Special Issue Article

Advances in Mechanical Engineering

2017, Vol. 9(9) 1–26
Ó The Author(s) 2017
Statistical wind prediction and fatigue DOI: 10.1177/1687814017724088
journals.sagepub.com/home/ade
analysis for horizontal-axis wind
turbine composite material blade
under dynamic loads

Pravin A Kulkarni1, Weifei Hu2, Ashwinkumar S Dhoble1 and

Pramod M Padole1

Abstract
The purpose of this article is to analyze various wind speed-forecasting methods, select the appropriate method for
developing synthetic wind speed for 1-year period at Salem in Tamilnadu state in India, and use it for the structural and
fatigue analysis of a small horizontal-axis wind turbine blade made of composite material. Various forecasting models
such as Markov chain, Kalman filter, and autoregressive integrated moving average are evaluated, and a long-term wind
speed pattern at Salem is developed using Markov chain. This wind pattern is used to create time-varying loads using the
blade element momentum on blade sections. Then, the fatigue analysis of the blade is carried out using the stress life
approach. The blade is found to have available life of about 20 years and the critical area for fatigue is found on the skin
near the root of the blade. Various blade skin materials are also compared for fatigue performance. A cohesive zone
model of the adhesively joined root joint is also developed and analyzed for fatigue at the metal–composite joint. Thus,
an integrated methodology involving high-fidelity modeling of the blade, wind forecasting, and static and fatigue analysis is
developed for horizontal-axis wind turbine blade for locations where historically wind speed measurements are available
for short time.

Keywords
Wind profile, blade element momentum, Markov chain, autoregressive integrated moving average, structural analysis,
wind speed prediction, blade fatigue life, cohesive zone modeling

Date received: 25 January 2017; accepted: 11 July 2017

Academic Editor: Aditya Sharma

Introduction also reduces operating costs and improves reliability of

power generation system as highlighted by Smith et al.5
The service life of wind turbine blades is highly influ- Researchers therefore focus on wind profile analysis
enced by the random and nonlinear nature of wind
leading to unsteady aerodynamic effects, wind turbu-
lence, and dynamic stall and causing fatigue loads as 1
Department of Mechanical Engineering, Visvesvaraya National Institute
highlighted by many researchers such as Mo et al.1 and of Technology, Nagpur, India
Liu et al.2,3 Wind speed forecast is closely associated 2
Cornell University, Ithaca, NY, USA
with power generation prediction, and high-precision
Corresponding author:
wind speed prediction is needed for effective and safe Pravin A Kulkarni, Department of Mechanical Engineering, Visvesvaraya
wind energy utilization for small power systems as National Institute of Technology, Nagpur 440010, India.
highlighted by Kaneko et al.4 Accurate wind prediction Email: pravin.kul@gmail.com

