WIND ENERGY

Wind Energ. 2015; 18:14331450

Published online 30 May 2014 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/we.1768

RESEARCH ARTICLE

Multi-hazard reliability assessment of offshore wind

turbines
Maryam Mardfekri1 and Paolo Gardoni2
1

Zachry Department of Civil Engineering, Texas A&M University, College Station, Texas 77843-3136, USA
Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, 3118 Newmark Civil Engineering
Laboratory, Urbana, Illinois 61801, USA
2

ABSTRACT
A probabilistic framework is developed to assess the structural performance of offshore wind turbines under multiple hazards. A
multi-hazard fragility surface of a given wind turbine support structure and the seismic and wind hazards at a specic site
location are incorporated into the probabilistic framework to assess the structural damage due to multiple hazards. A database
of virtual experiments is generated using detailed three-dimensional nite element analyses of a set of typical wind turbine
systems subject to extreme wind speeds and earthquake ground motions. The generated data are used to develop probabilistic
models to predict the shear and moment demands on support structures. A Bayesian approach is used to assess the model
parameters incorporating the information from virtual experiment data. The developed demand models are then used to estimate
the fragility of the support structure of a given wind turbine. As an example of the proposed framework, the annual probabilities
of the occurrence of different structural damage levels are calculated for two identical wind turbines, one located in the Gulf of
Mexico of the Texas Coast (prone to hurricanes) and one off the California Coast (a high seismic region). Copyright 2014
John Wiley & Sons, Ltd.
KEYWORDS
multi-hazard; reliability; probabilistic models; fragility; offshore wind turbines
Correspondence
P. Gardoni, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, 3118 Newmark Civil
Engineering Laboratory, 205 N. Mathews Ave. Urbana, Illinois 61801, USA.
E-mail: gardoni@illinois.edu
Received 16 November 2012; Revised 1 May 2014; Accepted 6 May 2014

1. INTRODUCTION
According to the annual reports by the Global Wind Energy Council,1 the global cumulative-installed wind capacity has
been doubling every 3 years, and it is projected to continue to grow at a similar rate. Offshore wind turbines installed
extensively around the world are subject to different hazards (e.g. earthquake, hurricane and typhoon) raising concerns
about the reliability of the wind turbine support structure. For instance, Japan is the worlds 13th largest producer of wind
power according to the World Wind Energy Association,2 despite having a considerably high occurrence rate of earthquakes and typhoons. Likewise, according to the National Renewable Energy Laboratory (NREL3), California, a highly
seismic region, is the third largest wind power producer in the nation. Furthermore, the wind industry is recently considering installing offshore wind farms in the south coast of the USA, and in particular in the Gulf of Mexico, because of the
superior wind resources available in this region.4 However, a considerably high hurricane occurrence rate in the Gulf of
Mexico raises a new concern about the safety of wind turbine support structures subject to hurricane. To investigate the
reliability of a wind turbine support structure, all possible hazards that can occur during the wind turbines life have to
be considered. To this end, a probabilistic framework is needed to evaluate the safety of the support structure under multiple
hazards and predict its expected structural damage. The results can assist the wind industry decision-makers choosing
optimum design and location for future wind energy projects. In addition, the assessment of the expected structural damage
can be used for an optimal design of wind turbines to maximize the power production and minimize manufacturing, operation and maintenance cost. Current standards for the design of wind turbines structural components (IEC 61400-15 and

Copyright 2014 John Wiley & Sons, Ltd.

1433

Multi-hazard reliability assessment of offshore wind turbines

M. Mardfekri and P. Gardoni

DNV-OS-J1016) are based on the partial safety factor methods, in which the target safety level is obtained by applying
safety factors to load and resistance terms in the design equation. The structural reliability analysis presented in this paper
can be valuable to calibrate the load and material factors to be used in the partial safety factor method and in particular to
update current standards such as IEC 61400-1 and IEC 61400-3. This paper develops estimates of the annual probabilities
of being in different structural damage states for offshore wind turbines installed in water depths less than 30 m considering
multiple hazards including seismic and hurricane.
Several probabilistic studies have been conducted on wind turbines. Walford7 and Tavner et al.8 investigated the reliability of operation and power production systems on the basis of historical data of failures and their associated costs. Walford7
also discussed the means for reducing operation and maintenance costs. Madsen et al.,9 Agarwal and Manuel,10 and
Manuel et al.11 employed probabilistic frameworks to predict the extreme and fatigue loads for the design of onshore
and offshore wind turbines on the basis of the dynamic response of the support structures. However, while the aeroelastic
interaction is successfully considered in the analyses, only limited work is available that incorporates the foundation
stiffness in the dynamic response of offshore wind turbines, including Andersen et al.12
Offshore wind turbine support structures installed in water depths less than 30 m are typically supported by monopile
foundations. Bush and Manuel13 investigated the effect that the use of alternative models for monopile foundation of shallow-water offshore wind turbines has on the design extreme loads. Their results showed the importance of incorporating
foundation stiffness in the simulations. The stiffness of monopile foundation in their study was obtained using py curves.
However, using a three-dimensional (3D) non-linear nite element (FE) model, Mardfekri et al.14 showed that, depending
on the pile diameter and soil type, using common simple models such as py curves and particularly modeling the pile
using 1D beamcolumn elements may result in inaccurate responses. This is true in particular for the pile sizes typical
of foundations of offshore wind turbines.
Aeroelastic simulators such as FAST,15 ADAMS16 and GH Bladed17 successfully include the aeroelastic interactions in
the analysis of dynamic response of the support structure. However, an important limitation of these simulators is that they
are not capable of continuous modeling of the non-linear foundation behavior and the dynamic soil-structure interaction. A
detailed non-linear FE analysis of the support structure and the foundation can be carried out to account for the non-linear
foundation behavior and the dynamic soil-structure interaction. However, a detailed non-linear FE analysis can be quite
expensive and time consuming both in developing and running it. In addition, assessing the reliability of a wind turbine requires
accounting for the uncertainties inherent in the structural material, soil and geometrical properties. To account for such uncertainties, a high number of FE analyses would need to be carried out making this approach too time consuming.
Mardfekri and Gardoni18 conducted detailed 3D non-linear FE analyses of the support structures of a suite of typical
offshore wind turbines supported by monopile foundations. The FE models included a continuous modeling of the pile
and the surrounding soil. As a result, the FE models successfully incorporated the dynamic soil-structure interaction into
the response of the support structure. In addition, Mardfekri and Gardoni18 used the data generated from the FE analysis
to calibrate simplied probabilistic models for the deformation, shear force and bending moment demands on the support
structure under day-to-day loading in operating conditions (i.e. day-to-day wind, wave and current loads.) The developed
probabilistic models were then used to assess the reliability of support structures in a more efcient way than by simulating
the structural responses using the FE models. However, Mardfekri and Gardoni18 only considered operating wind turbines
under day-to-day environmental loads and not extreme loading from earthquakes and hurricanes.
To address the concern related to the installation of wind farms in moderate and high seismic regions, a number of
researchers conducted studies on the seismic response of wind turbines. Early publications on the analysis of dynamic response
of wind turbines during earthquake19,20 were based on the simplied models that lumped the nacelle and rotor as a point mass at
the top of the tower. As a result, the aeroelastic interaction was not accounted for. More recently, Witcher21 and Prowell et al.22
developed more rened models that considered the aeroelastic interaction. Specically, Witcher21 studied the seismic response
of support structures for both operating and parked wind turbines. The results showed the importance of accounting for
aeroelastic interaction for operating wind turbines. Prowell et al.22 calibrated the aeroelastic interaction modeled in FAST using
experimental data from a shake-table test of a small onshore 65 kW wind turbine.23 Yet both studies fail to incorporate the
dynamic soil-structure interaction. To address this limitation, Mardfekri and Gardoni24 used the detailed 3D non-linear FE
models developed in Mardfekri and Gardoni18 to conduct time history analyses of offshore wind turbines subject to seismic
loading in addition to day-to-day operational loading accounting for the dynamic soil-structure interaction. Using the generated
data, Mardfekri and Gardoni24 also developed probabilistic models for the seismic demands on the support structure. The
developed probabilistic models were then used to assess the reliability of support structures conditioning on spectral acceleration and the mean wind speed acting on the structure. However, there is still a need for a probabilistic framework to assess the
multi-hazard reliability of wind turbine support structures and predict their annual failure probabilities.
This paper addresses this need by proposing a probabilistic framework to assess the multi-hazard reliability of offshore
wind turbines. As a rst step, we develop novel probabilistic models for the shear and moment demands on the support
structures subject to extreme wind loads in addition to seismic loads. The models are developed by updating available
probabilistic seismic demand models24 using additional virtual experiment data generated for support structures subject
to extreme wind loads like those experienced during hurricanes. The virtual experimental data are obtained by developing
1434

