Renewable Energy 35 (2010) 2732e2738

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Wind speed forecasting in three different regions of Mexico, using a hybrid

ARIMAeANN model
Erasmo Cadenas a, Wilfrido Rivera b, *
a
Facultad de Ingeniería Mecánica, Universidad Michoacana de San Nicolás de Hidalgo, Santiago Tapia No. 403, Centro, Mexico
b
Centro de Ivestigación en Energía, Universidad Nacional Autónoma de México, Apartado Postal 34, Temixco 62580, Morelos, Mexico

a r t i c l e i n f o a b s t r a c t

Article history: In this paper the wind speed forecasting in the Isla de Cedros in Baja California, in the Cerro de la Virgen
Received 5 May 2009 in Zacatecas and in Holbox in Quintana Roo is presented. The time series utilized are average hourly wind
Accepted 20 April 2010 speed data obtained directly from the measurements realized in the different sites during about one
Available online 21 May 2010
month. In order to do wind speed forecasting Hybrid models consisting of Autoregressive Integrated
Moving Average (ARIMA) models and Artificial Neural Network (ANN) models were developed. The
Keywords:
ARIMA models were first used to do the wind speed forecasting of the time series and then with the
Wind speed forecasting
obtained errors ANN were built taking into account the nonlinear tendencies that the ARIMA technique
ARIMA
ANN and Hybrid models
could not identify, reducing with this the final errors. Once the Hybrid models were developed 48 data
out of sample for each one of the sites were used to do the wind speed forecasting and the results were
compared with the ARIMA and the ANN models working separately. Statistical error measures such as
the mean error (ME), the mean square error (MSE) and the mean absolute error (MAE) were calculated to
compare the three methods. The results showed that the Hybrid models predict the wind velocities with
a higher accuracy than the ARIMA and ANN models in the three examined sites.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction Spatial correlation models take into account the spatial rela-
tionship of the wind speed in different sites. The wind speed time
Because of the intermittency and random character of the wind, series of the predicted points and its neighboring sites are
models for wind speed forecasting and power generation are employed to predict the wind speed in a new site. This kind of
valuable support tools for the operators of the Control Energy model has been used recently by a number of authors [7e9].
Centre (CEC) of the power utility [1]. Its importance has been Lalarukh and Yasmin [10] used an Autoregressive Moving
recognized and valued to the degree that in countries with a large Average (ARMA) model for wind speed forecasting using data of 2
wind power development like Spain it is a legal requirement for the years with non-Gaussian distribution and diurnal non-stationary.
producers to give the prediction of production to the CEC for the The conclusion was that forecast values of variance and wind speed
electricity supply at short and medium term [2]. with a confidence interval of 95% can be acceptable both for short-
To handle wind speed forecasting, different kinds of methods and long-term prediction.
have been proposed by various authors. Lei et al. [3] classify the Torres et al. [11] used the ARMA model to predict the hourly
forecasting methods into four categories: i) physical models, ii) averaged wind speed and compared the results with the persis-
spatial correlation models, iii) conventional statistical models, and tence model. They concluded that the ARMA model outperformed
iv) artificial intelligence and new models. the persistence model. For forecasting horizon of 1 h, the persis-
Physical models use physical data such as temperature, pressure tence had less errors than ARMA model, while for forecasts 10 h in
and topography information to predict the future wind speed advance, the errors of ARMA model were between 12% and 20%
[4e6]. These models are in general developed by meteorologists for smaller than those of persistence model.
large scale area weather prediction and they do not give accurate Statistical conventional models have been used by Cadenas and
results in short-term prediction. Rivera [12] who compared the Autoregressive Integrated Moving
Average (ARIMA) and the Artificial Neural Network (ANN) methods
applied to a time series consisting of seven years of wind speed
* Corresponding author. Tel.: þ52 55556229740. measurements in La Venta, Oaxaca, Mexico. The ARIMA models
E-mail address: [email protected] (W. Rivera). presented a better sensitivity to the adjustment and prediction of

0960-1481/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.renene.2010.04.022
 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738 2733

Fig. 1. Hourly wind speed time series of the Isla de Cedros in Baja California, México. Fig. 3. Hourly wind speed time series of Holbox in Quintana Roo, México.