Creative Commons CC-BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Advances in Mechanical Engineering
and forecasting to better predict the blade loads and
fatigue and optimize the blade design for improving
safety and reliability. With a focus on blade optimiza-
tion study which depends on wind, rotor control strat-
egy and generator type, Yi et al.6 studied the effects of
rotor control strategies such as combinations of fixed
and variable speed and pitch algorithms on optimal
blade shapes and rotor performance using optimized
blade designs. They also evaluated the effects of envi-
ronmental wind data and the optimization functions
using the NREL HARP_Opt tool which uses WT_Perf
code for performance evaluation based on blade ele-
ment momentum (BEM). Some of the performance
indices which the authors compared and are significant
for this study were thrust, torque, and roof-flap
moment forces. Ghasemi et al.7 studied the aeroelastic
stability aspects of composite blades of a 660 kW wind
turbine blade under flutter, that is, coupling of bending
and torsional vibrations. The blade section is made up
of NACA 632-415. Analytical techniques were used to
calculate flutter speed and safety factor, while finite ele-
ment model (FEM) was used for static and modal anal-
yses. In another study on blade optimization, Lee
et al.8 numerically analyzed effect of active load control
devices such as trailing-edge flaps for reducing fatigue
loading on wind turbine blades and found a 30%–50%
reduction in the standard deviation of the root bending
moment. The authors of this study have also used the
blade model with trailing-edge stiffeners for the same
purpose. Chen and Lee9 have used an adaptive input
estimation method which combines Kalman filter with-
out input term and adaptive weighted recursive least
square estimator to predict the wind load on a multi-
layer shearing stress system.
As pointed by Grujicic et al.,10 the main structural
performance requirements for wind turbines are (1)
flapwise bending strength to withstand extreme wind
loads and (2) flapwise bending stiffness to ensure a
minimum clearance between blade tip and the turbine
tower. To meet these conditions, Hu et al.11 developed
a fatigue analysis procedure including random wind
field simulation, aerodynamic analysis, FE stress analy-
sis, and fatigue damage simulation. Probability density
functions (PDFs) for 10-min mean wind speed and 10-
min turbulence intensity were developed to simulate a
1-year wind field and calculated the accumulated fati-
gue damage, which could facilitate reliability analysis
and reliability-based design optimization of composite
wind turbine blades considering wind load uncertainty.
Lecheb et al.12 have analyzed fatigue life of a 25-m-long
wind turbine blade made of glass/epoxy composite.
They first carried out modal and static analyses to find
out blade critical zone for stresses followed by crack
growth study, experimental fatigue life determination,
and fast fracture. In order to help meet needs such as
carrying out risk analysis, reliability-based design
optimization, and maintenance scheduling, An et al.13
have used Bayesian technique to obtain a fatigue life
prediction model for wind turbines. Their method
makes use of available field data, noise, and bias in
measurements on the distribution of fatigue life so that
difference between designer’s predicted life and field
observations can be reduced. They used Markov chain
Monte Carlo technique to obtain the posterior distribu-
tion. Significant amount of simulation and experiments
have been investigated on the life prediction of compo-
site materials in literature; however, experimental vali-
dations of fatigue failure of composite wind turbine
blades (made of composite materials) are often ham-
pered due to lack of enough structural health monitor-
ing data measured under realistic random wind loading
in the field. Hence, this article focuses on accurate wind
modeling for structural and fatigue analyses of compo-
site wind turbine blades.
There are various methods available for wind mod-
eling that could be classified as (1) physical models, (2)
spatial correlation models, (3) artificial intelligence
models, (4) stochastic statistical models, and (5) hybrid
models.
Spatial correlation models take the spatial relation-
ship of different sites’ wind speed into account for pre-
diction. Spatial correlation models need site-specific
local historic data for accurate prediction and are use-
ful when the nonlinear relationship between these local
variables is required to be modeled. However, these
models may not be appropriate to be used when little
historical information is available.
The use of artificial intelligence models such as artifi-
cial neural network (ANN) and fuzzy logic is increasing
over the years. Barbounis and Theocharis14 developed
a locally feedback dynamic fuzzy neural network (LF-
DFNN) for 15 min to 3 h ahead wind speed prediction
with spatial correlation and having superior accuracy
over other network models. Flores et al.15 supplied a
control algorithm based on ANN model using back-
propagation method for wind speed prediction. Surveys
on artificial intelligence models on wind speed predic-
tion has been carried out in Lei et al.,16 which con-
cludes that it is necessary to further study on artificial
intelligence methods and improve their training algo-
rithm aiming at more accurate results of wind speed
prediction.
Stochastic statistical models consist of a mathemati-
cal model based on historical data, patterns, and wind
parameters. Amezcua et al.17 developed an algorithm
using generalized random Fourier series to generate a
real-time wind speed time series from data logger
records containing the average, maximum, and mini-
mum values of the wind speed in a fixed interval, as
well as the standard deviation. They also incorporated
extreme values using the concept of constrained simula-
tion in post processing.
Kulkarni et al.
One typical group of stochastic statistical
approaches includes autoregressive (AR) models, inte-
grated models (I), and moving average (MA) models
which depend linearly on the previous data. The com-
bined methods such as autoregressive moving average
(ARMA) described by Box and Jenkins18 and autore-
gressive integrated moving average (ARIMA) models.
A review of the literature by Jung and Broadwater19
suggests that statistical models can be utilized for
short-, medium-, and long-term predictions quite well.
Cadenas and Rivera20 applied ARIMA and the ANN
methods and also compared the two techniques for
wind speed prediction in Mexico. They found that both
the models provided reasonable predictions, and seaso-
nal ARIMA model has a better sensitivity to wind
speed adjustment. They utilized mean absolute percent-
age error (MAPE) as a forecast error for comparison.
Another commonly used stochastic statistical model
for wind speed prediction is the Markov chain model.21
Synthetic wind speed generation for a specific site using
Markov transition matrix has been employed by vari-
ous researchers such as Negra et al.22 for short-term
prediction. These models are easy and inexpensive to
implement as compared to spatial correlation models
which need large geospatial data of various sites as well
as higher processing cost due to the multiple input vari-
ables representing physical properties of the sites such
as temperature, pressure, humidity, and altitude to pre-
dict the wind speed. Thus, in this article, the wind speed
simulation has been carried out by three stochastic sta-
tistical models such as ARIMA, Kalman filter, and
first-order Markov chain, and the suitable model is
selected for wind speed prediction and related struc-
tural fatigue analysis.
Hybrid models combine the benefits of individual
models and overcome limitations of using a single
model. They could be formed by combining physical
and statistical models or statistical and spatial models.
Huang and Xu23 have proposed a numerical simula-
tion procedure for predicting directional typhoon
design wind speeds and profiles for sites over complex
terrain by integrating typhoon wind field model,
Monte Carlo simulation technique, computational fluid
dynamics (CFD) simulation, and ANNs. Ke et al.24
utilized another hybrid forecast model called wavelet
packet transform (WPT) for wind prediction based on
daily average wind speed data from four wind farms in
China. The hybrid model although is more time-con-
suming, and it has better accuracy than the individual
models as indicated by the error criteria such as mean
absolute error (MAE), mean square error (MSE), and
MAPE.
The literature survey on wind modeling and wind
turbine blade fatigue analysis shows that although
there are various studies available on wind prediction
methods and blade structural and fatigue analyses, very
3
few are available for an integrated framework starting
from forecasting of long-term wind to aerodynamic
and structural analyses and to fatigue analysis of com-
posite horizontal-axis wind turbine (HAWT) blades
with limited or no raw wind data available for the site.
Hence, this study focuses on developing an integrated
procedure incorporating long-term wind forecasting
methods into blade fatigue analysis.
The unique contribution of this study is that (1) it
compares the various statistical wind forecasting mod-
els to select the one which is best suitable for fatigue
analysis applications, (2) develops long-term wind pro-
file prediction for sites which does not have enough his-
torical wind measurement data to a good level of
accuracy, (3) performs high-fidelity three-dimensional
(3D) modeling and analysis for random wind loads for
fatigue, and (4) extends the conventional fatigue analy-
sis by comparing the cohesive zone modeling (CZM)
approach for root joint of blade for fatigue. Thus, this
study presents an end-to-end framework for wind tur-
bine blade analysis.
For accurate structural analysis, a high-fidelity blade
model is generated referring from the work by Perkins
and Cromack.25 The structural analysis results are
compared with the published results so that the base-
line model is available. The predicted long-term wind
load is then applied to the fatigue analysis considering
wind load variation at the selected location of Salem,
Tamilnadu in India.
The blade analysis procedure consists of four steps
as shown in Figure 1. First, a detailed computer-aided
design (CAD) model for a wind turbine blade is devel-
oped in Catia V5R20 software based on the specifica-
tions. Pre-processing of the model is performed by
applying material properties and boundary conditions
and meshing the model. This is followed by natural fre-
quency and static structural analysis of the model as
per the loading specified in the reference and validating
the baseline model. In the third step, the available
monthly Weibull wind speed parameters at the selected
site are fitted with a stochastic model. Weibull prob-
ability distribution method is commonly used for wind
speed prediction on account of its accuracy as noted by
various authors.26,27 A forecast of the long-term wind
speeds is then carried out using stochastic wind-
generation models, and the suitable model is selected as
a basis for long-term load calculation. Finally, a soft-
ware program based on BEM theory with consider-
ation for tip and hub losses is developed to obtain the
variable loads on the turbine cross sections for 1-year
period. The static stresses obtained from the load
amplitudes on cross sections along with the variable
wind load spectrum are then used to predict the cumu-
lative fatigue damage, life, and safety factor. The
above-mentioned BEM method is one of the most com-
monly used methods to achieve rapid simulations for
4 Advances in Mechanical Engineering
Figure 1. Blade analysis procedure.
standard sets of aerodynamic calculation in both aca-
demia and industry. Ke et al.24 developed a procedure
to find the aerodynamic loads and aeroelastic responses
of a 5-MW large wind turbine in yaw condition. They
employed harmonic superposition method and modi-
fied BEM method to calculate aerodynamic loads while
fully accounting for wind shear, tower shadow, tower-
blade modal and aerodynamic interactions, and rota-
tional effects. The authors also analyzed the yaw effect
and aeroelastic effect on aerodynamic loads and wind-
induced responses. In their earlier study, Ke et al.28
developed an approach for predicting the wind
models and wind effects of large wind turbines and
the vibration effects due to wind on the 5 MW tower-
blade-coupled wind turbine. They calculated the wind-
induced static loads and responses considering
resonance coupling of modes and background
responses. Their study found that the resonant compo-
nent plays a significant role on equivalent static wind
loads and responses than background component at
the middle-upper part of the tower and blades. The
cross terms between background and resonant compo-
nents affected the total fluctuation responses. Other
notable work in this area includes those of Burton
et al.29 and Hansen.30
Structural and aerodynamic analysis of the wind tur-
bine blades is studied by many researchers as the size of
wind blades is increasing due to growing power
demands. Tenguria et al.31 analyzed a composite mate-
rial HAWT blade subjected to flapwise loading. They
utilized ANSYS software and focused on root section
and transition section where the cross section changes
to NACA 634-221. The current authors of this article
also have performed structural analysis of a similar
blade as a first step and used it as a base for fatigue
analysis due to wind loads. Rafiee and Fakoor32 ana-
lyzed a 23-m HAWT composite blade having a mixed
airfoil geometry (NACA, FFA-W3, etc.) across its
length for modal, aeroelastic, and aerodynamic analy-
ses using FEM. The authors carried out structural
analysis at startup, standby, power production, and
shutdown events and optimized the geometry of the
blade as per observed deformations. One of their
important recommendations is to consider static
aeroelasticity for blade design for fatigue since this phe-
nomenon increases induced stresses in blade laminates.
Vardar and Eker33 analyzed six NACA profiles and
found a general trend that blade area and volume are
directly proportional to starting speed and inversely to
the rotor performance. Spagnoli and Montanari34 car-
ried out time-domain along-wind dynamic analysis of
upwind HAWT considering coupling effect of blades
and tower. The wind speed field is considered with
mean and fluctuating components. Their study brings
out the importance of analyzing blade and tower as a
coupled system and serves as a tool for preliminary
design of wind turbine–tower systems.
This article is organized as follows: section ‘‘Blade
modeling and structural analysis’’ presents the details
of the blade model, structural analysis procedure, and
comparison with the available reference solution.
Section ‘‘Wind field modeling and forecasting’’ presents
the wind field modeling and forecasting procedure.
Section ‘‘Blade fatigue analysis’’ extends the procedure
to fatigue analysis and presents the fatigue life, damage,
and safety factor results for the blade. Section ‘‘CZM
of blade root joint’’ discusses the CZM of the root joint
of the blade. Finally, the conclusions and future work
directions are provided.
Blade modeling and structural analysis
In this section, details of blade specifications, 3D mod-
eling, and static and natural frequency analyses of the
WF-1 blade are explained along with comparison to
available results.
Blade specifications and modeling
The basic geometry of the blade used in this study is
referred from the Technical Report by Perkins and
Cromack25 at University of Massachusetts Amherst.
The blade called as WF-1 (Wind Furnace-1) belongs to
a three-bladed prototype downwind wind turbine. The
reason for selecting this blade is because this is a small
HAWT blade which is the focus of this study.
Furthermore, the blade geometry and modal and static
analyses information are available for this blade which
Kulkarni et al. 5

Figure 2. A sample blade section at 0.495 m from root as defined by Perkins and Cromack.25

Table 1. WF-1 blade thickness across section from Perkins and Cromack.25

Blade section Skin thickness (mm) Spar thickness (mm) Spar web thickness (mm) Stiffener thickness (mm)

Section 1 at 0.495 m 0.9144 6.0452 6.0452 51.435

Section 2 at 0.99 m 0.9144 6.0452 6.0452 55.5625
Section 3 at 1.49 m 0.9144 5.207 5.588 47.9425
Section 4 at 1.98 m 0.9144 5.207 3.81 38.735
Section 5 at 2.48 m 0.9144 5.207 1.27 32.385
Section 6 at 2.97 m 0.6858 4.7244 1.27 27.94
Section 7 at 3.47 m 0.6858 3.7592 1.27 24.13
Section 8 at 3.96 m 0.4572 2.794 1.27 20.955
Section 9 at 4.46 m 0.4572 1.8288 1.27 17.145
Section 10 at 4.95 m 0.4572 1.3462 1.27 13.335

