the latter one uses Fourier like expansions to approximate the output.
Recently, a novel fast deterministic method, named URQ [32], has been developed and
1 The bias of an estimator, in statistics, is the difference between this estimator’s expected value and the true
value of the parameter being estimated.
48 Robust design optimization

presented. As already mentioned in Chapter 1, this method was used to perform the robust
design of airfoils [32] and wind turbine blades [19, 20] under probabilistic uncertainty. The
rest of this section will be devoted to the presentation of the aforementioned determinis-
tic uncertainty propagation algorithms, namely the Taylor-based method of moments, the
stochastic expansion method and the URQ technique.

Taylor-based method of moments

Taylor-based method of moments represents a generic stochastic output QoI, f (zz), by using a
Taylor expansion with respect to the uncertain input variable set around a reference solution
(e.g., the mean value). The Taylor series of f (zz) developed around the mean of the uncertain
input variable vector µ z can be written as:

n
1 n n
   2 
∂f ∂ f
f (zz) = f (µ
µ z) + ∑ ∆z p + ∑ ∑ ∆z p ∆zq
p=1 ∂ z p 2 p=1 q=1 ∂ z p ∂ zq
1 n n n ∂3 f
 
+ ∑ ∑∑ ∆z p ∆zq ∆zr
6 p=1 q=1 r=1 ∂ z p ∂ zq ∂ zr (3.17)
1 n n n n ∂4 f
 
+ ∑ ∑∑∑ ∆z p ∆zq ∆zr ∆zs
24 p=1 q=1 r=1 s=1 ∂ z p ∂ zq ∂ zr ∂ zs
+O(∆zz5 )

where ∆z p = z p − µz p , and the partial derivatives of f with respect to the input variables are
computed at z = µ z . The remainder of the series is denoted by O(∆zz5 ), including all terms of
order five and higher. Considering independent input variables, µ f and σ 2f can be determined
by substituting Eq. 3.17 into Eq. 3.13 and Eq. 3.14, as follows:
!
1 n ∂2 f
Z +∞
µf = f (zz)pz (zz)dzz = f (µ µ z) + ∑ 2
σz2p
−∞ 2 p=1 ∂ z p
! !
n 3 n 4
1 ∂ f 1 ∂ f
+ ∑ 3
γz p σz3p + ∑ 4
Γz p σz4p (3.18)
6 p=1 ∂ z p 24 p=1 ∂ z p
!
1 n n ∂4 f
+ ∑ ∑ 2 2
σz2p σz2q + O(σ σ 5z )
8 p=1 q=1 ∂ z p ∂ zq
q̸= p
3.4 Methods for uncertainty propagation 49

Z +∞ n  
∂f
σ 2f
= 2
[ f (zz) − µ f ] pz (zz)dzz = ∑ σz2p
−∞ p=1 ∂ z p
! !
n n n
∂2 f ∂3 f
  
∂f 3 ∂f
+∑ γ z p σz p + ∑ ∑ σz2p σz2q
p=1 ∂ z2p ∂ zp p=1 q=1 ∂ z2∂z
p q ∂ zq
q̸= p
! (3.19)
1 n n ∂2 f 1 n ∂3 f
  
∂f
+ ∑ ∑ σz2p σz2q + ∑ Γz p σz4p
2 p=1 q=1 ∂ z p ∂ zq 3 p=1 ∂ z3p ∂ zp
q̸= p
!2
1 n ∂2 f
+ ∑ (Γz p − 1)σz4p + O(σ
σ 5z )
4 p=1 ∂ z2p

where the term σ 5z denotes all monomials of order five or higher. The partial derivatives of f
with respect to the input variables are computed at z = µ z . The skewness and kurtosis of the
variable z p can be defined as:
E[(z p − µz p )3 ]
γz p = (3.20)
σz3p

E[(z p − µz p )4 ]
Γz p = (3.21)
σz4p
The Taylor-based method of moments relies on a model approximation, achieved by
truncating the Taylor series to a lower order. The most common and simple Taylor-based
method of moments used in the literature is first-order2 . This means that only the first-order
derivatives are taken into account. The approximation of the stochastic outputs is associated
with a reduction in computational costs. However, even for relatively small spread of the
input variables, the accuracy of the method may be severely spoiled by nonlinearities in
the system response. To increase the accuracy of this method, one can retain higher-order
derivatives. The computational efficiency of this method depends on the available methods
for calculating derivatives. Several techniques have been devised to calculate derivatives of
the simulation outputs, such as automatic differentiation and the complex variable method.

Stochastic expansion method

Stochastic expansion methods expand the QoIs in a series of random variables. Polynomial
chaos expansion [108] and stochastic collocation [109] are two related techniques belonging
2 Orders of approximation refer to formal or informal terms for how precise an approximation is.
50 Robust design optimization

to this category. Once the series is determined, the statistical moments can be calculated
deterministically. Stochastic expansion methods have been defined in terms of intrusive
and non-intrusive formulations. The latter formulation allows one to use the original code,
treating it as a “black box”. The former formulation instead involves defining a new problem.
Focusing on the non-intrusive version of this category of methods, f , supposed square
integrable, can be written as:

∞ ∞ i1
f (zz) = a0 H0 + ∑ ai1H1(ui1) + ∑ ∑ ai1,i2H2(ui1, ui2)
i1=1 i1=1 i2=1
(3.22)
∞ i1 i2
+ ∑ ∑ ∑ ai1,i2,i3H3(ui1, ui2, ui3) + ...
i1=1 i2=1 i3=1

where u is a vector of standard normal variables (which can be derived from z by resorting to
the Rosenblatt transformation [48]), Hq is a qth-order multidimensional Hermite polynomial,
and the ai are suitable coefficients. For further details on stochastic expansion methods, the
reader can refer to [31, 103, 108, 109]

URQ technique

URQ method is a simplified technique belonging to the quadrature method family [32].
According to this method, µ f and σ 2f can be determined by the following equations:
" #
n f (zz+
p) f (zz−
p)
µ f = W0 f (µ
µ z) + ∑ Wp − (3.23)
p=1 h+
p h−
p

( " #2 " #2
n f (zz+
p ) − f (µ
µ z) f (zz−
p ) − f (µ
µ z)
σ 2f = ∑ Wp+ +Wp−
p=1 h+
p h−
p
) (3.24)
±
[ f (zz+ µ z )][ f (zz−
p ) − f (µ p ) − f (µ
µ z )]
+Wp −
h+
p hp

where n is the number of the uncertain variables or parameters. The sampling points z +
p and
−
z p are found using the following equation:

z± ±
p = µ z + h p σz p e p (3.25)
3.5 Optimization algorithms 51

where e p is the pth vector of the identity matrix of size n, and h±

p are given by:

s
γz p 3γz2p
h±
p = ± Γz p − (3.26)
2 4

The weights appearing in Eqs. 3.23 and 3.24 are determined as follows:
n
1 1
W0 = 1 + ∑ h+p h−p ; Wp = −
p=1 h+
p − hp

(h+ 2 + − (h− 2 + −
p ) − hp hp − 1 p ) − hp hp − 1
Wp+ = − 2 ; Wp− = − 2
(3.27)
(h+
p − hp ) (h+
p − hp )
2
Wp± = +
(h p − h−p)
2

This method requires 2n + 1 function evaluations. Hence, the computational cost is compara-
ble to the first-order Taylor-based method of moments, however, as demonstrated in [32], the
accuracy of URQ is higher.

3.5 Optimization algorithms

The choice of the optimizer depends mainly on the problem characteristics and the type
of solution sought. For problems involving smooth objectives and constraints (i.e., twice
continuously differentiable with respect to the design variables), a gradient-based solver [110]
might be the best suited solution, as they converge to local minima quickly. If a global solution
is desired, multi-start gradient-based solvers [111] should be instead used because they are
able to provide multiple local solutions. Gradient-based solvers, however, may be ineffective
when the problem is not smooth, as they use derivatives to determine the search direction.
In this circumstance, derivative-free methods, such as evolution-based algorithms [112] or
pattern search solvers [113], may be a better choice. This section will introduce derivative-
based algorithms, as well as genetic and pattern search algorithms. In the robust design
optimization framework developed in this work evolution-based and pattern search solvers
(reviewed below) are included.

3.5.1 Derivative-based algorithms

In mathematical optimization, derivative-based or gradient methods are algorithms used to
solve optimization problems by means of gradients. More specifically, these methods define
52 Robust design optimization

the search direction by means of derivatives of the objective function with respect to the
design variables. For these reason the smoothness of the objective function is a mandatory
prerogative when using these methods. One of the simplest gradient methods is the gradient
descent method. This method is a first-order optimization algorithm, used to find a local
minimum of a function using gradient descent. This method starts with an initial guess of the
solution, and then the gradient of the function at that point is taken. The solution is iteratively
approached using the negative direction of the gradients. The descent method converges
eventually where the gradient is zero, which corresponds to the local minimum. This method
is first-order because it considers only the first derivative of the function. Let’s assume we
want to find the minimum of a given objective function f (x). Using the gradient descent
method we give some initial value x0 for x. We then take the gradient (or derivative) of f
with respect to x, denoted by ∇ f . The derivative, calculated at x0 , gives the slope of the
curve at that point, and its direction points to an increase in the function. Therefore, in order
to approach the minimum value of f , we calculate the value of x for the next step moving
towards the opposite direction of the gradient. Hence:

xk+1 = xk − λ ∇ f (xk ) (3.28)

where xk and xk+1 represent the value of x at the iteration steps k and k + 1, respectively.
∇ f (xk ) is the derivative of f with respect to x, calculated at point xk . λ is a number, greater
than zero and normally lower than one, that forces the algorithm to make small jumps,
keeping the algorithm numerically stable. For more details about the numerical stability of
the gradient descent method, one is referred to [114].

In order to better understand the gradient descent method, an example is hereby provided.
Assuming to have the following equation:

f (x) = x2 (3.29)

and to provide an initial guess of the solution x0 = 1. The derivative of f with respect to x,
∇ f , is:
f (x) dx2
∇f = = = 2x (3.30)
dx dx
The value of ∇ f calculated at the point x = x0 is then:

∇ f (x0 ) = 2 · x0 = 2 · 1 = 2 (3.31)
3.5 Optimization algorithms 53

Assuming λ = 0.1, the point for the next iteration step is then computed as follows:

x1 = x0 − λ ∇ f (x0 ) = 1 − 0.1 · 2 = 0.8 (3.32)

In order to calculate the following points x2 , x3 , ..., Eq. 3.31 and Eq. 3.32 are used recursively.
The iterative process continues until the local minimum is achieved (i.e., when the gradient
calculated at the current iteration point becomes zero).

3.5.2 Genetic algorithms

GAs are used for solving both constrained and unconstrained, single- or multi-objective
optimization problems. They are primarily based on natural selection, which is the process
on which the biological evolution is based. A GA, in fact, continuously modifies a population
made up of individual solutions. At each iteration step, the algorithm selects individuals
making up the current population to be parents, and combine them in order to produce the
children for the next generations. The population therefore evolves towards an optimal
solution over successive generations. GAs can be used to solve a wide range of problems, in-
cluding those in which the objective function is discontinuous, non-differentiable, stochastic,
or highly nonlinear. In order to find next generations, the main mechanisms on which the GA
is based include:

• The selection mechanism. This mechanism selects the best individuals form the current
population to be parents, contributing to the definition of the population of the next
generation.

• The crossover mechanism. This mechanism combines two parents to form the children
for the next generation.

• The mutation mechanism. This mechanism applies random changes to individual

parent to form children.

As explained in [115], the GA starts by creating a random initial population. A population is

made up of a number of individuals. Practically, an individual represents a given combination
of the design variables, also called genome. For each of these individuals the fitness (or
objective) function is calculated. The GA then performs a series of modifications of the
individuals’ genomes in order to produce a new population. Each successive population is
called a new generation. In order to create a new generation, the GA performs the following
steps:

• Calculate the objective function of each individual of the current generation;

54 Robust design optimization

• Score these individuals according to their objective function values.

• Select some members of the current generation, called parents. This selection is based
on the individuals’ objective function values.

• Some of the best members, called elite, that have the lowest value of the objective
function, are selected. Elite members are passed without modification to the next
generation.

• Produce children from the parents. Children are produced either by making random
changes to a single parent (i.e., mutation), or by combining the genomes of a pair of
parents (i.e., crossover).

• Replaces the current population with the children to form the next generation.

The iterative process normally proceeds until the GA is not able to improve the objection
function for a given user-defined number of consecutive generations.
GAs have been extensively used to solve optimization problems with inequality con-
straints, such as that given by Eq. 3.1. The penalty function approach has been often used in
conjunction with GAs due to its simplicity. Penalty functions consist of penalty parameters
added to the objective function, giving a measure of violation of the constraints. The measure
of violation is nonzero when the constraints are violated and is zero in the region where
constraints are not violated. In particular, the larger the constraint violation, the higher
the penalty parameter. In this way, the optimizer is able to determine the search direction
that could lead the solution to regions where the constraints are not violated. The use of
penalty functions, however, has several functional and operational drawbacks. For example,
penalty functions do not take into account the mathematical form of the constraints. More-
over, the addition of penalty terms to the objective functions can create distortions (such
as discontinuities) in the objective functions. Therefore, more sophisticated methods have
been devised to treat inequality constraints in the context of GAs. One of these methods is
the augmented Lagrangian genetic algorithm [116], belonging to the family of augmented
Lagrangian algorithms [117]. These algorithms attempt to solve optimization problems with
inequality constraints by formulating minimization subproblems. In such subproblems, a
function depending on the objective function, the constraint functions and penalty parameters
is minimized, using the genetic algorithm, such that all the constraints are verified. Practically,
a minimization subproblem is carried out for each individual that does not satisfy one or
more inequality constraints. This subproblem aims to satisfy all constraints by varying the
values of the design variables of the solution under consideration. The subproblems stop
when all constraints are verified. When using the augmented Lagrangian genetic algorithm,
3.5 Optimization algorithms 55

the number of function evaluations per generation is therefore much higher (and normally
unknown) than that required when treating inequality constraints through penalty functions.

3.5.3 Pattern search algorithms

Pattern search algorithms belongs to the family of direct search algorithms. Like GAs, these
methods do not require information regarding the gradient of the objective function and
constraints, and therefore they can be used to optimize problems the objective function and
constraints of which are not differentiable or non-continuous. Pattern search algorithms
approach an optimal solution computing a sequence of points. At each iteration step, the
algorithm generates a set of points, called the mesh, around the current point, which is the
result of the previous iteration step. The mesh is determined by adding the current point
to a scalar multiple of a set of vectors called a pattern. If a point in the mesh improves the
objective function of the current point, then that point becomes the current point at the next
iteration step. The mechanism of generating the mesh is named polling. After a polling, the
algorithm changes the value of the mesh size. In case of a successful polling the mesh size
contracts, while if the polling has not found a better point the mesh size expands. Like GAs,
pattern search algorithm can handle inequality constraints by means of augmented Lagrangian
algorithms, such as the augmented Lagrangian pattern search [118]. This algorithm solves
optimization problems with inequality constraints by formulating minimization subproblems
such that all the constraints are verified.
In order to show how this algorithm works, a simple example (adapted from that in [115])
is given. Let’s assume we want to optimize a given system by minimizing an objective
function, f , depending on the design variables x = {xa , xb }. Considering an unconstrained
problem, this optimization can be formulated as follows:

Find: x = {xa , xb }
(3.33)
to minimize: f (xx)

Pattern search finds a sequence of points x 0 , x 1 , x 2 , ..., that approach an optimal solution.
The value of the objective function either decreases (i.e., improves) or remains the same
from each point in the sequence to the next. The pattern search begins at the initial point x 0
provided. In this example, xa is equal to 2.1 and xb is equal to 1.7. Therefore the initial point
can be expressed as x 0 = {2.1, 1.7}. At this point, the value of the objective function is equal
to 4.6381. At the first iteration, the mesh size is 1 and the algorithm adds the pattern vectors
56 Robust design optimization

to the initial point x 0 to compute the following four mesh points:

[1 0] + x 0 = [3.1 1.7]
[0 1] + x 0 = [2.1 2.7]
(3.34)
[−1 0] + x 0 = [1.1 1.7]
[0 − 1] + x 0 = [2.1 0.7]

The algorithm then computes the objective function at these mesh points, as show in Fig. 3.1.
After the first polling, point [1.1 1.7] (depicted in Fig. 3.1) improves the value of the

3
initial point x0
6.6347 mesh points

2.5

4.1146 4.6381 4.582

1.5
polled point which improves the objective function the most

5.6347

0.5
1 1.5 2 2.5 3 3.5

Fig. 3.1 Objective function values at initial point and mesh points after the first polling.

objective function the most (with respect to that of x 0 ), achieving a value of 4.1146. In this
case, the first polling is successful. The pattern search algorithm then sets the current point
for the following iteration equal to x 1 = {1.1, 1.7}. After a successful poll, the current mesh
size is multiplied by a factor of 2 (i.e., the mesh size is doubled). Therefore, because the
initial mesh size was 1, at the second iteration the mesh size is 2. Thus, the mesh for the
3.5 Optimization algorithms 57

second pooling will be:

2 · [1 0] + x 1 = [3.1 1.7]
2 · [0 1] + x 1 = [1.1 3.7]
(3.35)
2 · [−1 0] + x1 = [−0.9 1.7]
2 · [0 − 1] + x1 = [1.1 − 0.3]
As seen before, the algorithm then calculates the values of the objective function at each of
the points making up the current mesh. The mesh point that improves the objective function
the most will become the current point, and the mesh size will be doubled. In case that
none of the mesh points can improve the objective function, the poll is unsuccessful. In this
case, the algorithm does not change the current point at the next iteration, and the algorithm
multiplies the current mesh size by 0.5 (i.e., the mesh size is halved). The iterative process
continues until the mesh size becomes smaller than a given user-defined tolerance.
Chapter 4

Multidisciplinary and robust design

optimization framework

The research reported in this thesis focuses on the development of a computational framework
which combines the multidisciplinary and the robust design optimization of HAWT rotors.
The adopted design optimization approach uses a FIO architecture, based on an integrated
aeroservoelastic analysis model that analyzes the aerodynamic and structural characteris-
tics of the rotor, directly incorporating the control strategy, and assesses the economical
performance of the rotor by means of a cost model. The analysis model is coupled with
the optimizer through the uncertainty propagation algorithm, needed to propagate the input
uncertainties throughout the design system, characterizing the objective function and con-
straints probabilistically. The rest of this chapter deals with the description of the developed
design framework, highlighting the main features of its various parts. The presented design
system has been used to perform the optimization of a multi-megawatt HAWT subject to
geometric and environmental uncertainties. The environmental uncertainties considered in
this work refer exclusively to the mean value of the wind speed frequency distribution. The
results of the environmental uncertainty application will be reported in Chapter 6, while the
results of the geometric uncertainty application have been published in [19].