the wind speed for this case in particular. Nevertheless, the authors this kind of model could help the producers and utilities to obtain
mentioned that it is probable when increasing the number of the maximum economic benefits.
training vectors of the ANN model, its performance to improve, Mabel and Fernández [17] used wind speed, relative humidity
giving a better accuracy. and generation hours as input variables to train an ANN-based
Kavasseri and Seetharaman [13] use the fractional-ARIMA or f- network. The forecasting data were compared with actual data
ARIMA models to model, and forecast wind speeds and power giving good results.
production one day and two days ahead in North Dakota. The Hamzaçe [18] proposed an artificial neural network structure for
forecast errors in wind speed/power were analyzed and compared four seasonal time series forecasting. The proposed structure
with the persistence model. The results indicated that significant considers the seasonal period in time series in order to determine
improvements in forecasting accuracy are obtained with the the number of input and output neurons. The results obtained with
proposed models compared to the persistence method. the model were compared with the results of traditional statistical
Artificial neural networks (ANNs) are some of the most widely models and other ANN architectures. The comparison showed that
used models in the last decade for wind speed forecasting and the proposed model has lower prediction error than other
other disciplines where time series are used. Sfetsos [14] developed methods. It was concluded that the proposed model is especially
some training algorithms including the back-propagation (BP) convenient when the seasonality in time series is strong; however,
algorithm, the Levenberg Marquardt (LM) algorithm to achieve the if the seasonality is weak, different network structures may be
minimal value of network error. more suitable.
Sfetsos [15] presented a novel approach based on ANN model for Salcedo-Sanz et al. [19] presented a hybrid system including
the forecasting of mean hourly wind speed time series. Ten weather forecast models and artificial neural networks for short-
minutes’ data were used to carry out multi-step forecasting, and the term wind speed prediction. The model for short-term wind speed
average results were used to generate the mean hourly predictions. forecast integrates a global numerical weather prediction model
The model was tested on two independent data sets. It produced and observations at different heights as initial and boundary
root mean square errors about four times lower than other models conditions for the MM5 model. Then, the outputs of this model are
which were based on mean hourly data. processed using a neural network to obtain the wind speed forecast
Flores et al. [16] presented a control algorithm for wind speed in specific points of the wind park. The authors mentioned that the
and power prediction. The algorithm was based on ANN model methodology have been applied for short-term wind speed
using the back-propagation method. Two different types of data prediction in a wind park in Spain with very good results.
sets were used to test the algorithm. The authors concluded that Monfared et al. [20] proposed a new strategy for wind speed
forecasting based on fuzzy logic and artificial neural networks. The
new strategy for fuzzy logic not only provides significantly fast
learning process but also has a better wind speed forecasting
compared to traditional methods.
Kalogirou [21] published a paper about various applications of
neural networks mainly in renewable energy problems such as:
solar steam generator, solar water heating systems, photovoltaic
systems and solar radiation and wind speed prediction among
others. The errors reported in the models were within acceptable
limits, reason why the author concluded that artificial neural

Table 1
Statistical measures for the three time series.

Sites Mean Std. dev Min. vel. Max. vel.

(m/s) (m/s) (m/s) (m/s)
Isla de Cedros (B.C) 5.33 4.32 0.00 17.26
Cerro de la Virgen (Zac.) 8.13 3.62 0.00 21.15
Holbox (Q. Roo) 6.40 2.28 0.40 13.18
Fig. 2. Hourly wind speed time series of the Cerro de la Virgen, Zacatecas, México.
2734 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738

Table 2
ARIMA models for the three sites.

Site ARIMA (p,d,q) Model

Isla de Cedros (B.C.) ARIMA (1,0,0) yt ¼ 0.9598yt1 þ et
Cerro de la Virgen (Zac.) ARIMA (2,0,0) yt ¼ 1.1554yt1  0.1644yt2 þ et
Holbox (Q. Roo.) ARIMA (2,0,0) yt ¼ 1.1824yt1  0.1824yt2 þ et