In order to maintain the aerodynamic configuration

Table 2. WF-1 blade specifications from Perkins and
Cromack.25
Blade length 16.25 ft (4.95 m)
Section profile NACA4415
Maximum chord length 1.35 ft (0.411 m)
Minimum chord length 0.35 ft (0.107 m)
Maximum blade twist 45° near root
is helpful to create a baseline 3D model, and the study
can be extended to carry out fatigue analysis due to
wind loads.
The blade section and planform are defined in
Figures 2 and 3. The blade geometry details are shown
in Tables 1 and 2, and material properties of blade
parts are shown in Table 3.
The blade is made up of four parts: skin, spar, spar
web, and stock. The skin composition is of fiberglass
epoxy matrix having low bending modulus. The spar
and the web are made of high-modulus fiberglass epoxy
matrix and the stock is surrounded by a steel sleeve.
The purpose of spar and spar web is to provide reinfor-
cements to the hollow interiors of the blade. A trailing-
edge stiffener is provided to prevent the top and bottom
faces from attaching to each other near the trailing edge
in case of heavy loads.
as per the specification, a high-fidelity 3D CAD model
of the blade airfoil section and the blade is created using
CATIA V5 3D modeling software. The airfoil sections
are drawn as per the chord lengths and span locations
defined in Figure 1 and the twists were applied to them
as specified. The airfoil sections were connected by
spline curves and a surface model is created. Figure 4
shows the airfoil sections and assembly of the blade
CAD model.
Blade structural analysis
FEM of the blade
The blade assembly is composed of four parts, which
are the handle, skin, stiffener, and the shear web. These
parts are assigned the thicknesses and material proper-
ties as defined in Table 2 and Figure 2, respectively.
The FE model of the blade was created using the
ANSYS software. The mesh was created using 71,212
elements and 141,385 nodes in the model. The root of
the blade handle was fixed for all degrees of freedom
(DOF) to create a fixed boundary condition for cantile-
ver to simulate the root joint with turbine hub. Figure
5(a) and (b) shows blade’s 3D CAD and meshed model,
respectively.
6 Advances in Mechanical Engineering

Table 3. WF-1 blade material characteristics from Perkins and Cromack.25

Component Material Moduli (GPa) Density (kg/m3)

Skin Low-modulus fiberglass epoxy Eskin = 15:168 2000

Gskin = 3:447
Spar, web, and stiffener High-modulus fiberglass epoxy Espar = 30:336 1500
Gspar = 2:068
Stock Fiberglass epoxy with steel sleeve E = 200 7850

Figure 3. Blade Planform as defined by Perkins and Cromack.25

Kulkarni et al. 7

Figure 4. Details of blade model: (a) NACA 4415 profile for cross section, (b) blade surface, (c) wireframe, and (d) blade spar,
web, and stiffener.

Figure 5. Blade model: (a) CAD model and (b) FE model.

Table 4. Blade natural frequencies.

Mode Mode Natural frequency From Perkins and Cromack25

Observed Measured Numerical prediction

1 Flapwise 31.57 rad/s (5.03 Hz) 25 rad/s (3.98 Hz) 28 rad/s (4.46 Hz)
2 Edgewise 61.01 rad/s (9.71 Hz) 56 rad/s (8.91 Hz) 65 rad/s (10.35 Hz)
3 Flapwise 93.43 rad/s (14.87 Hz) 84 rad/s (13.37 Hz) 93 rad/s (14.80 Hz)

Natural frequency analysis

In order to validate the correctness of the developed
simulation model of the blade, the natural frequencies
of the blade are calculated and compared with the ref-
erence solution by Perkins and Cromack.25 The block
Lanczos method is used to perform the modal analysis
as it is widely used. The vibration modes of first three
orders were extracted. Figure 6 shows the three modes
and Table 4 shows the observed natural frequencies
and comparison with the reference (experimental and
numerical predictions). As shown in Figure 6 and
Table 4, the most important top three modes of the
developed blade match well with the reference blade.
The difference between the observed natural fre-
quencies with the reference numerical prediction can be
attributed to uncertainty in the available geometry
details such as trailing-edge region, stiffener thickness,
and resin-steel sleeve at the handle. This uncertainty
8 Advances in Mechanical Engineering

Figure 6. Natural frequency and mode shapes: (a) mode I, (b) mode II, and (c) mode III.

Table 5. Blade deflections and maximum bending stress under applied reference load.

Current blade Reference blade25

Tip deflection (mm) U (vertical) 13.69 14.5

V (horizontal 72.3 76
Maximum bending stress (MPa) Top surface (compression) 23.73 23.88
Bottom surface (tensile) + 3.26 + 3.21

Figure 7. Deflections of blade free end: (a) vertical deflection and (b) horizontal deflection.

could have resulted in lesser flexibility in the model
especially for the second mode. However, since flapwise
bending is more critical for fatigue load,10 the variation
in edgewise mode can be accepted.
Static structural analysis
The blade is subjected to static loads to check whether
the deflection and bending stresses on top and bottom
surfaces are computed correctly as compared with ref-
erence solution by Perkins and Cromack.25 The load
applied is 68.95 N acting downward with blade in
horizontal position (15 lbs and 8 Oz as per reference) at
a location of 0.95R where R is the blade span radius.
The results for the deflection of the free end and the
maximum bending stresses on top and bottom surfaces
at a section located at 0.475R along with comparison
with reference is shown in Table 5. Figures 7 and 8
show these simulation results from ANSYS computer-
aided engineering (CAE) software.
The variation in the bending stress on the top (suc-
tion) and bottom (pressure) surfaces of the blade across
the chord length was found out at a specified location
of 0.475R in the reference. Figure 9 shows a good
Kulkarni et al.
Figure 8. Bending stresses on (a) top and (b) bottom surfaces at 0.475R.
Figure 9. Comparison of observed and reference Perkins and Cromack25 bending stresses on top and bottom surfaces of the
blade at a location of 0.475R.
agreement between the observed and the reference
distribution.
Thus, the baseline model of the blade is established
having structural response close to the reference. This
blade model is further analyzed for transient and fati-
gue analysis in next section.
Wind field modeling and forecasting
For the Salem location, the information available for
the mean wind speed and Weibull parameters for wind
distribution for each month of the year is referred from
the works of Sriram et al.35 The parameters are based
on the data taken every 4 h on Monday, Wednesday,
and Friday during 2001 and 2002 time period. The
Weibull distribution function is defined as
 
K V k1 ½V =C k
f ðV Þ = ࠟexpࠠ
C C
9
where K and C are shape and scale parameters, respec-
tively, and V is the wind speed. These parameters for
each month of the above-mentioned period are given in
Table 6.
As mentioned in section ‘‘Introduction,’’ the details
of the three stochastic forecasting models—first-order
Markov chain, ARIMA, and Kalman filter—and their
comparison for the wind prediction are discussed in the
following.
Markov chain
This is a discrete time model made of sequence of ran-
dom variables X1 , X2 , X3 , . . . such that the future state
Xn + 1 is not dependant on the past states Xn1 , Xn2 , . . .
but only on the present state Xn
Pr (Xn + 1 = xn + 1 jX1 = x1 , X2 = x2 , . . . , Xn = xn )
ð1Þ ð2Þ
= Pr (Xn + 1 = xn + 1 jXn = xn )
10 Advances in Mechanical Engineering

Table 6. Average monthly wind speed and Weibull parameters.

S. no. Month From Sriram et al.35

Average wind velocity (m/s) Shape parameter (K) Scale parameter (C)

1 April 2001 2.95 5.0467 3.2495

2 May 3.51 4.9083 3.7978
3 June 2.91 6.2790 3.2195
4 July 1.94 4.0466 2.6134
5 August 1.41 3.0215 1.5671
6 September 1.38 3.4941 1.5474
7 October 1.12 4.5830 1.2318
8 November 1.30 3.0281 1.4714
9 December 1.90 8.9593 2.0786
10 January 2002 2.25 5.8633 2.3088
11 February 2002 2.46 7.7020 2.6013
12 March 2002 1.98 4.6899 2.1418

2 3
The set of possible states of Xn S = fs1 , s2 , s3 , . . . , sk g 0:5 0:2 0:3
6 7
are called state space of the chain and the probability xð2Þ = ð0:15 0:45 0:4Þ4 0:15 0:45 0:4 5
of the chain to move from state si to sj , that is, pij is
0:35 0:35 0:3
called as transition probability from state i to j with
s1  s i , sj  sk = (0:2825 0:3725 0:385)

pij = Pr (X1 = sj jX0 = si ) ð3Þ If all the states in the matrix communicate, that is,
si ! sj and sj ! si (both states can be accessible from
The probability distribution of transitions from one each other with positive probability), then the matrix P
state to another can be represented into a transition is said to be irreducible.
matrix P = (pij )i, j . For K state, the first-order transition The availability of Weibull parameters and mean
probability matrix P has a size of K 3 K and has the wind speed for each month of the year as shown in
form Table 6 accounts for the seasonality of the wind. For
2 3 each month of the year, the wind speed vectors are gen-
p11 p12 ... p1k erated based on the mean wind speed (Table 6) and
6 p21 p22 ... p2k 7 maximum wind speed for Salem which falls in zone 4.36
P=6
4 ...
7
5
... ... These data are used to create a Weibull probability
pk1 pk2 ... pkk density for the vector values. The transition parobabil-
So, for a three-state chain, it can be represented as ity matrix for each month is then defined using the pro-
cedure for simulating non-Gaussian wind speed time
2 3 series.37 This matrix relates each hourly wind speed
p11 p12 p13
P = 4 p21 p22 p23 5 value with previous hourly average based upon (1)
p31 p32 p33 desired probability density as per monthly Weibull
parameters and (2) autocorrelation of wind speed.
or let us say with an example Hence, the transition matrix is modeled as a normal-
2 3 ized product of a diagonal probability matrix for prob-
0:5 0:2 0:3 ability density and a suitable symmetric decay matrix
P = 4 0:15 0:45 0:4 5 for autocorrelation.
0:25 0:35 0:4