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

detailed 3D non-linear FE models of wind turbines accounting for the dynamic soil-structure interaction. The probabilistic
demand models are calibrated using a Bayesian approach. The proposed probabilistic models provide unbiased predictions
for the shear and moment demands on the support structures, accounting for the inherent uncertainties, including the
statistical uncertainty (associated to the nite sample size) and the modeling errors (associated with the selection of the
variables in the models and the model forms.) The probabilistic models are then used to develop the fragility curves of wind
turbines for given intensity measures of the seismic and wind loading, namely, the spectral acceleration Sa and the mean
wind speed Ws. The fragility curves and site-specic hazard functions are then used to assess the annual probabilities of
structural damage states. The developed methodology proposes three structural damage state classications for wind
turbines. As an illustration, fragility curves and the annual probabilities of being in specied damage states are estimated
for two identical 5 MW offshore wind turbines one located in the Gulf of Mexico of the Texas Coast (prone to hurricanes)
and one off the California Coast (a high seismic region).
The next section introduces the probabilistic framework to assess the multi-hazard reliability of wind turbine support
structures. The third section discusses the proposed probabilistic demand models. In the fourth section, we present the formulation of fragility estimates for offshore wind turbine support structures along with the assessment of importance and
sensitivity measures. Finally, fragility estimates, and importance and sensitivity measures are developed for the examples
wind turbines along with their annual probabilities of three structural damage states.

2. MULTI-HAZARD ASSESSMENT
According to the total probability rule,25 the probability of failure to meet a specied performance level for a component or
system, Pf, can be written as
Pf

Fwf wdw

(1)

where w = vector of loading variables, f(w) = the joint probability density function (PDF) of occurrence of w and
F(w) = conditional probability of failure (or fragility) to meet a specied performance level given the occurrence of w.
The focus of this paper is on the two most signicant hazards for offshore wind turbines support structures: seismic and
wind. Given the intensity measures Sa = spectral acceleration at the natural period of the wind turbine, and Ws = mean wind
speed, equation (1) can be written as
Pf

FS ; W f S ; W dS dW
a

(2)

where F(Sa, Ws) = probability of failure conditioned on Sa and Ws, and f(Sa, Ws) = joint PDF of Sa and Ws. Given that the
occurrence, or non-occurrence, of earthquake does not affect the probability of occurrence of any particular level of wind
speed and vice versa, Sa and Ws, can be assumed to be statistically independent. Therefore, equation (2) can be written as
Pf

FS ; W f S f W dS dW
w

(3)

where f(Sa) and f(Ws) = marginal PDF of Sa and Ws, respectively.

2.1. Seismic contribution to probability of failure
To quantify the probability of future seismic activity at a particular location, we use the seismic hazard function, Q(Sa),
dened as the expected annual frequency of experiencing a spectral acceleration equal to Sa or greater. Assuming the arrival
of earthquakes at a site is a Poisson process,26 f(Sa) can be expressed in terms of Q(Sa) as



dQSa
f Sa expQSa  
dSa

(4)

The United States Geological Survey (USGS) provides annualized seismic hazard exceedance curves, containing
discrete values of Q(Sa) for locations throughout the USA, on the basis of the available information about past earthquakes,
deformation of the earth crust, geologic site conditions and seismic attenuation relationships.26
2.2. Wind contribution to probability of failure
To develop the annual PDF for wind hazard, we combine the PDF for day-to-day wind speed with the one for extreme wind
speed during hurricanes. Morgan et al.27 investigated annual probability distributions for offshore wind speeds on the basis
Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1435

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

statistical analysis of day-to-day 10 min average wind speed data. Wind speed data were recorded at 178 ocean buoy stations
around North America, by the National Data Buoy Center.28 On the basis of Morgan et al.,27 we use the bimodal Weibull
mixture distribution (BIW) to model the day-to-day wind speed. The BIW is a combination of two Weibull distributions and


has two different modes. Using the BIW, the conditional PDF of Ws given that there is no hurricane f W s jH is expressed as
"   #
"   #
  
b1 W s b1 1
W s b1
b2 W s b2 1
W s b2

exp

exp


1


f W s H
b
b
a1 1
a1
a2 2
a2

(5)

where shape b and scale a parameters have subscripts corresponding to the two different modes, = mixing parameter and H
indicates the event of non-occurrence of a hurricane.
Wang29 characterized the hurricane event on the basis of the statistical analysis of 4776 hurricanes simulated to occur in
10,000 years with landfall position assumed to occur with equal probability along the length of the Texas coastline.
Specically, using the existing hurricane tracking models, event-based simulation techniques and information extracted from
historical hurricane data, Wang29 developed a database of synthetic hurricanes simulated to occur during a 10,000 year period.
He then found the lognormal distributions that t the marginal cumulative distribution functions of the maximum wind speed
and the radius of maximum wind speed. On the basis of Wang,29 we use the lognormal distribution to model the extreme wind
speed associated to hurricanes. The PDF of Ws given the occurrence of a hurricane f(Ws|H) is written as
"

 #
1
1 lnW s  2
p exp 
f W s jH
2

W s 2

(6)

where the location parameter and scale parameter are the mean and standard deviation of the natural logarithm of Ws,
respectively. The PDF of Ws can now be written as using the total probability rule25 as
  
f W s f W s jH PH f W s H 1  PH 

(7)

where P(H) = annual probability of occurrence of a hurricane. With the assumption of arrival of hurricane being a Poisson
process, P(H) = 1  exp[T], in which = annual occurrence rate of hurricane and T = 1 year.

3. PROBABILISTIC DEMAND MODELS

In this paper, a model is an analytical expression or procedure that relates the demand on a structural component to the
properties x of the considered system (such as material properties, structural dimensions and boundary conditions) listed
in Table I and the loading variables w listed in Table II. To facilitate the use of the proposed probabilistic models in practice, Gardoni et al.30,31 suggested developing a demand model starting from a commonly accepted deterministic model or
procedure and adding correction terms and a model error to, respectively, correct for the inherent bias and capture the
uncertainty in the developed model. Accordingly, we write the general form of the proposed demand models as
Dk x; w; k d^ k x; w k x; w; k k k

k v; m

(8)

where Dk(x, w, k) = kth probabilistic demand model, in which k = v or m to indicate the shear and moment demand models,
respectively, d^ k x; w selected deterministic demand model, k(x, w, k) = correction term for the bias inherent in the
deterministic model, in which k = vector of unknown model parameters, kk = model error, in which k = random variable
with zero mean and unit variance, and k = unknown standard deviation of the model error, nally k = (k, k) = vector of
all the unknown parameters. We also dene the vector of all the unknown parameters = (v, m, ), where = the correlation between v and m. On the basis of a preliminary study of the data, we employ a variance stabilizing logarithmic transformation of the quantity of interest to satisfy the homoskedasticity assumption (k is constant), the normality assumption
(k follows the normal distribution) and the additive form used in equation (8).
3.1. Deterministic model
Being a popular wind turbine simulator in the wind industry around the world makes FAST a proper candidate for
deterministic predictions of the demands on the wind turbine support structures. FAST employs a combined modal and
multi-body dynamics formulation to simulate the aerodynamics and structural response of wind turbines.15 For given
values of the mean wind speed and turbulence intensity, a time history of wind speed is generated by TurbSim and used
1436

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Table I. Geometrical and mechanical parameters x used in experimental design.

