Professional Documents
Culture Documents
HAWT Fatiue Paper PDF
HAWT Fatiue Paper PDF
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
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 optimization, and maintenance scheduling, An et al.13
fatigue and optimize the blade design for improving have used Bayesian technique to obtain a fatigue life
safety and reliability. With a focus on blade optimiza- prediction model for wind turbines. Their method
tion study which depends on wind, rotor control strat- makes use of available field data, noise, and bias in
egy and generator type, Yi et al.6 studied the effects of measurements on the distribution of fatigue life so that
rotor control strategies such as combinations of fixed difference between designer’s predicted life and field
and variable speed and pitch algorithms on optimal observations can be reduced. They used Markov chain
blade shapes and rotor performance using optimized Monte Carlo technique to obtain the posterior distribu-
blade designs. They also evaluated the effects of envi- tion. Significant amount of simulation and experiments
ronmental wind data and the optimization functions have been investigated on the life prediction of compo-
using the NREL HARP_Opt tool which uses WT_Perf site materials in literature; however, experimental vali-
code for performance evaluation based on blade ele- dations of fatigue failure of composite wind turbine
ment momentum (BEM). Some of the performance blades (made of composite materials) are often ham-
indices which the authors compared and are significant pered due to lack of enough structural health monitor-
for this study were thrust, torque, and roof-flap ing data measured under realistic random wind loading
moment forces. Ghasemi et al.7 studied the aeroelastic in the field. Hence, this article focuses on accurate wind
stability aspects of composite blades of a 660 kW wind modeling for structural and fatigue analyses of compo-
turbine blade under flutter, that is, coupling of bending site wind turbine blades.
and torsional vibrations. The blade section is made up There are various methods available for wind mod-
of NACA 632-415. Analytical techniques were used to eling that could be classified as (1) physical models, (2)
calculate flutter speed and safety factor, while finite ele- spatial correlation models, (3) artificial intelligence
ment model (FEM) was used for static and modal anal- models, (4) stochastic statistical models, and (5) hybrid
yses. In another study on blade optimization, Lee models.
et al.8 numerically analyzed effect of active load control Spatial correlation models take the spatial relation-
devices such as trailing-edge flaps for reducing fatigue ship of different sites’ wind speed into account for pre-
loading on wind turbine blades and found a 30%–50% diction. Spatial correlation models need site-specific
reduction in the standard deviation of the root bending local historic data for accurate prediction and are use-
moment. The authors of this study have also used the ful when the nonlinear relationship between these local
blade model with trailing-edge stiffeners for the same variables is required to be modeled. However, these
purpose. Chen and Lee9 have used an adaptive input models may not be appropriate to be used when little
estimation method which combines Kalman filter with- historical information is available.
out input term and adaptive weighted recursive least The use of artificial intelligence models such as artifi-
square estimator to predict the wind load on a multi- cial neural network (ANN) and fuzzy logic is increasing
layer shearing stress system. over the years. Barbounis and Theocharis14 developed
As pointed by Grujicic et al.,10 the main structural a locally feedback dynamic fuzzy neural network (LF-
performance requirements for wind turbines are (1) DFNN) for 15 min to 3 h ahead wind speed prediction
flapwise bending strength to withstand extreme wind with spatial correlation and having superior accuracy
loads and (2) flapwise bending stiffness to ensure a over other network models. Flores et al.15 supplied a
minimum clearance between blade tip and the turbine control algorithm based on ANN model using back-
tower. To meet these conditions, Hu et al.11 developed propagation method for wind speed prediction. Surveys
a fatigue analysis procedure including random wind on artificial intelligence models on wind speed predic-
field simulation, aerodynamic analysis, FE stress analy- tion has been carried out in Lei et al.,16 which con-
sis, and fatigue damage simulation. Probability density cludes that it is necessary to further study on artificial
functions (PDFs) for 10-min mean wind speed and 10- intelligence methods and improve their training algo-
min turbulence intensity were developed to simulate a rithm aiming at more accurate results of wind speed
1-year wind field and calculated the accumulated fati- prediction.