networks can be used for modeling in other fields of renewable

energy production and use.
As can be seen from the literature review, Hybrid models based
on ARIMA and ANN have not been proposed for wind speed fore-
Fig. 5. Artificial neural network for the forecasting errors of the Isla de Cedros.
casting. In this paper the wind speed forecasting in the Isla de
Cedros in Baja California, in the Cerro de la Virgen in Zacatecas and
in Holbox in Quintana Roo is presented. The time series utilized are Figs. 1, 2 and 3 show the hourly wind speed time series in the Isla
average hourly wind speed data obtained directly from the de Cedros, in Baja California, the Cerro de la Virgen in Zacatecas and
measurements realized in the different sites during about one in Holbox in Quintana Roo, respectively.
month. In order to carry out the wind speed forecasting in the three Table 1 shows the mean, the standard deviation and the
sites Hybrid models consisting of ARIMA models and ANN models minimum and maximum velocities for the three wind speed time
were developed. The ARIMA models were first used to do the wind series. In this table it can be observed that the statistical measures
speed forecasting of the time series and then with the obtained of the time series are considerably different among them which is
errors ANN were built taking into account the nonlinear tendencies convenient in order to see if the proposed methodology can be
that the ARIMA technique could not identify, reducing with this the applied for different conditions.
final errors. Once the Hybrid models were developed, 48 data out of
the samples (not considered previously during their training stage)
for each one of the sites were used to do the wind speed forecasting
and the results were compared with the ARIMA and the ANN 3. Mathematical models
models working separately. Statistical error measures such as the
mean error (ME), the mean square error (MSE) and the mean 3.1. ARIMA models
absolute error (MAE) were calculated to compare the three
methods. A great number of ARIMA models exist. The general no seasonal
model is known as ARIMA (p,d,q), where: p ¼ order of the autore-
gressive part, d ¼ degree of first differencing involved, q ¼ order of
moving average part.
2. Data measurement The linear expression that defines the previous notation is the
following:
The analyzed data correspond to the wind velocity measure-
ments done by the Comisión Federal de Electricidad (CFE) in three X
P X
q
different regions of México: the Isla de Cedros in Baja California, the yt ¼ fi yti þ qj etj þ 3t (1)
Cerro de la Virgen in Zacatecas and Holbox in Quintana Roo. The i¼1 j¼1
time series correspond to 24 hourly mean data for about one month
where fi ¼ ith autoregressive parameter, qj ¼ jth moving average
of measurements. In order to validate the developed model, 48 data
parameter, and 3t ¼ the error term at time t.
were reserved for each one of the three sites for the wind speed
In order to model the time series the BoxeJenkins methodology
forecasting.
was followed according to the procedure described by Cadenas and
Rivera [12].
According to the development of the analysis it was found that
the representative models for the Isla de Cedros in Baja California,
the Cerro de la Virgen in Zacatecas and Holbox in Quintana were
ARIMA(1,0,0), ARIMA(2,0,0) and ARIMA(2,0,0), respectively. Which
can be also written as AR(1), AR(2) and AR(2). Table 2 shows details
of the specific models for each one of the three sites.

Fig. 4. Comparison of the forecasting in the Isla de Cedros using the proposed models. Fig. 6. Artificial neural network for the forecasting errors of the Cerro de la Virgen.
 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738 2735

Fig. 7. Artificial neural network for the forecasting errors of Holbox.

3.2. ANN models

For the generation of the ANN models, it is convenient to know

Fig. 9. Residuals from the ARIMA model for the Isla de Cedros.
the following parameters: i) the number of input vectors, ii) the
number of layers, iii) the number of output vectors, iv) the number
of neurons, etc. In general rules do not exist, nevertheless, it is 2) Model 2: Three layers, three input neurons, two hidden ones
important to start with the simplest configurations to see if they are and one output neuron (3L-6N).
acceptable and if not, then more sophisticated configurations 3) Model 3: Two layers, three input neurons and one output
should be proposed. neuron (2L-4N).
4) Model 4: Two layers, two input neurons and one output neuron
3.2.1. Iterative method (2L-3N).
The intuitive technique, denominated as iterative method, has
been used by several researchers to generate forecasting [22e24]. Fig. 4 shows the results obtained of the forecasting in the Isla de
An example of this is to assume a model of prediction with two lags, Cedros using the proposed models. It can be seen that the simplest
represented in the following expression: configuration (Model 4) is the configuration with the highest
accuracy. Furthermore, it was decided to use this configuration for
ytþ1 ¼ f ðxt ; xt1 Þ (2) the three different places.
In order to consider two periods in advanced, the model Figs. 5, 6 and 7 show the artificial neural networks obtained
would be: from the resulting errors of the forecasting in the Isla de Cedros, the
Cerro de la Virgen and Holbox, respectively.
ytþ2 ¼ f ðytþ1 ; xt Þ (3)
In the same way, for three periods we have: 3.3. Hybrid models

ytþ3 ¼ f ðytþ2 ; ytþ1 Þ (4) Since the ARIMA is a lineal methodology and assuming that the
generated errors by the use of this technique still conserve the non-
It is commonly used for short forecast horizons in situations where
lineal behaviour, an artificial neural network (which is a non-lineal
the one-lag correlation is dominant over other lag correlations [25].
methodology) is developed to model the errors and to obtain the
Based on the previous paragraphs it was decided to begin with
non-lineal behaviour, and the result of this is added to the final
the following models:
forecasting. Furthermore the final forecasting errors will be lower
than those generated by the ANN.
1) Model 1: Three layers, three input neurons, three hidden ones
and one output neuron (3L-7N).

