Prediction of Extreme Wind Speed Using Artificial Neural Network Approach

Scientific Review
ISSN(e): 2412-2599, ISSN(p): 2413-8835
Vol. 2, No. 1, pp: 8-13, 2016
URL: http://arpgweb.com/?ic=journal&journal=10&info=aims
8
Academic Research Publishing Group
Prediction of Extreme Wind Speed Using Artificial Neural
Network Approach
N. Vivekanandan Central Water and Power Research Station, Pune, Maharashtra, India
1. Introduction
Wind is renewable source of energy and is almost the fastest growing energy resource in the world, which offers
many benefits to human beings. Wind energy has been getting immense attention because renewable energies have
got tremendous focus. Due to increase in cost of fossil fuel and the various environmental problems, it is important
to appreciate the potential of electricity generation from nonconventional sources. The effective use of wind energy
is the conversion of wind power into valuable forms of electricity. The distribution of wind speed is important for
power generators [1]. For the purpose, the predicted variations of meteorological parameters such as wind speed,
relative humidity, solar radiation, air temperature, etc. are needed in the renewable industry for design, performance
analysis, and running cost estimation of the systems. Moreover, for proper and efficient utilization of wind power, it
is important to know the statistical characteristics, persistence, availability, diurnal variation and prediction of wind
speed. The wind characteristics are needed for site selection, performance prediction and planning of wind turbines.
Of these characteristics, the prediction of mean monthly and daily wind speed is very important. For the purpose,
number of methods has been employed for the wind power forecasting. The wind power forecasting methods can be
generally categorized into six groups such as persistence, physical, statistical, spatial correlation, artificial
intelligence network and hybrid [2]. In this paper, application of ANN method in prediction of wind speed is
discussed with illustrative example.
2. Theoretical Description of ANN
ANN modelling procedures adapt to complexity of input-output patterns and accuracy goes on increasing as
more and more data become available. Figure 1 presents the architecture of ANN that consists of input layer, hidden
layer, and output layer [3]. In turn, these layers have a certain number of neurons or units, so the units are called as
input units, hidden units and output units. From ANN structure, it can be easily understood that input units receive
data from external sources to the network and send them to the hidden units, in turn, the hidden units send and
receive data only from other units in the network, and output units receive and produce data generated by the
network, which goes out of the system. In this process, a typical problem is to estimate the output as a function of
the input. This unknown function may be approximated by a superposition of certain activation functions such as
tangent, sigmoid and polynomial. A common threshold function used in ANN is the sigmoid function (f(S))
expressed by Eq. (1), which provides an output in the range of 0≤f(S)<1.
Abstract: Prediction of an accurate wind speed of wind farms is necessary because of the intermittent nature
of wind for any region. Number of methods such as persistence, physical, statistical, spatial correlation, artificial
intelligence network and hybrid are generally available for prediction of wind speed. In this paper, ANN based
methods viz., Multi Layer Perceptron (MLP) and Radial Basis Function (RBF) neural networks are used. The
performance of the networks applied for prediction of wind speed is evaluated by model performance indicators
viz., Correlation Coefficient (CC), Model Efficiency (MEF) and Mean Absolute Percentage Error (MAPE).
Meteorological parameters such as maximum and minimum temperature, air pressure, solar radiation and
altitude are considered as input units for MLP and RBF networks to predict the extreme wind speed at Delhi.
The study shows the values of CC, MEF and MAPE between the observed and predicted wind speed (using
MLP) are computed as 0.992, 95.4% and 4.3% respectively while training the network data. For RBF network,
the values of CC, MEF and MAPE are computed as 0.992, 95.9% and 3.0% respectively. The model
performance analysis indicates the RBF is better suited network among two different networks studied for
prediction of extreme wind speed at Delhi.
Keywords: Artificial neural network; Multi layer perception; Correlation coefficient; Mean absolute percentage error;
Model efficiency; Radial basis function; Wind speed.

