This article was downloaded by: [M. A. Behrang]On: 27 February 2012, At: 21:58Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Energy Sources, Part A: Recovery,Utilization, and Environmental EffectsPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/ueso20

An Analysis of Wind Speed PredictionUsing Artificial Neural Networks: A CaseStudy in Manjil, IranE. Assareh a , M. A. Behrang a , M. Ghalambaz a , A. R. Noghrehabadib & A. Ghanbarzadeh ba Department of Mechanical Engineering, Dezful Branch, IslamicAzad University, Dezful, Iranb Department of Mechanical Engineering, Engineering Faculty,Shahid Chamran University of Ahvaz, Ahvaz, Iran

Available online: 24 Feb 2012

To cite this article: E. Assareh, M. A. Behrang, M. Ghalambaz, A. R. Noghrehabadi & A. Ghanbarzadeh(2012): An Analysis of Wind Speed Prediction Using Artificial Neural Networks: A Case Study in Manjil,Iran, Energy Sources, Part A: Recovery, Utilization, and Environmental Effects, 34:7, 636-644

To link to this article: http://dx.doi.org/10.1080/15567036.2011.551915

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representationthat the contents will be complete or accurate or up to date. The accuracy of anyinstructions, formulae, and drug doses should be independently verified with primarysources. The publisher shall not be liable for any loss, actions, claims, proceedings,demand, or costs or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.

Energy Sources, Part A, 34:636–644, 2012

Copyright © Taylor & Francis Group, LLC

ISSN: 1556-7036 print/1556-7230 online

DOI: 10.1080/15567036.2011.551915

An Analysis of Wind Speed Prediction

Using Artificial Neural Networks:

A Case Study in Manjil, Iran

E. ASSAREH,1 M. A. BEHRANG,1 M. GHALAMBAZ,1

A. R. NOGHREHABADI,2 and A. GHANBARZADEH2

1Department of Mechanical Engineering, Dezful Branch, Islamic Azad

University, Dezful, Iran2Department of Mechanical Engineering, Engineering Faculty, Shahid

Chamran University of Ahvaz, Ahvaz, Iran

Abstract In this study, air temperature, relative humidity, and vapor pressure datacollected from Manjil station between 1993–2004, were used for wind speed pre-

dictions in a future time domain using artificial neural networks. The followingcombinations of data are considered for this study: (i) month of the year, monthly

mean daily air temperature, and relative humidity as inputs, and monthly mean dailywind speed as output; (ii) month of the year, monthly mean daily air temperature,

relative humidity, and vapor pressure as inputs, and monthly mean daily wind speed asoutput. The generalized regression neural networks, multilayer perceptron, and radial

basis function neural networks were used in this study. The measured data between1993 and 2003 is applied for training and the data for 2004 is used for testing. The

data for testing were not applied for training the neural networks. Obtained resultsshow that neural networks are well capable of estimating wind speed from simple

meteorological data. These results indicate that using vapor pressure along with themonth of the year, monthly mean daily air temperature, and relative humidity based

on a multilayer perceptron network has better performance than the other cases withthe mean absolute percentage error of 7.03% (2004).

Keywords generalized regression neural networks, multi-layer perceptron neuralnetworks, prediction, radial basis function neural networks, wind speed

1. Introduction

The power of wind is a clean, endless, and free source of energy, which has helped

mankind for centuries to move the ships and run the turbines in order to mill the grains

and pump water. Because an ample supply of petroleum was cheap and accessible (pre-

1970s), the high cost and uncertainty of wind placed it at an economic disadvantage.

However, after the 1973 oil embargo, it became clear that the oil supplies would not

remain forever and so other kinds of energy sources must be developed (Mohandes et al.,

2004).

Address correspondence to Ehsanolah Assareh, Department of Mechanical Engineering,Dezful Branch, Islamic Azad University, Dezful, Iran. E-mail: Ehsanolah.assareh@gmail.com

636

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

Analysis of Wind Speed Prediction Using ANN: A Case Study 637

Manjil, which is located between N 36ı1401800–N 36ı4104200 and E 49ı230600–E

49ı3104800 (about 80 km south of the Caspian Sea) in Gilan province, is considered to

be a windy city of Iran. The wind speed average of Manjil is about 6 m/s (which was

measured at a height about 10 m above the ground) in winter and the wind conditions are

great, especially in the summer. The strong north wind blowing from May to September,

whose average wind speed is 14 m/s (at 10 m height), can be explained by the local

climate and geographical conditions (Hagenkort, 2005; Mostafaeipour and Abarghooei,

2008). It is estimated that the wind energy potential is about 12 GW in Iran (IORE,

2006).