gue damage, which could facilitate reliability analysis Stochastic statistical models consist of a mathemati-
and reliability-based design optimization of composite cal model based on historical data, patterns, and wind
wind turbine blades considering wind load uncertainty. parameters. Amezcua et al.17 developed an algorithm
Lecheb et al.12 have analyzed fatigue life of a 25-m-long using generalized random Fourier series to generate a
wind turbine blade made of glass/epoxy composite. real-time wind speed time series from data logger
They first carried out modal and static analyses to find records containing the average, maximum, and mini-
out blade critical zone for stresses followed by crack mum values of the wind speed in a fixed interval, as
growth study, experimental fatigue life determination, well as the standard deviation. They also incorporated
and fast fracture. In order to help meet needs such as extreme values using the concept of constrained simula-
carrying out risk analysis, reliability-based design tion in post processing.
Kulkarni et al. 3
One typical group of stochastic statistical few are available for an integrated framework starting
approaches includes autoregressive (AR) models, inte- from forecasting of long-term wind to aerodynamic
grated models (I), and moving average (MA) models and structural analyses and to fatigue analysis of com-
which depend linearly on the previous data. The com- posite horizontal-axis wind turbine (HAWT) blades
bined methods such as autoregressive moving average with limited or no raw wind data available for the site.
(ARMA) described by Box and Jenkins18 and autore- Hence, this study focuses on developing an integrated
gressive integrated moving average (ARIMA) models. procedure incorporating long-term wind forecasting
A review of the literature by Jung and Broadwater19 methods into blade fatigue analysis.
suggests that statistical models can be utilized for The unique contribution of this study is that (1) it
short-, medium-, and long-term predictions quite well. compares the various statistical wind forecasting mod-
Cadenas and Rivera20 applied ARIMA and the ANN els to select the one which is best suitable for fatigue
methods and also compared the two techniques for analysis applications, (2) develops long-term wind pro-
wind speed prediction in Mexico. They found that both file prediction for sites which does not have enough his-
the models provided reasonable predictions, and seaso- torical wind measurement data to a good level of
nal ARIMA model has a better sensitivity to wind accuracy, (3) performs high-fidelity three-dimensional
speed adjustment. They utilized mean absolute percent- (3D) modeling and analysis for random wind loads for
age error (MAPE) as a forecast error for comparison. fatigue, and (4) extends the conventional fatigue analy-
Another commonly used stochastic statistical model sis by comparing the cohesive zone modeling (CZM)
for wind speed prediction is the Markov chain model.21 approach for root joint of blade for fatigue. Thus, this
Synthetic wind speed generation for a specific site using study presents an end-to-end framework for wind tur-
Markov transition matrix has been employed by vari- bine blade analysis.
ous researchers such as Negra et al.22 for short-term For accurate structural analysis, a high-fidelity blade
prediction. These models are easy and inexpensive to model is generated referring from the work by Perkins
implement as compared to spatial correlation models and Cromack.25 The structural analysis results are
which need large geospatial data of various sites as well compared with the published results so that the base-
as higher processing cost due to the multiple input vari- line model is available. The predicted long-term wind
ables representing physical properties of the sites such load is then applied to the fatigue analysis considering
as temperature, pressure, humidity, and altitude to pre- wind load variation at the selected location of Salem,
dict the wind speed. Thus, in this article, the wind speed Tamilnadu in India.