4.1 Fully integrated multidisciplinary and robust design

optimization framework
An overview of the developed fully integrated multidisciplinary and robust design optimiza-
tion framework is schematically shown in Fig. 4.1. As depicted in this figure, the design
variables, b , generated by the optimizer, are directly passed on to the uncertainty propagation
60 Multidisciplinary and robust design optimization framework

OPTIMIZATION
Optimization
Optimization
algorithm
algorithm

UNCERTAINTY PROPAGATION
b = (b1, …, bn) Uncertainty F(b, p, z), Gj(b, p, z)
Uncertainty propagation
propagation
algorithm
algorithm

f(b+zb, p+zp), gj(b+zb, p+zp)

Aerodynamics
Aerodynamics Control
Control

Structure
Structure Cost
Cost

INTEGRATED MULTIDISCIPLINARY ANALYSIS MODEL

Fig. 4.1 Overview of the fully integrated multidisciplinary and robust design optimization
framework. b represents the vector of the design variables, bi , i = 1, n, p is the vector
of the design parameters, pi , i = 1, m, while z b and z p are the vector of the uncertainties
associated with the design variables and parameters, respectively. F(bb, p , z ) and G j (bb, p , z )
are the probabilistic definition of the objective function and constraints, respectively. f (bb +
z b , p + z p ) and g j (bb + z b , p + z p ) represent the stochastic objective function and constraints,
respectively.

module. This module then evaluates the probabilistic form of the objective function, F, and
constraints, G j . F and G j depend on the design variables, the design parameters, p , and
the uncertainties associated with the design variables and parameters, denoted by z b and
z p , respectively. As discussed in Chapter 3, F and G j are generally defined in terms of the
statistical moments of the stochastic objective function and constraints, respectively denoted
by f and g j . Practically, the probabilistic definition of the objective function and constraints
is determined by the uncertainty propagation algorithm by means of a given number of
deterministic analyses of the system. In each of these analyses, f and g j are determined
by means of the integrated multidisciplinary analysis model described in Sect. ??. Once
the probabilistic objective function and constraints are determined, they are sent back and
assessed by the optimizer. The iterative process continues until convergence is reached.
4.2 Integrated multidisciplinary analysis model 61

4.2 Integrated multidisciplinary analysis model

The optimization problems reported in Chapter 6 were formulated considering the LCOE as
performance metric. Such optimizations included constraints on blade deformations, normal
and buckling stresses, fatigue damage and blade natural frequencies. A detailed description
of the adopted objective function and constraints will be given in Chapter 6.
The assessment of the aeroservoelastic behavior of HAWT rotors, and thus the evaluation
of the design objective function and constraints required the integration of several simulation
tools, shown schematically in Fig. 4.2. These computational tools were linked together

XFOIL
XFOIL
AeroDyn
AeroDyn
AERODAS
AERODAS BModes
BModes

AIRFOIL Control
Control FAST
FAST
AERODYNAMICS
Co-Blade
Co-Blade

Cost
Cost MLife
MLife

COST MODEL ROTOR AEROSERVOELASTICITY

INTEGRATED MULTIDISCIPLINARY ANALYSIS MODEL

Fig. 4.2 Overview of the integrated multidisciplinary analysis model, showing the computa-
tional tools used and their mutual interactions.

in a MATLAB environment, which managed the data flow between all the codes. The
multidisciplinary analysis framework for the integrated rotor design combines three main
modules, encompassing: a) a module for the aerodynamic analysis of the blade airfoils, b) a
module for the aeroservoelastic analysis of the rotor, and c) a cost model for the assessment
of the economical performance of the rotor. All modules of the multidisciplinary analysis
system are presented below, describing in details the inputs and outputs of the integrated
computational tools.
62 Multidisciplinary and robust design optimization framework

4.2.1 Airfoil aerodynamics

In this study, the aeroelastic analysis of the rotor was based on the BEM theory, which requires
the airfoil aerodynamic coefficients. As seen in Chapter 2, 2D aerodynamic characteristics of
airfoils can be determined using panel codes, such as XFOIL and RFOIL, or by means of
CFD. Run-times of CFD, even in 2D simulations, are still excessive for their use in design
optimizations, and RFOIL is not an open-source tool and could thus not be used in this
study. Therefore, in this research work, 2D airfoil lift and drag coefficients were calculated
using the airfoil analysis code XFOIL. In this code, transition from laminar to turbulent flow
along the airfoil is simulated by the eN method, through the parameter NCRIT. A suitable
value of this parameter depends on the environmental disturbance1 level in which the airfoil
operates, and mimics the effect of such disturbances on laminar-to-turbulent transition. For
all optimizations carried out in this research work, NCRIT was fixed to the standard value
of 9, corresponding to average wind tunnel conditions. The 2D aerodynamic data thus
calculated were then corrected to account for the complex 3D physics occurring over rotating
blades, in both pre- and post-stall regions, by means of AERODAS.

As shown in Fig. 4.3, XFOIL was used to compute 2D airfoil lift and drag coefficients
for a Re ranging from 2 · 106 to 14 · 106 , over values of the AoA ranging from -6 to 26 deg.
2D lift and drag coefficients were then fed into AERODAS, evaluating the extended 3D lift
and drag coefficients. AERODAS polars for AoA ranging from -180 to 180 deg were derived
from a small number of parameters of the XFOIL polars. These parameters were the zero-lift
AoA, the maximum lift and drag coefficients, the AoA at maximum lift and drag, the slope
of the linear part of the lift curve, and the minimum drag coefficient.

In order to provide a practical example, Fig. 4.4 shows the comparison between the
experimental polars of the 21-percent-thick S809 airfoil [119], and those determined by
means of XFOIL and AERODAS for the same airfoil, for a Reynolds number of 1 · 106 . It
is noted that the XFOIL polars match the experimental ones up to 9 deg. At AoAs above
9 deg, XFOIL over-predicts the lift coefficient, while it under-predicts the drag coefficient.
Figure 4.4 also shows that the AERODAS correction takes into account an indicative blade
aspect ratio of 15.3 by lowering the lift curve with respect to that obtained for an infinite
blade aspect ratio.

1 In
this context, environmental disturbance is referred to both acoustic (sound) and vortical (turbulence)
disturbances.
4.2 Integrated multidisciplinary analysis model 63

XFOIL
XFOIL

2D
2D airfoil
airfoil
lift
lift and
and drag
drag coefficients
coefficients

extended
extended 3D
3D airfoil
airfoil
AERODAS
AERODAS lift
lift and
and drag
drag coefficients
coefficients

AIRFOIL
AERODYNAMICS

Fig. 4.3 Details of the module for the aerodynamic analysis of the blade airfoils. Blue boxes
indicate the computational codes integrated in this module, while yellow boxes represent
their inputs/outputs.

4.2.2 Rotor aeroservoelasticity

The aeroservoelastic simulations of the turbine rotor were performed by means of FAST,
which was linked to an external library defining the control system. A detailed description of
the control strategy implemented in the aeroservoelastic model for this work is reported in
Chapter 5. For a given rotor geometry and wind conditions, FAST was used to predict the
rotor speed and power, as well as blade mass and ultimate and fatigue loads. FAST inputs
include distributed blade properties, such as the blade section mass per unit length and blade
section stiffness [40]. The blade mass per unit length is then integrated along the blade length
in order to determine the total blade mass. As shown in Fig. 4.5, using the aerodynamic
characteristics of the airfoils determined by XFOIL and AERODAS, the rotor aerodynamics
was analyzed with AeroDyn, a library implementing the BEM theory and several corrections,
including the Prandtl’s correction to account for tip and hub losses, and the correction
accounting for axial induction factors exceeding the maximum theoretical limit of 0.5. In the
analyses reported in this thesis, however, Prandtl’s correction was disabled because the 2D
airfoil data were already corrected to account for finite aspect ratio effects in AERODAS.
FAST employs a linear modal representation to characterize blade flexibility. Blade modes
depend on the rotor speed, blade geometry and span-variant structural properties, such as
mass, moments of inertia, and stiffnesses. Such blade span-variant structural properties
64 Multidisciplinary and robust design optimization framework

1.6
XFOIL
1.4 AERODAS (AR = ∞)
AERODAS (AR = 15.3)
Exp
1.2

0.8
CL, CD

0.6

0.4

0.2

−0.2

−0.4
−5 0 5 10 15 20 25 30
AoA [deg]

Fig. 4.4 Comparison between the S809 aifoil experimental polars and those determined
by means of XFOIL and AERODAS, for a Reynolds number of 1 · 106 . The AERODAS
correction was implemented by using an infinite blade aspect ratio and an indicative aspect
ratio equal to 15.3.

were computed by means of the structural analysis code Co-Blade, and subsequently fed
into the preprocessor BModes in order to determine the blade modes. BModes was also
used to determine the blade natural frequencies. The structural properties and blade modes
determined by means of Co-Blade and BModes were then used by FAST to perform the
aeroservoelastic calculations, determining the blade ultimate and fatigue loading, as well
as the rotor power curve and blade mass. During such aeroelastic calculations, FAST was
coupled with a control routine, defining the rotor control strategy. The obtained ultimate
blade loads were then assessed by means of Co-Blade, which was used to compute blade
deformations and material normal and buckling stresses. The fatigue loads were instead
post-processed by MLife, providing the resulting total fatigue damage.

4.2.3 Cost model

LCOE is the figure of merit used to assess the economical performance of each design, and
the LCOE model adopted herein is based on the DOE/NREL scaling model. According to
4.2 Integrated multidisciplinary analysis model 65

extended
extended 3D
3D airfoil
airfoil
lift
AeroDyn
AeroDyn
lift and
and drag
drag coefficients
coefficients blade
blade natural
natural
frequencies
frequencies
rotor
rotor aerodynamics
aerodynamics

blade
blade modes
modes BModes
BModes
blade
blade span-variant
span-variant
Control
Control control
control strategy
strategy FAST
FAST structural
structural properties
properties

blade
blade deflections
deflections
power
power curve,
curve,
blade
ultimate
ultimate loads
loads Co-Blade
Co-Blade buckling
buckling stresses
stresses
blade mass
mass
normal
normal stresses
stresses

fatigue
fatigue loads
loads MLife
MLife fatigue
fatigue damage
damage

ROTOR AEROSERVOELASTICITY

Fig. 4.5 Details of the module for the aeroservoelastic analysis of the rotor. Blue boxes
indicate the computational codes integrated in this module, while yellow boxes represent
their inputs/outputs. Green boxes instead represent the constraints.

this model, the LCOE of an offshore HAWT is given by Eq. 2.35. Although only the rotor
design is considered in the optimization study of this thesis, the LCOE of each design refers
to the cost of the whole turbine, including drive train, nacelle, tower and foundations. This is
accomplished by including in ICC the cost of all turbine components, which results in this
variable being the sum of a term proportional to the blade mass varying with the particular
rotor configuration, and other terms referring to the costs of all other components [102]. In the
LCOE equation, also AEP varies for each rotor being analyzed. AEP is defined as the amount
of electricity that a wind turbine can generate in a year at a given site. Therefore, the AEP
of a wind turbine depends on its power curve, as well as on the annual wind characteristics
of the site in which it has been installed. At a specific site, wind speed variations during
one year are described in terms of a PDF, that define the so-called wind speed distribution.
Figure 4.6 shows the power curve of a typical multi-megawatt HAWT, and the indicative
wind speed distribution of a generic site. The power curve defines the electrical power Pe
that a turbine can generate at a given wind speed u, while the wind speed PDF, pu , describes
the probability for the wind speed to take on a given value during one year.
66 Multidisciplinary and robust design optimization framework

6000 0.12

5000 0.1

4000 0.08
Pe [kW]

pu [−]
3000 0.06

2000 0.04

1000 0.02

0 0
0 5 10 15 20 25
u [m/s]

Fig. 4.6 Typical multi-megawatt HAWT power curve (in black) and indicative wind speed
distribution of a generic site (in red).

The power curves of all the HAWT configurations analyzed in this work were evaluated
by means of FAST. More specifically, the power curve, Pe (u), of a given HAWT was
reconstructed by running FAST under increasing discrete hub height wind speeds, from
cut-in to cut-out. These wind conditions were represented by AeroDyn hub height wind files,
defining steady wind speeds, perpendicular to the rotor plane, and characterized by a 1/7
power law profile. The power law shear profile is used to determine the wind speed, uZ , at
any height, Z, based on a reference height, Zr , and the wind speed at the reference height, ur ,
using the following equation:  α p
Z
uZ = ur (4.1)
Zr
where the exponent α p is an empirical coefficient that dependents on a large number of
variables including atmospheric stability, temperature, terrain surface roughness, changes
in the terrain surface conditions and terrain shape. A review of methods for determining
representative power law exponents can be found in [42]. In this work, neutral stability
conditions were considered, corresponding to a value of α p equal to 1/7 (representing indeed
a 1/7 power law profile).

The wind speed distribution of a given site can be well characterized by the Weibull
distribution. The Weibull distribution is characterized by two parameters: the shape parameter
kW (dimensionless) and the scale parameter cW (expressed in m/s). The PDF of the Weibull
4.3 Uncertainty propagation 67

distribution is given by:

kW −1 "   #
u kW

kW u
pu = exp − (4.2)
cW cW cW

The average wind speed of the Weibull distribution can be expressed as:
 
1
ū = cW Γ 1 + (4.3)
kW

where Γ is the gamma function given by:

Z ∞
Γ(n) = e−x xn−1 dx (4.4)
0

In this research, kW was always set equal to 2, corresponding to the Rayleigh distribution.
As given by Eq. 4.5, from cut-in wind speed ucut-in to cut-out wind speed ucut-out , AEP
can be calculated by integrating the turbine power curve, Pe (u), against the wind speed
distribution PDF, pu . The value of AEP is one of the inputs required by the cost model
represented by Eq. 2.35.
Z ucut-out
AEP = 8670 · Pe (u)pu du (4.5)
ucut-in

where 8670 is the number of hours in one year.

As depicted in Fig. 4.7, in the presented design framework, the power curve and the blade
mass, evaluated by means of FAST, were used as input values for the cost model, eventually
evaluating the LCOE.
The costs of the turbine components indicated in the DOE/NREL scaling model are based
on 2002 dollars. Therefore, these costs need to be updated to account for the year-over year
change in products, material and labor costs. Cost escalation was based on the producer price
index (PPI). The PPI, maintained and updated on a monthly basis by the U.S. Department
of Labor, Bureau of Labor Statistics, tracks the price change of a wide range of product (or
commodity) categories. Commodity escalation factors were determined by using the PPI
categories provided by the DOE/NREL scaling model for each turbine component.

4.3 Uncertainty propagation

In the probabilistic design framework presented herein, the uncertainty propagation algorithm
was implemented in a MATLAB computational interface, connecting the optimizer to the
68 Multidisciplinary and robust design optimization framework

power
power curve,
curve,
LCOE
LCOE Cost
Cost blade
blade mass
mass

COST MODEL

Fig. 4.7 Details of the cost model.

analysis model described in Sect. 4.2. The optimization strategies integrated in the presented
design system will be described in Sect. 4.4. This section focuses on the techniques used
to perform uncertainty propagation. During the optimization process, at each iteration, the
optimizer generates a vector of design variables, defining a given HAWT configuration to
be assessed probabilistically. Subsequently, the uncertainty propagation algorithm, using
the design variables and the input uncertainties, determines the probabilistic form of the
objective function and constraints. In this work, the objective function and constraints are
defined probabilistically by calculating their mean and standard deviation. Within a robust
design optimization framework, uncertainty propagation is a computationally onerous step,
usually involving a high number of analyses of HAWT configurations. Therefore, the choice
of the uncertainty propagation algorithm is a critical task, involving a trade-off between
accuracy and computation time.
In the robust design optimization reported in [19], the uncertainty affecting 10 geometric
parameters of HAWT blades were considered, encompassing blade thickness, chord and twist
distributions. In this work, uncertainty propagation was efficiently achieved by using the
URQ approach. According to this method, for each HAWT configuration, 21 rotor analyses
were necessary to determine the mean and the standard deviation of the objective function
and constraints. In this particular application, nonlinearities in the system response were not
4.3 Uncertainty propagation 69

too important, and therefore, under these circumstances, the accuracy of the URQ approach
was not affected.
In the design optimization under environmental uncertainty reported in Chapter 6, a
single uncertain variable was taken into account in the robust optimization process of a
multi-megawatt HAWT, namely the mean value of the wind speed distribution, denoted by
the symbol ū. In this optimization problem, the mean and the standard deviation of LCOE,
denoted respectively by µLCOE and σLCOE , were minimized concurrently. The stochastic
variable ū was considered uniformly distributed within a given range of variability from ūL
to ūU . Figure 4.8 shows the uniform distribution of ū, denoted by pū , spanning the range of ū
from ūL to ūU . The probability of the random variable falling within a particular range of

pū

1/(ūU-ūL)

ūL ū1 ū2 ūi ūm-1 ūm ūU ū

Fig. 4.8 Uniform PDF of the uncertain variable, ū, denoted by pū , defined within the lower
bound, ūL , and the upper bound, ūU . The range of variability of ū was divided in m intervals
with equal probability for subsequent Latin hypercube sampling.

values is given by the integral of this variable’s PDF over that range, that is, it is given by the
area under the PDF but above the x-axis and between the lowest and greatest values of the
range. The area under the PDF of ū between ūL to ūU is equal to 1 (all values of ū falls within
ūL to ūU ). In this robust optimization problem, the uncertainty propagation was accomplished
by using the Latin hypercube sampling. As shown in Fig. 4.8, the range of variability of the
input stochastic variable was divided in m intervals with equal probability. Therefore, each
70 Multidisciplinary and robust design optimization framework

of these intervals is characterized by the same height and thickness. Subsequently, the metric
LCOE was sampled for ūi , i = 1, m, where ūi is the center value of the i-th interval. Once
LCOE was calculated for all m sampling points, its mean and the standard deviation were
computed as follows:
1 m
µLCOE = ∑ LCOE(ūi ) (4.6)
m i=1
s
1 m
q
2
σLCOE = σLCOE = ∑ [LCOE(ūi) − µLCOE]2
m − 1 i=1
(4.7)

4.4 Optimization
The optimizations carried out in this work aimed to find a global minimum over a large
design space. As explained in Chapter 6, some of the solutions included in this space are
considered geometrically infeasible or prevent some of the analysis codes from converging
(e.g., XFOIL fails to converge for some airfoil shapes at given flow conditions). In this
work, these solutions are treated with penalty functions. The aerostructural constraints of the
optimization problems, also described in Chapter 6, are treated in the same way. As seen in
Chapter 3, the use of penalty functions introduces discontinuities in the objective function,
and thus affect its smoothness. For this reason, as explained in Chapter 3, gradient-based
algorithms could not be used in this study, as they use derivatives to determine the search
direction. Therefore, in the framework of this research, gradient-free codes have been used,
namely the multi-objective parzen-based estimation of distribution (MOPED) [120] and
the inflationary differential evolution algorithm (IDEA) [121], two GAs, and the MATLAB
patternsearch function [115], based on a pattern search algorithm. MOPED and IDEA codes
were used to perform the probabilistic optimization under geometric uncertainty reported
in [19], while the MATLAB patternsearch function was used to carry out the probabilistic
optimization under environmental uncertainty described later on in this thesis, in Chapter 6.