Fig. 8. ARIMA model and residuals for the Isla de Cedros. Fig. 10. ARIMA and Hybrid residuals for the Isla de Cedros.
2736 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738

Fig. 11. Hourly wind speed time series of the reserved data for the Isla de Cedros.

In a simple way and considering the assumptions mentioned Fig. 13. Sample data forecasting utilizing the Hybrid and ANN models for the Isla de
before, the actual observation with the Hybrid model can be esti- Cedros.

mated as
model comparison: the mean error (ME), the mean square error
yt ¼ Lt þ nLt þ et (5)
(MSE) and the mean absolute error (MAE).
where yt ¼ actual observation, Lt ¼ lineal part, nLt ¼ non-lineal part, If yt is the actual observation for a time period t and Ft is the
and et ¼ final error. forecast for the same period, then the error is defined as:
The first term of the equation (Lt) is obtained from the ARIMA
models reported in Table 2, the second (nLt) from the ANN models et ¼ yt  Ft (6)
obtained shown in Figs. 5e7 and et are the final forecasting errors. Then the mean error can be calculated as:
Fig. 8 shows the fitting function and the forecasting errors
obtained by the ARIMA model for the Isla de Cedros. 1X n

Fig. 9 shows exclusively the residuals obtained from the ARIMA ME ¼ et (7)
n t ¼1
model in the Isla de Cedros. It can be observed that the wind speed
errors are considerably high reaching values up to 10 m/s. The mean square error is:
Fig. 10 shows the forecasting errors obtained with the ARIMA
and Hybrid models. As can be observed the final errors obtained 1Xn
MSE ¼ e2 (8)
from the Hybrid model are considerably lower than those obtained nt¼1 t
with the ARIMA methodology.
and the mean absolute error can be estimated as:
4. Statistical error measures
1Xn
MAE ¼ jet j (9)
In order to determine quantitatively the best model, three nt¼1
forecast error measures were employed for model evaluation and
where n is the number of periods of time.

5. Test data forecasting

A good fitting of the model with the historical data does not
necessarily imply good forecasting. These problems can be over-
come by measuring true out of sample (or validation data set)
forecast accuracy. That is, the total data are divided into an
“initialization” set and a “test” set or “holdout” set. Then, the
initialization set is used to estimate any parameters and to initialize
the method. Forecasts are made for the test set. Since the test set
was not used in the model fitting, these forecasts are genuine
forecast made without using the values of the observations for

Table 3
Statistical error measures for the ARIMA, ANN and Hybrid models for the out of
sample data of the Isla de Cedros.

Model ME MAE MSE

ARIMA 0.4537 1.4772 4.1606
ANN 0.2690 1.7638 5.6510
Fig. 12. Sample data forecasting utilizing the hybrid and ARIMA models for the Isla de
Hybrid 0.0373 0.5083 0.4904
Cedros.
 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738 2737

Fig. 14. Sample data forecasting utilizing the Hybrid and ARIMA models for the Cerro
de la Virgen, Zac. Fig. 16. Sample data forecasting utilizing the Hybrid and ARIMA models for Holbox, Q.
Roo.