simulation has been carried out by three stochastic sta- The blade analysis procedure consists of four steps
tistical models such as ARIMA, Kalman filter, and as shown in Figure 1. First, a detailed computer-aided
first-order Markov chain, and the suitable model is design (CAD) model for a wind turbine blade is devel-
selected for wind speed prediction and related struc- oped in Catia V5R20 software based on the specifica-
tural fatigue analysis. tions. Pre-processing of the model is performed by
Hybrid models combine the benefits of individual applying material properties and boundary conditions
models and overcome limitations of using a single and meshing the model. This is followed by natural fre-
model. They could be formed by combining physical quency and static structural analysis of the model as
and statistical models or statistical and spatial models. per the loading specified in the reference and validating
Huang and Xu23 have proposed a numerical simula- the baseline model. In the third step, the available
tion procedure for predicting directional typhoon monthly Weibull wind speed parameters at the selected
design wind speeds and profiles for sites over complex site are fitted with a stochastic model. Weibull prob-
terrain by integrating typhoon wind field model, ability distribution method is commonly used for wind
Monte Carlo simulation technique, computational fluid speed prediction on account of its accuracy as noted by
dynamics (CFD) simulation, and ANNs. Ke et al.24 various authors.26,27 A forecast of the long-term wind
utilized another hybrid forecast model called wavelet speeds is then carried out using stochastic wind-
packet transform (WPT) for wind prediction based on generation models, and the suitable model is selected as
daily average wind speed data from four wind farms in a basis for long-term load calculation. Finally, a soft-
China. The hybrid model although is more time-con- ware program based on BEM theory with consider-
suming, and it has better accuracy than the individual ation for tip and hub losses is developed to obtain the
models as indicated by the error criteria such as mean variable loads on the turbine cross sections for 1-year
absolute error (MAE), mean square error (MSE), and period. The static stresses obtained from the load
MAPE. amplitudes on cross sections along with the variable
The literature survey on wind modeling and wind wind load spectrum are then used to predict the cumu-
turbine blade fatigue analysis shows that although lative fatigue damage, life, and safety factor. The
there are various studies available on wind prediction above-mentioned BEM method is one of the most com-
methods and blade structural and fatigue analyses, very monly used methods to achieve rapid simulations for
4 Advances in Mechanical Engineering
standard sets of aerodynamic calculation in both aca- aeroelasticity for blade design for fatigue since this phe-
demia and industry. Ke et al.24 developed a procedure nomenon increases induced stresses in blade laminates.
to find the aerodynamic loads and aeroelastic responses Vardar and Eker33 analyzed six NACA profiles and
of a 5-MW large wind turbine in yaw condition. They found a general trend that blade area and volume are
employed harmonic superposition method and modi- directly proportional to starting speed and inversely to
fied BEM method to calculate aerodynamic loads while the rotor performance. Spagnoli and Montanari34 car-
fully accounting for wind shear, tower shadow, tower- ried out time-domain along-wind dynamic analysis of
blade modal and aerodynamic interactions, and rota- upwind HAWT considering coupling effect of blades
tional effects. The authors also analyzed the yaw effect and tower. The wind speed field is considered with
and aeroelastic effect on aerodynamic loads and wind- mean and fluctuating components. Their study brings
induced responses. In their earlier study, Ke et al.28 out the importance of analyzing blade and tower as a
developed an approach for predicting the wind coupled system and serves as a tool for preliminary
models and wind effects of large wind turbines and design of wind turbine–tower systems.
the vibration effects due to wind on the 5 MW tower- This article is organized as follows: section ‘‘Blade
blade-coupled wind turbine. They calculated the wind- modeling and structural analysis’’ presents the details
induced static loads and responses considering of the blade model, structural analysis procedure, and
resonance coupling of modes and background comparison with the available reference solution.
responses. Their study found that the resonant compo- Section ‘‘Wind field modeling and forecasting’’ presents
nent plays a significant role on equivalent static wind the wind field modeling and forecasting procedure.
loads and responses than background component at Section ‘‘Blade fatigue analysis’’ extends the procedure
the middle-upper part of the tower and blades. The to fatigue analysis and presents the fatigue life, damage,
cross terms between background and resonant compo- and safety factor results for the blade. Section ‘‘CZM
nents affected the total fluctuation responses. Other of blade root joint’’ discusses the CZM of the root joint
notable work in this area includes those of Burton of the blade. Finally, the conclusions and future work
et al.29 and Hansen.30 directions are provided.