4.4.1 MOPED and IDEA

As mentioned in Chapter 3, evolutionary algorithms solve optimization problems by making
a generation of individuals evolve subject to selection and search operators. In this study, an
individual denotes a given HAWT configuration. This iterative process eventually leads to
a population containing the fittest possible individuals (best HAWT configuration designs),
or individuals who are significantly fitter than those of the starting population. The role of
the selection operators is to identify the fittest or most promising individuals of the current
4.4 Optimization 71

population, whereas search operators such as crossover and mutation attempt to generate
better offspring starting from suitably selected individuals of the current generation. Each
individual is defined by genes, which correspond to design variables in design optimization.
The solution of the optimization problems reported in [19] is based on a two-stage approach
using MOPED and IDEA.
MOPED belongs to a subset of evolutionary algorithms and was developed to circum-
vent certain algorithmic problems of conventional evolutionary algorithms, which can be
ineffective when the problem at hand features a high level of interaction among the design
variables. This is mainly due to the fact that the recombination operators are likely to disrupt
promising sub-structures that may lead to optimal solutions. Additionally, the use of the
crossover and mutation operators may result in slow convergence to the solution of the opti-
mization; that is, it may require a large number of generations to obtain very fit individuals.
MOPED was developed to circumvent shortfalls of this kind. Its use of statistical tools
enables it to preserve promising sub-structures associated with variable interaction from
one generation to another (automatic linkage learning). Such statistical tools also replace
the crossover and mutation operators of conventional evolutionary algorithms, and they
allow a faster convergence of MOPED with respect to the latter class of optimizers. Starting
from the individuals of the current population, MOPED builds an approximate probabilistic
model of the search space. The role of the crossover and mutation operators is replaced by
sampling of this probabilistic model. There exist other similar evolutionary methods that
use the aforementioned strategy, and they are called estimation of distribution algorithms
(EDAs) [122]. MOPED is a multi-objective optimization EDA for continuous problems that
uses the Parzen method [123] to build a probabilistic representation of Pareto solutions, and
can handle multivariate dependencies of the variables [120, 124]. MOPED implements the
general layout and the selection techniques of the non-dominated sorting genetic algorithm
II (NSGA-II) [125], but traditional crossover and mutation search approaches of NSGA-II
are replaced by sampling of the Parzen model. NSGA-II was chosen as the base for MOPED
mainly due to its simplicity, and also for the excellent results obtained for many diverse
optimization problems [126, 127].
The Parzen method utilizes a non-parametric approach to kernel density estimation, and
results in an estimator that converges asymptotically to the true PDF over the whole design
space. Additionally, when the true PDF is uniformly continuous, the Parzen estimator can
also be made uniformly consistent. The Parzen method allocates Nind identical kernels (where
Nind is the number of individuals of the current population), each centered on a different
element of the sample. A probabilistic model of the promising search space portion is built
on the basis of the statistical data provided by the Nind individuals through their kernels, and
72 Multidisciplinary and robust design optimization framework

τE Nind new individuals (τE ≤ 1) are sampled. The variance of each kernel depends on (i) the
location of the individuals in the search space and (ii) the fitness value of these individuals,
and its construction leads to values that favour sampling in the neighborhood of the most
promising solutions.
The features of MOPED often prevent the true Pareto front from being achieved, particu-
larly when the front is broad and the individuals of the population are spread over different
areas, which are far apart from each other in the feasible space. This circumstance has
prompted coupling MOPED with another evolutionary algorithm, which has better conver-
gence properties. To this aim, IDEA was selected. IDEA was first developed for the design
optimization of interplanetary trajectories, and it is an improved variant of the differential evo-
lution (DE) algorithms [121]. The IDEA algorithm is based on a synergistic hybridization of
a standard DE algorithm and the strategy behind the monotonic basin hopping (MBH) [128].
The resulting algorithm was shown to outperform both standard DE optimizers and the
MBH algorithm in the solution of challenging space trajectory design problems, featuring a
multiple funnel-like structure. In this paper, a modified version of IDEA was used to move
the individuals of the approximate Pareto front obtained with MOPED closer to the true
front.
The main features of the original IDEA algorithm are reported in [121]. The IDEA
algorithm works as follows: a DE process is performed several times and each process is
stopped when the population contracts below a predefined threshold. At the end of each
DE step, a local search is performed in order to get closer to the local optimum. In the case
of non-trivial functions, where there is a high likelihood of converging to local optima, the
combined DE/local search is usually iterated several times, performing either a local or a
global restart on the basis of a predefined scheduling.
In the current implementation, MOPED and IDEA are used sequentially. When MOPED
has reached a given number of generations, its final population represents a first and good
approximation to the sought Pareto front. Then, clustered sub-populations of such a popu-
lation are used as initial solutions of the single-objective constraint IDEA optimizer. this
algorithm “pushes” the individuals of a sub-population of the MOPED front towards a better
local approximation of the sought Pareto front. The resulting two-stage optimizer blends
the exploratory capabilities of MOPED (global exploration) and the favorable convergence
characteristics of IDEA (exploitation of local information).

4.4.2 MATLAB patternsearch function

The MATLAB pattern search algorithm was used to carry out the optimizations reported
in Chapter 6. This function was parallelized over 32 processors through the MATLAB
4.4 Optimization 73

distributed computing server tool [115]. As seen in Chapter 3, a pattern search algorithm
searches the global minimum of a function by using a set of points, called pattern, which
expands or shrinks depending on whether any point within the pattern has a lower objective
function value than the current point. As a starting point for the optimizations carried out in
this study, a suitable reference turbine, described in Chapter 6, was used. For more details
about the patternsearch function and the distributed computing server tool used, one can refer
to the MATLAB documentation [115].
Chapter 5

Design space and reference turbine

definition

The aim of an optimization process is normally to improve the performance of a given system
by optimizing a specified set of parameters without violating some constraints. Therefore, a
preliminary step in design optimization is to characterize the system under consideration in
terms of numerical parameters, defining selected properties that drive the optimization. The
set of these parameters, denoted by design variables, and the numerical bounds in which they
can vary during the course of the optimization define the design space. In this work, the blade
external geometry, the rotor speed and the blade internal structural layup of a multi-megawatt
HAWT were optimized concurrently, and therefore these wind turbine elements needed to
be parametrized. The aim of this chapter is twofold. Firstly, it presents the parametrization
of the turbine’s blade external geometry, rotor speed and internal structural layup needed
for the subsequent optimization. Secondly, it defines a reference turbine by means of the
developed parametrization. The reference turbine is based on the NREL offshore 5-MW
baseline wind turbine [36]. The design specifications of this turbine are reported in [36]. As
explained below, the definition of a reference turbine was required to determine an opportune
starting point for the optimizer, and to provide a metric against which the effectiveness of the
optimizations could be assessed.

5.1 Turbine parametrization

This section deals with the parametrization of the turbine’s blade external geometry, rotor
speed and internal structural layup.
76 Design space and reference turbine definition

5.1.1 External geometry parametrization

The 3D external geometry of each rotor blade considered in the optimization process was
determined by the airfoil schedule, and the chord and twist distributions along its span. The
airfoil schedule was defined by 2 circular sections, respectively at 0% and 6.67% blade length,
and the geometry of 3 airfoils, namely that at 16.67% blade length (root airfoil), 50% blade
length (midspan airfoil) and 100% blade length (tip airfoil). The geometry of the sections
between the root and tip radii was obtained with linear interpolations. As reported below,
the geometry of the interpolated sections was needed to determine the structural properties
of the blade by means of Co-Blade. Root, midspan and tip airfoils were treated as part
of the optimization, as well as blade chord and twist distributions. In order to reduce the
computational burden associated with the optimization process, the rotor geometry ideally
should be parametrized by means of a small number of parameters, yet allowing for an
extensive flexibility needed to describe any shape with good accuracy. The rest of this section
deals with the description of the adopted parametrizations for the airfoil shapes and chord
and twist distributions.

Airfoil parametrization

The geometry of the three base airfoils is parametrized by means of a composite Bézier
curve based on 15 points, and made up of two 3rd-order Bézier curves and two 4th-order
Bézier curves. A parametrization based on Bézier curves was chosen because, as shown in
published applications such as those reported in [15, 18], it allows one to use a relatively
small number of parameters, yet allowing for a high flexibility needed to describe any airfoil
shape with good accuracy. As mentioned in Chapter 1, many other airfoil parametrizations
with attractive properties are available. However, in this work, only Bézier curves were
taken into account, as they were able to fulfill the aforementioned requirements. The airfoil
parametrization is illustrated in Fig. 5.1. The rear part of the airfoil upper side is based on
points 1, 2, 3, 4, and is defined by a 3rd order Bézier curve; the front part of the upper side is
based on points 4, 5, 6, 7, 8, and is defined by a 4th order Bézier curve; the front part of the
lower side is based on points 8, 9, 10, 11, 12, and is defined by a 4th order Bézier curve; the
rear part of the lower side is based on points 12, 13, 14, 15, and defined by a 3rd order Bézier
curve. Position (C0 ), tangent (C1 ) and curvature (C2 ) continuity was enforced at the junction
points 4, 8 and 12. In the optimization process, points 4, 5, 6, 7, 11 and 12 were actively
controlled in both x- and y-directions, while points 1 and 15 were controlled only along the
y-direction. This corresponds to 14 degrees of freedom. The leading edge position (point
8) is fixed at (x = 0, y = 0). Once the position of points 4, 5, 6, 7, 11 and 12 is assigned,
5.1 Turbine parametrization 77

the coordinates of points 2, 3, 9, 10, 13 and 14 are determined by enforcing C0 , C1 and C2

continuity at the junctions 4, 8 and 12. Thus, each of the 3 base airfoils is described by 14
independent variables, and therefore 42 design variables, labeled bi , i = 1, 2, ..., 42, are used
to define the blade airfoils from root to tip. For more details about composite Bézier curves
and their continuity, one is referred to [129, 130].

Detail of the trailing edge

0.8
0.015

0.01
1
0.6 0.005

−0.005
0.4 15
−0.01
y/c [−]

5 −0.015
6 4 3
0.2 2 0.99 1 1.01

7
0 8
14
9

−0.2
10 13
12
11
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
x/c [−]

Fig. 5.1 Airfoil parametrization. The airfoil shape is defined by a composite Bézier curve
controlled by 15 points. Horizontal and vertical arrows denote the actual degrees of freedom.

Parametrization of chord and twist distributions

The radial profiles of blade chord and twist angle were parametrized by means of a shape
preserving piecewise cubic interpolation technique implemented in the pchip MATLAB
function, using 5 control points for the chord distribution and 4 control points for the twist
distribution. The only requirement associated to the parametrization of the chord and twist
78 Design space and reference turbine definition

angle parametrization was the continuity of the first derivative of the curve connecting
the control points. The pchip MATLAB function was chosen because it is able to fulfill
the first derivative continuity requirement, and available in the MATLAB package used to
carry out the work reported in this thesis. The control points of the chord profile 1, 3, 4,
5 are fixed at 0%, 50%, 83.33% and 100% blade span, respectively, whereas point 2 is
actively controlled also in the radial direction. Hence, 6 independent design variables, labeled
bi , i = 43, 44, ..., 48, are used to define the chord profile. The active degrees of freedom
for the parametrization of the chord profile are reported in the left subplot of Fig. 5.2. The
control points of the twist profile 1, 2, 3, 4 are fixed at 16.67%, 50%, 83.33% and 100%
blade span, respectively. Hence, 4 independent design variables, labeled bi , i = 49, 50, ..., 52,
are used to define the twist profile. The active degrees of freedom for the parametrization of
the twist profile are reported in the right subplot of Fig. 5.2.

6 16
1
2 14
5
12
3
4 10
ϑ [deg]

4 8 2
c [m]

3
6
1 5
2 4
3
2 4
1
0
0 −2
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
r [−] r [−]

Fig. 5.2 Parametrization of chord distribution (left subplot) and twist distribution (right
subplot). The chord distribution is defined by a cubic spline based on 5 points, and the twist
distribution is defined by a cubic spline based on 4 points. Horizontal and vertical arrows
denote the actual degrees of freedom.

5.1.2 Parametrization of the rotor speed

Turbine power control is accomplished by means of variable-speed and variable-pitch.
This control strategy is based on two independent control systems: a generator-torque
controller and a full-span rotor-collective blade-pitch controller [36]. The former controller
5.1 Turbine parametrization 79

aims at maximizing the rotor power capture below the rated wind speed, while the latter
one is used to regulate the rotor speed above the rated wind speed. As is typical in utility-
scale multi-megawatt wind turbines, both the generator-torque and blade-pitch controllers
use the generator speed measurement as the sole feedback input [36]. In this work, the
design specifications include that the cut-in and cut-out wind speeds are 3.0 and 25.0 m/s
respectively, and the rated electrical power is 5 MW. A typical rotor speed ω curve is reported
in the left subplot of Fig. 5.3. The right subplot of Fig. 5.3 shows the corresponding tip-speed
ratio λ curve. λ is defined as follows:
ωR
λ= (5.1)
u
where R is the rotor radius and u is the wind speed. The ω and λ curves consist of 5 regions

14 16
1 1/2 2 2 1/2 3
14
12
12
ω [rpm]

10
λ [−]

10
8

6
8
4
1 1/2 2 2 1/2 3
6 2
0 5 10 15 20 25 0 5 10 15 20 25
u [m/s] u [m/s]

Fig. 5.3 Rotor speed (left subplot) and tip-speed ratio (right subplot) against wind speed. Left
and right subplots report different control regions.

denoted respectively by 1 (not shown), 11/2, 2, 21/2 and 3. Region 1 is a region below the cut-in
wind speed in which no power is generated, instead the wind is used to accelerate the rotor.
In region 2, below the rated wind speed, the power capture is optimized by making the rotor
work at a constant optimal tip-speed ratio, allowing it to operate at the peak power coefficient.
In region 3, above the rated wind speed, the generator speed, and hence the generator power,
are maintained constant by varying the blade pitch. Region 11/2 is a transition region between
regions 1 and 2, while region 21/2 is a transition region between regions 2 and 3. Region 11/2 is
used to place a lower limit on the generator speed, limiting the wind turbine’s operational
80 Design space and reference turbine definition

speed range. Region 21/2 is needed to limit the tip speed (and therefore noise emissions) at
rated power and above the rated wind speed. Control in regions 11/2, 2, 21/2 is accomplished
by the generator-torque controller, while in region 3, the blade-pitch control system takes
over. The generator torque versus generator speed response of the variable-speed controller
is depicted in Fig. 5.4. Region 11/2 is defined to span the range of generator speeds between
670 rpm and 871 rpm, while the transitional generator speed between regions 21/2 and 3 is
the 99% of the rated generator speed, or 1162 rpm. The generator-slip percentage in Region
21/2 is taken to be 10%. Throughout region 2, the control torque (i.e., generator torque) τc is

50
region 1 1 1/2 2 2 1/2 3
40
τ [kN ⋅ m]

30

20
c

10

0
0 200 400 600 800 1000 1200 1400
ωg [rpm]

Fig. 5.4 Generator torque versus generator speed response of the variable-speed controller.
Different control regions are shown.

used to maintain a constant λ . In this region, the control of the rotor speed is accomplished
by varying τc proportionally to the square of the generator speed ωg , that is by imposing:

τc = Kωg2 (5.2)

where K is a constant of proportionality.

A typical power curve corresponding to the control strategy described above is reported
in Fig. 5.5. In the design optimization reported below, the tip-speed ratio in region 2 is
controlled by means of the torque control parameter K in Eq. 5.2, which is therefore treated
as a design variable, and is labeled b53 .
5.1 Turbine parametrization 81

6000

5000

4000
Pe [kW]

3000

2000

1000

0
0 5 10 15 20 25
u [m/s]

Fig. 5.5 Turbine electrical power against wind speed.

5.1.3 Internal structural layup parametrization

The blades were assumed to be fabricated from fiber-reinforced polymer, a composite material
made of a polymer matrix reinforced with fibers. These types of composite blades are made
up of stacks of laminas, the number and the thickness of which are piecewise constant along
the cross-section periphery. A stack of laminas is commonly denoted by the term laminate.
Figure 5.6 shows an example of a spar-cap type layup in which the primary bending loads
are supported by the thick mid-sections, called spar caps, connected by shear webs, and
the aerodynamic shape is maintained by sandwich panels around the blade periphery. The
number of laminates along the section periphery, the number of laminas in each laminate,
and the thickness of the laminas vary along the blade length. The webs may begin at any
section and end at any other section on the blade. Each web is assumed straight and normal
to the chord at each cross-section, and its cross-sectional dimensions and composites layup
vary along the blade. Along the blade span, at each radial section, the internal structural
layup of the blade was specified in terms of the number of laminates, lamina schedule of
each laminate, orientation of fibers in each lamina, and the lamina constituent properties.
In this work, the internal structural layup of the blades is based on the Sandia 100-m
all-glass baseline wind turbine blade [33]. Figure 5.7 illustrates graphically the laminate
placement and the lamina schedule of each laminate. Each blade consists of nine laminates,
that constitute the root build-up, leading edge panel (LEP), spar cap, trailing edge panel
(TEP), shear webs, and blade tip. The principal two shear webs extend from 2.2 to 95.5% of
the blade span. The third shear web begins at 14.3% of the blade span and ends at 58% of
the blade span. Each of these laminates shown in Fig. 5.7 is characterized by a particular
laminate schedule. The constituent material of each laminate is depicted in Table 5.1. The
82 Design space and reference turbine definition

Fig. 5.6 Representative composite blade cross-section using two shear webs to define a
spar-cap type layup. The upper and lower surfaces are each defined by three laminates
organized in three peripheral sectors. Each shear web is instead defined by a single laminate.
Lamina schedules of the top spar cap and the right shear web are shown. Image reproduced
from [65].

Fig. 5.7 Planform view of the turbine blade configuration used in the optimization study,
showing the laminate schedules in the root build-up, LEP, spar cap, TEP, shear webs, and
blade tip. For each laminate the constituent material are shown. Image reproduced from [69].

structural configuration uses a combination of composite materials and a structural foam.

NCT307-D1-E300 and NCT307-D1-34-600-G300 are composite materials incorporating
fiberglass in an epoxy resin matrix. NCT307-D1-34-600-G300 is a composite material
made up of carbon fibers in a epoxy resin matrix. The Gurit Corecell M-foam M200 is a
structural foam developed for marine applications. Mechanical properties of these materials
5.2 Design space 83

material
blade-root NCT307-D1-E300
blade-shell NB307-D1-7781-497A
spar-uni NCT307-D1-34-600-G300
spar-core Gurit Corecell M-foam M200
LEP-core Gurit Corecell M-foam M200
TEP-core Gurit Corecell M-foam M200
web-shell NB307-D1-7781-497A
web-core Gurit Corecell M-foam M200

Table 5.1 Blade laminate constituent materials.

are depicted in Table 5.2 [131]. The Sandia layup is characterized by a given laminate

E11 [GPa] E22 [GPa] G12 [GPa] ν12 [-] ρ [kg/m3 ]

NCT307-D1-E300 35.5 8.33 4.12 0.33 1780
NB307-D1-7781-497A 19.2 19.2 3.95 0.13 1670
NCT307-D1-34-600-G300 123 8.2 4.71 0.31 1470
Gurit Corecell M-foam M200 0.21 0.21 0.098 0.33 200

Table 5.2 Mechanical properties of the materials used to define the composite laminates of
the blades. E11 and E22 represent the principal and the lateral Young’s moduli, respectively.
G12 is the shear modulus. ν12 is the Poisson ratio. ρ is the density.

thickness distribution along the blade span. During the optimizations performed in this
thesis, all characteristics of the Sandia layup were kept constant, except the thickness of
the laminates making up the blade. During the optimization, the thickness of each laminate
is controlled with a constant of proportionality s multiplying the laminate thickness of the
Sandia turbine blade. Therefore, the parameter s is the design variable controlling the blade
internal structure and is labeled b54 .