Predicting the wind speed is very important in order to use the wind power suf-

ficiently. It is required for selecting the site, to predict the optimum size of the wind

machine that is used for a special site (Mohandes et al., 1998).

Several studies are presented for predicting the wind speed. In a study by Njau

(1994a), an electronic system to predict air temperature and wind speed was developed.

He found a good agreement between the predicted and actual values of wind speed and

temperature.

Njau (1994b) also has considered a semi-empirical correlation to predict the hourly,

daily, and monthly average values of wind speed in Dares Salaam, Tanzania. Rehman and

Halawani (1994) used stochastic time series analysis to predict the hourly wind speed

of nine cities of Saudi Arabia and found a good agreement between the predicted and

actual values. Also, Mohandes et al. (2004) developed support vector machines models to

predict wind speed and compared their performance with multi-layer perceptron (MLP)

neural networks. Cadenas and Rivera (2007) compared autoregressive integrated moving

average and artificial neural networks (ANN) techniques to predict the wind speed in the

south coast of Oaxaca, Mexico. However, many researchers have presented their papers

in this regard based on different techniques (Cellura and Cirrincione, 2008). In this study,

three types of ANNs are applied to predict wind speed in Manjil station using measured

air temperature, relative humidity, and vapor pressure.

2. Neural Networks

Neural networks are computational models of the biological brain. Like the brain, a neural

network comprises a large number of interconnected neurons. Each neuron is capable of

performing only simple computation (Pham et al., 2006a). Consequently, the architecture

of an artificial neuron is simpler than a biological neuron. ANNs are constructed in a layer

that connects to one or more hidden layers where the factual processing is performance

through weighted connections. Each neuron in the hidden layer joins to all neurons in the

output layer. The results of the processing are acquired from the output layer. Learning in

ANNs is achieved through particular training algorithms that are expanded in accordance

with the learning laws, assumed to simulate the learning mechanisms of a biological

system (Yilmaz and Ozer, 2009).

However, as an assembly of neurons, a neural network can learn to perform complex

tasks including pattern recognition, system identification, trend prediction, and process

control (Pham et al., 2006a).

2.1. Generalized Regression Neural Network (GRNN)

The GRNN can solve any function approximation problem. The GRNN proposed by

Specht (1991) does not require an iterative training procedure. It approximates any

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

638 E. Assareh et al.

arbitrary function between input and output vectors, drawing the function estimate directly

from the training data. In addition, it is consistent that as the training set size becomes

large, the estimation error approaches zero. The GRNN is used for the estimation of

continuous variables, as in standard regression techniques. It is related to the radial

basis function network, and is based on a standard statistical technique called kernel

regression. By definition, the regression of a dependent variable y on an independent

x estimates the most probable value for y, given x and a training set. The regression

method will produce the estimated value of y, which minimizes the mean-squared error

(MSE). The principal advantages of the GRNN are fast learning and convergence to the

optimal regression surface as the number of samples becomes very large. The GRNN

is particularly advantageous with sparse data in a real-time environment, because the

regression surface is instantly defined everywhere.

The schematic diagram of GRNN architecture is presented in Figure 1. As it can

be seen from Figure 1, the GRNN is organized using an input layer, a pattern layer,

a summation layer, and an output layer. The relation between input and output can be

expressed as:

y D

mX

j D1

wj 'j .X/

mX

j D1

'j .X/

�˛

ˇ; (1)

where X D Œx1; x2; : : : ; xn�T is a n-dimensional input vector, wj is the weight between

the jth pattern layer node and summation layer node, and ' is the Gaussian function.

Layer 1, the input layer, accepts the input signals into the GRNN. The nodes at layer 1

represent linguistic variables (namely the driving frequency f and the phase difference '

in the TWUSM drive system (Celikoglu, 2006). Layer 2, the pattern layer, possessed a

nonlinear transformation applied on the data from the input space to the pattern space.