Structural and aerodynamic analysis of the wind tur-
bine blades is studied by many researchers as the size of
Blade modeling and structural analysis
wind blades is increasing due to growing power
demands. Tenguria et al.31 analyzed a composite mate- In this section, details of blade specifications, 3D mod-
rial HAWT blade subjected to flapwise loading. They eling, and static and natural frequency analyses of the
utilized ANSYS software and focused on root section WF-1 blade are explained along with comparison to
and transition section where the cross section changes available results.
to NACA 634-221. The current authors of this article
also have performed structural analysis of a similar
Blade specifications and modeling
blade as a first step and used it as a base for fatigue
analysis due to wind loads. Rafiee and Fakoor32 ana- The basic geometry of the blade used in this study is
lyzed a 23-m HAWT composite blade having a mixed referred from the Technical Report by Perkins and
airfoil geometry (NACA, FFA-W3, etc.) across its Cromack25 at University of Massachusetts Amherst.
length for modal, aeroelastic, and aerodynamic analy- The blade called as WF-1 (Wind Furnace-1) belongs to
ses using FEM. The authors carried out structural a three-bladed prototype downwind wind turbine. The
analysis at startup, standby, power production, and reason for selecting this blade is because this is a small
shutdown events and optimized the geometry of the HAWT blade which is the focus of this study.
blade as per observed deformations. One of their Furthermore, the blade geometry and modal and static
important recommendations is to consider 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)
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.
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)
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.
Figure 7. Deflections of blade free end: (a) vertical deflection and (b) horizontal deflection.
could have resulted in lesser flexibility in the model horizontal position (15 lbs and 8 Oz as per reference) at
especially for the second mode. However, since flapwise a location of 0.95R where R is the blade span radius.
bending is more critical for fatigue load,10 the variation The results for the deflection of the free end and the
in edgewise mode can be accepted. 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
Static structural analysis show these simulation results from ANSYS computer-
The blade is subjected to static loads to check whether aided engineering (CAE) software.
the deflection and bending stresses on top and bottom The variation in the bending stress on the top (suc-
surfaces are computed correctly as compared with ref- tion) and bottom (pressure) surfaces of the blade across
erence solution by Perkins and Cromack.25 The load the chord length was found out at a specified location
applied is 68.95 N acting downward with blade in of 0.475R in the reference. Figure 9 shows a good
Kulkarni et al. 9
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 where K and C are shape and scale parameters, respec-
distribution. tively, and V is the wind speed. These parameters for
Thus, the baseline model of the blade is established each month of the above-mentioned period are given in
having structural response close to the reference. This Table 6.
blade model is further analyzed for transient and fati- As mentioned in section ‘‘Introduction,’’ the details
gue analysis in next section. 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
Wind field modeling and forecasting following.
For the Salem location, the information available for
the mean wind speed and Weibull parameters for wind Markov chain
distribution for each month of the year is referred from
the works of Sriram et al.35 The parameters are based This is a discrete time model made of sequence of ran-
on the data taken every 4 h on Monday, Wednesday, dom variables X1 , X2 , X3 , . . . such that the future state
and Friday during 2001 and 2002 time period. The Xn + 1 is not dependant on the past states Xn1 , Xn2 , . . .
Weibull distribution function is defined as but only on the present state Xn
Pr (Xn + 1 = xn + 1 jX1 = x1 , X2 = x2 , . . . , Xn = xn )
K V k1 ½V =C k
f ðV Þ = ࠟexpࠠ ð1Þ ð2Þ
C C = Pr (Xn + 1 = xn + 1 jXn = xn )
10 Advances in Mechanical Engineering
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
Table 7. RMSE, MAE, and Weibull parameter comparison of Markov prediction with measured wind.