5.2 Design space

The design space b in which the optimum is sought is defined by the set of design variables
presented above, and could be expressed as follows:

b = {b1 , ..., b54 } (5.3)

84 Design space and reference turbine definition

The ranges of variability, or bounds, within which the design variables are allowed to vary
are shown in Tables 5.3, 5.4, 5.5 and 5.6.

design variable range of variability variable name

b1 ∈ [−0.4975, 0.5025] [-] y/c(p1 )
b2 ∈ [−0.1508, 0.8492] [-] x/c(p2 )
b3 ∈ [−0.3123, 0.6877] [-] y/c(p2 )
b4 ∈ [−0.2914, 0.7086] [-] x/c(p3 )
b5 ∈ [−0.3001, 0.6999] [-] y/c(p3 )
b6 ∈ [−0.4218, 0.5782] [-] x/c(p4 )
b7 ∈ [−0.3274, 0.6726] [-] y/c(p4 )
b8 ∈ [−0.5000, 0.5000] [-] x/c(p5 )
b9 ∈ [−0.4155, 0.5845] [-] y/c(p5 )
b10 ∈ [−0.2953, 0.7047] [-] x/c(p6 )
b11 ∈ [−0.7524, 0.2476] [-] y/c(p6 )
b12 ∈ [−0.1058, 0.8942] [-] x/c(p7 )
b13 ∈ [−0.7006, 0.2994] [-] y/c(p7 )
b14 ∈ [−0.5036, 0.4964] [-] y/c(p8 )

Table 5.3 Design variable bounds of the root airfoil (see Fig. 5.8).

design variable range of variability variable name

b15 ∈ [−0.4979, 0.5021] [-] y/c(p9 )
b16 ∈ [−0.1974, 0.8026] [-] x/c(p10 )
b17 ∈ [−0.3729, 0.6271] [-] y/c(p10 )
b18 ∈ [−0.2570, 0.7430] [-] x/c(p11 )
b19 ∈ [−0.3766, 0.6234] [-] y/c(p11 )
b20 ∈ [−0.2500, 0.7500] [-] x/c(p12 )
b21 ∈ [−0.3735, 0.6265] [-] y/c(p12 )
b22 ∈ [−0.4981, 0.5019] [-] x/c(p13 )
b23 ∈ [−0.4231, 0.5769] [-] y/c(p13 )
b24 ∈ [−0.0481, 0.9519] [-] x/c(p14 )
b25 ∈ [−0.6147, 0.3853] [-] y/c(p14 )
b26 ∈ [0.1030, 1.1030] [-] x/c(p15 )
b27 ∈ [−0.5617, 0.4383] [-] y/c(p15 )
b28 ∈ [−0.5023, 0.4977] [-] y/c(p16 )

Table 5.4 Design variable bounds of the midspan airfoil (see Fig. 5.9).
5.3 Reference turbine definition 85

design variable range of variability variable name

b29 ∈ [−0.4994, 0.5006] [-] y/c(p17 )
b30 ∈ [−0.1259, 0.8741] [-] x/c(p18 )
b31 ∈ [−0.3819, 0.6181] [-] y/c(p18 )
b32 ∈ [−0.3005, 0.6995] [-] x/c(p19 )
b33 ∈ [−0.3842, 0.6158] [-] y/c(p19 )
b34 ∈ [−0.4609, 0.5391] [-] x/c(p20 )
b35 ∈ [−0.4393, 0.5607] [-] y/c(p20 )
b36 ∈ [−0.4992, 0.5008] [-] x/c(p21 )
b37 ∈ [−0.4736, 0.5264] [-] y/c(p21 )
b38 ∈ [−0.2712, 0.7288] [-] x/c(p22 )
b39 ∈ [−0.5745, 0.4255] [-] y/c(p22 )
b40 ∈ [−0.0681, 0.9319] [-] x/c(p23 )
b41 ∈ [−0.5581, 0.4419] [-] y/c(p23 )
b42 ∈ [−0.4994, 0.5006] [-] y/c(p24 )

Table 5.5 Design variable bounds of the tip airfoil (see Fig. 5.10).

design variable range of variability variable name

b43 ∈ [0.1000, 5.5000] [m] c(c1 )
b44 ∈ [0.1000, 5.5000] [m] c(c2 )
b45 ∈ [1.0000, 50.0000] [m] r(c2 )
b46 ∈ [0.1000, 5.5000] [m] c(c3 )
b47 ∈ [0.1000, 5.5000] [m] c(c4 )
b48 ∈ [0.1000, 5.5000] [m] c(c5 )
b49 ∈ [−1.7507, 28.2493] [deg] ϑ (ϑ1 )
b50 ∈ [−8.5235, 21.4765] [deg] ϑ (ϑ2 )
b51 ∈ [−13.5416, 16.4584] [deg] ϑ (ϑ3 )
b52 ∈ [−15.0968, 14.9032] [deg] ϑ (ϑ4 )
b53 ∈ [1.0000, 2.8526] [N · m/s2 ] K
b54 ∈ [0.1000, 2.5000] [-] s

Table 5.6 Design variable bounds of chord and twist distributions (see Fig. 5.11), torque
control parameter and thickness parameter.

5.3 Reference turbine definition

The reference turbine to which the optimal designs determined in Chapter 6 are compared is
based on the NREL offshore 5-MW baseline wind turbine [36]. This turbine is a conventional
three-blade variable-speed variable blade-pitch-to-feather-controlled HAWT with upwind
rotor. The blade tip and root diameters are 126 and 3 m, respectively, the rotor hub is at 90
86 Design space and reference turbine definition

m above the sea level, and the cut-in and cut-out speeds are 3.0 and 25.0 m/s, respectively.
From root to tip, each blade is modeled with seventeen sections, including two circular
sections near the blade root, one section transitioning from the outermost circular section to
the innermost airfoil section, and fourteen airfoil sections over the remainder of the blade.
For each of these seventeen sections, airfoil polars are given for a single Re and an AoA
range from -180 to 180 deg (more details on how these polars were determined can be found
in [36]). This turbine belongs to a IEC standard class I1 [17]. A turbine belonging to a IEC
standard class I is designed to withstand climates for which the extreme 10 min average wind
speed with a recurrence period of 50 years at turbine hub height is lower than or equal to
50.0 m/s. In several published papers and reports, like in [17], it is stated that the NREL
offshore 5-MW baseline wind turbine is a class I turbine, defining the IEC reference wind
speed. However, none of these references defines the IEC turbulence characteristics to be
considered when dealing with this turbine. For this reason, in this thesis a medium turbulence
intensity, denoted by the subscript “B”, was assumed.
The NREL test case provides also the structural properties (e.g., stiffness and inertia
properties) of various turbine components, such as nacelle, hub and tower. These properties
are in fact fundamental to characterize the dynamic response of the whole turbine, and
therefore they must be included in the aeroelastic modeling. The design of these structural
components depends generally on the loading acting on the rotor. In particular, the rotor
loads and therefore the sizing of the nacelle, hub and tower are related to the rotor swept
area, hub height and machine rating [102]. In this work such properties are not varied in the
course of the optimization, and therefore, the structural properties of the nacelle, hub and
tower are considered constant in the aeroelastic simulations.
Before starting the design optimization, the geometry of the blade of the NREL 5-MW
turbine given in [36] as well as its internal structural layup and the rotor speed have been
redefined using the parametrization defined above. The resulting parametrized turbine was
taken as the reference configuration. The NREL offshore 5-MW baseline test case does not
provide detailed information regarding its internal structural layup. Structural characteristics
of the blades are indeed described only in terms of span-variant properties, such as mass,
moments of inertia, and stiffnesses. For such reason, in this work, an internal structural layup

1 Accordingto the IEC standard [28], the external conditions to be considered for the design of wind turbines
depend on the characteristics of the intended installation site. Wind turbine IEC standard classes define these
characteristics in terms of wind speed and turbulence parameters. For example, a turbine belonging to a IEC
standard class IIC is designed to withstand climates for which the extreme 10 min average wind speed with
a recurrence period of 50 years at turbine hub height is lower than or equal to 42.5 m/s. The subscript “C”
indicates that a hub height turbulence intensity, at a 10 min average wind speed of 15 m/s, equal to 0.12 is
considered. In this context, turbulence intensity is defined as the ratio of the wind speed standard deviation to
the mean wind speed.
5.3 Reference turbine definition 87

based on the Sandia 100-m all-glass baseline wind turbine blade was adopted. As shown
below, the use of this layup resulted in different blade span-variant structural properties
with respect to those of the original NREL test case. However, in this study, the purpose of
defining a reference turbine was not to replicate the same performance and characteristics
of the NREL offshore 5-MW baseline wind turbine. Instead, the definition of a reference
turbine was needed both to build a convenient initial solution for the optimizations, and to
determine the performance of a turbine configuration using the same modeling set-up of
the multidisciplinary analysis used to redesign the turbine, and determine the performance
of new designs. Therefore, the differences in performance and characteristics between the
reference and the NREL turbines do not have any practical implications in the optimizations
reported later in this thesis. Additional information on the design specifications and turbine
definition of the original NREL 5-MW turbine are available in [36].

5.3.1 Definition of the reference turbine’s airfoils

Parametrizing the external geometry of the reference turbine involved finding the best
geometric fit for the NREL offshore 5-MW baseline wind turbine’s root, midspan and tip
airfoils, which are the DU 99-W-405, DU 91-W2-250 NACA 64-618 airfoils, respectively.
These fittings were performed using the MATLAB pattern search algorithm, minimizing
the root-mean-square error (RMSE) between the parametric airfoils and the NREL offshore
5-MW baseline wind turbine’s ones. In practice, for each airfoil, the fitting problem was
formulated and solved as follows:

Find: p
(5.4)
to minimize: RMSE(pp)

where the symbol p represents the vector of the airfoil parametrization’s independent vari-
ables, as depicted in Fig. 5.1. The RMSE was computed as follows:
r
n
∑t=1 (ŷt (pp) − yt )2
RMSE(pp) = (5.5)
n

where (ŷt (pp) − yt ) represents the distance, measured along the y/c direction (see Fig. 5.1),
between the parametric airfoil and the reference one. This distance was evaluated at n = 200
different locations, along the x/c direction, on both the suction and pressure sides of the
airfoils. Figures 5.8, 5.9 and 5.10 show the comparison between the parametric reference
turbine’s airfoils and those of the NREL offshore 5-MW baseline wind turbine. These figures
highlight the high flexibility of the adopted airfoil parametrization, describing with good
88 Design space and reference turbine definition

accuracy three different airfoil shapes. The RMSEs of the reference turbine’s root, midspan
and tip airfoils are equal to 4.337 · 10−4 , 3.915 · 10−4 and 2.964 · 10−4 , respectively.

Detail of the trailing edge

0.8
0.015

0.01
p1
0.6 0.005

−0.005
0.4 p8
−0.01
y/c [−]

p3 p2 −0.015
p4
0.2 0.99 1 1.01

p5

−0.2

p7
NREL 5−MW
p6
ref.
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
x/c [−]

Fig. 5.8 Comparison between the NREL 5-MW baseline turbine’s root airfoil and the
parameterized reference one, determined as a best fit to the NREL 5-MW baseline turbine’s
root airfoil.

5.3.2 Definition of the reference turbine’s chord and twist distributions

Control points of the reference turbine’s chord and twist distributions were determined as
the best fit to the NREL offshore 5-MW baseline wind turbine’s ones. These fittings were
carried out using the MATLAB pattern search algorithm, minimizing the RMSE between the
parametric turbine’s and the NREL offshore 5-MW baseline wind turbine’s chord and twist
5.3 Reference turbine definition 89

1
Detail of the parametrization
0.145 Detail of the trailing edge
0.8 0.14 0.015
0.135 0.01
p12 p9
0.13 p11 0.005
0.6
0.125 0
0.12 −0.005 p16
0.4 0.115 −0.01
y/c [−]

0.11 −0.015
0.24 0.25 0.26
0.2 0.99 1 1.01

p13
p
10
0

p15
−0.2 p14

NREL 5−MW
ref.
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
x/c [−]

Fig. 5.9 Comparison between the NREL 5-MW baseline turbine’s midspan airfoil and the
parameterized reference one, determined as a best fit to the NREL 5-MW baseline turbine’s
midspan airfoil.

distributions. For the chord distribution, the fitting problem was formulated as follows:

Find: c
(5.6)
to minimize: RMSE(cc)

where the symbol c represents the vector of the chord parametrization’s design variables, as
depicted in the left subplot of Fig. 5.2. The RMSE was computed as follows:
r
n
∑t=1 (ĉt (cc) − ct )2
RMSE(cc) = (5.7)
n
90 Design space and reference turbine definition

Detail of the trailing edge

0.8
0.015
0.01
p17
0.6 0.005
0
p24
−0.005
0.4
−0.01
y/c [−]

−0.015

0.2 p p18
19 0.99 1 1.01
p
20

p21
0

p22 p23
−0.2

NREL 5−MW
ref.
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
x/c [−]

Fig. 5.10 Comparison between the NREL 5-MW baseline turbine’s tip airfoil and the param-
eterized reference one, determined as a best fit to the NREL 5-MW baseline turbine’s tip
airfoil.

where (ĉt (cc) − ct ) represents the difference between the parametric chord and the reference
one, evaluated at n = 100 different radial locations. For the twist distribution, the fitting
problem was formulated as follows:

Find: ϑ
(5.8)
ϑ)
to minimize: RMSE(ϑ
5.3 Reference turbine definition 91

where the symbol ϑ represents the vector of the twist parametrization’s design variables, as
depicted in the right subplot of Fig. 5.2. The RMSE was computed as follows:
s
n ϑ ) − ϑt )2
∑t=1 (ϑ̂t (ϑ
ϑ) =
RMSE(ϑ (5.9)
n

ϑ ) − ϑt ) represents the difference between the parametric twist and the reference
where (ϑ̂t (ϑ
one, evaluated at n = 100 different radial locations. Figure 5.11 shows the comparison be-
tween the parametric reference turbine’s chord and twist distributions and those of the NREL
offshore 5-MW baseline wind turbine. This figure shows that the developed parametrizations
for chord and twist distributions are able to describe with accuracy the original shapes. The
RMSEs of the reference turbine’s chord and twist distributions are equal to 7.129 · 10−2 m
and 9.622 · 10−2 deg, respectively.

6 16
NREL 5−MW ϑ1
c2 ref. 14
5
12
c3
4 10
ϑ [deg]

c4 8
c [m]

ϑ
3 2
c 6
1
2 c5 4
ϑ3
2
1 ϑ4
0
0 −2
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
r [−] r [−]

Fig. 5.11 Comparison between the reference turbine’s chord and twist distributions and the
NREL offshore 5-MW baseline wind turbine’s ones.

5.3.3 Definition of the reference turbine’s rotor speed and internal struc-
tural layup
In order to make a fair comparison between the optimized turbines and the reference one, both
the torque control parameter, K, and the thickness parameter, s, of the reference turbine were
optimized for minimum LCOE subject to the structural constraints described in Chapter 6 for
92 Design space and reference turbine definition

NREL 5-MW ref.

K [N · m/s2 ] 2.33 (λ = 7.55) 2.13 (λ = 7.67)
s - 0.964

Table 5.7 Torque control parameter with resulting tip-speed ratio in region 2 in parentheses,
and thickness parameter of the reference turbine and the NREL offshore 5-MW baseline one.

a mean wind speed of 10.0 m/s. During this optimization, the aerodynamic forces required to
determine the turbine power and the aerodynamic loads were determined as follows. Lift
and drag coefficients of the two circular sections near the hub were assumed to be 0 and 0.5,
respectively. Lift and drag coefficients of the sections between the outermost circular section
and the root airfoil section were obtained by linear interpolation of the force coefficients of
abovesaid limiting sections. 3D extended force coefficient of the root, midspan and tip airfoils
were calculated by means of XFOIL/AERODAS. Lift and drag coefficients of all airfoils
between the root airfoil and the tip airfoil were determined by linearly interpolating the force
coefficients of the root, midspan and tip airfoils. The force coefficients thus determined
were used as input variables by AeroDyn, the BEM libraries used by FAST to determine
the radial distribution of the aerodynamic forces and the aerodynamic power. Co-Blade
was used to determine the blade span-variant structural properties required by FAST. Co-
Blade along with the definition of the internal structure of the blade, also requires a detailed
geometric definition of the outer shape of the blade to determine its structural properties.
The outer shape is defined by the chord and twist profiles, and the airfoil geometry at 34
radial positions, determined by linearly interpolating the blade sections defined by the hub
circles, and the root, midspan and tip airfoils. The aerodynamic loads determined by FAST
were then used to evaluate the objective function and constraints as explained in Chapter 4.
The optimized value of K and s for the reference turbine are depicted in Table 5.7. In this
table, K of the reference turbine and that of the NREL offshore 5-MW baseline turbine are
compared. Table 5.7 reports also the resulting tip-speed ratio in region 2 in parentheses,
showing that, in region 2, the reference turbine rotates at a speed 1.56% higher than the
NREL one. As explain above, the parameter s cannot be associated with the NREL turbine,
as its internal layup is not known. Instead, the parameter s refers to the internal structural
layup based on the Sandia 100-m all-glass baseline wind turbine blade. At the beginning of
the optimization, this parameter was set equal to 1 (i.e., the laminate thickness was equal to
that of the Sandia turbine blade). As mentioned above, the use of an internal structural layup
based on the Sandia turbine blade led to different structural properties with respect to the
NREL turbine. This is particularly evident when comparing the blade weight of these two
5.3 Reference turbine definition 93

NREL 5-MW ref.

wb [kg] 17.740 ·103 44.216 ·103

Table 5.8 Blade mass of the reference turbine and the NREL offshore 5-MW baseline one.

turbines (Table 5.8). The reference blade is about 150% heavier that the NREL one. Such a
difference could be reduced by adopting other structural layups, such as that reported in [34],
or by optimizing the current layup by minimizing such differences.
The rotor speed and tip-speed ratio curves depicted respectively in the left and right
subplots of Fig. 5.12 refer to the NREL 5-MW turbine and the modified reference turbine.
These pictures shows that, from cut-in to rated wind speed, the rotor speed of the reference
turbine is higher than that of the NREL 5-MW turbine. Fig. 5.13 shows a comparison

14 16
NREL 5−MW
ref. 14 8.2
8
12 7.8
12 7.6
7.4
ω [rpm]

10 8 9 10
λ [−]

10
8
9.6
6
8 9.4
4
7.5 8 8.5
6 2
0 5 10 15 20 25 0 5 10 15 20 25
u [m/s] u [m/s]

Fig. 5.12 Rotor speed (left subplot) and tip-speed ratio (right subplot) as a function of wind
speed of the reference turbine and the NREL offshore 5-MW baseline wind turbine. Different
control regions are shown in both subplots.

between the power curves determined for the NREL 5-MW and reference turbines from
cut-in wind speed to cut-out wind speed. From cut-in to rated wind speed the NREL turbine
is able to extract more energy than the reference one. This difference is primarily due to the
fact that the reference turbine’s power curve was calculated by considering a 1/7 wind profile
power law, while the NREL turbine’s power curve was evaluated considering a zero vertical
wind shear.
94 Design space and reference turbine definition

6000
NREL 5−MW
ref.
5000

5200
4000
5100
Pe [kW]

3000
5000

2000 4900

1000 4800
11 11.2 11.4 11.6 11.8

0
0 5 10 15 20 25
u [m/s]

Fig. 5.13 Electrical power as a function of wind speed of the reference turbine and the NREL
offshore 5-MW baseline wind turbine.
Chapter 6

HAWT design optimization under

environmental uncertainty

In this chapter, the multidisciplinary and robust design optimization framework presented in
Chapter 4 is demonstrated by performing the design optimization of a multi-megawatt HAWT
rotor under environmental uncertainty. More specifically, the considered problem consisted
of optimizing the design of the reference turbine defined in Chapter 5 so as to minimize the
mean and the standard deviation of its LCOE when used at sites characterized by wind power
density class from 3 to 71 , in which the mean wind speed, based on a Rayleigh distribution,
at 50 m above the ground ranges from 6.4 to 11.9 m/s2 . At the actual reference turbine’s hub
height of 90 m, the mean wind speed, ū, determined using a vertical extrapolation based on
the 1/7 power law, varies between 7.0 and 13.0 m/s. The design specifications are that the
optimal turbine has the same rotor diameter, rated power, cut-in and cut-out speeds as the
reference turbine.
The 54 blade geometry and control variables defined in Chapter 5 are optimized assuming
the mean wind speed to be uniformly distributed in the abovesaid speed range. For com-
parison purposes, a deterministic optimization was performed as well. This optimization
was carried out considering a fixed mean wind speed of 10 m/s. The aim of this chapter

1 According to [132], the wind power density class defines the wind power density, expressed in W/m2 ,
and the mean wind speed, based on a Rayleigh speed distribution, of a given site. For example, in a site
characterized by a wind power density class of 2, the wind power density varies between 200 and 300 W/m2 ,
with a mean wind speed, based on a Rayleigh speed distribution, ranging from 5.6 to 6.4 m/s at 50 m above the
ground.
2 As mentioned in Chapter 1, the probabilistic design scenario considered in this thesis was described by

Ning et al. [17]. This scenario, as stated by Ning et al., may be simplistic. However, in the framework of this
research work, it provided a plausible and feasible scenario to demonstrate the effectiveness of the developed
optimization framework.
96 HAWT design optimization under environmental uncertainty

is to describe the deterministic and robust optimization problem formulations in terms of

objective functions and constraints, and summarize the optimization results.