these times. The accuracy measures are computed for the errors in best model, the three forecast error measures indicated above were
the test set only [26]. employed for model evaluation and model comparison i.e., the
As it was mentioned above, in order to validate the developed mean error (ME), the mean square error (MSE) and the mean
models, it is necessary to do the forecasting with data not utilized to absolute error (MAE).
generate the models. For this purpose 48 data were reserved for Table 3 shows the statistical error measures for the ARIMA, ANN
each one of the three sites. Fig. 11 shows the validation data set time and Hybrid models. It can be observed that the ME, MAE and the
series for the Isla de Cedros. MSE values obtained with the Hybrid model are considerably lower
Fig. 12 compares the forecasting obtained by using the Hybrid than those obtained with the other two models.
and ARIMA models for the Isla de Cedros. It can be seen that From the forecasting analysis made to the wind speed velocities
although both models follow the same tendency of the real data, in the Isla de Cedros in Baja California, it was clear that the Hybrid
the hybrid model fits better to the real data since the ARIMA model model showed to be a better alternative than the ARIMA and ANN
shows a delay during the whole time series. models used separately, however it is not clear until now that the
Fig. 13 compares the forecasting obtained by using the Hybrid Hybrid model is in general a better alternative or if it was only for
and ANN models for the Isla de Cedros. In a similar way to Fig. 12, in this specific data. Because of this the same methodology was fol-
the present figure it can be seen that both models follow the same lowed for the wind velocities time series in the Cerro de la Virgen in
tendency of the real data, however, the Hybrid model fits better to Zac. and in Holbox in Quintana Roo.
the real data. The ANN model shows a similar delay than that Fig. 14 compares the results obtained with the ARIMA and
observed with the ARIMA model. Hybrid models with the real data for Cerro de la Virgen in Zac. In
From Figs. 11e13 it was qualitatively observed that the Hybrid this figure it can be observed how both models follow the same
model fits better to the real data than the ARIMA and ANN models tendency of the real data, however a higher difference between the
working independently. In order to determine quantitatively the real data and the data predicted with the ARIMA model is observed.

Fig. 15. Sample data forecasting utilizing the Hybrid and ANN models for the Cerro de
la Virgen, Zac. Fig. 17. Sample data forecasting Hybrid and ANN, Holbox, Q. Roo.
2738 E. Cadenas, W. Rivera / Renewable Energy 35 (2010) 2732e2738