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 ARIMA model
predicted wind speed at ith time instant and N(t) is the ARIMA is a type of statistical method for time series–
total number of observations (Figure 10). These RMSE based forecasting which combines the ‘‘autoregression’’
and MAE error ranges match with other published (AR), that is, modeling the linear dependence of the
results such as Zastrau et al.40 output on its past ‘‘p’’ values, and ‘‘moving average’’
(MA), that is, the linear dependence of the error/resi-
Application of Markov method for Salem dual term on the current and past ‘‘q’’ values of stochas-
tic error terms. ARIMA is a special case of the ARMA
location
model where the non-stationary time series is differen-
The Markov method which is tested with reference tiated ‘‘d’’ times to make it stationary. Thus, the general
wind is now applied for the Salem city in India having equation for ARIMA (p,d,q) model for a variable Y is
the Weibull parameters and mean wind speed as 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.
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.
ARIMA (2,1,2) ARIMA (1,1,2) ARIMA (2,1,1) ARIMA (3,1,3) ARIMA (1,1,1)
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 Advances in Mechanical Engineering
Figure 16. Loads and moments on some blade sections for 1 month.
loads for the 10 blade sections for all the months of the Table 12. Parameters for dynamic loading.
year. Blade rotational speed is taken as 12 r/min. Thus,
a duty cycle for 1 year is created. Figure 16 shows the R, L (blade span) 4.953 m
v (angular speed) 1.257 rad/s
loads and moments on some sections (1, 2, 5, 6, 9 and
m (blade mass) 24 kg
10) for 1 month duration.
Comparison of static load calculation by BEM area of blade, and CL CD are the lift and drag coeffi-
method with analytical method cients (0.86 and 0.0126, respectively) for the airfoil con-
sidering angle of attack at 4.1° and a Reynolds number
The major loads on the wind turbine blade are due to
of 1 million as referred from Hoffman et al.45 which is
wind itself along with gravitational force in horizontal
shown in Figure 17.
position. The centrifugal forces are comparatively low
The gravity force Fg = mg, where m is the blade
due to low rotational speed. The lift and drag forces on
mass and the Fr = mRv2 , where R is the blade span
the blade as given by Yahya44 are
radius and v is the angular speed of rotor. Table 12
1 2 shows the important parameters for dynamic loads on
FLift = rV ACL ð14Þ the current blade and Figure 18 shows the lift and drag
2
forces on blade.
1
FDrag = rV 2 ACD ð15Þ Thus, the BEM-based program calculation for static
2 structural loads matches the analytical solution with
where r is the density of air (1.225 kg/m3), V is the mean acceptable tolerance as seen from Table 13. The resul-
wind velocity which is taken as 10 m/s, A is the swept tant thrust loads closely match. Many authors such as
Kulkarni et al. 17
D Salimi-Majd et al.47 have used this approach to find calculation’’ are used as basis for the fatigue analysis.
approximate blade loads. 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
Blade fatigue analysis
order to create the time-varying signal, first static anal-
The time-varying wind loads and moments on blade ysis is carried out by taking the peak values of load and
calculated in section ‘‘Aerodynamic wind load 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.
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. 19
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 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.
Property Epoxy S-Glass UD54 QQ1 (E-Glass UD)49 P2B (Carbon/Glass-hybrid Epoxy)49
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. 23
Figure 29. Maximum principal stress and deflection at the interlaminar region.
and static analysis. The modal frequencies and static compared to ARIMA and Kalman spectrum
structural results based on BEM approach were found because of higher variation in wind speeds lead-
to be very close to the reference, thus validating the ing to higher alternating stresses which influence
model. After this, wind pattern for Salem city in fatigue. This effectiveness of Markov chain is
Tamilnadu, India was calculated using three stochastic found to be more in the months of higher mean
statistical methods—ARIMA, Markov chain, and speed.
Kalman filter. These models were developed based on 4. The fatigue analysis results show that the critical
the mean wind speed and the Weibull distribution data area for blade failure is near the joint of the han-
available in the reference for Salem city. Markov chain dle and the blade body. This is in line with vari-
approach was selected due to its accuracy, and wind ous available studies on blade fatigue analysis.
loads for 1 year were calculated for the predicted wind 5. CZM of the blade root is an effective method to
pattern using BEM method. These loads were used to analyze fatigue-induced debonding without the
perform fatigue analysis to calculate the fatigue life, need to have initial crack.57 However, in the cur-
damage, and safety factor. The variable stress vector rent model, this method provided less conserva-
due to wind loads was used to calculate rainflow matrix tive results, that is, higher available life. This is
of stress amplitudes. highly dependent on the properties of the inter-
The authors conclude following: face elements and needs further study.