6.1 Objective function

The performance parameter considered in this study was LCOE, and the goal of the RDO
exercise reported below was to minimize both the expectation and the standard deviation
of LCOE considering the uncertainty affecting the mean wind speed, uniformly distributed
within a given range of variability. This is a multi-objective design optimization problem, and
it can be solved using either ad-hoc bi-objective optimization approaches [133], including
evolution-based algorithms [20], or single-objective approaches [133]. Optimization of the
mean often conflicts with minimization of the standard deviation, and, in this circumstance,
a Pareto set of optimal solutions is encountered. As seen in Chapter 3, one of the simplest
single-objective approaches to robust optimization is the weighted sum method, whereby a
single objective function is obtained by considering a weighted average of mean and standard
deviation of the performance parameter of interest. When a Pareto front exists, each of its
points could be determined by varying the weights and solving a new optimization problem.
Although rather simple to implement and often quite robust, this approach to the calculation
of the entire Pareto front suffers from some limitations, including its inadequacy to determine
non-convex regions of the front [134]. Making use of this approach, the objective function to
be minimized in the considered HAWT design problem is:
   
µLCOE (bb, ū) σLCOE (bb, ū)
frob (bb, ū) = αw ∗ + (1 − αw ) ∗ (6.1)
µLCOE σLCOE

where b is the array of 54 design variables defined in Chapter 5, µLCOE and σLCOE are
respectively the mean and the standard deviation of LCOE of the current HAWT design,
∗
and µLCOE ∗
and σLCOE denote the values of these two parameters for the reference turbine.
These two values are thus used to normalize the probabilistic estimate of LCOE of all
considered HAWT rotor designs. The constants αw and (1 − αw ) define respectively the
weight assigned to µLCOE and σLCOE . The values of µLCOE and σLCOE for each considered
HAWT configuration were evaluated by means of the Latin hypercube sampling, as explained
in Chapter 4. Practically, the range of variability of the mean wind speed was divided into
100 intervals with equal probability3 . Subsequently, LCOE was evaluated for the central

3 Theaccuracy of the Latin hypercube sampling depends on the number of the selected equal probability
intervals. The more intervals are considered, the more accurately this method can estimate the statistical
6.2 Constraints 97

mean wind speed of each interval, and then µLCOE and σLCOE were calculated by means of
Eqs. 4.6 and 4.7.

6.2 Constraints
The structural verification of wind turbines requires a large number of aeroelastic simulations
in order to evaluate the loads that their components experience under design-driving scenarios.
The purpose of these analyses is to make sure that the loads acting on the wind turbine over its
lifetime do not lead to structural failure. The IEC 61400-3 [28] standard specifies an essential
set of design requirements to ensure the structural integrity of offshore wind turbines. This
standard prescribes a number of design load cases (DLCs) which define design scenarios to
be considered during the structural design of wind turbine components. Considering these
DLCs, the aim is to assess the turbine response with respect to the following analyses:

• ultimate load analysis;

• fatigue load analysis.

Ultimate load analysis refers to the assessment of material ultimate strength, blade tip
deflection and structural stability (i.e., buckling), while fatigue load analysis concerns fatigue
strength. The full set of DLCs includes, among others, normal, fault, and parked operating
conditions, and considers both normal and extreme wind conditions. The IEC 61400-3
standard specifies also marine conditions (waves, sea currents, water level, etc.) which should
be included in the design of offshore wind turbines. In this work, however, marine conditions
were not considered. A few representative DLCs were taken into account in the present study,
namely those denoted by 1.1, 1.2 and 6.1b. These load cases have shown to be the most
likely design drivers for the majority of turbine blades [16, 33, 34]. In the framework of the
presented optimization problem, each HAWT configuration generated during the course of
the optimization process was assessed structurally by performing ultimate and fatigue load
analyses, considering the DLCs 1.1, 1.2 and 6.1b. According to the conditions prescribed
by these DLCs, described in detail in Sect. 6.2.1, ultimate and fatigue loads were evaluated
through a series of aeroelastic simulations performed by means of FAST. As explained in
Sect. 6.2.2, ultimate and fatigue loads were then augmented by means of safety factors in
order to account for uncertainties and variabilities of various natures (such as in the evaluation
of loads and materials) in the design procedure. Eventually, ultimate load analysis was carried
out by means of Co-Blade, while fatigue loads were analyzed through MLife. Sect. 6.2.3 and
moments. A preliminary study has show that, for this particular application, 100 intervals allowed the method
to converge.
98 HAWT design optimization under environmental uncertainty

Sect. 6.2.4 will outline the procedure followed for ultimate load analysis and fatigue load
analysis, respectively.

6.2.1 Design load cases

The design scenarios prescribed by the DLCs depend primarily on the considered IEC
standard class, which is defined in terms of wind speed and turbulence parameters of the
intended installation site. In this study, a turbine class of IB , defining a site reference wind
speed of 50.0 m/s, and medium turbulence characteristics, was selected. According to the
IEC standard, a turbine belonging to this class is designed to withstand climates for which
the extreme 10 min average wind speed with a recurrence period of 50 years at turbine hub
height, denoted by Vref , is lower than or equal to 50.0 m/s, with a hub height turbulence
intensity, at a 10 min average wind speed of 15 m/s, denoted by Iref , equal to 0.14. Table 6.1
provides details of the DLCs 1.1, 1.2 and 6.1b for a IB turbine class, showing the relative
operation and wind conditions and the associated type of analysis. These DLCs are described

DLC operating condition wind conditions type of analysis

NTM
1.1 normal ultimate
3 m/s < Vhub < 25 m/s
NTM
1.2 normal fatigue
3 m/s < Vhub < 25 m/s
steady EWM
6.1b parked ultimate
Vhub = 70 m/s

Table 6.1 IEC DLCs considered in the structural verification of each wind turbine generated
during the optimization process. The table shows the operation and wind conditions and the
type of analysis related to each DLC. Vhub represents the wind speed at the hub height.

in detail below.

Design load cases 1.1 and 1.2

Considering normal operating conditions, the DLCs 1.1 and 1.2 involve the analysis of
ultimate and fatigue loads, respectively, using the normal turbulence model (NTM) to define
wind conditions. According to the NTM, a series of aeroelastic simulations were carried out
considering different turbulent wind speed conditions, the mean value of which ranged from
cut-in to cut-out wind speed. More specifically, from cut-in to cut-out, wind speed at the hub
height was defined by 11 bins characterized by a size of 2.0 m/s, and mean equal to 4.0, 6.0,
8.0, 10.0, 12.0, 14.0, 16.0, 18.0, 20.0, 22.0 and 24.0 m/s, respectively. As recommended
6.2 Constraints 99

by the standard, loads were estimated by 6 10-minute stochastic simulations for each of
the 11 mean wind speeds considered. TurbSim was used to create the 10-minute turbulent
wind files, using 6 different random seeds for each of the 11 mean wind speeds considered.
According to the standard, all of these wind files were generated considering a power law
shear profile with an exponent of 0.14. Turbulence intensity, I ,was given by [28]:
σ1
I= (6.2)
Vhub

where σ1 is the turbulence standard deviation, and Vhub represents the mean wind speed at
the hub height. σ1 can be evaluated as follows [28]:

σ1 = Iref (0.75Vhub + b) (6.3)

where Iref = 0.14 (for a class IB turbine) and b = 5.6 m/s. These values are derived from [28].
With this configuration, 66 10-minute simulations were carried out for the DLCs 1.1 and 1.2.
Ultimate and fatigue loads were evaluated concurrently, using the same turbulent wind files.

Design load case 6.1b

The DLC 6.1b considers the analysis of ultimate loads in parked conditions using the steady
extreme wind speed model (EWM) for wind conditions. Parked conditions can be achieved
in several ways, such as by pitching the blades towards feather, or by applying a high-speed
shaft (HSS) brake control. In this work, the parked conditions were achieved by making the
following assumption:

• The turbine’s HSS brake is engaged for a parked configuration so that the rotor speed
is fixed to zero.

• The turbine was parked with a blade pitch angle equal to 0 deg (i.e., the blade is flat to
the wind).

• Computation of the induced velocities is turned off in AeroDyn because the rotor is
stationary. This means that the calculation of the axial and tangential induction factors
is bypassed during BEM computations, considering all induction factors equal to zero.

According to the EWM, a steady extreme wind speed, Ve50 , with a recurrence period of 50
years of 70 m/s was considered. Ve50 was calculated by using the following expression [28]:

Ve50 = 1.4Vref (6.4)

100 HAWT design optimization under environmental uncertainty

where Vref = 50 m/s (for a class IB turbine). The analysis associated with the DLC 6.1b
involved a single 100-second aeroelastic simulation4 , using a steady AeroDyn hub height
wind file created by means of IECWind. More specifically, this file prescribed a horizontal
wind, perpendicular to the rotor plane, with a speed of 70 m/s at the hub height. With reference
to the wind speed at the hub height, wind speed at any other heights was determined by
means of a power law shear profile with an exponent of 0.14. With this configuration, a
100-second simulation was carried out for the DLC 6.1b.

6.2.2 Safety factors

As already stated in the preceding sections, ultimate and fatigue analyses for the structural
verification of HAWTs are carried out by means of aeroelastic simulations. These simulations
can be affected by uncertainties related for instance to errors in the loading model or variations
of the actual material strengths from the theoretical ones. In the ultimate and fatigue analyses,
such uncertainties are taken into account by means of safety factors. Practically, these factors
increase the ultimate and fatigue loads calculated by means of the aforementioned aeroelastic
simulations. The IEC 61400-3 standard prescribes a set of specific safety factors, accounting
for different types of uncertainty and variability. This set is made up of the following partial
safety factors:

• partial safety factors for loads, δ f ;

• partial safety factors for materials, δm ;

• partial safety factor for consequence of failure and component classes, δn .

δ f takes into account possible unfavorable deviations/uncertainties of the load from the char-
acteristic value, and uncertainties in the loading model. δm instead takes into account possible
unfavorable deviations/uncertainties of the strength of material from the characteristic value,
possible inaccurate assessment of the resistance of sections or load-carrying capacity of parts
of the structure, and uncertainties in the geometrical parameters. The value of δn depends
on the structural components making up wind turbines, and, in particular, on whether their
failure may lead to the failure of a major part of the turbine. More specifically, according to
the IEC 61400-3 standard, wind turbines are distinguished into classes that define whether
a failure of one of their structural components can result in major turbine failures. In this
work, a component class of 2 was used, meaning that the wind turbine failure may be caused
by the failure of one of its components. The aforementioned partial safety factors, depicted
4 Although steady wind conditions are considered by this DLC, a time-variant aeroelastic simulation was
needed to assess the time-variant dynamic response of the blades.
6.2 Constraints 101

in Table 6.2, are combined in total safety factors, δT , used in the load analyses. This table

DLC δf δm δn δT
1.1 1.25 1.3 1 1.625
ultimate strength 1.2 - - - -
6.1b 1.35 1.3 1 1.755
1.1 1.25 1.1 1 1.375
deflection 1.2 - - - -
6.1b 1.35 1.1 1 1.485
1.1 1.25 1.2 1 1.5
buckling 1.2 - - - -
6.1b 1.35 1.2 1 1.62
1.1 - - - -
fatigue strength 1.2 1 1.2 1.15 1.38
6.1b - - - -

Table 6.2 Total safety factor, δT , used in the ultimate and fatigue load analyses. δ f , δm and
δn are respectively the partial safety factors for loads, materials, and consequence of failure
and component classes.

shows that the value of the partial safety factors depend on both the considered DLC and
the specific structural analysis carried out (e.g., ultimate strength, deflection, buckling and
fatigue strength)

6.2.3 Ultimate load analysis

For each ultimate load analysis, carried out for both the DLCs 1.1 and 6.1b, blade laminate
(tensile, σT , and compressive, σC ) normal stress, panel (i.e., laminate) buckling and out-of-
plane tip deflection, δ , were evaluated by means of Co-Blade. As explained in Chapter 2,
panel buckling is treated by means of the dimensionless buckling criteria, R [69]. Maximum
tensile and compressive stresses and buckling criteria were specifically calculated on the
upper and lower surfaces, and webs of the blade at 4 radial sections fixed at 0%, 14.3%,
34% and 63.3% of its length. Maximum stresses, buckling criteria and tip deflections were
evaluated on a blade in the 3 o’clock azimuth position, when looking downwind5 . Due to the
blade weight forces, such conditions represent the worst case for the edgewise loads.
Thus, regarding the ultimate load analysis, the objective function reported above was
minimized subject to constraints on maximum allowable tensile and compressive stresses,
buckling criteria and out-of-plane tip deflection. As depicted in Table 6.2, normal stresses,
5 The rotor of the considered turbine rotates anticlockwise when looking downwind.
102 HAWT design optimization under environmental uncertainty

calculated by means of FAST, were augmented considering a total safety factor of 1.625
for the DLC 1.1, and a total safety factor of 1.755 for the DLC 6.1b. Laminate tensile and
compressive yield strength were both considered equal to 325 MPa for simplicity6 . Therefore,
the maximum allowable tensile and compressive stresses were equal to 325 MPa and -325
MPa, respectively. Buckling loads were instead augmented by a total safety factor of 1.5 for
the DLC 1.1, and by a total safety factor of 1.62 for the DLC 6.1b. The maximum allowable
buckling criteria R was set equal to 1. Indeed, for this value of R, buckling loads become
critical (i.e., lead to structural failure by buckling). Maximum allowable out-of-plane tip
deflection was determined by dividing the total blade clearance in unloaded conditions, equal
to 10.23 m, by a total safety factor of 1.375 for the DLC 1.1, and by a total safety factor of
1.485 for the DLC 6.1b. This resulted in a maximum allowable out-of-plane tip deflection of
7.44 m and 6.89 m for the DLC 1.1 and DLC 6.1b, respectively.
The total blade clearance in unloaded conditions is the distance between the tower and
the blade tips in undeflected conditions. As shown in Fig. 6.1, the minimum clearance of a
given blade is achieved when the blade in the 6 o’clock azimuth position. In this work, the
tower diameter, T , was assumed constant along the tower span, and equal to 6 m. The rotor
shaft overhang, O, of the turbine under investigation was equal to 5.019 m. The rotor shaft
tilt angle, θ , and the blade precone angle7 , β , were equal to 5 and 2.5 deg, respectively. The
distance between the blade tip and the rotor axis, R, was equal to 63 m. Therefore, the total
blade clearance in unloaded conditions, C, of a blade in the 6 o’clock azimuth position can
be calculated as follows:
T
C = O cos(θ ) + R sin (θ + β ) − =
2 (6.5)
6
= 5.019 · cos(5 · π/180) + 63 · sin (30 · π/180) − = 10.23 m
2

The constraints associated with the ultimate load analysis can be expressed by means the
following nonlinear inequalities:

σT ≤ 350 MPa (6.6)

6 This approach has been followed also by Ashuri et al. [16]. These authors explain that the layup and the
volumetric mixture of the elements that the composite sample is made of depends strongly on the blade yield
stress. Exceeding this yield stress in some elements of a composite material does not necessarily mean the
entire structure will fail, since the experienced load of that element will be distributed among other elements.
Considering these facts, the defined value in this work is just a representative value suitable for the design
applications.
7 One way of increasing the distance between the blade tip and the tower is to orientate the blades with a

so-called cone angle. In a HAWT, the rotor blades mounted on a hub rotate about a substantially horizontal
axis, and the cone angle is formed by mounting the blades such that a longitudinal axis of the blade is not
perpendicular to the rotational axis of the rotor [135].
6.2 Constraints 103

rotor blade

rotor hub nacelle

O
wind
θ
R
rotor shaft
C
tower
β θ
T

Fig. 6.1 Total blade clearance in unloaded conditions, C, of a blade in the 6 o’clock azimuth
position. In this picture, T is the tower diameter (assumed constant along the tower length),
O is the rotor shaft overhang, θ is the rotor shaft tilt angle, β is the blade precone angle and
R is the distance between the blade tip and the rotor axis.

σC ≤ −350 MPa (6.7)

R≤1 (6.8)

δ ≤ 7.44 m (DLC 1.1) (6.9)

δ ≤ 6.89 m (DLC 6.1b) (6.10)

6.2.4 Fatigue load analysis

The fatigue load analysis is carried out considering the fluctuations of the in-plane and out-
of-plane moments (i.e., moments caused by in-plane and out-of-plane forces, respectively)
acting on the blade root, for a design lifetime of 20 years. In-plane and out-of-plane moment
histories were obtained through a series of FAST simulations, according to the DLC 1.2 (as
104 HAWT design optimization under environmental uncertainty

explained in Sect. 6.2.1). In-plane and out-of-plane forces act respectively along the x- and
y-direction of the FAST’s coned coordinate framework. This coordinate system, depicted in
Fig. 6.2, rotates with the rotor, but does not pitch with the blades. The origin of the FAST’s

Fig. 6.2 FAST’s coned coordinate system. Image reproduced from [136].

coned coordinate system is at the intersection of the rotor axis and the plane of rotation, for
non coned rotors, or the apex of the cone of rotation, for coned rotors. y-axis points towards
the trailing edge of blade along the chord line if the pitch and twist were zero. x-axis is
orthogonal to the y-axis and the blade pitch axis. z-axis points along the pitch axis towards
the tip of blade. In-plane and out-of-plane moment histories were processed by means of
MLife (as explained in Chapter 2), which computed fatigue cycles for each time-series using
rainflow counting, and applied the Miner’s rule to eventually estimate the lifetime damage D.
For design purposes, fatigue failure is assumed to occur when the lifetime damage is equal to
1. The relationship between load range and cycles to failure (S-N curve) was modeled by
means of a Whöler exponent of 10.
In order to calculate the lifetime damage associated with the in-plane and out-of-plane
moments acting on the blade root, the ultimate design in-plane and out-of-plane moments
on the blade root needed to be specified. Practically, these ultimate moments lead the blade
root to structural failure. Ultimate design blade root in-plane and out-of-plane moments, in
particular, depend respectively on the in-plane and out-of-plane stiffnesses at the blade root.
Since the shape of the blade at the root is circular, and the internal structural layup is constant
along the cross-section periphery, the in-plane and out-of-plane stiffnesses at this section
are assumed equal. Therefore also the ultimate design blade root in-plane and out-of-plane
moments are constant. The blade root stiffness of the rotors generated during the course of
the optimization varies. In particular, it depends directly on the chord length at the blade
root, c(c1 ), and the value of the thickness parameter, s, defining the internal structural layup
6.2 Constraints 105

thickness. Therefore, in a preprocessing step, Table 6.3 was generated, providing the value
of the ultimate design blade root moments as a function of the chord length at the blade
root, and the thickness parameter. Within this table interpolations were performed by means
bilinear interpolation.
❳❳
❳❳❳ c(c1 ) [m]
❳❳❳ 0.1 1.9 3.7 5.5
s [-] ❳❳❳
❳
0.1 28.367 16,453 64,298 139,940
0.9 283.67 119,140 510,600 1,172,500
1.7 3,782.3 179,660 869,920 2,042,400
2.5 1,4562 213,700 1,134,700 2,798,900

Table 6.3 Ultimate design blade root moment, expressed in kN·m, as a function of the chord
length at the blade root, c(c1 ), and the thickness parameter, s.