Property

Symbol

Ranges

Unit

Pile diameter
Pile penetration
Pile diameter to wall thickness ratio
Soil modulus of elasticity
Friction between pile and soil

RD
HH
dt
t
HWr
ST

Tn
Ts
dp
Hp
p
Esoil
frp  s

40126
4090
1.94.0
100200
2030
S235, S275, S355
0.05
0.911.9
0.53.6
3.06.0
1050
50100
13200
0.20.3

m
m
m

s
s
m
m

MPa

Soil type
Soil cohesion
Soil friction angle

Csoil
soil

Rotor diameter
Tower height
Tower top diameter
Tower diameter to wall thickness ratio
Water depth
Steel type
Material damping ratio
Support structure vibration period

First mode
Second mode

Clay
10200
1025

Sand
080
3545

kPa

Table II. Loading parameters w used in experimental design.

Property

Symbol

Ranges

Unit

Mean wind speed

Turbulence intensity
Signicant wave height
Wave peak period
Rated wind speed
Earthquake moment magnitude
Distance between earthquake record and rupture zone

Ws
ITw
Hs
Tp
Ws  rated
Meq
Req

3.075
00.16
1.010
p
p
3:6 H s  5:0 H s
10.311.7
5.87.0a
1.060.0a

ms

m
s
1
ms

km

1

These ranges are not used in the experimental design. Earthquake records are selected using the bin approach within these ranges.

as an input for the dynamic analysis in FAST. TurbSim uses a statistical model to numerically simulate time series of three-component wind speed vectors.32 TurbSim supports the Kaimal spectrum33 to simulate the turbulence, while using the standard IEC
categories of turbulence characteristics to estimate turbulence. More details on the turbulence modeling can be found in
Mardfekri and Gardoni.18 In addition, FAST supports the JONSWAP/PiersonMoskowitz spectrum34 to model linear irregular
sea states (stochastic waves), which represents the superposition of a large number of periodic and parallel wave components. A
sea state is specied by a wave frequency spectrum with a given signicant wave height and wave peak period, which are
considered stationary within periods of typically 3 h duration, according to DNV-OS-J101.6 It then uses the Morisons equation
to determine the hydrodynamic forces on the tower. Current loading is also incorporated in the Morisons equation. Additional
details on the modeling of wave and current loading can be found in Mardfekri and Gardoni.18
In this paper, we use FAST to predict deterministic seismic demands on the support structure of wind turbines. The
currently in-practice version of FAST does not include the seismic module. However, FAST allows forces and moments
to be applied at the tower base platform. The platform has six degrees of freedom, and it provides the possibility to model
the earthquake ground motions in a time marching simulation.
In a practice consistent with those conducted for calculation of wind and wave loads, for given intensity and duration
parameters and frequency content of the ground motion, Mardfekri and Gardoni24 generated synthetic ground motions
using a stochastic model proposed by Rezaeian and Der Kiureghian.35 Generated synthetic ground motion time histories
are then applied as a time history of force F(t) to the platform. Using an articially large mass for the support platform,
the force F t Mx t produces the desirable acceleration x t at the base of the turbine support structure, where M is
the total mass of the support platform and the wind turbine.
3.2. Model correction
The correction term k(x, w, k) is added to incorporate the missing terms in the deterministic model into the developed
demand models and to correct for the potential bias in d^ k x; w. It is written as
Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1437

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

k x; w; k

N
X

kp hkp x; w

(9)

p1

where k = [kp], hk1(x, w), , hkN(x, w) normalized explanatory functions that might be signicant in correcting d^ k x; w
and N = the number of unknown model parameters. In order to facilitate the use in practice of the probabilistic model,
we identify the most parsimonious model for k(x, w, k) by removing the unimportant explanatory function, which is
identied by examining the posterior statistics of k.
The same candidate explanatory functions as those of Mardfekri and Gardoni24 considered in their study are selected in
this paper. Table III presents the candidate explanatory functions. Specically, hk1(x, w) = 1 captures the potential bias in
the model that is independent of x and w, hk2 x; w d^ k x; w captures any possible under-estimation or over-estimation
of the deterministic models, functions hk3  hk6 characterize the possible inuence of wind and wave parameters, hk7  hk13
incorporates the possible inuence of the characteristic of the ground motion, hk14 brings in the size-related characteristics
of the wind turbine and nally functions hk15  hk18 incorporate the effects of the foundation stiffness in the model.
The proposed probabilistic models are calibrated using a Bayesian updating rule36 on the basis of experimental observations. However in this study, because of the lack of data from eld observations or laboratory experiments, virtual data
are generated by conducting dynamic analyses of detailed 3D non-linear FE models of the support structures.
3.3. Virtual data
A set of representative congurations are selected to generate virtual data. We select the representative congurations using
a space lling experimental design technique to ensure that the congurations have a good coverage of the design space.
Tables I and II present, respectively, the ranges of the variables considered to characterize the wind turbine congurations
and loading parameters. The proposed ranges for the rst and second modes of vibration are estimated on the basis of modal
analyses on the support structures of wind turbines with rated powers of 500 kW to 5 MW. The upper limit of the range for
the mean wind speed Ws reaches 75 m s1 to incorporate the extreme wind velocities during hurricanes.
The FE models are developed in ABAQUS37 to simulate the dynamic response of the support structure of a broad range
of offshore wind turbines, subject to different load cases including seismic excitations in addition to day-to-day environmental loads on operating and parked wind turbines, and extreme wind velocities due to hurricanes on parked wind
turbines. It is noteworthy that no specic design load cases are considered in the analyses. However, as mentioned earlier
and presented in Table II, ranges for loading parameters are considered wide enough to develop a comprehensive demand
model for the structural behavior of wind turbines.
Witcher21 conducted time domain simulations of a 2 MW wind turbine with 80 m diameter rotor mounted on a 60 m
high tubular steel tower in different load cases including continuous operation throughout earthquakes, emergency
shutdown initiated during an earthquake and parked throughout earthquakes. The results showed a signicant difference

Table III. Explanatory functions for demand models.

Explanatory function

1438

Formula

hk1
hk2
hk3

1
d^ k
ln(Ws  Tn/HH)

hk4
hk5
hk6
hk7
hk8
hk9
hk10
hk11
hk12
hk13
hk14
hk15
hk16
hk17
hk18

ln(ITw)
ln(Hs/HH)
ln(Tp/Tn)
ln(Sa/g)
ln(Sd/HH)
ln(PGA/g)
ln(PGV  Tn/HH)
ln(PGD/HH)
ln[2PGV/(PGA  Tn)]
ln[2PGD/(PGV  Tn)]
ln(RD/HH)
ln(Cs/Cs max)
ln(Csoil/Esoil)
ln[tan( soil)]
ln(kt/kf)

Parameter
k = v or m
d^ k = deterministic shear or moment demand
Ws = mean wind speed; HH = Hub height
Tn = natural period of the support structure
ITw = wind turbulence intensity
Hs = signicant wave height
Tp = wave peak period
Sa = spectral acceleration; g = ground acceleration
Sd = spectral displacement
PGA = peak ground acceleration
PGV = peak ground velocity
PGD = peak ground displacement

RD = rotor diameter
1
Cs = soil shear wave velocity; Cs max = 194.594 m s
Csoil = soil cohesion; Esoil = soil modulus of elasticity
soil = soil friction angle
kt = tower stiffness; kf = foundation stiffness