Scientific Review, 2016, 2(1): 8-13
9
   1
ii Sexp1)S(f

 and )S(f)S(F i
'
i  … (1)
i
N
1i
ijii OWIS  

, j=1,2,3,…..M … (2)
where Si is the characteristic function of ith
layer, Ii is the input unit of ith
layer, Oi is the output unit of ith
layer, Wij is
the synaptic weights between ith
input and jth
hidden layers, N is the number of observations and M is the number of
neurons in the hidden layer [4]. The sigmoid function is chosen for mathematical convenience because it resembles a
hard-limiting step function for extremely large positive and negative values of the incoming signal and also gives
sufficient information about the response of the processing unit to inputs that are close to the threshold value.
Figure-1. Architecture of ANN
3. Literature Review of ANN
With the development of Artificial Intelligence (AI), a number of various AI methods have been developed for
prediction of wind speed. The new developed methods include Artificial Neural Network (ANN), Adaptive Neuro-
Fuzzy Inference System, Fuzzy Logic, Support Vector Machine, Neuro-Fuzzy Network (NFN) and Evolutionary
Optimization Algorithm. Out of these methods, ANN could deal with non-linear and complex problems in terms of
classification or forecasting. The ANN models can represent a complex nonlinear relationship and extract the
dependence between variables through the training process [5]. In ANN, number of training algorithms such as Multi
Layer Perceptron (MLP), Recurrent (REC), Radial Basis Function (RBF), Ridgelet and Adaptive Linear Element
(ALE) are used for training the network. Sfetsos [6] applied ANN method for forecasting of mean hourly wind speed
data using time series analysis. The proposed methodology has an additional benefit for utility that have significant
wind penetration level and use hourly interval for power system operational procedures such as economic dispatch
and unit commitment. Chang [7] described wind power forecasting methodologies using MLP network. The
developed model for short-term wind forecasting showed a very good accuracy to be used by a 2400kw wind energy
conversion system in Taichung coast for the energy supply. More and Deo [8] presented two wind forecasting
methodologies based on MLP and REC networks. They also found that the results obtained from ANN methods are
more accurate than traditional statistical time series analysis.
Chang [9] adopted a method to do time series prediction of wind power generation using RBF network. They
expressed that the good agreement between the realistic values and forecasting values are obtained; and the
numerical results showed that the proposed forecasting method is accurate and reliable. Li and Shi [10] compared
three types of networks namely, ALE, MLP and RBF to forecast the wind speed. They found that no single ANN
model outperforms others universally in terms of all evaluation metrics even for the same wind data set. Moreover,
the selection of the type of ANN for best performance is also dependent upon the data sources. However, the
research reports indicated that there is a general agreement in applying ANN based method for prediction of wind
speed. Therefore, in the present study, ANN based methods viz., MLP and RBF networks are used to predict the
wind speed with illustrative example.
3.1. Multi-Layer Perceptron Network
MLP network is the most widely used for prediction of wind speed and its architecture with single hidden layer
is shown in Figure 1. Gradient descent is the most commonly used supervised training algorithm in MLP in which

10
each input unit of the training data set is passed through the network from the input layer to output layer [11]. The
network output is compared with the desired target output and output error ( E ) is computed using Eq. (3).
  

N
1i
2*
ii XX
2
1
E … (3)
where, iX is the observed wind speed for ith
sample and *
iX is the predicted wind speed for ith
sample.
 1MW
W
E
)M(W ij
ij
ij 


 … (4)
where, ijW is the synaptic weights between input and hidden layers, )M(Wij is the weight increments between ith
and jth
units during M neurons (units) and )1M(Wij  is the weight increments between ith
and jth
units during 1M 
neurons. In MLP network, momentum factor ( α ) is used to speed up training in very flat regions of the error surface
to prevent oscillations in the weights and learning rate ( ε ) is used to increase the chance of avoiding the training
process being trapped in local minima instead of global minima [12].
3.2. Radial Basis Function Network
RBF network is supervised and three-layered feed forward neural network. The hidden layer of RBF network
consists of a number of nodes and a parameter vector, which can be considered the weight vector. In RBF, the
standard Euclidean distance is used to measure the distance of an input vector from the center. The design of neural
networks is a curve-fitting problem in a high dimensional space in RBF. Training the RBF implies finding the set of
basis nodes and weights. Therefore, the learning process is to find the best fit to the training data [13]. The transfer
functions of the nodes are governed by nonlinear functions that is assumed to be an approximation of the influence
that data points have at the center. The transfer function of a RBF is mostly built up of Gaussian rather than sigmoid.
The Gaussian functions decrease with distance from the center. The transfer functions of the nodes are governed by
nonlinear functions that is assumed to be an approximation of the influence that data points have at the center.
The Euclidean length is represented by jr that measures the radial distance between the datum
vector )X,...X,X(X M21
and the radial center )W,...W,W(X Mjj2j1
)j(
 can be written as:
 