Regarding the fatigue load analysis, the objective function reported above was minimized
subject to constraints on the lifetime damage associated with the in-plane and out-of-plane
moments, denoted respectively by D(RootMxc1) and D(RootMyc1) . For each fatigue load analysis,
carried out for the DLCs 1.2, in-plane and out-of-plane moments were augmented by a total
safety factor of 1.38. The lifetime damage maximum allowed value was 1. Indeed, for this
value of D, fatigue loads become critical (i.e., lead to structural failure by fatigue).
The constraints associated with the fatigue load analysis can be expressed by means the
following nonlinear inequalities:

D(RootMxc1) ≤ 1 (6.11)

D(RootMyc1) ≤ 1 (6.12)

6.2.5 Constraint on maximum rotor speed

Wind turbines are generally designed taking into account noise. The blade tip speed is the
most significant parameter affecting the turbine aerodynamic noise. In particular, aerody-
namic noise can be controlled by lowering the blade tip speed. The latest generation of
offshore wind turbines can operate at tip speeds up to 80 m/s [137]. Accordingly, in this
work, as a surrogate for noise constraint, the maximum rotor speed was set equal to 12.1
rpm, leading to a tip speed of 80.0 m/s. Practically, this constraint was implemented by
conveniently setting the controller of all turbines generated during the optimization in order
to achieve a constant maximum rotor speed equal to 12.1 rpm.
106 HAWT design optimization under environmental uncertainty

6.2.6 Blade natural frequency analysis

In order to avoid blade resonance, a constraint on the minimum value of the blade first natural
frequency, ω1 , was also enforced. Among all blade natural frequencies, the first natural
frequency is the lowest, and therefore it was the only one considered here. Considering a
safety factor of 1.1, the minimum allowable blade first natural frequency was set equal to the
maximum blade passing frequency (calculated at the maximum rotor speed of 12.1 rpm), fB .
fB was evaluated as follows:
fB = 1.1 · Nb fR (6.13)

where Nb is the number of blades (equal to 3), and fR is the frequency corresponding to the
maximum angular speed of the rotor (12.1 rpm) equal to

12.1
fR = = 0.202 Hz (6.14)
60
Therefore, for each of the turbines generated during the optimization, the minimum allowable
blade first natural frequency, at a rotor speed of 12.1 rpm, was equal to 0.666 Hz.

The constraint associated with the blade first natural frequency can be expressed by
means the following nonlinear inequality:

ω1 ≥ 0.666 Hz (6.15)

6.2.7 Geometric feasibility checks

Each airfoil generated during the course of the optimization process was subject to a number
of feasibility checks, aiming to avoid unphysical or unconventional shapes. These checks
identified and discarded airfoils with self-intersecting shapes, and airfoils respectively with
more than 1 and 2 changes in the curvature sign of the suction and pressure sides. Airfoil
relative thicknesses and their chordwise locations are also subject to feasibility checks.
Relative thickness of root, midspan and tip airfoils are enforced to vary within the ranges
[38%, 42%], [23%, 27%] and [16%, 20%], respectively, while, for all airfoils, the location of
maximum thickness was constrained to range between 20% and 40% of chord.
6.3 Problem formulation 107

6.3 Problem formulation

Denoting by g j (bb, ū) the set of nonlinear inequality constraints described by Eqs. 6.6, 6.7,
6.8, 6.9, 6.10, 6.11, 6.12 and 6.15, the robust optimization problem is:

Find: b = {b1 , ..., b54 }

to minimize: frob (bb, ū)
where: ū ∼ U(ūL , ūU ) (6.16)
subject to: g j (bb, ū) ≤ 0
and: bL ≤ b ≤ bU

where frob (bb, ū) is the objective function of the robust problem defined by Eq. 6.1. The
symbols ū ∼ U(ūL , ūU ) indicate that the mean wind speed ū is uniformly distributed in the
range from ūL , equal to 7.0 m/s, to ūU , equal to 13.0 m/s. bL and bU are respectively the
lower and upper bounds of the design variables’ variability ranges, as shown in Chapter 5.
The deterministic optimization problem is instead formulated as follows:

Find: b = {b1 , ..., b54 }

to minimize: fdet (bb, ū)
where: ū = ūdet (6.17)
subject to: g j (bb, ū) ≤ 0
and: b L ≤ b ≤ bU

where ūdet is equal to 10.0 m/s, and fdet (bb, ū) is objective function of the deterministic
problem expressed as follows:

fdet (bb, ū) = LCOE(bb, ū) (6.18)

All the constraints, g j (bb, ū), appearing in Eq. 6.16 and Eq. 6.17 are treated deterministically
(i.e., not considering the variability of the wind speed distribution mean). These constraints
are treated in such a way primarily because they have been calculated by augmenting the
loads by safety factors that consider various types of uncertainty. Moreover, these constraints
do not depend on the wind speed distribution mean. Therefore, these constraints were
evaluated for a single wind speed distribution mean equal to 10.0 m/s
108 HAWT design optimization under environmental uncertainty

6.4 Results and discussion

The purpose of this section is to illustrate and assess the results obtained for the robust
and deterministic optimization problems formulated in Sect. 6.3. The aim of the robust
optimization was to minimize both the expectation and the standard deviation of the LCOE
of the reference 5-MW HAWT defined in Chapter 5, considering the uncertainty affecting
the mean wind speed of the installation site. In the attempt to determine the Pareto front
of optimal solutions using the weighted sum method illustrated in Sect. 6.1, several robust
optimizations were performed by varying the weights given to µLCOE and σLCOE in Eq. 6.1.
This assessment indicated that the Pareto front associated with the design problem under
investigation is rather small. In this circumstance, variations of the weights do not result
in designs yielding significantly different values of µLCOE and σLCOE . Therefore, in this
section, a representative robust solution obtained giving equal weights to µLCOE and σLCOE ,
i.e., αw = 0.5, is considered. For comparison purposes, a deterministic optimization was
performed along with the robust one. The goal of the deterministic optimization was to
minimize the LCOE, considering a fixed mean wind speed of 10 m/s. The rest of this section
deals with the comparison between the reference turbine and the robust and deterministic
ones.
As reported in Table 6.4, the robust and deterministic optimizations led to the design
of improved rotors with respect to the reference one, achieving a reduction in both µLCOE
and σLCOE . With respect to the reference turbine, the robust optimization achieved a 7.69%
and 12.5% reduction in µLCOE and σLCOE , respectively. The lower µLCOE was due primarily
to the lower blade weight (- 44.01%), and the enhanced aerodynamic characteristics of the
robust rotor, leading to a greater mean of AEP (+ 1.52%). The standard deviation of AEP
of the robust configuration with respect to that of the reference turbine decreases by 0.70%.
With respect to the reference configuration, the deterministic optimization achieves a 7.16%
and 12.0% reduction in µLCOE and σLCOE , respectively. The lower µLCOE is due to the lower
blade weight (- 38.81%), and the greater mean of AEP (+ 1.91%). The standard deviation of
AEP of the deterministic configuration with respect to that of the reference turbine decreases
by 0.93%. With reference to the deterministic optimum, the robust one is characterized
by better overall performance, with lower µLCOE (- 0.57%) and σLCOE (- 0.59%). The
enhancement of the robust rotor performance with respect to that of the deterministic one are
related to the lower blade weight (- 8.50%), which compensates for a slightly lower mean of
AEP (- 0.38%). The standard deviation of AEP of the robust configuration is 0.23% greater
with respect to that of the deterministic turbine.
As depicted in Table 6.4, the deterministic and robust designs reported in this thesis
have obtained by keeping the rotor diameter, rated power, cut-in and cut-out speeds constant
6.4 Results and discussion 109

ref. det. rob.

rated power [MW] 5 5 5
rotor diameter [m] 126 126 126
cut-in wind speed [m/s] 3 3 3
cut-out wind speed [m/s] 25 25 25
µLCOE [$ / kWh] 0.1132 0.1051 0.1045
σLCOE [$ / kWh] 0.0192 0.0169 0.0168
µAEP [kWh] 23.03 ·106 23.47 ·106 23.38 ·106
σAEP [kWh] 4.31 ·106 4.27 ·106 4.28 ·106
wb [kg] 44.216 ·103 27.055 ·103 24.755 ·103

Table 6.4 Comparison of the gross design specifications and overall performance of the
reference, deterministic and robust designs. µLCOE and σLCOE are mean and standard
deviation of LCOE, respectively; µAEP and σAEP are mean and standard deviation of AEP,
respectively; wb is the blade weight.

and equal to those of the reference turbine. As explained above, with reference to the
deterministic optimum, the robust one is characterized by better overall performance, with
0.57% lower LCOE mean. Considering the same probabilistic scenario reported in the thesis,
in the work of Ning et al. [17] the robust design has achieved a 1.2% lower average LCOE
than that of a deterministically designed rotor. With respect to the results highlighted in
this thesis, the higher reduction of the LCOE mean (of the robust rotor with respect to the
deterministic one) achieved by Ning at al. is due to the fact that these authors considered,
in addition to the design variables considered in the thesis, the diameter and rated power as
design variables, allowing for a higher design space. Indeed, in the optimization process,
the inclusion of the rotor diameter and turbine rated power as design variables allows
greater freedom in the design search, allowing the generation of significantly diverse wind
turbine configurations. Rotor diameter and turbine rated power, however, affect directly
the design of several structural components of the turbine, such as the tower. Therefore,
an optimization system that considers variations in rotor diameter and turbine rated power
should be able to structurally assess turbine structural components, such as the tower, for
each configuration generated during the course of the optimization process. At present, the
optimization system presented in this thesis does not include suitable tools for the structural
analysis of turbine structural components, focusing purely on the aeroservoelastic design
of the rotor. Therefore, in this context, the rotor diameter and turbine rated power are not
included as design variables, that is they are kept constant, which has allowed us to keep the
same turbine structural components for each turbine.
110 HAWT design optimization under environmental uncertainty

The purpose of the rest of this chapter is to present and compare the characteristics of
the reference, deterministic and robust design in detail. Table 6.5 compares the structural
properties at selected wind speeds and radial positions for the reference, deterministic and
robust blades. The ultimate load analysis of these turbines highlights that the maximum
stresses, buckling criteria and tip deflections are achieved for the DLC 6.1b. As seen in
Sect. 6.2.1, this DLC considers the turbine in parked conditions, subject to an extreme
wind speed of 70 m/s. In such conditions, stresses, buckling criteria and tip deflections are
evaluated on a blade in the 3 o’clock azimuth position, representing the worst case for the
edgewise loads due to the blade weight forces. Under such conditions, the reference and the
robust blades have an active constraint on the maximum allowable buckling criteria acting on
the top surface of the blade at 63.3% span, while the deterministic and robust turbines have
an active constraint on the maximum allowable tensile stress, on the bottom surface of the
blade at 14.3% span. Table 6.5 reports also the maximum compressive stress σC acting on the
top surface at the blade root section of the three blade designs, at a wind speed of 70.0 m/s
for a blade in the 3 o’clock azimuth position. One sees that σC has increased considerably
for the deterministic and robust designs, and it is respectively 19.43% and 3.71% lower than
the allowed maximum values. At a wind speed of 70.0 m/s, the maximum tip deflections of
the deterministic and robust blades are about 35.85% and 19.16% lower than the allowed
maximum. Calculated from the cut-in and cut-out wind speeds, the means of the lifetime
damage caused by the in-plane and out-of-plane moments at the blade root are well below the
maximum allowed values for both the deterministic and robust turbines. The deterministic
and robust blade first natural frequencies are 37.84% 34.38% higher than the minimum
allowed value.

Figures 6.3 and 6.4 present respectively the comparison of the airfoils and force coeffi-
cients of the reference, deterministic and robust blade designs. The root, midspan and tip
airfoils of the deterministic and robust designs feature a higher lift-to-drag ratio than that of
the corresponding reference turbine airfoils. It is observed that, while the maximum airfoil
thickness and the chordwise position where such maximum is achieved vary fairly little
between corresponding airfoils of the three blade designs, the airfoils of the deterministic and
robust blades are more cambered than those of the reference blade. The camber line, defined
as the locus of the points midway between the suction and pressure sides, plays a crucial
role in the improvements of the robust airfoils’ aerodynamic performance. Indeed, the larger
amount of camber of the deterministic and robust blade airfoils enables them to achieve a
higher lift coefficient. This is particularly evident for the root airfoil, the maximum CL of
which is about 40% higher than that of the reference blade root airfoil. It is also observed
that, with respect to the robust midspan and tip airfoils, for an AoA ranging from -5 to about
6.4 Results and discussion 111

u [m/s] r [-] surf. ref. det. rob. const.

σT [MPa] 70.0 0.143 bot 138 350 350 ≤ 350
σC [MPa] 70.0 0 top -50 -282 -337 ≥ -350
R 70.0 0.633 top 1.000 0.919 1.000 ≤1
δ [m] 70.0 1 - 2.77 4.42 5.57 ≤ 6.89
D(RootMxc1) (3.0, 25.0) 0 - 1.4 ·10−9 1.5 ·10−2 4.9 ·10−2 ≤1
D(RootMyc1) (3.0, 25.0) 0 - 2.5 ·10−11 1.2 ·10−2 4.7 ·10−2 ≤1
ω1 [Hz] - - - 1.173 0.918 0.895 ≥ 0.666

Table 6.5 Comparison of the structural characteristics of the reference, deterministic and
robust blade designs at selected wind speeds and radial positions. σT and σC denote tensile
and the compressive stresses, respectively; R is the buckling criteria; δ is the tip deflection;
D(RootMxc1) and D(RootMyc1) are the lifetime damages caused by the in-plane and out-of-plane
moments at the blade root, respectively; ω1 is the blade first natural frequency at a rotor
speed of 12.1 rpm.

20 deg, the deterministic midspan and tip airfoils achieve higher lift-to-drag ratios. This is
due to greater cambers. The optimization of the airfoil shapes is also driven by structural con-
siderations. In particular, the midspan and tip airfoils have a lower mean radius of curvature
over the aft portion of the suction side with respect to the reference airfoils. This results in an
increase of the buckling strength of the blade laminates at a location characterized by a high
buckling stress concentration (i.e., over the aft part of the top surface, at around half-span of
the blade). Moreover, the root airfoil of the deterministic and robust turbines have a thicker
trailing edge, improving in this way both the flapwise and edgewise bending stiffnesses.
Radial chord and twist profiles of the two designs are reported respectively in the left
and right top subplots of Fig. 6.5. Due to smaller chord lengths of the deterministic and
robust blade designs (top left subplot of Fig. 6.5), these rotors have a lower solidity at most
radii, and this occurrence contributes significantly to decrease the blade weight. The blade
weight reduction is also a result of the 6.95% and 13.69% decrease of the internal structure
thickness parameter s of the deterministic and robust turbines, respectively (see Table 6.6).
The deterministic and robust blade designs also have less twist than the reference design
along the inboard part of the blade and more twist along the outboard part. The consequences
of this trend are discussed below.
The left and right subplots of Fig. 6.6 report respectively the rotor speed and the tip-speed
ratio of the three turbines against the wind speed. Figure 6.7 shows instead the power curve
of the reference, deterministic and robust turbines. As expected, the power extracted by the
deterministic and robust HAWT designs is higher than that of the reference turbine in regions
11/2, 2 and 21/2. The power extraction of the deterministic design in this wind speed range is
112 HAWT design optimization under environmental uncertainty

root
0.3
0.2
0.1
y/c [−] 0
−0.1
−0.2 ref.
det.
−0.3 rob.
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
x/c [−]
midspan
0.3
0.2
0.1
y/c [−]

0
−0.1
−0.2
−0.3
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
x/c [−]
tip
0.3
0.2
0.1
y/c [−]

0
−0.1
−0.2
−0.3
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
x/c [−]

Fig. 6.3 Comparison of the base airfoil shapes of the reference, deterministic and robust
designs.

higher than that of the robust one. It is also noted that the rotor speed of the robust turbine in
region 2 is higher than that of the reference turbine, while the rotor speed of the deterministic
turbine in region 2 is slightly lower than that of the reference turbine.
The torque control parameter, the tip-speed ratio in region 2, and the thickness parameter
of the three turbines are provided in Table 6.6. These results highlight that the robust design
features an increment of 3% of the tip-speed ratio in region 2, and a reduction of 13.69% of
the thickness of the internal structure over the reference turbine. The deterministic design
instead is characterized by a slight decrement of 0.26% of the tip-speed ratio in region 2, and
6.4 Results and discussion 113

root root
3 0.6
2.5 0.5
2
0.4

CD [−]
CL [−]

1.5
0.3
1
ref. 0.2
0.5
det. 0.1
0 rob.
−0.5 0
−5 0 5 10 15 20 25 30 −5 0 5 10 15 20 25 30
AoA [deg] AoA [deg]
midspan midspan
3 0.6
2.5 0.5
2
0.4
CD [−]
CL [−]

1.5
0.3
1
0.2
0.5
0 0.1
−0.5 0
−5 0 5 10 15 20 25 30 −5 0 5 10 15 20 25 30
AoA [deg] AoA [deg]
tip tip
3 0.6
2.5 0.5
2
0.4
CD [−]
CL [−]

1.5
0.3
1
0.2
0.5
0 0.1
−0.5 0
−5 0 5 10 15 20 25 30 −5 0 5 10 15 20 25 30
AoA [deg] AoA [deg]

Fig. 6.4 Comparison of lift and drag coefficients of the reference, deterministic and robust
designs’ base airfoils, for an indicative Reynolds number of 12 · 106 .

a reduction of 6.95% of the thickness of the internal structure with respect to the reference
turbine.
The higher mean of AEP of the deterministic and robust turbines over the reference one
is determined by a higher power production of the former turbines in regions 11/2, 2 and
21/2, as highlighted above. This improvement is due to the reported alterations of the outer
blade shapes, and, in the case of the robust turbine, a higher rotor speed. All these alterations
have conflicting effects on the net power production, and a discussion on their interactions in
regions 11/2, 2 and 21/2 for both the deterministic and robust turbines is provided below.
114 HAWT design optimization under environmental uncertainty

6 16
ref.
det. 14
5
rob. 12
4 10

ϑ [deg]
8
c [m]

3
6
2 4
2
1
0
0 −2
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
r [−] r [−]

Fig. 6.5 Comparison of chord and twist distributions (left and right subplots, respectively) of
the reference, deterministic and robust designs.

14 16
ref.
det. 14 8.2
rob. 8
12 7.8
12 7.6
7.4
ω [rpm]

10 8 9 10
λ [−]

10
8
9.6
6
8 9.4
4
7.5 8 8.5
6 2
0 5 10 15 20 25 0 5 10 15 20 25
u [m/s] u [m/s]

Fig. 6.6 Comparison of rotor speed and tip-speed ratio against wind speed (left and right
subplots, respectively) of the reference, deterministic and robust designs.

At the inboard sections, the twist reduction of the deterministic design visible in the top
right subplot of Fig. 6.5 yields an increment of the aerodynamic load, as one can see in the
right subplot of Fig. 6.8. This is a consequence of the expected higher AoA, and also the
higher lift coefficient curves at the inboard blade sections. At a wind speed of 9.0 m/s, the
6.4 Results and discussion 115

6000
ref.
det.
5000 rob.
5200
4000
5100
Pe [kW]

3000
5000

2000 4900

1000 4800
11 11.2 11.4 11.6 11.8

0
0 5 10 15 20 25
u [m/s]

Fig. 6.7 Electrical power as a function of wind speed of the reference, deterministic and
robust designs.

ref. det. rob.

K [N · m/s2 ] 2.13 (λ = 7.67) 2.26 (λ = 7.65) 2.03 (λ = 7.90)
s 0.964 0.897 0.832

Table 6.6 Torque control parameter with resulting tip-speed ratio in region 2 in parentheses,
and thickness parameter of the reference, deterministic and robust designs.