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

in the responses of operating and parked wind turbines. He concluded that this difference is due to the absence of
aerodynamic damping in the parked condition. Comparing the peak loads resulted from time domain analyses with
those obtained using frequency domain procedure, Witcher21 showed that the aerodynamic damping experienced by
an operational wind turbine is close to the typical 5% value used for the design spectra in building codes. Prowell
et al.23 estimated the structural damping of a 65 kW wind turbine in idling (parked) condition through a full-scale
test on a wind turbine mounted on the Network for Earthquake Engineering Simulation (NEES) shake table at the
University of California, San Diego. They suggested a value of 0.6% for the structural damping of a parked wind
turbine. In another study on the seismic response of wind turbines, Ishihara and Sarwar38 suggested a structural
damping of 0.5% for parked wind turbines with rated powers of 400 kW and 2 MW. In this paper, we account
for the aerodynamic damping of an operating wind turbine by considering a 5% structural damping for the steel
tower. The structural damping for parked wind turbines is considered to be 0.5%.
Foundation non-linearities are considered explicitly in dening non-linear behavior of the soil and soil-structure
interaction. We use the MohrCoulomb plasticity model to dene the non-linear behavior of the soil. The Rayleigh
damping formulation is used to include frequency-dependent soil damping in the structural response. Soil-pile
interaction is modeled using contact pair, a formulation used in ABAQUS to dene the non-linear contact properties
of two bodies. Forces at the top of the tower due to the wind and the rotation of the rotor in case of operating wind
turbine are obtained using FAST. The resulted time histories are then used in the FE model of the support structure
as an external loading in addition to wave, current and earthquake. Ground motion records are selected from the Pacic
Earthquake Engineering Research Center Next Generation Attenuation (NGA) database.39 We select the ground motion
records on the basis of the bin approach, proposed by Shome and Cornell.40 We use ve bins on the basis of the moment
magnitude (Meq) and the closest distance between the record location and the rupture zone (Req) to capture all possible
characteristics of the earthquake. More details on the analytical modeling to generate virtual experiment data are
presented in Mardfekri and Gardoni.24
3.3.1. Equality and lower bound data
The accuracy of the results of FE analyses is sensitive to how the solution method handles large displacements and
second-order effects. To include the data from the analyses that lead to large deformations without letting inaccurate values
wrongfully inuence the model parameters, the data from virtual experiments are divided into equality and lower bound
data.30,41 A threshold of 5% is considered for the drift ratio, such that if the maximum drift ratio during one time history
analysis is less than 5%, then the shear and moment data are considered to be accurate and taken as equality data. If an
analysis produces a drift ratio that exceeds 5%, then we consider the maximum shear and moment that occurred prior to
reaching the 5% drift ratio as lower bound data for the shear and moment, respectively.
3.4. Bayesian model updating
A Bayesian formulation is used to calibrate the proposed probabilistic models with the virtual data. In a Bayesian approach,
the unknown parameters are estimated using the following updating rule:36
f Lp

(10)

where p() = the prior distribution of that reects the state of knowledge about available before generating the virtual
data, L() = the likelihood function that represents the objective information on contained in the virtual experiment data,
= a normalizing factor and f() = the posterior distribution of that represents the updated state of knowledge about .
The posterior distribution f() incorporates both the information about included in p() and L(). Application of the
updating rule in equation (10) can be repeated to update our present state of knowledge as new information on becomes
available. Because of lack of prior information on the unknown parameters, we use a non-informative prior that reects that
no knowledge about is available a priori. Gardoni et al.30 showed that a non-informative prior for the parameters can
be written as



2
 1 3=2 Y
1

p

1 
i1 i

(11)

3.4.1. Likelihood function

The likelihood is a function that is proportional to the conditional probability of observing the results from the virtual
experiments for given values of the model parameters. Under the assumption of statistically independent observations,
following Gardoni et al.,30 L() is written as
Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1439

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

8
9
<
=
k ki r ki k 
k ki > rki k 
L
P
Lower bound
:Equality
;
Observation i

(12)

data k

data k

where r ki k Dki  d^ k xi ; wi  k xi ; wi ; k and Dki = ith observed value for the kth demand for given xi and wi.
However, the events in the aforementioned expression are dependent because of the correlation between ki for different
demands k. To consider this dependence, the probability term for each observation in equation (12) can be computed using
the multi-normal probability density and cumulative distribution functions for both shear and moment demand observations. Table IV lists the formulation for probability terms proposed by Gardoni et al.30 for bivariate probabilistic models.
The likelihood function is formulated on the basis of the type and form of the available data. In this paper, we divide the
virtual experiments into equality and lower bound data.

3.5. Model selection

To develop parsimonious probabilistic demand models (i.e. with only the explanatory functions that are strictly needed), a
model selection process is used to identify the important explanatory functions among the candidates presented in Table III.
For the model selection, we use a stepwise deletion procedure developed by Gardoni et al.30 for the case in which data are
in the form not only of equality data but also of lower (or upper) bound data. Starting with a comprehensive candidate form
of k(x, w, k), unnecessary terms are deleted in a stepwise manner on the basis of the posterior statistics of model parameters. At each step, the term hkj whose coefcient kj has the largest posterior coefcient of variation (COV) is deleted.
Model reduction is continued until an unacceptable increase is seen in the value of k.
Figure 1 summarizes the stepwise deletion process for shear demand model. At each step, the solid dots show the posterior COVs of the model parameters vi, and the open circle shows the posterior mean of v. It is seen that after 15 steps,
further model reduction deteriorates the accuracy of the model (i.e. v increases signicantly.) Stopping at this step, the
model is left with four terms.

Table IV. Probability terms for likelihood function with equality and lower bound data.30
Moment model
Shear model

Equality datum

h
i
h
i
rvi vjm 1
r mi
1

vjm
m
m
vjm

Equality datum

h
i
h
i
r 
 mi mjv mjv 1v rviv

Lower bound

r mi

Lower bound
rvi vjm
vm

1
m

rmi
m




r vi  vj 1

d


m m
vjm

p2 , k, l = v, m and v| = vm(v/m).
k|l = kl (k/l)rli, 
k lk
1
kl

Figure 1. Stepwise deletion process for shear demand model, where () indicates term to be removed.

1440

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Upon carrying out the model selection process, nal probabilistic seismic shear demand model is written as
 


PGA
Kt
v v
v18 ln
(13)
Dv x; w; v d^ v x; w v1 v2 d^ v x; w v9 ln
g
Kf
where d^ v the natural logarithm of the deterministic shear demand at the base of the tower normalized by the mean value of the

 

yield shear force, dened as V^ y ^f y A3=4 R2 r 2 = R2 Rr r 2 , in which ^f y expected yield stress of steel, and A, R
and r = tower base cross section area, outer and inner diameter, respectively; PGA = peak ground acceleration, g = gravitational
acceleration, Kt = tower stiffness and Kf = foundation stiffness.
Likewise, the model selection process for the moment model is carried out. Figure 2 shows the results of the
stepwise deletion process. As in Figure 1, solid dots show the posterior COVs of the model parameters mi, and open
circles show the posterior mean of m at each step. Figure 2 shows that after 13 steps of model reduction, the largest COV
(for parameter m15) is close in magnitude to m, and further reduction deteriorates the quality of the model (m increases).
Stopping at this step, the moment demand model is left with six correction terms and is written as




PGD
PGD
Dm x; w; m d^ m x; w m1 m2 d^ m x; w m11 ln
m13 ln 2


 
HH
PGVT n
(14)
Cs
Kt
m18 ln
m m
m15 ln
C s max
Kf
^ y ^f y S, in which
where d^ m the natural logarithm of the deterministic moment demand at the tower base normalized by M
S = elastic section modulus at the tower base; PGD = peak ground displacement, HH = hub height, PGV = peak ground velocity,
Tn = natural period of the support structure, Cs = soil shear wave velocity and Cs max = maximum shear wave velocity considered
in the analyses (194.6 m s1.)
3.6. Proposed probabilistic models
Table V gives the developed posterior statistics of the parameters v = (v1, v2, v9, v18, v). Figure 3 shows a comparison
between measured and predicted shear demands on the support structure on the basis of the deterministic (left) and probabilistic (right) models. For the probabilistic model, the median predictions are shown. Solid dots and open triangles indicate equality and lower bound data, respectively. The dashed lines in the Figure 3(b) delimit the region within one standard
deviation of the model.