This means that p12 = 0.2 is the probability of system

Validation of Markov approach with a
to go from state 1 to state 2, that is, X1 = s2 if it is cur-
rently at X0 = s1 . Similarly, the probability distribution reference wind measurement
of states after first step and second step if it currently in The above method is first validated against a measured
state X0 = s1 is wind data,38,39 and the comparison for the predicted
2 3 and the measured wind is shown in Table 7 in terms of
0:5 0:2 0:3 root mean square error (RMSE), MAE error metrics,
xð1Þ = ð010Þ4 0:15 0:45 0:4 5 = ð0:15 0:45 0:4Þ and the Weibull parameters. The measured wind data
0:35 0:35 0:3 considered here are for the Iowa, USA location every
Kulkarni et al. 11

Table 7. RMSE, MAE, and Weibull parameter comparison of Markov prediction with measured wind.

RMSE MAE Measured wind Predicted wind

Weibull shape (K) Weibull scale (C) Weibull shape (K) Weibull scale (C)

March 2008 2.0076 0.61 9.141 2.229 9.194 2.278

July 2008 0.9949 0.50 8.03 2.062 8.107 2.083
November 2008 1.1658 0.79 10.489 2.279 10.419 2.24

Figure 10. Comparison of observed and predicted wind for the month of March 2008.

10 min from 1 January 2008 till 31 December 2008. defined in Table 6. The approach of including seasonal-
There are 52,704 observations and the wind speed range ity for long-term wind prediction where the actual mea-
is from 0.35 to 19.22 m/s. The RMSE is defined as surements are limited is also used by researchers like
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Karatepe and Corscadden41 with good results. The
u wind prediction using Markov and Kalman methods
u 1 X N ðt Þ
 2
RMSEðVt Þ = t Vmeasured, i  Vpredicted, i ð4Þ for months having maximum and minimum of mean
N ðt Þ i = 1 wind speed is shown in Figure 11. Figure 12 shows the
predicted wind speed for Salem for 1 year and its PDF
The benefit of using RMSE as metric for comparison is calculated using MATLAB.
is that it brings in more weightage on larger error terms The transition probability matrix from Markov
due to the square of the difference term. approach is shown in Tables 8 and 9 for November
Another commonly used error metric called MAE is and April months, respectively, which have minimum
calculated. MAE is defined as and maximum mean speeds, respectively. This matrix is
based on 12 bins for wind speed ranging from 0 to
N ðt Þ
1 X   12 m/s with 1 m/s interval.
MAEðVt Þ = Vmeasured, i  Vpredicted, i  ð5Þ
N ðt Þ i = 1

where Vmeasured, i and Vpredicted, i are the measured and
predicted wind speed at ith time instant and N(t) is the
total number of observations (Figure 10). These RMSE
and MAE error ranges match with other published
results such as Zastrau et al.40
Application of Markov method for Salem
location
The Markov method which is tested with reference
wind is now applied for the Salem city in India having
the Weibull parameters and mean wind speed as
ARIMA model
ARIMA is a type of statistical method for time series–
based forecasting which combines the ‘‘autoregression’’
(AR), that is, modeling the linear dependence of the
output on its past ‘‘p’’ values, and ‘‘moving average’’
(MA), that is, the linear dependence of the error/resi-
dual term on the current and past ‘‘q’’ values of stochas-
tic error terms. ARIMA is a special case of the ARMA
model where the non-stationary time series is differen-
tiated ‘‘d’’ times to make it stationary. Thus, the general
equation for ARIMA (p,d,q) model for a variable Y is
constructed as follows.
12 Advances in Mechanical Engineering

Figure 11. Hourly wind prediction based on Markov chain and Kalman filter for Salem city for (a) May and (b) October months of
the year which have maximum and minimum mean wind speed, respectively.

Figure 12. (a) Wind speed for 1 year generated by Markov chain and (b) Weibull probability distribution.

Table 8. Markov probability transition matrix for November having mean wind speed of 1.3 m/s.

1 2 3 4 5 6 7 8 9 10 11 12

1 0.758 0.200 0.036 0.005 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000
2 0.080 0.762 0.135 0.020 0.003 0.000 0.000 0.000 0.000 0.000 0.000 0.000
3 0.012 0.116 0.744 0.111 0.015 0.002 0.000 0.000 0.000 0.000 0.000 0.000
4 0.002 0.021 0.136 0.732 0.096 0.011 0.001 0.000 0.000 0.000 0.000 0.000
5 0.000 0.004 0.029 0.153 0.720 0.084 0.009 0.001 0.000 0.000 0.000 0.000
6 0.000 0.001 0.007 0.036 0.169 0.706 0.074 0.007 0.000 0.000 0.000 0.000
7 0.000 0.000 0.002 0.009 0.044 0.184 0.690 0.065 0.000 0.000 0.000 0.000
8 0.000 0.000 0.001 0.003 0.013 0.053 0.199 0.671 0.057 0.004 0.000 0.000
9 0.000 0.000 0.000 0.001 0.004 0.017 0.063 0.214 0.653 0.046 0.002 0.000
10 0.000 0.000 0.000 0.000 0.002 0.006 0.024 0.080 0.244 0.616 0.028 0.001
11 0.000 0.000 0.000 0.000 0.001 0.003 0.012 0.040 0.121 0.307 0.501 0.015
12 0.000 0.000 0.000 0.000 0.000 0.002 0.008 0.026 0.081 0.204 0.332 0.347

Bold value shows the probability of wind speed staying in same bin/range.

If Y is the dth differential of Y such that yt = ðYt  Yt1 Þ  ðYt1  Yt2 Þ

ð6cÞ
= ðYt  Yt1 + Yt2 Þ for d = 2
yt = Y t when d = 0 ð6aÞ
yt = u1 yt1 +    + up ytp + 2t
yt = Yt  Yt1 for d = 1 ð6bÞ ð7aÞ
+ u1t1 +    + uq tq
Kulkarni et al. 13

Table 9. Markov probability transition matrix for April having mean wind speed of 2.95 m/s.

1 2 3 4 5 6 7 8 9 10 11 12

1 0.714 0.227 0.048 0.009 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000
2 0.091 0.723 0.154 0.027 0.004 0.001 0.000 0.000 0.000 0.000 0.000 0.000
3 0.017 0.132 0.703 0.126 0.020 0.003 0.000 0.000 0.000 0.000 0.000 0.000
4 0.004 0.029 0.154 0.688 0.108 0.015 0.002 0.000 0.000 0.000 0.000 0.000
5 0.001 0.007 0.039 0.172 0.674 0.094 0.012 0.001 0.000 0.000 0.000 0.000
6 0.000 0.002 0.011 0.048 0.189 0.657 0.082 0.009 0.001 0.000 0.000 0.000
7 0.000 0.001 0.003 0.015 0.059 0.205 0.638 0.072 0.007 0.000 0.000 0.000
8 0.000 0.000 0.001 0.005 0.020 0.070 0.219 0.615 0.062 0.005 0.000 0.000
9 0.000 0.000 0.000 0.002 0.008 0.027 0.083 0.234 0.593 0.050 0.003 0.000
10 0.000 0.000 0.000 0.001 0.003 0.012 0.037 0.103 0.262 0.550 0.030 0.001
11 0.000 0.000 0.000 0.000 0.002 0.007 0.021 0.059 0.150 0.315 0.430 0.015
12 0.000 0.000 0.000 0.000 0.001 0.005 0.016 0.044 0.110 0.231 0.315 0.277

Bold value shows the probability of wind speed staying in same bin/range.

Table 10. ARIMA models’ comparison.