AoA at the root airfoil of the deterministic turbine increases by about 2.5 deg, slightly less
than 3 deg, which is the twist reduction at this radius. Moreover, the chords of the inboard
sections of the deterministic design are smaller than those of the reference turbine. As a
consequence, the increment of the aerodynamic forces due to higher AoA and the higher lift
coefficient curve is partly compensated by the reduction of these forces caused by a smaller
blade planform. As a consequence, the tangential force, and thus the aerodynamic torque,
of the inboard sections of the deterministic blade design is significantly higher than that of
the reference blade design. At 40% blade length the tangential force of the former blade
achieves increments of up to 10% over the latter blade design. The higher power extracted
by the inboard sections of the deterministic design is thus a result of the higher torque.
Like at the inboard sections, the deterministic blade is more heavily loaded than the
reference one at the outboard sections. This is a consequence of a higher lift coefficient
distribution and larger chords. The increase of the aerodynamic loads due to higher lift
coefficient distribution and larger chords is partially compensated by the lower AoA, which is
a consequence of the twist increase at near-tip sections shown in the right subplot of Fig. 6.5.
116 HAWT design optimization under environmental uncertainty

60 2
ref.
det.
50
rob. 1.5

40
AoA [deg]

FT [kN]
14
30
13
0.5
12
20
0.15 0.2
0
10

0 −0.5
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
r [−] r [−]

Fig. 6.8 Comparison of the angle of attack and tangential force distributions along the blade
span (left and right subplots, respectively) of the reference, deterministic and robust designs,
for a wind speed of 9.0 m/s.

At a wind speed of 9.0 m/s, the AoA at the near-tip airfoils of the robust turbine decreases by
about 2.6 deg, slightly more than the 2.2 deg twist increase at the blade tip. As a result, at the
outboard sections, the deterministic turbine extracts more power than the reference one due
to greater tangential forces, and thus a higher aerodynamic torque.

At the inboard sections, the twist reduction of the robust design shown in the right subplot
of Fig. 6.5 yields an increment of the aerodynamic load, as one can see in the right subplot
of Fig. 6.8. This is a consequence of the expected higher AoA, and also the higher lift
coefficient curves of the inboard blade sections. However, the AoA increment due to the
lower twist is reduced by the increment of about 3% of the rotational speed. Due to this,
at a wind speed of 9.0 m/s, the AoA at the root airfoil of the robust turbine increases by
about 1 deg (see the left subplot of Fig. 6.8), and not by more than 2 degrees, which is the
twist reduction at this radius. Moreover, the chords of the inboard sections of the robust
design are smaller than those of the reference turbine. As a consequence, the increment of
the aerodynamic forces due to slightly higher AoA and the higher lift coefficient curve is
partly compensated by the reduction of these forces caused by a smaller blade planform. As
a consequence, the tangential force, and thus the aerodynamic torque, of the inboard sections
of the robust blade design is noticeably higher than that of the reference blade design. At
30% blade length the tangential force of the former blade achieves increments of up to 2.5%
6.4 Results and discussion 117

over the latter blade design. The higher power extracted by the inboard sections of the robust
design is thus a result of both the higher torque and the higher rotational speed.
At the outboard sections, the twist increment of the robust design yields a reduction of
the aerodynamic load. This is a consequence of the expected lower AoA. The AoA at high
blade radii is further reduced by the higher rotational speed of the robust rotor. Therefore,
at a wind speed of 9.0 m/s, the AoA at the near-tip airfoils of the robust turbine decreases
by about 2.5 deg, slightly more than the 1.9 deg twist increase at the blade tip visible in the
right subplot of Fig. 6.5. As for the power production of the outboard sections, it is observed
that the chords of the outboard sections of the robust design are larger than those of the
reference turbine. In addition, the lift coefficient distribution along the outboard part of the
robust blade is higher than that of the reference design. As a consequence, the reduction of
the aerodynamic forces due to lower AoA is compensated by the increment of these forces
caused by a larger blade planform and higher lift coefficient. Thus, the tangential force and
the aerodynamic torque of the outboard sections of the two blade designs are fairly similar,
and the outboard sections of the robust blade design produce more power due primarily to
the higher rotational speed.
Figures 6.9, 6.10 and 6.11 compare the blade displacements and the applied forces of the
reference, deterministic and robust blade designs, respectively, at the 3 o’clock azimuthal
position for a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm. This figure highlights
a) the lighter weight of the inboard part of the deterministic and robust designs, and b) the
larger deflections of the deterministic and robust designs resulting from the larger normal
components of the aerodynamic force, and lower stiffness with respect to the reference blade.

Figures 6.12, 6.13 and 6.14 report the laminate normal stresses acting on the reference,
deterministic and robust blades, respectively, at the 3 o’clock azimuthal position for a wind
speed of 12.0 m/s and a rotor speed of 12.1 rpm. It is noted that the maximum tensile stress
occurs on the blade suction side close to its root for all configurations, and that the maximum
tensile stress of the deterministic and robust designs is higher than that of the reference blade,
as previously indicated. This points to better utilization of the material strength, since the
maximum stress of the deterministic and robust blade designs is still lower than the maximum
value enforced by the structural constraints.
Figures 6.15, 6.16 and 6.17 report the buckling criteria on the reference, deterministic
and robust blades, respectively, at the 3 o’clock azimuthal position for a wind speed of 12.0
m/s and a rotor speed of 12.1 rpm. These figures highlight that, for all the three blades, the
maximum buckling stress concentration acts over the aft part of the top surface, at around
half-span of the blade.
118 HAWT design optimization under environmental uncertainty

aerodynamic forces
0
weight forces
−2
−4
−6
z [m]

−8
−10
−12
−14

0
−10
−20
−30
−40
−50
y [m] −60
−2 0 2 4 6 8

x [m]

Fig. 6.9 Displacement and applied forces of the reference blade in the 3 o’clock azimuth
position, for a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.

aerodynamic forces
0
weight forces
−2
−4
−6
z [m]

−8
−10
−12
−14

0
−10
−20
−30
−40
−50
y [m] −60
−2 0 2 4 6 8

x [m]

Fig. 6.10 Displacement and applied forces of the deterministic blade in the 3 o’clock azimuth
position, for a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.
6.4 Results and discussion 119

aerodynamic forces
0
weight forces
−2
−4
−6
z [m] −8
−10
−12
−14

0
−10
−20
−30
−40
−50
y [m] −60
−2 0 2 4 6 8

x [m]

Fig. 6.11 Displacement and applied forces of the robust blade in the 3 o’clock azimuth
position, for a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.

Fig. 6.12 Laminate normal stress of the reference blade in the 3 o’clock azimuth position, for
a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.
120 HAWT design optimization under environmental uncertainty

Fig. 6.13 Laminate normal stress of the deterministic blade in the 3 o’clock azimuth position,
for a wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.

Fig. 6.14 Laminate normal stress of the robust blade in the 3 o’clock azimuth position, for a
wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.
6.4 Results and discussion 121

buckling criteria, R
0 0.2 0.4 0.6 0.8 1

1.5
1
0.5
y [m]
0
−0.5
−1
−1.5

0
10
20
30
40
50

z [m] 60
−1.5−1−0.5 0 0.5 1 1.5
x [m]

Fig. 6.15 Buckling criteria of the reference blade in the 3 o’clock azimuth position, for a
wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.

buckling criteria, R
0 0.2 0.4 0.6 0.8 1

1.5
1
0.5
y [m]

0
−0.5
−1
−1.5

0
10
20
30
40
50

z [m] 60
−1.5−1−0.5 0 0.5 1 1.5
x [m]

Fig. 6.16 Buckling criteria of the deterministic blade in the 3 o’clock azimuth position, for a
wind speed of 12.0 m/s and a rotor speed of 12.1 rpm.
122 HAWT design optimization under environmental uncertainty

buckling criteria, R
0 0.2 0.4 0.6 0.8 1

1.5
1
0.5
y [m]

0
−0.5
−1
−1.5

0
10
20
30
40
50

z [m] 60
−1.5−1−0.5 0 0.5 1 1.5
x [m]

Fig. 6.17 Buckling criteria of the robust blade in the 3 o’clock azimuth position, for a wind
speed of 12.0 m/s and a rotor speed of 12.1 rpm.
Chapter 7

Conclusions and suggestions for further

work

7.1 Summary and concluding remarks

A novel integrated computational framework for the multidisciplinary and robust design
optimization of HAWTs has been developed, assessed and demonstrated by performing
the aeroservoelastic design optimization of a multi-megawatt HAWT under environmental
uncertainty.

7.1.1 Conclusions on the development of the design system

The multidisciplinary and robust design optimization system has been presented, highlighting
the interconnection and main characteristics of its constituent modules. An overview of the
topics which are fundamental to understanding the aeroservoelastic behavior of HAWTs,
uncertainty propagation and mathematical optimization has been also provided. The key
features of the developed framework include:

• the integration of a method for uncertainty propagation to account for uncertainty in

the design optimization process, and

• the concurrent optimization of the blade airfoils, chord and twist radial distributions,
rotor speed and internal structural layup, and

• the enforcement of a comprehensive set of aerostructural constraints to ensure the

blade structural integrity during the course of the optimization.
124 Conclusions and suggestions for further work

The main element of novelty associated with the presented HAWT MDO system is that it
simultaneously includes uncertainty propagation, the optimization of the blade airfoil shapes
within the aeroservoelastic optimization of the rotor, and the enforcement of a comprehensive
set of aerostructural constraints. Airfoil optimization has been made possible by developing
a suitable parametrization based on a composite Bézier curve, enabling high flexibility, yet
keeping a low number of degrees of freedom. Another important feature of the presented
system is that it is made up of integrated modules. In principle, this allows one to replace one
or more current modules (for instance with higher-fidelity ones) without the need to modify
significantly the rest of the system.

7.1.2 Conclusions on the application of the design system

The developed computational framework has been successfully used to optimize the design of
a HAWT rotor subject to the environmental uncertainty associated with the mean value of the
wind speed frequency distribution. The probabilistic design exercise has involved minimizing
the expected value and standard deviation of LCOE by optimizing the turbine’s blade external
geometry, rotor speed and internal structural layup. The main motivation for analyzing this
scenario has been to demonstrate the effectiveness of the developed probabilistic design
system taking into account a realistic problem.
The achieved robust optimum turbine design has been found to achieve a reduction of
about 7.7% of the expectation of LCOE, and a reduction of about 12.5% of the standard
deviation of this objective function with respect to the initial baseline reference turbine. This
improvement has been made possible primarily by both increasing the lift-to-drag ratio of
the rotor airfoils and decreasing the rotor blade weight, still satisfying the aerostructural
constraints.
The better airfoil aerodynamic performance of the robust airfoils with respect to the
reference ones are due to greater airfoil cambers. It has also been shown that structural
considerations have driven the optimization of the airfoils. In particular, the robust airfoils
have a lower mean radius of curvature over the aft portion of the suction side with respect
to the reference airfoils. This results in an increase of the buckling strength of the blade
laminates at a location characterized by a high buckling stress concentration.
The probabilistically designed turbine has also achieved more favorable probabilistic
performance than those of a turbine designed deterministically, with a reduction of about
0.6% of the expectation of LCOE, and a reduction of about 0.6% of the standard deviation of
LCOE. These results highlight that taking into account the environmental uncertainty in the
optimization process, has enabled us to achieve a robust design with lower expectation and
7.2 Future work 125

sensitivity of LCOE to the mean value of the wind speed distribution with respect to those of
a deterministically designed turbine.
The obtained optimal solutions may require further verifications due to possible in-
accuracy of the low-fidelity aerodynamic module XFOIL for stall and post-stall airfoil
performance analysis, and also the high level of empiricism of the 3D correction model
used by the BEM analysis. The main objective of the probabilistic turbine design opti-
mization exercise presented herein, however, has been to highlight the potential of robust
fully integrated MDO of improving HAWT design technologies, rather than proposing a
ready-for-installation design, and this objective has been accomplished.

7.2 Future work

On the methodology front, future improvements of the developed design framework include
the incorporation of higher-fidelity analysis modules, such as transitional NS CFD for stall
and post-stall airfoil aerodynamics, and possibly a full 3D finite element stress analysis
code for rotor structural analysis. Another planned extension of the current design system
encompasses the integration of a module for the tower structural analysis and design. This
will allow one to extend the design search by including turbine rated power and rotor diameter
as design variables.
On the application side, planned and partly already ongoing work includes using the
developed methodology to carry out the design optimization of multi-megawatt wind tur-
bine rotors considering other sources of uncertainty, such as those associated with rotor
manufacturing, assembly and blade internal structural layup errors.
References

[1] Global Wind Energy Council. Global Wind Report 2013 - Annual market update.
Technical report, GWEC, Brussels, Belgium, 2013.
[2] P. Brøndsted. Introduction. In P. Brøndsted and R. P. L. Nijssen, editors, Advances in
wind turbine blade design and materials, volume 47 of Energy. Woodhead Publishing
Limited, 2013. ISBN 978-0-85709-426-1.
[3] T. Ashuri. Beyond Classical Upscaling: Integrated Aeroservoelastic Design and
Optimization of Large Offshore Wind Turbines. PhD thesis, TU Delft, 2012.
[4] G. Sieros, P. Chaviaropoulos, J. D. Sørensen, B. H. Bulder, and P. Jamieson. Upscaling
wind turbines: theoretical and practical aspects and their impact on the cost of energy.
Wind Energy, 15(1):3–17, 2012.
[5] P. Jamieson. Innovation in Wind Turbine Design. Wiley, 2011.
[6] J. R. R. A. Martins and A. B. Lambe. Multidisciplinary design optimization: A survey
of architectures. AIAA Journal, 51(9):2049–2075, 2013.
[7] P. Fuglsang and H. A. Madsen. Optimization method for wind turbine rotors. Journal
of Wind Engineering and Industrial Aerodynamics, 80(1–2):191–206, 1999.
[8] E. Benini and A. Toffolo. Optimal design of horizontal-axis wind turbines using blade-
element theory and evolutionary computation. Journal of Solar Energy Engineering,
124:357–363, 2002.
[9] Juan Méndez and D. Greiner. Wind blade chord and twist angle optimization by using
genetic algorithms. In Fifth International Conference on Engineering Computational
Technology, Las Palmas de Gran Canaria, Spain, 2006.
[10] G. Kenway and J. R. R. A. Martins. Aerostructural shape optimization of wind
turbine blades considering site-specific winds. In 12th AIAA/ISSMO Multidisciplinary
Analysis and Optimization Conference, MAO, Victoria, British Columbia, Canada,
2008.
[11] W. Xudong, W. Z. Shen, W. J. Zhu, J. N. Sørensen, and C. Jin. Shape optimization of
wind turbine blades. Wind Energy, 12(8):781–803, 2009.
[12] G. Petrone, C. de Nicola, D. Quagliarella, J. Witteveen, and G. Iaccarino. Wind
turbine optimization under uncertainty with high performance computing. In 29th
AIAA Applied Aerodynamics Conference 2011, Honolulu, Hawaii, USA, 2011.
128 References

[13] J. Jeong, K. Park, S. Jun, K. Song, and D. Lee. Design optimization of a wind turbine
blade to reduce the fluctuating unsteady aerodynamic load in turbulent wind. Journal
of Mechanical Science and Technology, 3(26):827–838, 2012.

[14] C. L. Bottasso, F. Campagnolo, and A. Croce. Multi-disciplinary constrained opti-

mization of wind turbines. Multibody System Dynamics, 27(1):21–53, 2012.

[15] R. W. Vesel and J. J. McNamara. Performance enhancement and load reduction of a 5

MW wind turbine blade. Renewable Energy, 66:391–401, 2014.

[16] T. Ashuri, M. B. Zaaijer, J. R. R. A. Martins, G. J. W. van Bussel, and G. A. M. van

Kuik. Multidisciplinary design optimization of offshore wind turbines for minimum
levelized cost of energy. Renewable Energy, 68:893–905, 2014.

[17] S. A. Ning, R. Damiani, and P. J. Moriarty. Objectives and constraints for wind turbine
optimization. Journal of Solar Energy Engineering, Transactions of the ASME, 136(4),
2014.

[18] C. L. Bottasso, A. Croce, L. Sartori, and F. Grasso. Free-form design of rotor blades.
Journal of Physics: Conference Series, 524(1), 2014.

[19] M. Caboni, E. Minisci, and M. S. Campobasso. Robust aerodynamic design opti-

mization of horizontal axis wind turbine rotors. In D. Greiner, B. Galván, J. Periaux,
N. Gauger, K. Giannakoglou, and G. Winter, editors, Advances in Evolutionary and
Deterministic Methods for Design, Optimization and Control in Engineering and
Sciences, volume 36 of Computational Methods in Applied Sciences. Springer, 2014.
ISBN 978-3-319-11540-5.

[20] M. S. Campobasso, E. Minisci, and M. Caboni. Aerodynamic design optimiza-

tion of wind turbine rotors under geometric uncertainty. Wind Energy, 2014. DOI:
10.1002/we.1820.

[21] P. Fuglsang and K. Thomsen. Site-specific design optimization of 1.5–2.0 MW wind

turbines. Journal of Solar Energy Engineering, Transactions of the ASME, 123(4):296–
303, 2001.

[22] P. Fuglsang, C. Bak, J. G. Schepers, B. Bulder, T. T. Cockerill, P. Claiden, A. Olesen,

and R. van Rossen. Site-specific design optimization of wind turbines. Wind Energy,
5(4):261–279, 2002.

[23] T. Diveux, P. Sebastian, D. Bernard, J. R. Puiggali, and J. Y. Grandidier. Horizontal

axis wind turbine systems: optimization using genetic algorithms. Wind Energy,
4(4):151–171, 2001.

[24] K. Lee, W. Joo, K. Kim, D. Lee, K. Lee, and J. Park. Numerical Optimization using
Improvement of the Design Space Feasibility for Korean Offshore Horizontal Axis
Wind Turbine Blade. In EWEC 2007, Milan, Italy, 2007.