Figure 2. Stepwise deletion process for moment demand model, where () indicates term to be removed.
Table V. Posterior statistics of the parameters in the shear demand model.
Correlation coefcient
Parameter
v1
v2
v9
v18
v

Mean

Standard deviation

3.050
0.737
0.259
0.228
0.508

0.480
0.067
0.044
0.046
0.033

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

v1
1
0.74
0.15
0.74
0.07

v2

1
0.41
0.15
0.11

v9

1
0.12
0.04

v18

1
0.01

1441

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Figure 3. Measured versus predicted shear demands based on (a) deterministic and (b) probabilistic models.

The deterministic model on the left is strongly biased on the non-conservative side, because almost all equality data and
most of the lower bound data lie below the 1:1 line. For a perfect model, all the equality data should be lined up along the
1:1 line, and all the lower bound data should lie above the 1:1 line. The proposed probabilistic demand model provides
unbiased estimates as shown in Figure 3(b).
Similarly, Table VI lists the posterior statistics of the parameters m = (m1, m2, m11, m13, m15, m18, m) for the
moment demand model. Figure 4 shows plots of predicted versus measured moment demands on the basis of the deterministic
(left) and probabilistic (right) models. The same comments as those already made in relation to Figure 3 apply. We note that

Table VI. Posterior statistics of the parameters in the moment demand model.

Parameter
m1
m2
m11
m13
m15
m18
m

Mean

Standard
Deviation

0.568
0.584
0.132
0.134
0.387
0.226
0.517

0.694
0.071
0.047
0.088
0.192
0.069
0.034

Correlation coefcient
m1

m2

m11

m13

m15

m18

1
0.16
0.40
0.54
0.62
0.88
0.01

1
0.48
0.49
0.40
0.33
0.09

1
0.68
0.11
0.01
0.02

1
0.05
0.15
0.01

1
0.69
0.05

1
0.03

Figure 4. Measured versus predicted moment demands based on (a) deterministic and (b) probabilistic models.

1442

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

whereas the deterministic model is strongly biased on the non-conservative side, the proposed probabilistic model corrects
the bias. In addition to the values of v and m, it is found that has a mean and standard deviation equal to 0.719 and
0.041, respectively.

4. FRAGILITY AND MEASURES OF IMPORTANCE AND SENSITIVITY

With the developed demand models, we can estimate the fragility of wind turbine support structures considering different
failure modes. We dene the fragility as the conditional probability that the shear or moment demands exceeds specied
capacity levels for given value of the vector w. According to the conventional notation in structural reliability theory,42
a limit state function gkj(x, w, k) is dened such that the event {gkj(x, w, k) 0} denotes the exceedance of the jth capacity level of kth limit state. Using the probabilistic demand model described earlier, gkj(x, w, k) is formulated as
gkj x; w; k C kj x  Dk x; w; k

(15)

in which Ckj(x) represents the capacity corresponding to Dk(x, w, k). In this paper, we consider yield and ultimate capacity
levels identied as j = y and u, respectively.
Table VII presents proposed damage states and the corresponding performance levels. The shear capacity is dened as the
allowable yield/ultimate shear force in the hollow cross section of the steel tubular tower Vj = fj A(3/4)(R2 + r2)/(R2 + Rr + r2),
normalized using V^ y , in which fj = fy or fu, for yield and ultimate stresses. Therefore, by simplifying the expression, the shear
^ y , which is
capacity can be written as C vj f j =^f y . Similarly, the yield-bending moment capacity is dened as C my M y =M
^
the resultant of the allowable yield-bending moment My = fyS, normalized using M y . Finally, the ultimate moment capacity is
^ y, in which Mu = maximum moment in the moment-curvature diagram, constructed for a tubular cross section
dened as M u =M
of tower, on the basis of the stressstrain curve for structural steel of grade S235, to calculate the ultimate bending moment
capacity.
The fragility is then formulated as



F j w; P
gkj x; w; k 0 w;
(16)

where P[B|w] denotes the conditional probability of event B for the given values of w. The uncertainty in the event for
given w arises from the inherent randomness in the structural and material properties, the inexact nature of the limit state
model and the uncertainty inherent in the model parameters.
4.1. Predictive estimates of fragility
Following Gardoni et al.,30 a predictive estimate of the fragility is formulated as

e j w F j w; f d
F

(17)

which incorporates the uncertainties in the model parameters by considering as random variables and taking the
expected value of Fj(w, ) over the posterior distribution of . Following Ditlevsen and Madsen,42 the corresponding


ej .
reliability index can be dened as e
j F
4.2. Sensitivity measures
A sensitivity analysis can be carried out to identify which parameter(s) inuences the most the reliability of wind turbine
support structures. Sensitivity measures can provide insight into the behavior of support structures and are useful for

Table VII. Proposed damage states.

Damage state

Description

No signicant damage (ND)

No structural damage

Permanently out-of-service (PO)

Support structure yields. Permanent

excessive deformations

Complete (C)

Support structure is unable to carry additional loads

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

Performance level
Tower base shear or moment
exceeds yield limit
Tower base shear or moment
exceeds ultimate limit

1443

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

optimal design and resource allocation. Hohenbichler and Rackwitz43 dene the sensitivity of a reliability index as the
gradient of with respect to the parameters g in the limit state function
g



1
g z*; g
kGk g

(18)

where G(u) = g(z(u)) = limit state function expressed in terms of the standard normal variables.42 Once g is known, the gradient of the rst-order reliability approximation of the failure probability is obtained using the chain rule of differentiation as
g p1 g

(19)

where () = standard normal PDF.

4.3. Importance measures
Among several random variables in the limit state function, some have larger effect on the variance of the limit state
function and thus are more important. Der Kiureghian and Ke44 proposed the measure of importance dened as
T Ju*;z* SD

T 

 T
 Ju*;z* SD 

(20)

where = unit vector at the design point directed toward the failure set, z = vector of random variables, Ju *,z * = Jacobian of
the probability transformation from the original space z to the standard normal space u with respect to the coordinates of the
design point z* and nally, SD = standard deviation matrix of equivalent normal variables z dened by the linearized
inverse transformation z = z * + Jz *,u *(u  u *) at the design point. The elements of SD are the square roots of the
corresponding diagonal elements of the covariance matrix = z * + Jz *,u *JTz *,u * of z.