The most popular choice for the function ' is a multivariate Gaussian function with an

Figure 1. Schematic diagram of GRNN.

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

Analysis of Wind Speed Prediction Using ANN: A Case Study 639

appropriate mean and auto covariance matrix. The Gaussian function is:

'j .X/ D exp

�kX � Cj k2

�2j

!

D

nY

iD1

exp

2

4

xi � xji

�j

!23

5 ; (2)

where Cj D Œxj

1 ; xj

2 ; : : : ; xjn �T and �j are the center vector and the standard deviation

of the Gaussian function, respectively.

Layer 3, the summation layer, executes the sum operation in Eq. (1). The outputs of

the pattern layer nodes are multiplied with appropriate interconnection weights to sum

up for producing the output of the network. Layer 4 is the output layer, where the nodes

are represented by a GRNN individual output.

2.2. Multi-layer Perceptron (MLP)

MLPs are perhaps the most common type of feedforward networks. Neurons in an input

layer only act as buffers for distributing the input signals xi to neurons in the hidden

layer. Each neuron j in the hidden layer sums up its input signals xi after weighting them

with the strengths of the respective connections wj i from the input layer and computes

its output yj as a function f of the sum, viz.:

yj D f�

X

wj ixi

�

; (3)

where f can be a simple threshold function or a sigmoidal, hyperbolic tangent, or radial

basis function (RBF).

The weight updating process can take place after the presentation of each training

pattern (pattern-based training) or after the presentation of the whole set of training

patterns (batch training). In either case, a training epoch is said to have been completed

when all training patterns have been presented once to the MLP.

For all but the most trivial problems, several epochs are required for the MLP to be

properly trained (Pham and Liu, 1995). For more details about different types of neural

networks and training algorithms, the reader is referred to Behrang et al., 2010; 2011a,

b, c, d, e; Assareh et al., in press, 2012.

2.3. Radial Basis Function (RBF)

The RBF network is a popular type of network that is very useful for pattern classification

problems (Bishop, 1995). The input layer neurons receive the input pattern (x1 to xN ).

The hidden layer neurons provide a set of activation functions that constitute an arbitrary

“basis” for the input patterns in the input space to be expanded into the hidden space

by way of non-linear transformation. At the input of each hidden neuron, the distance

between the center of each activation or basis function and the input vector is calculated.

Applying the basis function to this distance produces the output of the hidden neuron.

The RBF network outputs y1 to yp are formed by the neurons in the output layer as

weighted sums of the hidden layer neuron activations (Pham et al., 2006b).

The training of an RBF network involves the minimization of an error function. The

error function defines the total difference between the actual output and the desired output

of the network over a set of training patterns (Jain and Dubes, 1988). Training proceeds

by presenting to the network a pattern of a known class taken from the training set.

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

640 E. Assareh et al.

The error component associated with that pattern is the sum of the squared differences

between the desired and actual outputs of the network corresponding to the presented

pattern. The procedure is repeated for all the patterns in the training set and the error

components for all the patterns are summed to yield the value of the error function for an

RBF network with a given set of basis function centers, spreads, and neuron connection

weights (Pham et al., 2006b).

3. Problem Definition and Results

In this study, air temperature, relative humidity, and vapor pressure data, collected by

Manjil station from 1993 to 2004, were used for wind speed prediction using ANNs. The

wind speed data is measured at 10 m above ground.

The following combinations of data are considered for this study:

1. month of the year, monthly mean daily air temperature, and relative humidity as

inputs and wind speed as output; and

2. month of the year, monthly mean daily air temperature, relative humidity, and

vapor pressure as inputs and wind speed as output.

The measured data between 1993 and 2003 were applied for training and the 12 months

of data of 2004 were used for testing. The data for testing were not applied for training

the neural networks.

In this article, wind speed, air temperature, relative humidity, and vapor pressure are

normalized in the (0, 1) range.

The Xmin and Xmax values for each variable are shown in Table 1.

Figure 2 shows the measured values of air temperature, relative humidity, vapor

pressure, and monthly wind speed for Manjil city (1993–2004).