2/1
M
1i
2
iji
)j(
j WXXXr 



 

… (5)
where jr is the Euclidean norm, () is the activation function and ijW is the connecting weight between the
ith
hidden unit and jth
output unit. A suitable transfer function is then applied to rj to give )k(
j XXΦ)r(Φ  .
Finally, the output layer (k-1) receives a weighted linear combination of )r( j .
   
 
N
1j
N
1j
)j()k(
j0j
)k(
j0
)k(
XXΦcW)r(ΦcWX … (6)
where, )r( j is the response of the jth
hidden unit and 0W is the bias term. For both MLP and RBF networks, the
units (or) neurons in the hidden layer are decided by the following equation:
N
PP
M S
2
OI
H 

 … (7)
where, MH is the number of neurons in hidden layer, NS is the number of data samples used in MLP and RBF
networks, PI is the number of input parameter and PO is the number of output parameter [14].
3.3. Normalization of Data
By considering the nature of sigmoid function adopted in ANN, the training data set values are normalized
between 0 and 1 using Eq. (8) and passed into the network [15]. After the completion of training, the output values
are denormalized to provide the results in original domain.
   
   ii
ii
i
XMinXMax
XMinX
XNOR


 … (8)
where,  iXNOR is the normalized value of iX ,  iXMin is the series minimum value of iX and  iXMax is the series
maximum value of iX .
3.4. Model Performance Analysis
The performance of predicted wind speed using MLP and RBF networks are evaluated by Model Performance
Indicators (MPIs) viz., Correlation Coefficient (CC), Model Efficiency (MEF) and Mean Absolute Percentage Error
(MAPE), and are:

11
CC =   
    





 





 
 


N
1i
2
**
i
N
1i
2
i
N
1i
**
ii
XXXX
XXXX … (9)
MEF (%) =
 
 
100*
XX
XX
1 N
1i
i
N
1i
2*
ii
2












 
 


 … (10)
MAPE(%) = 100*
X
XX
N
1 N
1i i
*
ii
 






 

… (11)
where X is the average observed wind speed and *
X is the average predicted wind speed [16].
4. Application
In this paper, a study on prediction of annual extreme wind speed (km/hr) for Delhi region was carried out. ANN
based methods viz., MLP and RBF networks were used for training the network. For MLP and RBF networks, the
meteorological parameters viz., minimum and maximum temperature, solar radiation, air pressure and altitude were
considered as the input units and predicted annual extreme wind speed was the desired output unit. Figure 2 gives the
ANN architecture for prediction of annual extreme wind speed. The annual series of meteorological parameters was
derived from the daily data recorded at Delhi for the period 1969 to 2013 and used. The data for the period 1969 to
1998 was used for training the network and data for the period 1999 to 2013 was used for testing the network.
Figure-2. ANN architecture for prediction of wind speed
5. Results and Discussions
Statistical software, namely, SPSS Neural Connection was used to train the network data with different
combinations of parameters to determine optimum network architectures of MLP and RBF networks for prediction
of annual extreme wind speed at Delhi.
5.1. Prediction of Wind Speed using MLP and RBF Networks
In the present study, the input units were normalized and applied to the network as the units of the parameters
are different. The momentum factor ( α ) and learning rate ( ε ) were fixed as 0.8 and 0.07 while optimizing the
network architecture of MLP. The network data was trained with the optimum network architectures of MLP and
RBF, as given in Table 1. The networks were tested with model parameters for prediction of wind speed. The output
unit was denormalized to obtain the value of annual maximum extreme wind speed in km/hr. The model
performance of MLP and RBF networks were evaluated by MPIs and also given in Table 1 for the region under
study.