5. ILLUSTRATION
As an illustration of the proposed probabilistic models and framework, we assess the reliability of a typical 5 MW wind
turbine with the same characteristics as those of the NREL offshore 5 MW baseline wind turbine, introduced by Jonkman
et al.45 The considered wind turbine is installed in a 20 m water depth and supported by a monopile foundation that is
typical for this water depth. Table VIII lists all the relevant properties.
5.1. Predictive fragility
Monte Carlo simulations are used to estimate the predictive fragility for the example wind turbine support structure, where
the parameters in the demand models are considered to be random variables with a Nataf distribution46 such that v and m
are jointly normal, v and m are lognormal and follows a beta distribution. Statistical properties of the model parameters
Table VIII. Properties of the NREL offshore 5-MW baseline wind turbine.
Rating
Rotor diameter
Hub height
Cut-in, rated, cut-out wind speeds
Rotor mass
Nacelle mass
Tower mass
Natural period of the tower
Tower top diameter and wall thickness
Tower base diameter and wall thickness
Water depth
Turbulence intensity
Signicant wave height
Wave peak period

1444

5 MW
126 m
90 m
1
1
1
3 m s , 11.4 m s , 25 m s
110 000 kg
240 000 kg
347 460 kg
2.5 s
3.87 m, 0.019 m
6.00 m, 0.027 m
20 m
0.10
1.0 m
4.87 s

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

are given in Tables V and VI. The random variables that dene the capacity models are described in Table IX. A COV of
10% is assumed for yield and ultimate strength of steel.47 The statistics of Mu is obtained using moment-curvature diagrams
constructed for the tubular cross section of the tower base, considering the stressstrain curve of structural steel of Grade
S235. A COV of 30% is selected for soil strength properties on the basis of data collected from literature.48
Figure 5(a) presents the predictive fragility estimates plotted as a function of spectral acceleration Sa at the natural period
of the support structure (Tn = 2.5 s) within its linear elastic range, for both the yield and ultimate limit states. The signicant
wave height is set to Hs = 1 m. The dotted, solid and dashed lines in the gure show the fragilities for cut-in, rated and cutout wind speeds, respectively, where cut-in and cut-out wind speeds are the lower and upper limits of the range of wind
speeds in which a turbine is operating and producing power. The rated wind speed is the wind speed at which a control
system is activated to keep the power-generated constant by changing the blade pitch angle. The pitch control system limits
the aerodynamic forces on the blades and consequently the operational loadings on the support structure of the wind
turbine. As shown in the gure, the fragility at the rated wind speed is higher than the fragilities at the other two wind
speeds because of the higher wind speed than the cut-in wind speed and higher operational loading than at the cut-out wind
speed. However, the contribution of the wind loading in the operational range of wind turbines is not signicant compared
with the seismic excitation even for small earthquakes. In addition, the fragility in shear failure mode is found to be
negligible compared with that in the bending failure mode, as expected for slender elements like wind turbines towers.
Predictive fragility estimates due to yield and ultimate limit states are also plotted as a function of the mean wind speed, for
Hs = 1 m and in absence of earthquakes, Sa = 0 (Figure 5(b).) The gure shows how the fragility rapidly increases after the cutout wind speed due to the lack of aerodynamic damping for parked (idle) wind turbine in the presence of high wind speeds.
5.2. Sensitivity measures
We dene g = [E(kt/kf), E(Cs), E(Mu), E(fy), PGD, PGV], where E() = the expected value of the variable. Table X lists the
sensitivity measures for the moment failure mode for both yield and ultimate capacity levels. Results show that increasing
the mean of the tower to foundation stiffness ratio kt/kf is the most effective way of increasing the bending moment reliability (reducing the probability of failure.) Also, the shear wave velocity Cs happens to be the second most important parameter, whose increment (increasing the soil stiffness) will increase the reliability of the support structure.

Table IX. Distribution, mean and COV for the random variables in the capacities in the limit state
function.
Random variables

Distribution

Mean

COV (%)

fy
fu
Mu
Cs
kt/kf

Lognormal
Lognormal
Lognormal
Lognormal
Lognormal

300.0
410.0
390.6
109.2
0.0020

10
10
10.13
30
30

Figure 5. Fragility estimates for a typical 5 MW offshore wind turbine as a function of (a) spectral acceleration and (b) mean wind speed.

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1445

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Table X. Sensitivity measures for the moment failure mode for both yield and ultimate performance
levels.
xc
Parameter, xc

Symbol

Yield

Ultimate

Mean of tower to foundation stiffness ratio

Mean of shear wave velocity of soil
Mean of ultimate bending moment capacity
Mean of yield stress of steel
Peak ground displacement
Peak ground velocity

E(kt/kf)
E(Cs)
E(Mu)
E(fy)
PGD
PGV

220.1
1.333
0.000
0.000
0.001
0.294

124.0
1.542
0.000
0.000
0.003
0.062

5.3. Importance measures

Table XI shows the importance measures for the moment failure mode, for yield and ultimate capacity performance levels,
where z = (xp, m), in which xp = (kt/kf, Cs, Mu, fy, fu, m) are the random variables in the limit state function, in addition to
the model parameters m considered in a Bayesian approach as random variables. We see that in addition to the model error
m, some of the model parameters m1, m18 and m13 are also important random variables that affect the variance of the
limit state function. On the other hand, there are random variables in the limit state function, which are not important,
and we can ignore their uncertainty in developing fragility estimates without a signicant loss of accuracy. Therefore,
we partition z in a vector of constant parameters zc = (kt/kf, Cs, Mu, fy, fu, m, m2, m11, m15), which includes the point estimates
of unimportant random variables equal to their mean values, and a vector of random variables zp = (m, m1, m18, m13),
so that z can be written as z = (zc, zp). By reducing the number of random variables in the limit state function, we make
the computation of fragilities faster without signicant loss of accuracy. This observation can be helpful to expedite the
computation of the reliability of similar wind turbines.

5.4. Structural damage states

Once the fragilities and the annual PDFs for seismic and wind hazards are available, we can estimate the annual probabilities of being in any specied damage state for a wind turbine support structure at any particular locations using the total
probability rule (equation (3)).
In this paper, we select two locations: site I in the Gulf of Mexico, about 70 Km east of Galveston, Texas, with latitude of
2925N and longitude of 9403W, which is prone to hurricanes; and site II in the west coast, about 90 Km west of Santa
Babara, California; with latitude of 3416N and longitude of 12042W, which is in a high seismic region.
We use USGS seismic hazard exceedance curves for the Gulf of Mexico and the west coast of the USA. Figure 6 shows
a comparison of USGS seismic hazard curves at the two sites of interest. The gure clearly shows that the annual probability of occurrence of an earthquake at site II is signicantly larger than at site I.

Table XI. Importance measures for the moment failure mode for both yield and ultimate performance
levels.
i
Random Variable
Model parameter for hm1
Model error/m
Model parameter for hm18
Model parameter for hm13
Model parameter for hm15
Shear wave velocity of soil
Ultimate bending moment capacity
Model parameter for hm11
Tower to foundation stiffness ratio
Model parameter for hm2
Standard deviation of moment model error
Yield stress of steel

1446

Symbol

Yield

Ultimate

m1
m
m18
m13
m15
Cs
Mu
m11
kt/kf
m2
m
fy

0.674
0.501
0.419
0.275
0.104
0.115
0.000
0.071
0.066
0.014
0.029
0.097

0.673
0.500
0.421
0.275
0.116
0.110
0.098
0.071
0.064
0.014
0.001
0.000

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Figure 6. Annual probability density function for spectral acceleration at site I (dotted line) and site II (solid line).