The GRNN, MLP, and RBF neural networks by using the neural network toolbox of

Matlab 2007 (The MathWorks, Natick, MA), are used for the two combinations of data

above.

In order to determine the optimal network architecture, various network architectures

were designed; different training algorithms were used; the number of neuron and hidden

layer and transfer functions in the hidden layer/output layer were changed. For both

combinations based on a MLP network, logistic sigmoid transfer function (logsig) for

both hidden layers, linear transfer function (purelin) for output layer, and LM (Levenberg–

Marquardt) train were found to perform reasonably good.

Table 1

Minimum and maximum values of variables

Xmin Xmax

Wind speed (knots)a 1.6 23.6

Air temperature (c) 5.3 28.8

Relative humidity (%) 45 79

Vapor pressure (HPA)b 4.95 20.41

a1 knot D 0.5144 m/s.b1 Hectopascal (HPA) D 0.001 Atmosphere (atm).

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

Analysis of Wind Speed Prediction Using ANN: A Case Study 641

Figure 2. Air temperature, relative humidity, vapor pressure, and wind speed for Manjil city.

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

642 E. Assareh et al.

Figure 3. Estimated values of wind speed based on MLP, RBF, and GRNN networks for

combinations I and II on testing data. (color figure available online)

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

Analysis of Wind Speed Prediction Using ANN: A Case Study 643

Table 2

Training and testing errors of each combination

Combination

Error for

train, MSE

Error for

test, MAPE, %

GRNN-I 0.0078 16.93

GRNN-II 0.0064 13.90

RBF-I 0.0048 11.20

RBF-II 0.0044 10.39

MLP-I 0.0033 10.32

MLP-II 0.0034 7.03

In Table 2, training and testing errors for both combinations based on GRNN, RBF,

and MLP networks, are shown.

Figure 3 shows the results of both combinations based on MLP, RBF, and GRNN

networks. These results indicate that using vapor pressure along with month of the year,

monthly mean daily air temperature, and relative humidity based on the MLP network

has better performance than the other cases with a mean absolute percentage error of

7.03%.

4. Conclusion

This study shows the results of an effort made to forecast the monthly wind speed

according to measured values of air temperature, relative humidity, vapor pressure, and

wind speed. This is of great importance because the above parameters are commonly

accessible. Data for Manjil station, a city in north Iran, from 1993 to 2003 is used for

training ANNs. Data for 12 months of the year 2004 is used for testing the ANNs. For

one case, month of the year, monthly mean daily air temperature, and relative humidity

are considered as inputs and monthly mean daily wind speed as output. In the second

case, month of year, monthly mean daily air temperature, relative humidity, and vapor

pressure are considered as inputs and monthly wind speed as output. These cases were

used for prediction of monthly mean daily wind speed. These results indicate that using

vapor pressure along with the month of year, monthly mean daily air temperature, and

relative humidity based on a MLP network has better performance than the other cases

with a mean absolute percentage error of 7.03%.

References

Assareh, E., Behrang, M. A., Assari, M. R., and Ghanbarzadeh, A. 2010. Application of PSO

(particle swarm optimization) and GA (genetic algorithm) techniques on demand estimation

of oil in Iran. Energy 35:5223–5229.

Assareh, E., Behrang, M. A., and Ghanbarzadeh, A. Integration of artificial neural networks and

particle swarm optimization to forecast world green energy consumption. Energ. Source. Part

B (in press).

Behrang, M. A., Assareh, E., Ghanbarzadeh, A., and Noghrehabadi, A. R. 2010. The potential of

different artificial neural network (ANN) techniques in daily global solar radiation modeling

based on meteorological data. Solar Energy 84:1468–1480.

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012

644 E. Assareh et al.

Behrang, M. A., Assareh, E., Assari, M. R., and Ghanbarzadeh, A. 2011a. Using bees algorithm

(BA) and artificial neural network (ANN) to forecast world carbon dioxide emission. Energy

Source. Part A 33:1747–1759.

Behrang, M. A., Assareh, E., Assari, M. R., and Ghanbarzadeh, A. 2011b. Assessment of electricity

demand in Iran’s industrial sector using different intelligent optimization techniques. Appl. Art.

Intel. 25:292–304.