12
Table-1. Network architectures with MPIs of MLP and RBF networks
Network Architecture
and MPIs
MLP RBF
Training Testing Training Testing
Network Architecture 5-10-1 5-15-1
CC 0.992 0.990 0.992 0.990
MEF (%) 95.4 94.9 95.9 94.7
MAPE (%) 4.3 4.5 3.0 3.8
From Table 1, it may be noted that: (i) The results of MPIs obtained from RBF network is comparatively better
than the corresponding values of MLP network while training the network data and therefore the network
architecture of RBF is better suited network for prediction of wind speed for Delhi; (ii) The percentage of MEF is
computed as about 95% while testing the data set with RBF network; (iii) The percentages of MAPE obtained from
MLP and RBF networks are 4.5% and 3.8% respectively while testing the network data; and (iv) There is generally a
good correlation between the observed and predicted wind speed using MLP and RBF networks, with CC values
vary between 0.990 and 0.992.
Based on the results obtained from performance analysis with the aid of MPIs, it was observed that the RBF
network gave high prediction accuracy than MLP network for Delhi. Figure 3 give the plots of observed and
predicted wind speed (using MLP and RBF networks) for Delhi.
Figure-3. Plots of observed and predicted wind speed (using MLP and RBF networks)
5.2. Analysis Based on Descriptive Statistics
The descriptive statistics such as average, Standard Deviation (SD), Coefficient of Variation (CV), Coefficient
of Skewness (CS) and Coefficient of Kurtosis (CK) for the observed and predicted wind speed (using MLP and RBF
networks) were computed and given in Table 2.
Table-2. Descriptive statistics of observed and predicted wind speed for Delhi
Statistical parameters Observed
wind speed
Predicted wind speed
MLP RBF
Training Testing Training Testing Training Testing
Average (km/hr) 67.8 64.1 67.0 62.0 65.8 63.1
SD (km/hr) 15.5 15.6 12.8 12.6 14.0 13.6
CV (%) 22.9 24.3 19.1 20.3 21.3 21.6
CS -0.013 -0.006 0.068 0.120 0.103 0.226
CK -1.815 -1.319 -1.716 -1.022 -1.747 -1.095
From Table 2, it may be noted that the percentage of variation on the average predicted wind speed using MLP
network, with reference to average observed wind speed, is 1.2% during training period and 3.3% during testing
period. For RBF network, percentage of variation on the average predicted wind speed with reference to average
observed wind speed during training and testing periods were computed as 2.9% and 1.6% respectively.