ARIMA (2,1,2) ARIMA (1,1,2) ARIMA (2,1,1) ARIMA (3,1,3) ARIMA (1,1,1)

AIC –4232.32 –4223.21 –4223.442 –4228.105 –4219.627

BIC –4192.12 –4194.96 –4195.185 –4198.778 –4203.129

X
p X
q
y t = 2t + ui yti + uiti 2t1 ð7bÞ Table 11. ARIMA (2,1,2) parameter values.
i=1 i=1
Parameter Value Standard
where ui and ui are the autoregressive and MA para-
error
meters, respectively, to be determined. 2t are the error
terms (white noise) with normal distribution and zero Constant 0.000193222 0.22
mean. AR(1), AR(2) {–0.631623 0.159847} 0.12
The process to generate the forecast data is to first at Lags [1,2]
MA(1), MA(2) {–0.151268 –0.780751} 0.15
estimate p, q, and d by differencing the data and itera- at Lags [1,2]
tively checking autocorrelation function (ACF) plots Variance 0.379122 0.003
for q and partial autocorrelation function (PACF) plots
for p till stationary trend is obtained. Thus, the value of
p is determined as the last lag after which the PAC cuts compared to some of the other trials. The AIC and
off and the value of q is determined from the last lag BIC criteria are a standard criteria for ARIMA-based
after which ACF cuts off. This identification phase is model selection.42 The parameters of the ARIMA
followed by finding ui and ui parameters using maxi- (2,1,2) model are given in Table 11 showing the good-
mum likelihood estimation (MLE) techniques or back- ness of the fit. The ARIMA prediction 1 month ahead
casting methods which is the estimation phase. Once the wind speed is shown in Figure 13, and Figure 14 shows
model is fit, diagnostic checks are done on autocorrela- the ACF and PACF plots of the training input.
tion plots of the residuals to see whether any large auto-
correlation exist. The values of p and q are adjusted if
needed. This is the final diagnostic check phase. Kalman filter
For wind prediction using ARIMA model, the wind In this model, the unknown state vector Xt is defined
generated using Markov Chain is used as training data. over a range of time steps t as follows
A software program is developed in MATLAB to find
the autocorrelation and partial autocorrelation plots of Xt + 1=t = At=t 3 Xt=t + wt ð8Þ
the wind. Using the MATLAB functions ‘‘estimate,’’
the AR, MA, and integration parameters are found to And the correlation between measured vector and
be of ARIMA (2,1,2) series which has lowest Akaike unknown state vector is defined as
information criterion (AIC) and Bayesian information
criterion (BIC) values as seen from Table 10 when Zt + 1=t = Ct=t 3 t + 1=t + vt ð9Þ
14 Advances in Mechanical Engineering

Figure 13. Markov chain wind pattern and 1 month ahead ARIMA prediction.

Figure 14. Autocorrelation and partial autocorrelation plots for the Markov wind data.

where At , Ct are the matrices in the state-space equation; vector of Fourier coefficients Cn of length N of N/2
Xt=t , Xt + 1=t are the state of the system; Zt + 1=t is the out- coefficients an and bn are generated such that
put prediction from Kalman filter; and wt , vt is the sys- 
tem and measurement noise, respectively. TU
(Cn )0...N 1 = ;
The wind spectrum based on Kalman filter proposed 2p
by Kalman43 is expressed as a + ib ; . . . ; aN + ibN ; + aN + 0; ð11Þ
1 1 2 1 2 1 2 1
! aN2 1  ibN2 1 ; . . . ; a1  ib1
52:5z
Su ðvÞ = u2 U
ð10Þ
ð1 + 33 3 2pnÞ5=3 and the standard deviation is distributed equally among
the real and imaginary parts, that is
where U is the mean wind speed, z is the height above 1 2
ground, u is the friction velocity, n =fz=U ½ f =( fs =N ) s2an = s2bn = s ð12Þ
2 xn
(1 : N=2), fs is the sampling frequency, and N is the
total number of points. Then, inverse Fourier transform is applied on the
The standard deviation (s2Xt ) for each Fourier coeffi- generated Fourier coefficients to get the series values in
cient Xt = (T =2p)Su (v). The Fourier coefficients for the time domain
above spectrum are generated using complex numbers
to represent time series in real numbers. For this, a fs
Xm = 2p ifft ð13Þ
N
Kulkarni et al. 15

Figure 15. Flowchart for calculating induction factors and blade forces and moments.

Thus, looking at the accuracy of the Markov method Aerodynamic wind load calculation
to predict wind speed based on monthly Weibull wind
The wind prediction obtained for 1 year using Markov
speed characteristics (section ‘‘Validation of Markov
chain method from previous section is used to calculate
approach with a reference wind measurement’’) and the
the aerodynamic forces on the blade section using the
higher alternating wind speed prediction compared to
BEM method. A software program is developed in
Kalman and almost constant mean speed prediction by
MATLAB which utilizes the varying wind speeds from
ARIMA model, Markov chain model is selected for
Markov prediction and the flowchart in Figure 15 for
wind speed forecast and blade fatigue load calculation.
calculating the hourly time-varying thrust and moment
16
Figure 16. Loads and moments on some blade sections for 1 month.
loads for the 10 blade sections for all the months of the
year. Blade rotational speed is taken as 12 r/min. Thus,
a duty cycle for 1 year is created. Figure 16 shows the
loads and moments on some sections (1, 2, 5, 6, 9 and
10) for 1 month duration.
Comparison of static load calculation by BEM
method with analytical method
The major loads on the wind turbine blade are due to
wind itself along with gravitational force in horizontal
position. The centrifugal forces are comparatively low
due to low rotational speed. The lift and drag forces on
the blade as given by Yahya44 are
1 2
FLift = rV ACL ð14Þ
2
1
FDrag = rV 2 ACD ð15Þ
2 structural loads matches the analytical solution with
where r is the density of air (1.225 kg/m3), V is the mean
wind velocity which is taken as 10 m/s, A is the swept
Advances in Mechanical Engineering
Table 12. Parameters for dynamic loading.
R, L (blade span) 4.953 m
v (angular speed) 1.257 rad/s
m (blade mass) 24 kg
area of blade, and CL CD are the lift and drag coeffi-
cients (0.86 and 0.0126, respectively) for the airfoil con-
sidering angle of attack at 4.1° and a Reynolds number
of 1 million as referred from Hoffman et al.45 which is
shown in Figure 17.
The gravity force Fg = mg, where m is the blade
mass and the Fr = mRv2 , where R is the blade span
radius and v is the angular speed of rotor. Table 12
shows the important parameters for dynamic loads on
the current blade and Figure 18 shows the lift and drag
forces on blade.
Thus, the BEM-based program calculation for static
acceptable tolerance as seen from Table 13. The resul-
tant thrust loads closely match. Many authors such as
Kulkarni et al.
Figure 17. Lift and drag coefficients for NACA441545 at Re = 1E6.
Figure 18. Aerodynamic forces on blade.46
D Salimi-Majd et al.47 have used this approach to find
approximate blade loads.
Blade fatigue analysis
The time-varying wind loads and moments on blade
calculated in section ‘‘Aerodynamic wind load
17
calculation’’ are used as basis for the fatigue analysis.
The load cycle is made of 1 year of loads with data
point for each hour of a month for the selected site of
Salem. Thus, there are 8760 load points in the cycle. In
order to create the time-varying signal, first static anal-
ysis is carried out by taking the peak values of load and
moment. Then, a non-dimensional signal is created by
18 Advances in Mechanical Engineering

Figure 19. (a) Loading, (b) equivalent von Mises stress, and (c) maximum principal stress for static analysis.

are generated from the constant life diagrams (CLDs)

Table 13. Static load comparison of BEM and analytical
approaches.
By BEM-based By analytical
program (N) method44,46 (N)
Lift force 1784.66 1841
Drag force 314.68 41
Total thrust 1812.2 1841.5
BEM: blade element momentum.
dividing actual loads by the peak loads used in the sta-
tic analysis. The equivalent von Mises stress obtained
from static analysis is used as a basis for fatigue analy-
sis. Figure 19(a) shows the applied loads in static analy-
sis and Figure 19(b) shows the Von Mises stress
distribution.
Since the reference material by Perkins and
Cromack25 does not contain enough information for
fatigue analysis, equivalent materials for skin, spar,
web, and stiffener are chosen as shown in Table 14 for
carrying out static analysis as first step. The material
properties for these materials are obtained from SNL/
MSU/DOE Composite Material Fatigue Database.48
The S-N curves for skin and spar, web, and stiffener
materials are shown in Figures 20 and 21, respectively,
which are used for fatigue analysis. These S-N curves
referred from Mandell et al.49
After performing the static analysis, the time-varying
load signal is used to calculate the fatigue life, damage,
and safety factor for the blade. The stress life approach
is used with Goodman’s criteria for mean stress correc-
tion and von Mises stresses are used as stress compo-
nent. It is seen that the blade has the available life of
around 28 years and corresponding damage of 3.58 for
the load cycle shown in Figure 16. The load scale factor
is then varied to see the effect on fatigue. Figure 22
shows the variation of available life (years), accumu-
lated damage, and safety factor at the fatigue critical
location (FCL) with the load scaling from 0.25 to 2.5.
Figure 23(a)–(d) shows the available life, factor of
safety, damage, and the rainflow cycle plots, respec-
tively, at the FCL which is on the skin near the root of
the blade. These plots are for a 1-year load cycle made
of 8760-h signal. The life units are in hours and stress
ratio R is 21. The design life is taken as 40 years (350
and 400 h).
Effect of stress ratio
The importance of studying the effect of stress ratio R
on the fatigue life of composite materials was high-
lighted by authors such as Huh et al.50 and Roundi
et al.51 Most of the studies have focused either on

Table 14. Properties of equivalent materials for static analysis.