Behrang, M. A., Assareh, E., Assari, M. R., and Ghanbarzadeh, A. 2011c. Total energy demand

estimation in Iran using bees algorithm. Energ. Source. Part B 6:294–303.

Behrang, M. A., Assareh, E., Ghalambaz, M., Assari, M. R., Noghrehabadi, A. R. 2011d. Forecast-

ing future oil demand in Iran using GSA (Gravitational Search Algorithm). Energy 36:5649–

5654.

Behrang, M. A., Assareh, E., Noghrehabadi, A. R., and Ghanbarzadeh, A. 2011e. New sunshine-

based models for predicting global solar radiation using PSO (particle swarm optimization)

technique. Energy 36:3036–3049.

Bishop, C. M. 1995. Neural Networks for Pattern Recognition. Oxford: Clarendon Press.

Cadenas, E., and Rivera, W. 2007. Wind speed forecasting in the South Coast of Oaxaca, Mexico.

Renew. Energy 32:2116–2128.

Celikoglu, H. B. 2006. Application of radial basis function and generalized regression neural

networks in non-linear utility function specification for travel mode choice modeling. Math.

Comput. Model. 44:40–658.

Cellura, M., and Cirrincione, G. 2008. Wind speed spatial estimation for energy planning in Sicily:

A neural kriging application. Renew. Energy 33:1251–1266.

Chen, T. C., Yu, C. H., and Tsai, M. C. 2008. A novel driver with adjustable frequency and phase

for traveling-wave type ultrasonic motor. J. Chinese Inst. Eng. 31:709–713.

Hagenkort, B. 2005. Information on the Lahmeyer International Group, No. 49. Available from

www.lahmeyer.dewww.lahmeyer.deS.

Iran Organization for Renewable Energy (IORE). 2006. Prospects of Renewable Energy in Iran.

Tehran, Iran: Iran Organization for Renewable Energy.

Jain, A. K., and Dubes, R. C. 1988. Algorithms for Clustering Data. Englewood Cliffs, NJ: Prentice

Hall.

Mohandes, M., Helawani, T. O., Rehman, S., and Hussain, A. 2004. A support vector machine for

wind speed prediction. Renew. Energy 29:939–947.

Mohandes, M., Rehman, S., and Halawani, T. O. 1998. A neural networks approach for wind speed

prediction. Renew. Energy 13:345–354.

Mostafaeipour, A., and Abarghooei, H. 2008. Harnessing wind energy at Manjil area located in

north of Iran. Renew. Sustain. Energy Rev. 12:1758–1766.

Njau, E. C. 1994a. An electronic system for predicting air temperature and wind speed patterns.

Renew. Energy 4:793–805.

Njau, E. C. 1994b. Predictability of wind speed patterns. Renew. Energy 4:261–263.

Pham, D. T., and Liu, X. 1995. Neural Networks for Identification, Prediction and Control. London:

Springer Verlag.

Pham, D. T., Koc, E., Ghanbarzadeh, A., and Otri, S. 2006a. Optimisation of the weights of multi-

layered perceptrons using the bees algorithm. 5th International Symposium on Intelligent

Manufacturing Systems, Sakarya University, Department of Industrial Engineering, Sakarya,

Turkey, May 29–31, pp. 38–46.

Pham, D. T., Ghanbarzadeh., A., Koc., E., and Otri, S. 2006b. Application of the Bees algorithm to

the training of radial basis function networks for control chart pattern recognition. 5th CIRP

International Seminar on Intelligent Computation in Manufacturing Engineering (CIRP ICME

’06), Ischia, Italy, July 25–28, pp. 711–716.

Rehman, S., and Halawani, T. O. 1994. Statistical characteristics of wind in Saudi Arabia. Renew.

Energy 4:949–956.

Specht, D. F. 1991. A general regression neural network. IEEE Trans. Neural Networks 2:568–576.

Yilmaz, A. S., and Ozer, Z. 2009. Pitch angle control in wind turbines above the rated wind

speed by multi-layer perceptron and radial basis function neural networks. Expert Sys, Applic.

36:9767–9775.

Dow

nloa

ded

by [

M. A

. Beh

rang

] at

21:

58 2

7 Fe

brua

ry 2

012