Table 4 work, and specially to the Ing. Arturo Nava Hernández for its
Statistical error measures for the three models in the different sites of Mexico. valuable aid.
Data Method ME (m/s) MAE (m/s) MSE (m/s)
Isla de Cedros ARIMA 0.4537 1.4772 4.1606
References
ANN 0.2690 1.7638 5.6510
Hybrid 0.0373 0.5083 0.4904
[1] Roulston MS, Kaplan DT, Hardenberg J, Smith LA. Using medium-range
Cerro de la Virgen ARIMA 0.0696 0.9434 1.6290 weather forecasts to improve the value of wind energy production. Renew-
ANN 0.6046 0.9881 2.0196 able Energy 2003;28:585e602.
Hybrid 0.4377 0.4377 0.2594 [2] Ministerio de Economía. Real Decreto 436/2004, España; 2004.
[3] Lei M, Shiyan L, Chuanwen J, Hongling L, Yan Z. A review on the forecasting of
Holbox ARIMA 0.0246 0.6016 0.6457 wind speed and generated power. Renewable and Sustainable Energy Reviews
ANN 0.3659 0.6748 0.7687 2009;13:915e20.
Hybrid 0.0012 0.0684 0.0089 [4] Landberg L. Short-term prediction of the power production from wind
farms. Journal of Wind Engineering and Industrial Aerodynamics
1999;80:207e20.
[5] Alexiadis MC, Dokopoulos PS, Sahsamanoglou HS, Manousaridis IM. Short-
In Fig. 15 it can be observed how the Hybrid model fits very well term forecasting of wind speed and related electrical power. Solar Energy
1998;63(1):61e8.
with the real data for Cerro de la Virgen in Zac., meanwhile a consid- [6] Negnevitsky M and Potter CW. Innovative short-term wind generation
erable deviation can be observed between the real data and the prediction techniques. In: Proceedings of the power systems conference and
predicted data utilizing the ANN model. exposition; 2006. p. 60e5.
[7] Focken U, Lange M, Monnich K, Waldl H-P, Georg Beyer H, Luig A. Short-term
Fig. 16 compares the real data with the Hybrid and ARIMA
prediction of the aggregated power output of wind farms e a statistical
models for Holbox, Q. Roo. It can be seen that the Hybrid model analysis of the reduction of the prediction error by spatial smoothing effects.
adjusts very well to the real data, meanwhile a delay can be Journal of Wind Engineering and Industrial Aerodynamics 2002;90
(3):231e46.
observed with the data obtained with the ARIMA model.
[8] Barbounis TG, Theocharis JB. A locally recurrent fuzzy neural network with
In Fig. 17 it can be seen again that the Hybrid model fits in a very application to the wind speed prediction using spatial correlation. Neuro-
reasonable way to the real data for Holbox, Q. Roo. The data obtained by computing 2007;70(7/9):1525e42.
using the ANN models follow the tendency of the real data but with [9] Barbounis TG, Theocharis JB. Locally recurrent neural networks for wind speed
prediction using spatial correlation. Information Sciences 2007;177
a considerable delay. (24):5775e97.
In order to determine quantitatively the best model for the data [10] Lalarukh K, Yasmin ZJ. Time series models to simulate and forecast
of the Cerro de la Virgen in Zac., and in Holbox, Q. Roo, the three hourly averaged wind speed in Quetta, Pakistan. Solar Energy 1997;61
(1):23e32.
forecast error measures were calculated. [11] Torres JL, García A, De Blas M, De Francisco A. Forecast of hourly average
Table 4 compares the statistical error measures for the ARIMA, ANN wind speed with ARMA models in Navarre. Solar Energy 2005;79
and Hybrid models for the three different sites. It can be observed in all (1):65e77.
[12] Cadenas E, Rivera W. Wind speed forecasting in the South Coast of Oaxaca,
the cases that the ME, MAE and the MSE values obtained with the México. Renewable Energy 2007;32(12):2116e28.
Hybrid model are considerably lower than those obtained with the [13] Kavasseri RG, Seetharaman K. Day-ahead wind speed forecasting using f-
other two models, which clearly show that the Hybrid model predicts ARIMA models. Renewable Energy 2009;34:1388e93.
[14] Sfetsos A. A comparison of various forecasting techniques applied to mean
with a high accuracy the behaviour and the wind velocities than the
hourly wind speed time series. Renewable Energy 2000;21(1):23e35.
other models. The positive and negative values for ME indicate that [15] Sfetsos A. A novel approach for the forecasting of mean hourly wind speed
there is a systematic under e or over e forecasting. time series. Renewable Energy 2002;27(2):163e74.
[16] Flores P, Tapia A, Tapia G. Application of a control algorithm for wind speed
prediction and active power generation. Renewable Energy 2005;30
(4):523e36.
6. Conclusions
[17] Mabel MC, Fernández E. Analysis of wind power generation and prediction
using ANN: a case study. Renewable Energy 2008;33(5):986e92.
Hybrid models consisting of Autoregressive Integrated Moving [18] Hamzaçe C. Improving artificial neural networks’ performance in seasonal
Average model and an Artificial Neural Network model were time series forecasting. Information Sciences 2008;178:4550e9.
[19] Salcedo-Sanz S, Pérez-Bellido AM, Ortiz-García EG, Portilla-Figueras A,
proposed and developed for the wind speed forecasting in three Prieto L, Paredes D. Hybridizing the fifth generation mesoscale model with
different regions of México. The results showed that the three artificial neural networks for short-term wind speed prediction. Renewable
models predict in a reasonable way the behaviour of the time series Energy 2009;34:1451e7.
[20] Monfared M, Rastegar H, Kojabadi HM. A new strategy for wind speed fore-
of the different sites, however with the evaluation of the statistical casting using artificial intelligent methods. Renewable Energy
error measures it was clear that the Hybrid models predict with 2009;34:845e8.
a higher accuracy the wind velocities than the ARIMA and ANN [21] Kalogirou SA. Artificial neural networks in renewable energy systems appli-
cations: a review. Renewable and Sustainable Energy Reviews
models working separately in the three examined sites. Further- 2001;5:273e401.
more, it can be concluded, that the proposed Hybrid model could be [22] Hwarn H. Insights into neural-network forecasting of time series corre-
a better alternative for the wind speed forecasting where linear and sponding to ARMA (p,q) structures, Omega. The International Journal of
Management Science 2001;29:273e89.
nonlinear tendencies are found.
[23] Balkin SD. The value of nonlinear models in the M3-competition. International
Journal of Forecasting 2001;17:545e6.
[24] Nelson M, Hill T, Remus W, O’Conor M. Time series forecasting using neural
Acknowledgments networks: should the data be deseasonalized first? Journal of Forecasting
1999;18:359e67.
We are thankful to Subgerencia de Proyectos Geo- [25] De Lurgio S. Forecasting principles and applications. New York: Irwing
McGraw-Hill; 1998.
termoeléctricos of the Comisión Federal de Electricidad (CFE), [26] Makridakis S, Wheelwright SC, Hyndman RJ. Forecasting methods and
for the support for the accomplishment of the present applications. 3rd ed. Wiley & Sons, Inc, ISBN 0-471-53233-9; 1998.