Component Equivalent material Moduli (GPa) Poisson ratio Density (kg/m3)

for fatigue

Skin Epoxy S-Glass UD E11 = 50 E22 , E33 = 8 g12 , g 13 = 0:3 g23 = 0:4 2000
G11 , G22 = 5 G33 = 3:9
Spar, web, Epoxy carbon E11 = 121 E22 , E33 = 8:6 g12 , g 13 = 0:27 g23 = 0:4 1500
and stiffener UD 230 GPa G11 , G22 = 4:7 G33 = 3:1
Stock Steel E = 200, G = 77 g12 = 0:3 7850
Kulkarni et al.
Figure 20. S-N curve for skin material at stress ratio R = –1
(Epoxy S-Glass UD).
Figure 21. S-N curve for spar, web and stiffener material at
stress ratio R = –1 (Epoxy Carbon UD 230 GPa).
Figure 22. Variation in available life, damage, and safety factor
with load scaling at stress ratio R = –1 and S-Glass as skin
material.
tension–tension or compression-only loadings.
Gambone,52 however, has analyzed a wide range of
stress ratios 21  R  0.5 and its effect on the mode-
II delamination of carbon/epoxy composites in terms
of strain energy release rate range (DGII ), maximum
strain energy release rate (GIImax ), and stress intensity
factor range (DKII ). He concluded that increase in R
ratio leads to decrease in crack growth rate (da/dN)
over a given DGII as well as when evaluated as a func-
tion of GIImax .
19
Table 15. Fatigue results considering two static stress criteria
at a stress ratio of 21.
Criteria Life (years) Safety factor Damage
Von Mises stress 27.9 0.76 3.58
Maximum 19.7 0.61 4.7
principal stress
Hence, present authors have carried out fatigue anal-
ysis of the blade for stress ratios of 21, 20.5, 0.1, and
0.5, and the results are shown in Figure 24. It shows the
effect of stress ratio on the fatigue parameters. It is seen
that the stress ratio for fully reversed condition of 21 is
the least favorable while with R = 0.5 is the most favor-
able. These findings are in line with the similar studies
mentioned above and can be explained by the fact that
stress ratio of 21 means a high stress amplitude as com-
pared to positive ratio.
The von Mises stress variation at the node where
maximum static stress is found is calculated from the
load signal and the static analysis results. A software
program in MATLAB is created to extract the rainflow
cycles from the variable nodal stresses. The rainflow
matrix in Figure 23(d) shows that the smaller stress
amplitudes up to 4 MPa are more in number and con-
tribute more to the fatigue failure than the large stress
amplitudes. Note that due to some modeling error far
away from root zone, there are some areas where high
damage and low life are shown by the ANSYS tool
which can be ignored as it is not in the FCL zone near
the blade root.
Effect of stress criteria
The fatigue analysis is also carried out by considering
the maximum principal stress criteria. Table 15 shows
the difference in the fatigue life and safety factor for
maximum principal stress and von Mises stress criteria
and Figure 25 shows these plots. It is seen that the max-
imum principal stress accounts for lower fatigue life
and safety factor than von Mises criteria due to higher
principal stresses as seen in other studies.53
Analysis with different skin materials
The fatigue life, damage, and factor of safety were also
calculated by doing analysis with glass fiber–reinforced
epoxy laminate (QQ1) and carbon/glass-hybrid-fiber-
reinforced epoxy (P2B) for skin material. The material
properties are shown in Table 16 which are first used to
do static analysis.
QQ1 is the glass-fiber-reinforced epoxy laminate
made up of Vantico TDT 177-155 Epoxy Resin,
Saertex U14EU920-00940-T1300-100000 0’s, and VU-
20
Figure 23. (a) Available life in FCL zone (2.443E05 h = 27.9 Years), (b) safety factor in FCL zone, (c) damage in FCL zone, and (d)
rainflow amplitudes and mean values for stress ratio R = –1 and von Mises criteria.
Figure 24. Variation in available life, damage, and safety factor
with stress ratio (R) and S-Glass as skin material.
90079-00830-01270-000000 45’s fabrics. P2B is a car-
bon/glass-hybrid-fiber-reinforced epoxy laminate made
up of Newport carbon NB307-D1-34-600 G300 prepreg
at 0° and glass NB307-D1-7781-497A at 6 45°. The
S-N curves for QQ1 and P2B were calculated using the
CLD referred from Mandell et al.49 and shown in
Figure 26. The S-N curves are shown in Figure 27(b)
and (c). These properties are used for the fatigue analy-
sis. The fatigue results for various load scales are
shown in Figure 28 for all the three skin materials.
Advances in Mechanical Engineering
CZM of blade root joint
The root joint of the current blade consists of epoxy/
resin composite core enclosed inside the steel sleeve.
This adhesively bonded root joint is a critical area
which can lead to delamination. The failure of the wind
turbine root joint due to delamination has been studied
by many researchers using CZM. The benefit of this
method is the possibility of modeling debonding with-
out requirement of crack and thus can be used to pre-
dict damage initiation in a healthy material.
Kalkhoran et al.55 analyzed the debonding initiation
and progression at the adhesive root joint of 23-m-long
blade of a 660-kW HAWT using CZM approach. They
implemented a cohesive zone constituent law based on
dissipated energy and used double cantilever beam
(DCB), end notch flexure (ENF), and mix mode bend-
ing (MMF) tests under static and fatigue loading for
developing root joint material properties. The static
analysis was effectively used to find the critical blade
position for doing fatigue analysis.
Salimi-Majd et al.47 analyzed the loading conditions
at the root joint of a large wind turbine blade through-
out a full rotation of the blade for the CZM-based fati-
gue analysis. They concluded that bending loads are
Kulkarni et al. 21

Figure 25. (a) Available life in FCL zone (1.72E5 h = 19.7 years) and (b) safety factor in FCL zone for stress ratio R = –1 and
maximum principal stress criteria.

Table 16. Material properties for various skin materials.

Property Epoxy S-Glass UD54 QQ1 (E-Glass UD)49 P2B (Carbon/Glass-hybrid Epoxy)49

E11 56 GPa 33.1 GPa 101 GPa

g12 0.3 0.27
g13 , g22 0.4 0.44
E22, E33 11.4 GPa 11.93 GPa 13.7 GPa
G12 4.48 GPa 4.33 GPa 4.33 GPa
Ultimate strength (tensile) 1.98 GPa 0.843 GPa 1.546 GPa
Ultimate strength (compression) 0.626 GPa 0.687 GPa 1.047 GPa
Tensile yield strength 0.2 GPa 0.149 GPa 0.08 GPa
Compressive yield strength 0.4 GPa 0.274 GPa 0.24 GPa
Density 1800 kg/m3 1919 kg/m3 1570 kg/m3

delamination index is a measure of relative strains at

Figure 26. Constant life diagrams (CLDs) for carbon and glass
fibers for different stress ratios.49
more effective than the axial forces for failure of adhe-
sively bonded root joint with blade horizontal position
being critical due to high delamination index. This
position and load ratio are taken as inputs for fatigue
analysis. The load ratio R was defined as the ratio of
minimum to maximum delamination index, while
the beginning of delamination with those of the cohe-
sive interface element.
Hosseini-Toudeshky et al.56 analyzed the progressive
fatigue debonding of the adhesively bonded root joint
of a 660-kW wind turbine blade under cyclic loading
using CZM modeling. They could verify the results of
the numerical analysis for mode I, mode II, and mixed-
mode crack initiation and propagation with experimen-
tal tests. They noted that the maximum length of the
deboned area after a typical overloading can go up to
40% of the height of the root joint in 1 month of opera-
tion under fatigue loading and thus highlighted the sig-
nificance of debonding analysis. Khoramishad et al.57
developed a numerical damage model for adhesively
bonded joints under fatigue loading that is only depen-
dent on the adhesive system and not on joint configura-
tion. They integrated a strain-based fatigue damage
model with CZM model to simulate the damaging
influence of the fatigue loading on the bonded joints.
Bak et al.58 developed a new method to simulate
fatigue-driven delamination based on CZM which can
predict both quasi-static and fatigue crack propagation.
Their method directly links crack growth and damage
22 Advances in Mechanical Engineering

Figure 27. S-N curve for skin materials: (b) QQ1 and (c) P2B for stress ratio R = –1.