13
6. Conclusions
The paper described the procedures involved in prediction of wind speed using MLP and RBF networks for
Delhi. From the results of data analysis, the following conclusions were drawn from the study:
i) Optimum network architectures such as 5-10-1 of MLP and 5-15-1 of RBF were used for training the
network data.
ii) The values of CC, MEF and MAPE between the observed and predicted wind speed (using MLP network)
were computed as 0.992, 95.4% and 4.3% respectively while training the network data. For RBF network,
the values of CC, MEF and MAPE were found to be 0.992, 95.9% and 3.0% respectively.
iii) The values of CC, MEF and MAPE between the observed and predicted wind speed (using MLP network)
were found to be 0.990, 94.9% and 4.5% respectively while testing the network data. For RBF network, the
values of CC, MEF and MAPE were computed as 0.990, 94.7% and 3.8% respectively.
iv) Analysis based on MPIs and descriptive statistics showed that the RBF network is comparatively better than
MLP network for the data under study. MLP and RBF networks were tested by predicting wind speed for
Delhi for which measured data are available; and the results indicated the developed RBF network gives
high prediction accuracy.
v) The percentage of variation on the average predicted wind speed with reference to average observed wind
speed was found to be 1.6% while testing the network data with RBF network, which is comparatively less
than the corresponding value of MLP network.
vi) The results presented in the paper would be helpful to the stakeholders for planning, design and
management of hydraulic and civil structures in Delhi region.
Acknowledgements
The author is grateful to Dr. M. K. Sinha, Director, Central Water and Power Research Station, Pune, for providing
the research facilities to carry out the study. The author is also thankful to M/s Nuclear Power Corporation of India
Limited, Mumbai for supply of temperature data used in the study.
References
[1] Sahin, A. D., 2004. "Progress and recent trends in wind energy." Progress in Energy and Combustion
Science, vol. 30, pp. 501-543.
[2] Chang, W. Y., 2014. "A literature review of wind forecasting methods." Journal of Power and Energy
Engineering, vol. 2, pp. 161-168.
[3] Pallavi, M., Swaptik, C., Sangeeta, R., Nikhil, B., and Roshan, S., 2012. "Dual artificial neural network for
rainfall-runoff forecasting." Journal of Water Resource and Protection, vol. 4, pp. 1024-1028.
[4] Amr, H. E., El-Shafie, A., Hasan, G. E., Shehata, A., and Taha, M. R., 2011. "Artificial neural network
technique for rainfall forecasting applied to Alexandria." International Journal of the Physical Sciences, vol.
6, pp. 1306-1316.
[5] Wu, Y. K., Lee, C. Y., Tsai, S. H., and Yu, S. N., 2010. "Actual experience on the short-term wind power
forecasting at Penghu-From an Island perspective." In Proceedings of the 2010 International Conference on
Power System Technology.Hangzhou. pp. 1-8.
[6] Sfetsos, A., 2002. "A novel approach for the forecasting of hourly wind speed time series." Renewable
Energy, vol. 27, pp. 163-174.
[7] Chang, W. Y., 2013. "Wind energy conversion system power forecasting using radial basis function neural
network." Applied Mechanics and Materials, vol. 284-287, pp. 1067-1071.
[8] More, A. and Deo, M. C., 2003. "Forecasting wind with neural networks." Marine Structures, vol. 16, pp. 35-
49.
[9] Chang, W. Y., 2013. "Application of back propagation neural network for wind power generation
forecasting." International Journal of Digital Content Technology and its Application, vol. 7, pp. 502-509.
[10] Li, G. and Shi, J., 2010. "On comparing three artificial neural networks for wind speed forecasting." Applied
Energy, vol. 87, pp. 2313-2320.
[11] Wang, Y. M., Traore, S., Kerh, T., and Leu, J. M., 2011. "Modelling reference evapotranspiration using feed
forward back propagation algorithm in arid regions of Africa." Irrigation Drainage, vol. 60, pp. 404-417.
[12] Jalal, S., Özgur, K., Oleg, M., Abbas-Ali, S., and Bagher, N., 2012. "Forecasting daily stream flows using
artificial intelligence approaches." ISH Journal of Hydraulic Engineering, vol. 18, pp. 204-214.
[13] Ma, L., Luan, S. Y., Jiang, C. W., Liu, H. L., and Zhang, Y., 2009. "A Review on the forecasting of wind
speed and generated power." Renewable and Sustainable Energy Reviews, vol. 13, pp. 915-920.
[14] Chow, S. K. H., Lee, E. W. M., and Li, D. H. W., 2012. "Short-term prediction of photovoltaic energy
generation by intelligent approach." Energy Build, vol. 55, pp. 660-667.
[15] Sudheer, K. P., Srinivasan, K., Neelakantan, T. R., and Srinivas, V., 2008. "A nonlinear data-driven model for
synthetic generation of annual streamflows." Hydrological Processes, vol. 22, pp. 1831-1845.
[16] Chen, J. and Adams, B. J., 2006. "Integration of artificial neural networks with conceptual models in rainfall-
runoff modelling." Journal of Hydrology, vol. 318, pp. 232-249.

Prediction of Extreme Wind Speed Using Artificial Neural Network Approach

More Related Content

What's hot (20)

Similar to Prediction of Extreme Wind Speed Using Artificial Neural Network Approach (20)

More from Scientific Review SR (20)

Recently uploaded (20)

Prediction of Extreme Wind Speed Using Artificial Neural Network Approach