6. The smaller stress amplitudes contribute to the
fatigue damage accumulation more than the
1. This article presents an integrated methodology
high stress amplitudes.
consisting of blade modeling, comparisons of
7. The Epoxy S-Glass UD material is found to
long-term wind forecasting methods followed
have better fatigue performance than E-Glass
by forecasting of long-term wind speeds, aero-
UD (QQ1) and carbon/glass-hybrid-fiber-rein-
dynamic load calculation and static, and fatigue
forced epoxy (P2B) materials as it gives the
analyses of composite HAWT blades for a site
highest fatigue life, better safety factor, and the
where limited or no raw wind data available for
lowest damage among the three materials for
the site. It also analyzes conventional fatigue
blade skin. It is also seen that Epoxy S-Glass
and CZM-based approach for root joint of
UD material has least fatigue sensitivity for life
blade. This end-to-end methodology is a unique
and safety factor when load scale is varied from
contribution from authors in HAWT blade
0.25 to 2 as compared to QQ1 and P2B. P2B
analysis.
has the maximum sensitivity.
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 As an extension of this work, the authors propose to
which is more suitable for the short-term wind extend the modeling and analysis methodology devel-
prediction. When the raw wind data are not oped in this article to large wind turbine blade with
available, Markov chain approach performs complex material layup in different parts of the blade
quite well in generating the synthetic wind as it such as the NREL 5 MW blade.59 The authors plan to
does not need historical wind data as training consider multiple input parameters such as pressure,
data. humidity, and temperature and use newer methods such
3. For fatigue analysis, the Markov chain–based as ANN to predict mean wind speeds for long term.
wind prediction is found to be more effective as 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 Declaration of conflicting interests
approach. Finally, the authors would like to study the The author(s) declared no potential conflicts of interest with
effect of variation of blade geometry, blade weight, and respect to the research, authorship, and/or publication of this
geospatial variation of wind loads on the fatigue life of article.
this large blade.
Kulkarni et al. 25
34. Spagnoli A and Montanari L. Along-wind simplified 48. Mandell JF and Samborsky DD. SNL/MSU/DOE com-
analysis of wind turbines through a coupled blade-tower posite material fatigue database. Albuquerque, NM: San-
model. Wind Struct 2013; 17: 589–608. dia National Laboratories, 2014.
35. Sriram G, Krishnamohan N and Gopalasamy V. Weibull 49. Mandell JF, Ashwill TD, Wilson TJ, et al. Analysis of
parameters for wind speed distribution in Salem city. J SNL/MSU/DOE fatigue database trends for wind tur-
Ind Pollut Contr 2004; 20: 53–58. bine blade materials. Sandia report, SAND2010-7052.
36. Lakshmanan N, Gomathinayagam S, Harikrishna P, et Albuquerque, NM: Sandia National Laboratories, 2010.
al. Basic wind speed map of India with long-term hourly 50. Huh YH, Lee JH, Kim DJ, et al. Effect of stress ratio on
wind data. Curr Sci India 2009; 96: 911–922. fatigue life of GFRP composites for WT blade. J Mech
37. McNerney GM and Veers PS. A Markov method for simu- Sci Technol 2012; 26: 2117–2120.
lating non-Gaussian wind speed time series. Albuquerque, 51. Roundi W, El Mahi A, El Gharad A, et al. Experimental
NM: Sandia National Laboratories, 1985. and numerical investigation of the effects of stacking
38. Hu W, Choi KK and Cho H. Reliability-based design sequence and stress ratio on fatigue damage of glass/
optimization of wind turbine blades for fatigue life under epoxy composites. Compos Part B: Eng 2017; 109:
dynamic wind load uncertainty. Struct Multidiscip O 64–71.