Figure 28. (a) Variation in fatigue life and (b) safety factor for different skin materials for stress ratio R = –1.

evolution of cohesive zone to Paris law. This makes Table 17. Cohesive zone properties.
their method distinct with other available methods
which have structural dependency on strain energy Property Symbol Value
release rate. They also found good numerical conver- Critical strain energy GIc 0.26 kJ/m2
gence even at coarse meshes which is a significant bene- release rate, mode I
fit for computational efficiency to analyze complex Critical strain energy GIIC 1.002 kJ/m2
turbine blades. release rate, mode II
Considering the significance of the CZM-based Penalty stiffness K 10E6 N/mm3
Interface strength in mode I and II TI0 , TII0 30 MPa
damage analysis as seen from the above literature, the
Benzeggagh–Kenane law parameter h 2.73
authors of this study carried out the CZM of the root
joint with fatigue analysis to study its behavior under
cyclic loads. The interface properties of the cohesive and 1.96E9 (62 years) at 10E6 cycles is higher than the
layer for mixed-mode loading are shown in Table 17. life of 17.9 years found earlier without CZM analysis
The static analysis is first carried out on the root joint for 3.5E05 cycles.
modeled as cohesive zone. The results for the maximum
principal stress and deflection are shown in Figure 29
which show the critical area in the interface zone. Conclusion and future work
Fatigue analysis is then carried out by applying fully
reversed load cycles for various load cycles. The results The objective of this is to examine the structural
for the life and damage at the critical location found response of a HAWT blade made of composite mate-
from fatigue analysis of the CZM model for various rial to random wind loads. A reference wind turbine
load cycles are shown in Figure 30. It is seen that the blade from Perkins and Cromack25 was modeled in 3D
damage at the interface area increases and available life CAD software and its natural frequency and static
decreases as the number of cycles are increased. structural analysis results were first compared with the
However, the life of 2.72E8 s (8.6 years) at 10E8 cycles available results to check the accuracy of the modeling
Kulkarni et al.
Figure 29. Maximum principal stress and deflection at the interlaminar region.
and static analysis. The modal frequencies and static
structural results based on BEM approach were found
to be very close to the reference, thus validating the
model. After this, wind pattern for Salem city in
Tamilnadu, India was calculated using three stochastic
statistical methods—ARIMA, Markov chain, and
Kalman filter. These models were developed based on
the mean wind speed and the Weibull distribution data
available in the reference for Salem city. Markov chain
approach was selected due to its accuracy, and wind
loads for 1 year were calculated for the predicted wind
pattern using BEM method. These loads were used to
perform fatigue analysis to calculate the fatigue life,
damage, and safety factor. The variable stress vector
due to wind loads was used to calculate rainflow matrix
of stress amplitudes.
The authors conclude following:
1. This article presents an integrated methodology
consisting of blade modeling, comparisons of
long-term wind forecasting methods followed
by forecasting of long-term wind speeds, aero-
dynamic load calculation and static, and fatigue
analyses of composite HAWT blades for a site
where limited or no raw wind data available for
the site. It also analyzes conventional fatigue
and CZM-based approach for root joint of
blade. This end-to-end methodology is a unique
contribution from authors in HAWT blade
analysis.
2. It is found that the Markov chain model is an
effective method to predict long-term wind
speeds as compared to the ARIMA model
which is more suitable for the short-term wind
prediction. When the raw wind data are not
available, Markov chain approach performs
quite well in generating the synthetic wind as it
does not need historical wind data as training
data.
3. For fatigue analysis, the Markov chain–based
wind prediction is found to be more effective as
23
compared to ARIMA and Kalman spectrum
because of higher variation in wind speeds lead-
ing to higher alternating stresses which influence
fatigue. This effectiveness of Markov chain is
found to be more in the months of higher mean
speed.
4. The fatigue analysis results show that the critical
area for blade failure is near the joint of the han-
dle and the blade body. This is in line with vari-
ous available studies on blade fatigue analysis.
5. CZM of the blade root is an effective method to
analyze fatigue-induced debonding without the
need to have initial crack.57 However, in the cur-
rent model, this method provided less conserva-
tive results, that is, higher available life. This is
highly dependent on the properties of the inter-
face elements and needs further study.
6. The smaller stress amplitudes contribute to the
fatigue damage accumulation more than the
high stress amplitudes.
7. The Epoxy S-Glass UD material is found to
have better fatigue performance than E-Glass
UD (QQ1) and carbon/glass-hybrid-fiber-rein-
forced epoxy (P2B) materials as it gives the
highest fatigue life, better safety factor, and the
lowest damage among the three materials for
blade skin. It is also seen that Epoxy S-Glass
UD material has least fatigue sensitivity for life
and safety factor when load scale is varied from
0.25 to 2 as compared to QQ1 and P2B. P2B
has the maximum sensitivity.
As an extension of this work, the authors propose to
extend the modeling and analysis methodology devel-
oped in this article to large wind turbine blade with
complex material layup in different parts of the blade
such as the NREL 5 MW blade.59 The authors plan to
consider multiple input parameters such as pressure,
humidity, and temperature and use newer methods such
as ANN to predict mean wind speeds for long term.
Authors would like to analyze effect of interface
24 Advances in Mechanical Engineering

Figure 30. Life (seconds) and damage at the root joint for 10E2, 10E6, and 10E8 repeating cycles.