For the day-to-day wind, we use the BIW as presented in equation (5), with the distribution parameters estimated by
Morgan et al.27 for sample ocean buoy stations at both sites I and II. Table XII lists the BIW distribution parameters for
the two locations. Depending on the type of buoy, the wind speed recorded by National Data Buoy Center28 is measured
at either 5 or 10 m above sea level.27 However, we are interested in wind speed data at the height of turbine hub. To obtain
the wind speed at the turbine hub height, we use an empirical approximation of wind speed prole Ws(h) written as49

W s h V r

h
Hr


(21)

In which, Vr = wind velocity at a reference height, where a common choice for the reference height is Hr = 10 m. The
Guidelines for Design of Wind Turbines (DNV/Riso)49 suggests a value of = 0.12 for offshore winds.
The PDFs of the wind speed at both sites I and II are then modied to represent the probability density of wind speeds at
the wind turbine hub height. As mentioned earlier, Wang29 found that the lognormal distribution provides the best t to
measured data for the annual PDF of the gradient wind speeds during hurricane along the Texas coastline. The gradient
level is generally taken as between 500 and 2000 m. Lee and Rosowsky50 summarized the gradient-to-surface wind speed
conversion factor for 10 min sustained wind speeds for different locations. They suggested a value of 0.65 for offshore
sites. The surface wind speed is the value of wind velocity at 10 m height above the ground or sea level. However, the wind
speed at a wind turbine hub height is of interest in this study. We convert the gradient wind speed to hub height wind speed
by rst bringing it down to surface by applying the gradient-to-surface conversion factor and then taking it up to hub height
using equation (21). Finally for site I, we use the lognormal distribution function presented in equation (6) for the hub
height wind speed during the hurricane with location and scale parameters of = 3.348 and = 0.34, respectively.
The occurrence of hurricane is modeled as a Poisson process with an annual occurrence rate of hurricane to be = 0.1689
for site I and = 0 for site II, on the basis of the Historical Hurricane Tracks database at the National Oceanic and
Atmospheric Administration.51 Wind hazard curves are then developed using equation (7). Figure 7 shows the wind hazard
curves for the two particular locations of interest. It can be seen that site II has generally higher day-to-day wind speeds and
is not expected to experience hurricanes.
Table XIII lists the annual probabilities of being in the damage states specied in Table VII for the NREL offshore 5
MW wind turbine subject to different hazards. Table XIII shows that even though the occurrence rate of hurricane is
signicantly larger at site I than site II, the wind hazard alone happens to result in the same failure probabilities for the

Table XII. BIW distribution parameters.

Value

Parameter
Site I
1
1
2
2

4.455
5.809
2.067
6.368
0.141

Site IIb
1.694
5.515
4.405
9.848
0.467

Buoy 42035.29
Buoy 46063.29

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1447

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

Figure 7. Annual probability density function for wind speeds at (a) site I and (b) site II.

Table XIII. Annual probabilities of specied damage states.

Annual probabilities of being in specied damage states

Damage state
ND
PO
C

Performance
level
Yield limit
exceedance
Ultimate limit
exceedance

Wind hazard

Seismic hazard

Multi-hazard

Site I

Site II

Site I

Site II

Site I

Site II

0.9869

0.9861

0.9902

0.9771

0.9857

0.9733

0.0116

0.0123

0.0088

0.0196

0.0126

0.0230

0.0015

0.0016

0.0010

0.0033

0.0017

0.0037

ND, no signicant damage; PO, permanently out-of-service; C, complete.

two sites of interest. Figure 7(b) can explain the reason for the relatively high probabilities of exceeding the performance
levels at site II due to wind hazard. The PDF of day-to-day wind speed for site II has a considerably higher density around
the rated wind velocity than site I, which has most of its PDF density at the low wind speeds (where the fragility is small).
For the case of seismic hazard in the presence of day-to-day wind for operating wind turbine, the wind turbine installed at
site II has considerably larger probabilities of exceeding the performance levels than the one installed at site I, as expected
on the basis of the seismic hazard curves. Finally with the consideration of multiple hazards (wind and seismic), it can be
seen that site II results in an overall higher failure probability for the wind turbine considered in this study.
It is noteworthy to mention that the obtained probabilities of complete damage (damage state C) are generally higher than
the nominal annual failure probability of 104, taken by DNV-OS-J1016 as target safety level for structural design of support
structures and foundations for wind turbines to the normal safety class. Given the fact that the DNV-OS-J101 guideline6 only
considers wave loads and normal and extreme wind conditionsexcluding conditions experienced in tropical storms6in
addition to permanent loads, the higher values show the importance of considering hurricane and seismic loading.

6. CONCLUSIONS
A probabilistic framework is proposed to evaluate the multi-hazard structural reliability of offshore wind turbines. Probabilistic
models were developed for shear and moment demands on the support structure of wind turbines using the information obtained
from detailed 3D non-linear FE models. The FE models incorporated the aeroelastic interaction as well as the inuence of soilstructure interaction in the dynamic response of the support structures. The developed demand models are used to assess the
fragilities of an example offshore wind turbine subject to day-to-day and extreme wind speeds in addition to seismic loading.
No specic design load cases are considered in the simulations. However, one could use the methodology described in this
paper for specic design load cases to assess the corresponding probability of failure. Incorporating the hazard information
for two particular locations in the USA (one in the Gulf of Mexico and one off the California Coast), the annual probabilities
of specied structural damage states are evaluated for a typical 5 MW offshore wind turbine subject to day-to-day and extreme
wind loads during hurricane in the presence of the seismic risk. The fact that obtained values for annual probability of failure
considering seismic and tropical storm hazards are higher than target safety level recommended by guidelines for industry practice, excluding those hazards, shows the importance of considering seismic and hurricane loading for design purposes, although
with the wind turbines operating at their maximum rate of power production at the rated wind speed, the West Coast could be an
1448

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

M. Mardfekri and P. Gardoni

Multi-hazard reliability assessment of offshore wind turbines

ideal location for an offshore wind farm on the power production side. However, on the basis of the resulted annual probabilities
of failure, there is a higher failure risk for wind turbines installed in the West Coast of the USA, because of high seismicity.
Given that choosing a location for a wind farm is always a trade-off between multiple considerations, including possible loss,
power production potential and construction cost, this paper proposes a valuable framework to estimate the failure probability to
be used in estimation of possible losses.

ACKNOWLEDGEMENT
The authors acknowledge the Texas A&M Supercomputing Facility (http://sc.tamu.edu/) for providing computing
resources useful in conducting the research reported in this paper.

REFERENCES
1. Sawyer S, Rave K. Annual market update2011. Global Wind Energy Council (GWEC), March 2012. Available: http://
www.gwec.net/leadmin/documents/NewsDocuments/Annualreport_2011_lowres.pdf (Accessed 10 September 2012).
2. Gsnger S, Pitteloud JD. World wind energy report 2011. World Wind Energy Association (WWEA), May 2012.
Available: http://www.wwindea.org/webimages/WorldWindEnergyReport2011.pdf (Accessed 10 September 2012).
3. Flowers L. Wind energy update. National Renewable Energy Laboratory (NREL), Jan. 2012. Available: http://www.
windpoweringamerica.gov/pdfs/wpa/wpa_update.pdf. (Accessed 12 September 2012)
4. Schwartz M, Heimiller D, Haymes S, Musial W. Assessment of offshore wind energy resources for the United States.
Technical Report NREL/TP-500-45889, NREL, Golden, CO, 2010.
5. IEC TC88-MT1 (ed.). IEC 61400-1 Ed.3. Wind Turbines. Part 1: Design Requirements. International Electrotechnical
Commission: Geneva, 2005.
6. Det Norske Veritas. Offshore standard DNV-OS-J101: design of offshore wind turbine structures, October 2007.
7. Walford CA. Wind turbine reliability: understanding and minimizing wind turbine operation and maintenance costs.
Report Number SAND2006-1100, Sandia National Laboratories, Albuquerque, NM and Livermore, CA, 2006.
8. Tavner PJ, Xiang J, Spinato F. Reliability analysis for wind turbines. Wind Energy 2007; 10: 118.
9. Madsen PH, Pierce K, Buhl M. Predicting ultimate loads for wind turbine design. NREL/CP-500-25787, NREL,
Golden, CO, 1998. Proceedings of AIAA/ASME Wind Energy Symposium, Reno, Nevada, 1114 Jan. 1999; 355364.
10. Agarwal P, Manuel L. Extreme loads for an offshore wind turbine using statistical extrapolation from limited eld data.
Wind Energy 2008; 11: 673684.
11. Manuel L, Veers PS, Winterstein SR. Parametric models for estimating wind turbine fatigue loads for design. ASME
Journal of Solar Energy Engineering 2001; 123: 346355.
12. Andersen LV, Vahdatirad MJ, Sichani MT, Srensen JD. Natural frequencies of wind turbines on monopile foundations in
clayey soilsa probabilistic approach. Computers and Geotechnics 2012; 43: 111.
13. Bush E, Manuel L. Foundation models for offshore wind turbines. Proceedings of 47th AIAA Aerospace Sciences
Meeting Including The New Horizons Forum and Aerospace Exposition, Orlando, Florida, 58 January 2009.
14. Mardfekri M, Gardoni P, Roesset J. Modeling laterally loaded piles accounting for soilpile interactions. Journal of
Engineering 2013; 2013: 243179243186.
15. Jonkman JM, Buhl ML Jr. FAST users guide. Technical Report NREL/EL-500-38230, NREL, Golden, CO, 2005.
16. Laino DJ, Hansen AC. Users Guide to the Computer Software Routines AeroDyn Interface for ADAMS. Windward
Engineering LLC: Salt Lake City, UT, 2001.
17. Bossanyi EA. Bladed for Windows User Manual. Garrad Hassan and Partners: Bristol, 2000.
18. Mardfekri M, Gardoni P. Probabilistic demand models and fragility estimates for offshore wind turbine support
structures. Engineering Structures 2013; 52: 478487.
19. Bazeos N, Hatzigeorgiou G, Hondros I, Karamaneas H, Karabalis D, Beskos D. Static, seismic and stability analyses of
a prototype wind turbine steel tower. Engineering Structures 2002; 24: 10151025.
20. Lavassas I, Nikolaidis G, Zervas P, Efthimiou E, Doudoumis IN, Baniotopoulos CC. Analysis and design of the
prototype of a steel 1-MW wind turbine tower. Engineering Structures 2003; 25: 10971106.
21. Witcher D. Seismic analysis of wind turbines in the time domain. Wind Energy 2005; 8: 8191.
22. Prowell I, Elgamal A, Jonkman J. FAST simulation of wind turbine seismic response. NREL/CP-500-46225, NREL,
Golden, CO, 2010. Proceedings of the Asian-Paciic Network for earthquake Engineering Research (ANCER)
Workshop, Urbana-Champaign, Illinois, 1314 August 2009.
Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we