[25] K. Maki, R. Sbragio, and N. Vlahopoulos. System design of a wind turbine using a
multi-level optimization approach. Renewable Energy, 43:101–110, 2012.
References 129

[26] M. Jureczko, M. Pawlak, and A. M˛eżyk. Optimisation of wind turbine blades. Journal
of Materials Processing Technology, 167:463–471, 2005.
[27] E. Lund and J. Stegmann. On Structural Optimization of Composite Shell Structures
Using a Discrete Constitutive Parametrization. Renewable Energy, 8:109–124, 2005.
[28] International Standard. International Electrotechnical Commission, 2009-02. IEC
61400-3 Edition 1.0.
[29] G. J. Park, T. H. Lee, K. H. Lee, and K. H. Hwang. Robust design: An overview.
AIAA Journal, 44(1), 2006.
[30] INVESTOPEDIA. Variance. See also URL
http://www.investopedia.com/terms/v/variance.asp.
[31] G. Iaccarino. Quantification of Uncertainty in Flow Simulations Using Probabilistic
Methods. Technical report, Stanford University, Stanford, California, USA, 2008.
VKI Lecture Series.
[32] M. Padulo, M.S. Campobasso, and M.D. Guenov. A Novel Uncertainty Propagation
Method for Robust Aerodynamic Design. AIAA Journal, 49(3):530–543, 2011.
[33] D. T. Griffith and T. D. Ashwill. The Sandia 100-meter All-glass Baseline Wind Tur-
bine Blade: SNL100-00. Technical report, Sandia National Laboratories, Albuquerque,
New Mexico, USA, 2011. SAND2011-3779.
[34] B. R. Resor. Definition of a 5MW/61.5m Wind Turbine Blade Reference Model.
Technical report, Sandia National Laboratories, Albuquerque, New Mexico, USA,
2013. SAND2013-2569.
[35] M. H. Straathof. Shape Parameterization in Aircraft Design: A Novel Method, Based
on B-Splines. PhD thesis, TU Delft, 2012.
[36] J. Jonkman, S. Butterfield, W. Musial, and G. Scott. Definition of a 5-MW Refer-
ence Wind Turbine for Offshore System Development. Technical report, National
Renewable Energy Laboratory, Golden, Colorado, USA, 2009. NREL/TP-500-38060.
[37] S. Øye. FLEX4 – Simulation of Wind Turbine Dynamics. In 28th IEA Meeting
of Experts “State of the Art of Aeroelastic Codes for Wind Turbine Calculations”,
Lyngby, Denmark, 1996.
[38] E. A. Bossanyi. GH-Bladed User Manual. Technical report, Garrad Hassan and
Partners Limited, Bristol, UK, 2004. Issue 14.
[39] C. Lindenburg and J. G. Schepers. Phatas-IV Aero-elastic Modelling, Release. Tech-
nical report, ECN, Petten, Netherlands, 2000. ECN-CX–00-027.
[40] J. M. Jonkman and M. L. Buhl. FAST User’s Guide. Technical report, National
Renewable Energy Laboratory, Golden, Colorado, USA, 2005. NREL/EL-500-29798.
[41] T. J. Larsen, H. A. Madsen, A. M. Hansen, and K. Thomsen. Investigations of
stability effects of an offshore wind turbine using the new aeroelastic code HAWC2.
In Copenhagen Offshore Wind 2005, Copenhagen, Denmark, 2005.
130 References

[42] J. Manwell, J. McGowan, and A. Rogers. Wind Energy Explained. Theory, Design
and Application. John Wiley and Sons Ltd., 2002.
[43] J. D. Anderson. Computational Fluid Dynamics. McGraw-Hill Higher Education,
1995. Sixth Edition.
[44] D. J. Laino. AeroDyn: Aerodynamics analysis routines for horizontal-axis wind-
turbine dynamics analyses. See also URL https://nwtc.nrel.gov/AeroDyn.
[45] S. A. Ning. CCBlade. See also URL https://nwtc.nrel.gov/CCBlade.
[46] M. S. Campobasso and M. H. Baba-Ahmadi. Analysis of Unsteady Flows Past
Horizontal Axis Wind Turbine Airfoils Based on Harmonic Balance Compressible
Navier-Stokes Equations With Low-Speed Preconditioning. Journal of Turbomachin-
ery, 6(134), 2012.
[47] Numeca. Fine™/Turbo. See also URL
http://www.numeca.com/en/products/finetmturbo.
[48] A. Suzuki and A. Hansen. Generalized dynamic wake model for YawDyn. In AIAA-
99–0041, 37th Aerospace Sciences Meeting and Exhibit, Reno, Nevada, USA, 1999.
[49] H. Glauert. The analysis of experimental results in the windmill brake and vortex ring
states of an airscrew. H.M. Stationery Office, 1926.
[50] M. L. Buhl. A New Empirical Relationship between Thrust Coefficient and Induction
Factor for the Turbulent Windmill State. Technical report, National Renewable Energy
Laboratory, Golden, Colorado, USA, 2005. NREL/TP-500-36834.
[51] H. Glauert. A General Theory of the Autogyro. H.M. Stationery Office, 1928.
[52] C. Bak, A. Madsen, and J. Johansen. Influence from blade-tower interaction on fatigue
loads and dynamics (poster), pages 394–397. WIP Renewable Energies, 2001.
[53] B. D. Hibbs. HAWT Performance With Dynamic Stall. Technical report, Solar Energy
Research Institute, Golden, Colorado, USA, 1986. STR-2732.
[54] T. Burton, N. Jenkins, D. Sharpe, and E. Bossanyi. Wind Energy Handbook. John
Wiley and Sons Ltd., 2002.
[55] R. E. Gormont. A Mathematical Model of Unsteady Aerodynamics and Radial
Flow for Application to Helicopter Rotors. Technical report, US Army Air Mobility
Research and Development Laboratory, Philadelphia, Pennsylvania, USA, 1973. AD-
767240.
[56] A. Björck, M. Mert, and H. A. Madsen. Optimal parameters for the FFA-Beddoes
dynamic stall model. In 1999 European Wind Energy Conference, Nice, France, 1999.
[57] H. Snel and J. G. Schepers. Engineering models for dynamic inflow phenomena. In
1991 European Wind Energy Conference, Amsterdam, Netherlands, 1991.
[58] H. Snel and J. G. Schepers. Investigation and modelling of dynamic inflow effects. In
1993 European Wind Energy Conference, Lübeck, Germany, 1993.
References 131

[59] D. M. Pitt and D. A. Peters. Theoretical predictions of dynamic inflow derivatives.

Vertica, 1(5):21–34, 1981.
[60] ANSYS Inc. ANSYS. See also URL http://www.ansys.com/.
[61] Dassault Systèmes. Abaqus Unified FEA. http://www.3ds.com/products-
services/simulia/products/abaqus/.
[62] Dassault Systèmes. SolidWorks. http://www.solidworks.co.uk/.
[63] Sandia National Laboratories. NuMAD. http://energy.sandia.gov/energy/renewable-
energy/wind-power/rotor-innovation/numerical-manufacturing-and-design-tool-
numad/.
[64] R. C. Juvinall and K. M. Marshek. Fundamental of Machine Component Design. John
Wiley and Sons Ltd., 2012.
[65] G. Bir. Pre-Processor for Computing Composite Blade Properties. See also URL
https://nwtc.nrel.gov/PreComp.
[66] M. E. Tuttle. Structural Analysis of Polymeric Composite Materials. CRC Press, 2012.
Second Edition.
[67] R. M. Jones. Mechanics of Composite Materials. McGraw-Hill, 1975.
[68] O. A. Bauchau and J. I. Craig. Structural Analysis: With Applications to Aerospace
Structures. Springer, 2009.
[69] D. C. Sale. Co-Blade: Software for Analysis and Design of Composite Blades. See
also URL https://code.google.com/p/co-blade/.
[70] R. M. Rivello. Theory and Analysis of Flight Structures. McGraw-Hill, 1969.
[71] D. H. Allen and W. E. Haisler. Introduction to Aerospace Structural Analysis. John
Wiley and Sons Ltd., 1985.
[72] D. J. Peery and J. J. Azar. Aircraft Structures. McGraw-Hill, 1982. Second Edition.
[73] W. Young and R. Budynas. Roark’s Formulas for Stress and Strain. McGraw-Hill,
2001. Seventh Edition.
[74] T. Ashuri, M. Zaaijer, G. van Bussel, and G. van Kuik. An analytical model to extract
wind turbine blade structural properties for optimization and up-scaling studies. In
The Science of Making Torque from Wind, Crete, Greece, 2010.
[75] G. Bir. BModes: Software for Computing Rotating Blade Coupled Modes. See also
URL https://nwtc.nrel.gov/BModes.
[76] G. Hayman. MLife: A MATLAB-based Estimator of Fatigue Life. See also URL
https://nwtc.nrel.gov/MLife.
[77] Astm international. Standard Practices for Cycle Counting in Fatigue Analysis, 2011.
ASTM E1049 - 85(2011)e1.
132 References

[78] M. A. Miner. Cumulative Damage in Fatigue. Journal of Applied Mechanics,

12(1):A159–A164, 1945.
[79] M. Buhl. IECWind: A program to create IEC hub-height wind files for
InflowWind/AeroDyn-based simulators. See also URL https://nwtc.nrel.gov/IECWind.
[80] N. Kelley and B. Jonkman. TurbSim: A stochastic, full-field, turbulence simulator
primarily for use with InflowWind/AeroDyn-based simulation tools. See also URL
https://nwtc.nrel.gov/TurbSim.
[81] J. L. Tangler and D. M. Somers. NREL Airfoil Families for HAWTs. Technical report,
National Renewable Energy Laboratory, Golden, Colorado, USA, 1995. NREL/TP-
442-7109.
[82] W.A. Timmer and R.P.J.O.M. van Rooij. Summary of the Delft University Wind
Turbine Dedicated Airfoils. Journal of Solar Energy Engineering, Transactions of the
ASME, 125:488–496, 2003.
[83] N.N. Sørensen. CFD Modelling of Laminar-Turbulent Transition for Airfoils and
˜ θ Model. Wind Energy, 12:715–733, 2009.
Rotors Using the γ − Re
[84] C. Bak. Aerodynamic design of wind turbine rotors. In P. Brøndsted and R. P. L.
Nijssen, editors, Advances in wind turbine blade design and materials, volume 47 of
Energy. Woodhead Publishing Limited, 2013. ISBN 978-0-85709-426-1.
[85] A.C. Aranake, V.K. Lakshminarayan, and K. Duraysami. Computational analysis of
shrouded wind turbine configurations using a 3-dimensional RANS solver. Renewable
Energy, 75:818–832, 2015.
[86] J. Tangler and J. D. Kocurek. Wind Turbine Post-Stall Airfoil Performance Character-
istics Guidelines for Blade-Element Momentum Methods. Technical report, National
Renewable Energy Laboratory, Golden, Colorado, USA, 2004. NREL/CP-500-36900.
[87] S. J. Schreck, N. N. Sørensen, and M. C. Robinson. Aerodynamic Structures and
Processes in Rotationally Augmented Flow Fields. Wind Energy, 4(10):159–178,
2001.
[88] S. Schreck. Wind Turbine Aerodynamics Part B: Turbine Blade Flow Fields. In D. A.
Spera, editor, Wind Turbine Tecnology. ASME PRESS, 2009. Second Edition.
[89] H. Snel, R. Houwink, G. J. W. van Bussel, and A. Bruining. Sectional prediction of
3D effects for stalled flow on rotating blades and comparison with measurements. In
European Community Wind Energy Conference, Lübeck-Travemünde,Germany, 1993.
[90] Z. Du and M. S. Selig. A 3-D stall-delay model for horizontal axis wind turbine
performance prediction. In AIAA-98-0021, 36th AIAA Aerospace Sciences. Meeting
and Exhibit, 1998 ASME Wind Energy Symposium, Reno, Nevada, USA, 1998.
[91] P. K. Chaviaropoulos and M. O. L. Hansen. Investigating three-dimensional and
rotational effects on wind turbine blades by means of a quasi-3D Navier Stokes solver.
Journal of Fluids Engineering, (122):330–336, 2000.
References 133

[92] C. Lindenburg. Modelling of rotational augmentation based on engineering consid-

erations and measurements. In European Wind Energy Conference, London, UK,
2004.

[93] C. Bak, J. Johansen, and P. B. Andersen. Three-dimensional corrections of airfoil

characteristics based on pressure distributions. In European Wind Energy Conference
Exhibition (EWEC), Athens, Greece, 2006.

[94] A. J. Eggers, K. Chaney, and R. Digumarthi. An Assessment of Approximate Modeling

of Aerodynamic Loads on the UAE Rotor. In AIAA-2003-0868, 41st Aerospace
Sciences Meeting and Exhibit, Reno, Nevada, USA, 2003.

[95] W. A. Timmer and C. Bak. Aerodynamic characteristics of wind turbine blade airfoils.
In P. Brøndsted and R. P. L. Nijssen, editors, Advances in wind turbine blade design
and materials, volume 47 of Energy. Woodhead Publishing Limited, 2013. ISBN
978-0-85709-426-1.

[96] L. A. Viterna and R. D. Corrigan. Fixed Pitch Rotor Performance of Large Horizontal
Axis Wind Turbines. Technical report, NASA Lewis Research Center, Cleveland,
Ohio, USA, 1981. N83 19233.

[97] E. M. Jacobs and I. H. Abbot. The NACA Variable-Density Wind Tunnel. Technical
report, NASA Langley Research Center, Hampton, Virginia, USA, 1932. NACA TR
416.

[98] D. Spera. Models of Lift and Drag Coefficients of Stalled and Unstalled Airfoils in
Wind Turbines and Wind Tunnels. Technical report, National Aeronautics and Space
Administration, Cleveland, Ohio, USA, 2008. NASA/CR—2008-215434.

[99] M. M. Hand, D. A. Simms, L.J. Fingersh, D. W. Jager, J. R. Cotrell, S. Schreck, and

S. M. Larwood. Unsteady Aerodynamics Experiment Phase VI: Wind Tunnel Test
Configurations and Available Data Campaigns. Technical report, National Renewable
Energy Laboratory, Golden, Colorado, USA, 2001. NREL/TP-500-29955.

[100] C. Lindenburg. Investigation into Rotor Blade Aerodynamics. Technical report, ECN,
Petten, Netherlands, 2003. ECN-C–03-025.

[101] C. Hansen. AirfoilPrep: An Excel workbook for generating airfoil tables for Aerodyn
and WT_Perf. See also URL https://nwtc.nrel.gov/AirFoilPrep.

[102] L. Fingersh, M. Hand, and A. Laxson. Wind Turbine Design Cost and Scaling Model.
Technical report, National Renewable Energy Laboratory, Golden, Colorado, USA,
2006. NREL/TP-500-40566.

[103] R. Smith. Uncertainty Quantification: Theory, Implementation, and Applications.

Society for Industrial and Applied Mathematics, 2014.

[104] J. C. Helton and F. J. Davis. Latin Hypercube Sampling and the Propagation of
Uncertainty in Analyses of Complex Systems. Reliability Engineering and System
Safety, 81(1):23–69, 2003.
134 References

[105] J. M. Hammersley and D. C. Handscomb. Monte Carlo Methods. Chapman and Hall,
New York, 1964.
[106] M. Tari and A. Dahmani. Refined Descriptive Sampling: A better Approach to Monte
Carlo Simulation. Simulation Modelling Practice and Theory, 14(2):143–160, 2006.
[107] Lumina Blog. Latin Hypercube vs. Monte Carlo Sampling. See also URL
http://blog.lumina.com/2014/latin-hypercube-vs-monte-carlo-sampling/.
[108] R. G. Ghanema and P. D. Spanos. Stochastic Finite Elements: A Spectral Approach.
Dover, London, 1991.
[109] L. Mathelin, M. Y. Hussaini, and T. A. Zang. Stochastic Approaches to Uncertainty
Quantification in CFD Simulations. Numerical Algorithms, 38(1–3), 2005.
[110] P.E. Gill, W. Murray, and M.H. Wright. Practical Optimization. Academic Press,
1981.
[111] Z. Ugray, L. Lasdon, J. Plummer, F. Glover, J. Kelly, and R. Martí. Scatter Search and
Local NLP Solvers: A Multistart Framework for Global Optimization. INFORMS
Journal on Computing, 19(3):328–340, 2007.
[112] D. E. Goldberg. Genetic Algorithms in Search, Optimization & Machine Learning.
Addison-Wesley, 1989.
[113] C. Audet and J. E. Dennis. Analysis of Generalized Pattern Searches. SIAM Journal
on Optimization, 13(3):889–903, 2003.
[114] Hugo. Gradient descent. See also URL
http://www.onmyphd.com/?p=gradient.descent.
[115] Mathworks. MATLAB documentation. See also URL
http://www.mathworks.co.uk/help/matlab/.
[116] A. R. Conn, N. I. M. Gould, and Ph. L. Toint. A Globally Convergent Augmented
Lagrangian Algorithm for Optimization with General Constraints and Simple Bounds.
SIAM Journal on Numerical Analysis, 28(2):545–572, 1991.
[117] K. Deb and S. Srivastava. A Genetic Algorithm Based Augmented Lagrangian Method
for Accurate, Fast and Reliable Constrained Optimization. Technical report, Kanpur
Genetic Algorithms Laborator, Kanpur, UP, India, 2011. KanGAL Report Number
2011012.
[118] T. G. Kolda, R. M. Lewis, and V. Torczon. A generating set direct search augmented
lagrangian algorithm for optimization with a combination of general and linear con-
straints. Technical report, Sandia National Laboratories, Albuquerque, New Mexico,
USA, 2006. SAND2006-5315.
[119] D. M. Somers. Design and Experimental Results for the S809 Airfoil. Technical report,
National Renewable Energy Laboratory, Golden, Colorado, USA, 1997. NREL/SR-
440-6918.
References 135

[120] M. Costa and E. Minisci. MOPED: a multi-objective parzen-based estimation of

distribution algorithm. In EMO 2003, pages 282–294, Faro, Portugal, 2003. Springer.
[121] M. Vasile, E. Minisci, and M. Locatelli. An inflationary differential evolution algorithm
for space trajectory optimization. Evolutionary Computation, IEEE Transactions on,
15(2):267–281, 2011.
[122] J.A. Lozano, P. Larranaga, I. Inza, and E. Bengoetxea. Towards a New Evolution-
ary Computation: Advances on Estimation of Distribution Algorithms (Studies in
Fuzziness and Soft Computing). Springer, February 2006.
[123] K. Fukunaga. Introduction to statistical pattern recognition. Academic Press, 1972.
[124] G. Avanzini, D. Biamonti, and E.A. Minisci. Minimum-fuel/minimum-time maneuvers
of formation flying satellites. In Adv. Astronaut. Sci., pages 2403–2422, 2003.
[125] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A Fast and Elitist Multiobjective
Genetic Algorithm: NSGA-II. Evolutionary Computation, IEEE Transactions on,
6(2):182–197, 2002.
[126] D. Datta, K. Deb, C.M. Fonseca, F.G. Lobo, P.A. Condado, and J. Seixas. Multi-
objective evolutionary algorithm for land-use management problem. International
Journal of Computational Intelligence Research, 3(4):371–384, 2007.
[127] K. Deb. Scope of stationary multi-objective evolutionary optimization: A case study on
a hydro-thermal power dispatch problem. Journal of Global Optimization, 41(4):479–
515, 2008.
[128] R.H. Leary. Global optimization on funneling landscapes. Journal of Global Opti-
mization, 18(4):367–383, 2000.
[129] N. Dodgson. Bezier curves. See also URL
http://www.cl.cam.ac.uk/teaching/2000/AGraphHCI/SMEG/node3.html.
[130] MalinC. Bézier curves and continuity explained. See also URL
http://scratch.mit.edu/projects/11824563/.
[131] J. F. Mandell, D. D. Samborsky, P. Agastra, A. T. Sears, and T. J. Wilson. Analysis
of SNL/MSU/DOE Fatigue Database Trends for Wind Turbine Blade Materials.
Technical report, Sandia National Laboratories, Albuquerque, New Mexico, USA,
2010. SAND2010-7052.
[132] National Renewable Energy Laboratory. Classes of wind power density at 10 m and
50 m. See also URL http://rredc.nrel.gov/wind/pubs/atlas/tables/1-1T.html.
[133] C. Zang, M.I. Friswell, and J.E. Mottershead. A review of robust optimal design and
its application in dynamics. Computers and Structures, 83:315–326, 2005.
[134] I. Das and J.E. Dennis. A closer look at drawbacks of minimizing weighted sums of
objectives for pareto set generation in multicriteria optimization problems. Structural
Optimization, 14(1):63–69, 1997.
136 References

[135] A. Haahr, Z. Chen, and M. Rajagopal. A wind turbine rotor blade with a cone angle
and a method of manufacturing a wind turbine rotor blade with a cone angle. See also
URL http://www.google.com/patents/WO2013091635A1?cl=en.
[136] J. Jonkman. FAST: An aeroelastic computer-aided engineering (CAE) tool for hori-
zontal axis wind turbines. See also URL https://nwtc.nrel.gov/FAST.
[137] G. Leloudas, W. J. Zhu, J. N. Sørensen, W. Z. Shen, and S. Hjort. Prediction and
Reduction of Noise from a 2.3 MW Wind Turbine. Journal of Physics: Conference
Series, 75(1):1–9, 2007.