2016; 54: 953–970. 52. Gambone LR. The effect of R-ratio on the mode II fatigue
39. Hu W, Choi KK, Zhupanska O, et al. Integrating vari- delamination growth of unidirectional carbon/epoxy com-
able wind load, aerodynamic, and structural analyses posites. Doctoral Dissertation, The University of British
towards accurate fatigue life prediction in composite wind Columbia, Vancouver, BC, Canada, 1991.
turbine blades. Struct Multidiscip O 2016; 53: 375–394. 53. Rahman MM, Kadirgama K, Noor MM, et al. Fatigue
40. Zastrau D, Schlaak M, Bruns T, et al. Differences in wind life prediction of lower suspension arm using strain-life
forecast accuracy in the German North and Baltic Seas. approach. Eur J Sci Res 2009; 30: 437–450.
Int J Environ Sci Dev 2014; 5: 575–580. 54. Samborsky DD, Wilson TJ and Mandell JF. Comparison
41. Karatepe S and Corscadden KW. Wind speed estimation: of tensile fatigue resistance and constant life diagrams for
incorporating seasonal data using Markov chain models. several potential wind turbine blade laminates. J Sol
ISRN Renew Energ 2013: 657437 (9 pp.). Energy Eng 2009; 131: 011006.
42. Shamsnia SA, Shahidi N, Liaghat A, et al. Modeling of 55. Kalkhoran A, Salimi-Majd D and Mohammadi B. Fati-
weather parameters using stochastic methods (ARIMA gue life prediction for adhesively bonded root joint of
model) (Case Study: Abadeh Region, Iran). In: Interna- composite wind turbine blade using cohesive zone
tional conference on environment and industrial innovation approach. In: Attaf B (ed.) Recent advances in compo-
(IPCBEE), Kuala Lumpur, Malaysia, 17–19 June 2011, site materials for wind turbine blades, Hong Kong, 30
vol. 12. Singapore: IACSIT Press. January–1 February 2013, pp.221–232. The World
43. Kalman RE. A new approach to linear filtering and pre- Academic Publishing Co. Ltd.
diction problems. J Basic Eng: T ASME 1960; 1: 35–45. 56. Hosseini-Toudeshky H, Jahanmardi M and Goodarzi
44. Yahya SM. Turbines compressors and fans. New Delhi, MS. Progressive debonding analysis of composite blade
India: Tata McGraw-Hill Education, 2010. root joint of wind turbines under fatigue loading. Compos
45. Hoffman MJ, Ramsay RR and Gregorek GM. Effects of Struct 2015; 120: 417–427.
grit roughness and pitch oscillations on the NACA 4415 57. Khoramishad H, Crocombe AD, Katnam KB, et al. Pre-
airfoil. Golden, CO, USA: National Renewable Energy dicting fatigue damage in adhesively bonded joints using
Laboratory, 1996. a cohesive zone model. Int J Fatigue 2010; 32: 1146–1158.
46. Rajadurai JS, Christopher T, Thanigaiyarasu G, et al. 58. Bak BL, Turon A, Lindgaard E, et al. A simulation
Finite element analysis with an improved failure criterion method for high-cycle fatigue-driven delamination using
for composite wind turbine blades. Forsch Ingenieurwes a cohesive zone model. Int J Numer Meth Eng 2015; 106:
163–191.
2008; 72: 193–207.
59. Jonkman J, Butterfield S, Musial W, et al. Definition of a
47. Salimi-Majd D, Azimzadeh V and Mohammadi B. Load-
5-MW reference wind turbine for offshore system develop-
ing analysis of composite wind turbine blade for fatigue
ment. Technical report no. NREL/TP-500-38060.
life prediction of adhesively bonded root joint. Appl Com-
Golden, CO: National Renewable Energy Laboratory,
pos Mater 2015; 22: 269–287.
2009.