1449

Multi-hazard reliability assessment of offshore wind turbines

M. Mardfekri and P. Gardoni

23. Prowell I, Veletsos M, Elgamal A. Full scale testing for investigation of wind turbine seismic response. Proceedings of
the 7th World Wind Energy Conference, Kingston, Ontario, Canada, 2426 June 2008.
24. Mardfekri M, Gardoni P. Seismic risk analysis of wind turbine support structures. Chapter 26 in Seismic Risk Analysis
and Management of Civil Infrastructure Systems. Tesfamariam S, Goda K (eds). Woodhead Publishing Ltd.:
Cambridge, UK, 2012: 716738.
25. Ang AHS, Tang WH. Probability Concepts in Engineering: Emphasis on Applications to Civil and Environmental
Engineering (2nd edn). John Wiley & Sons, Inc: New York, 2007.
26. Frankel AD, Petersen MD, Mueller CS, Haller KM, Wheeler RL, Leyendecker EV, Wesson RL, Harmsen SC, Cramer
CH, Perkins DM, Rukstales KS. Documentation for the 2002 update of the National Seismic Hazard Maps. U.S.
Geological Survey Open-File Report 02-420, USGS, Denver, CO, 2002.
27. Morgan EC, Lackner M, Vogel RM, Baise LG. Probability distribution for offshore wind speeds. Energy Conversion
and Management 2011; 52: 1526.
28. NDBC. NOAAs national data buoy center, 2009. Available: http://www.ndbc.noaa.gov/. (Accessed July 2009)
29. Wang Y. Studies on hazard characterization for performance based structural design. Ph.D. Dissertation, Department of
Civil Engineering, Texas A&M University, College Station, TX, 2010.
30. Gardoni P, Der Kiureghian A, Mosalam KM. Probabilistic capacity models and fragility estimates for RC columns
based on experimental observations. Journal of Engineering Mechanics ASCE 2002; 128: 10241038.
31. Gardoni P, Mosalam KM, Der Kiureghian A. Probabilistic seismic demand models and fragility estimates for RC
bridges. Journal of Earthquake Engineering 2003; 7: 79106.
32. Jonkman BJ. TurbSim users guide. Technical Report NREL/EL-500-46198, NREL, Golden, CO, 2009.
33. Kaimal JC, Wyngaard JC, Izumi Y, Cote OR. Spectral characteristics of surface-layer turbulence. Quarterly Journal of
the Royal Meteorological Society 1972; 98: 563589.
34. Dean RG, Dalrymple RA. Water Wave Mechanics for Engineers and Scientists. World Scientic: Singapore, 1991.
35. Rezaeian S, Der Kiureghian A. Stochastic modeling and simulation of ground motions for performance-based
earthquake engineering. PEER 2010/02, University of California, Berkeley, CA, 2010.
36. Box GEP, Tiao GC. Bayesian Inference in Statistical Analysis. John Wiley & Sons, Inc: New York; 1992.
37. ABAQUS. Version 6.7-2. Hibbitt, Karlsson and Sorensen, Inc., 2007.
38. Ishihara T, Sarwar MW. Numerical and theoretical study on seismic response of wind turbines. Proceedings of
European Wind Energy Conference (EWEC), Brussels, Belgium, 31 March3 April 2008.
39. Pacic Earthquake Engineering Research Center: NGA Database. 1999. Available: http://peer.berkeley.edu/nga. (Accessed
September 2011)
40. Shome N, Cornell CA. Probabilistic seismic demand analysis of nonlinear structures. Reliability of Marine Structures
Report RMS-35, Department of Civil and Environmental Engineering, Stanford Univ., Palo Alto, CA, 1999.
41. Bisadi V, Gardoni P, Head M. Probabilistic demand models and fragility estimates for bridges elevated with steel
pedestals. Journal of Structural Engineering 2012. DOI: 10.1061/(ASCE)ST.1943-541X.0000741.
42. Ditlevsen O, Madsen HO. Structural Reliability Methods (2nd edn). John Wiley & Sons, Inc: New York, 2005.
43. Hohenbichler M, Rackwitz R. First-order concepts in system reliability. Structural Safety 1983; 1: 177188.
44. Der Kiureghian A, Ke JB. Finite-element based reliability analysis of frame structures. Proceedings of 4th International
Conference on Structural Safety and Reliability (ICOSSAR), Kobe, Japan, 1, 2729 May 1985; 395404.
45. Jonkman J, Buttereld S, Musial W, Scott G. Denition of a 5-MW reference wind turbine for offshore system
development. Technical Report NREL/TP-500-38060, NREL, Golden, CO, 2009.
46. Der Kiureghian A, Liu PL. Structural reliability under incomplete probability information. Journal of Engineering
Mechanics 1986; 112: 85104.
47. Mirza SA, MacGregor JG. Variability of mechanical properties of reinforcing bars. Journal of Structural Division
(ASCE) 1979; 105: 921937.
48. Jones AL, Kramer SL, Arduino P. Estimation of uncertainty in geotechnical properties for performance-based
earthquake engineering. PEER 2002/16, University of California, Berkeley, CA, 2002.
49. DNV/Ris. Guidelines for Design of Wind Turbines (2nd edn). Det Norske Veritas and Wind Energy Department, Ris
National Laboratory: Denmark, 2002.
50. Lee KH, Rosowsky DV. Synthetic hurricane wind speed records: development of a database for hazard analysis and
risk studies. ASCE Natural Hazards Review 2007; 8: 2334.
51. Historical Hurricane Tracks, National Oceanic and Atmospheric Administration (NOAA). Available: http://www.csc.
noaa.gov/hurricanes/# (Accessed 05 July 2012).
1450

Wind Energ. 2015; 18:14331450 2014 John Wiley & Sons, Ltd.
DOI: 10.1002/we