properties on the life and damage evolution using CZM
approach. Finally, the authors would like to study the
effect of variation of blade geometry, blade weight, and
geospatial variation of wind loads on the fatigue life of
this large blade.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.
Kulkarni et al.
Funding
The author(s) received no financial support for the research,
authorship, and/or publication of this article.
References
1. Mo W, Li D, Wang X, et al. Aeroelastic coupling analy-
sis of the flexible blade of a wind turbine. Energy 2015;
89: 1001–1009.
2. Liu X, Lu C, Liang S, et al. Influence of the vibration of
large-scale wind turbine blade on the aerodynamic load.
Enrgy Proced 2015; 75: 873–879.
3. Liu X, Lu C, Liang S, et al. Vibration-induced aerody-
namic loads on large horizontal axis wind turbine blades.
Appl Energ 2017; 185: 1109–1119.
4. Kaneko T, Uehara A, Senjyu T, et al. An integrated
control method for a wind farm to reduce frequency devia-
tions in a small power system. Appl Energ 2011; 88:
1049–1058.
5. Smith JC, Ahlstrom ML, Zavadil RM, et al. The role of
wind forecasting in utility system operation. In: IEEE
power & energy society general meeting, 2009 (PES’09),
Calgary, AB, Canadam, 26–30 July 2009, pp.1–5. New
York: IEEE.
6. Yi JH, Yoon GL and Li Y. Numerical investigation on
effects of rotor control strategy and wind data on opti-
mal wind turbine blade shape. Wind Struct 2014; 18:
195–213.
7. Ghasemi AR, Jahanshir A and Tarighat MH. Numerical
and analytical study of aeroelastic characteristics of wind
turbine composite blades. Wind Struct 2014; 18: 103–116.
8. Lee JW, Kim JK, Han JH, et al. Active load control for
wind turbine blades using trailing edge flap. Wind Struct
2013; 16: 263–278.
9. Chen TC and Lee MH. Inverse active wind load inputs
estimation of the multilayer shearing stress structure.
Wind Struct 2008; 11: 19–33.
10. Grujicic M, Arakere G, Pandurangan B, et al. Multidisci-
plinary design optimization for glass-fiber epoxy-matrix
composite 5 MW horizontal-axis wind-turbine blades. J
Mater Eng Perform 2010; 19: 1116–1127.
11. Hu W, Choi KK, Zhupanska O, et al. A new fatigue
analysis procedure for composite wind turbine blades. In:
32nd ASME wind energy symposium, AIAA SciTech
Forum, National Harbor, MD, 13–17 January 2014.
Reston, VA: AIAA.
12. Lecheb S, Nour A, Chellil A, et al. Dynamic prediction
fatigue life of composite wind turbine blade. Steel Com-
pos Struct 2015; 18: 673–691.
13. An D, Choi JH, Kim NH, et al. Fatigue life prediction
based on Bayesian approach to incorporate field data
into probability model. Struct Eng Mech 2011; 37:
427–442.
14. Barbounis TG and Theocharis JB. A locally recurrent
fuzzy neural network with application to the wind speed
prediction using spatial correlation Neurocomputing 2007;
70: 1525–1542.
15. Flores P, Tapia A and Tapia G. Application of a control
algorithm for wind speed prediction and active power
generation. Renew Energ 2005; 30: 523–536.
25
16. Lei M, Luan S, Jiang C, et al. A review on the forecasting
of wind speed and generated power. Renew Sust Energ
Rev 2009; 13: 915–920.
17. Amezcua J, Munoz R and Probst O. Reconstruction of
gusty wind speed time series from autonomous data log-
ger records. Wind Struct 2011; 14: 337–357.
18. Box GEP and Jenkins JM. Time series analysis: forecast-
ing and control. San Francisco, CA: Holden-Day, 1976.
19. Jung J and Broadwater RP Current status and future
advances for wind speed and power forecasting. Renew
Sust Energ Rev 2014; 31: 762–777.
20. Cadenas E and Rivera W. Wind speed forecasting in the
South Coast of Oaxaca, México. Renew Energ 2007; 32:
2116–2128.
21. Sansui N, Azami Z and Kamaruzzaman S. An initial
review: stochastic application in wind speed forecasting.
In: Zaharim A, Panagopoulos T, Zhang Y, et al. (eds)
Recent advances in energy, environment and development,
Cambridge, MA, 30 January-1 February 2013, pp.82–86.
World Scientific and Engineering Academy and Society
(WSEAS) Press..
22. Negra NB, Holmstrøm O, Bak-Jensen B, et al. Model of
a synthetic wind speed time series generator. Wind Energy
2007; 11: 193–209.
23. Huang WF and Xu YL. Prediction of typhoon design
wind speed and profile over complex terrain. Struct Eng
Mech 2013; 45: 1–18.
24. Ke ST, Wang TG, Ge YJ, et al. Aerodynamic loads and
aeroelastic responses of large wind turbine tower-blade
coupled structure in yaw condition. Struct Eng Mech
2015; 56: 1021–1040.
25. Perkins FW and Cromack DE. Wind turbine blade stress
analysis and natural frequencies. Wind Energy Center
Reports, Paper-11, 1978, http://scholarworks.umass.edu/
cgi/viewcontent.cgi?article=1010&context=windenergy_report
26. Carta JA, Ramirez P and Velázquezc S. A review of wind
speed probability distributions used in wind energy anal-
ysis: case studies in the Canary Islands. Renew Sust Energ
Rev 2009; 13: 933–955.
27. Yılmaz V and Cxelik HE. A statistical approach to esti-
mate the wind speed distribution: the case of Gelibolu
region. Doğusx Üniversitesi Dergisi 2008; 9: 122–132.
28. Ke ST, Wang TG, Ge YJ, et al. Wind-induced responses
and equivalent static wind loads of tower-blade coupled
large wind turbine system. Struct Eng Mech 2014; 52:
485–505.
29. Burton T, Jenkins N, Sharpe D, et al. Wind energy hand-
book. 2nd ed. West Sussex: John Wiley & Sons, Ltd.,
2011.
30. Hansen MOL. Aerodynamics of wind turbines. 2nd ed.
London: Earthscan, 2008.
31. Tenguria N, Mittal ND and Ahmed S. Structural analysis
of horizontal axis wind turbine blade. Wind Struct 2013;
16: 241–248.
32. Rafiee R and Fakoor M. Aeroelastic investigation of a
composite wind turbine blade. Wind Struct 2013; 17:
671–680.
33. Vardar A and Eker B. Mathematical modelling of wind
turbine blades through volumetric view. Wind Struct
2006; 9: 493–503.
26
34. Spagnoli A and Montanari L. Along-wind simplified
analysis of wind turbines through a coupled blade-tower
model. Wind Struct 2013; 17: 589–608.
35. Sriram G, Krishnamohan N and Gopalasamy V. Weibull
parameters for wind speed distribution in Salem city. J
Ind Pollut Contr 2004; 20: 53–58.
36. Lakshmanan N, Gomathinayagam S, Harikrishna P, et
al. Basic wind speed map of India with long-term hourly
wind data. Curr Sci India 2009; 96: 911–922.
37. McNerney GM and Veers PS. A Markov method for simu-
lating non-Gaussian wind speed time series. Albuquerque,
NM: Sandia National Laboratories, 1985.
38. Hu W, Choi KK and Cho H. Reliability-based design
optimization of wind turbine blades for fatigue life under
dynamic wind load uncertainty. Struct Multidiscip O
2016; 54: 953–970.
39. Hu W, Choi KK, Zhupanska O, et al. Integrating vari-
able wind load, aerodynamic, and structural analyses
towards accurate fatigue life prediction in composite wind
turbine blades. Struct Multidiscip O 2016; 53: 375–394.
40. Zastrau D, Schlaak M, Bruns T, et al. Differences in wind
forecast accuracy in the German North and Baltic Seas.
Int J Environ Sci Dev 2014; 5: 575–580.
41. Karatepe S and Corscadden KW. Wind speed estimation:
incorporating seasonal data using Markov chain models.
ISRN Renew Energ 2013: 657437 (9 pp.).
42. Shamsnia SA, Shahidi N, Liaghat A, et al. Modeling of
weather parameters using stochastic methods (ARIMA
model) (Case Study: Abadeh Region, Iran). In: Interna-
tional conference on environment and industrial innovation
(IPCBEE), Kuala Lumpur, Malaysia, 17–19 June 2011,
vol. 12. Singapore: IACSIT Press.
43. Kalman RE. A new approach to linear filtering and pre-
diction problems. J Basic Eng: T ASME 1960; 1: 35–45.
44. Yahya SM. Turbines compressors and fans. New Delhi,
India: Tata McGraw-Hill Education, 2010.
45. Hoffman MJ, Ramsay RR and Gregorek GM. Effects of
grit roughness and pitch oscillations on the NACA 4415
airfoil. Golden, CO, USA: National Renewable Energy
Laboratory, 1996.
46. Rajadurai JS, Christopher T, Thanigaiyarasu G, et al.
Finite element analysis with an improved failure criterion
for composite wind turbine blades. Forsch Ingenieurwes
2008; 72: 193–207.
47. Salimi-Majd D, Azimzadeh V and Mohammadi B. Load-
ing analysis of composite wind turbine blade for fatigue
life prediction of adhesively bonded root joint. Appl Com-
pos Mater 2015; 22: 269–287.
Advances in Mechanical Engineering
48. Mandell JF and Samborsky DD. SNL/MSU/DOE com-
posite material fatigue database. Albuquerque, NM: San-
dia National Laboratories, 2014.
49. Mandell JF, Ashwill TD, Wilson TJ, et al. Analysis of
SNL/MSU/DOE fatigue database trends for wind tur-
bine blade materials. Sandia report, SAND2010-7052.
Albuquerque, NM: Sandia National Laboratories, 2010.
50. Huh YH, Lee JH, Kim DJ, et al. Effect of stress ratio on
fatigue life of GFRP composites for WT blade. J Mech
Sci Technol 2012; 26: 2117–2120.
51. Roundi W, El Mahi A, El Gharad A, et al. Experimental
and numerical investigation of the effects of stacking
sequence and stress ratio on fatigue damage of glass/
epoxy composites. Compos Part B: Eng 2017; 109:
64–71.
52. Gambone LR. The effect of R-ratio on the mode II fatigue
delamination growth of unidirectional carbon/epoxy com-
posites. Doctoral Dissertation, The University of British
Columbia, Vancouver, BC, Canada, 1991.
53. Rahman MM, Kadirgama K, Noor MM, et al. Fatigue
life prediction of lower suspension arm using strain-life
approach. Eur J Sci Res 2009; 30: 437–450.
54. Samborsky DD, Wilson TJ and Mandell JF. Comparison
of tensile fatigue resistance and constant life diagrams for
several potential wind turbine blade laminates. J Sol
Energy Eng 2009; 131: 011006.
55. Kalkhoran A, Salimi-Majd D and Mohammadi B. Fati-
gue life prediction for adhesively bonded root joint of
composite wind turbine blade using cohesive zone
approach. In: Attaf B (ed.) Recent advances in compo-
site materials for wind turbine blades, Hong Kong, 30
January–1 February 2013, pp.221–232. The World
Academic Publishing Co. Ltd.
56. Hosseini-Toudeshky H, Jahanmardi M and Goodarzi
MS. Progressive debonding analysis of composite blade
root joint of wind turbines under fatigue loading. Compos
Struct 2015; 120: 417–427.
57. Khoramishad H, Crocombe AD, Katnam KB, et al. Pre-
dicting fatigue damage in adhesively bonded joints using
a cohesive zone model. Int J Fatigue 2010; 32: 1146–1158.
58. Bak BL, Turon A, Lindgaard E, et al. A simulation
method for high-cycle fatigue-driven delamination using
a cohesive zone model. Int J Numer Meth Eng 2015; 106:
163–191.
59. Jonkman J, Butterfield S, Musial W, et al. Definition of a
5-MW reference wind turbine for offshore system develop-
ment. Technical report no. NREL/TP-500-38060.
Golden, CO: National Renewable Energy Laboratory,
2009.