Energies 09 00
Energies 09 00
Energies 09 00
Article
Wind Speed Prediction Using a Univariate ARIMA
Model and a Multivariate NARX Model
Erasmo Cadenas 1 , Wilfrido Rivera 2,† , Rafael Campos-Amezcua 2, *,† and Christopher Heard 3,†
1 Facultad de Ingenieria Mecanica, Universidad Michoacana de San Nicolas de Hidalgo,
Santiago Tapia No. 403, Col. Centro, CP 58000 Morelia, Michoacan, Mexico; [email protected]
2 Instituto de Energias Renovables, Universidad Nacional Autonoma de Mexico, Apartado postal 34,
CP 62580 Temixco, Morelos, Mexico; [email protected]
3 Division de Ciencias de la Comunicacion y Diseno, Departamento de Teoria y Procesos del Diseno,
Diseno Ambiental, Universidad Autonoma Metropolitana Unidad Cuajimalpa, Torre III, 5to. piso,
Av. Vasco de Quiroga 4871, Col. Santa Fe Cuajimalpa, Del. Cuajimalpa, Mexico D.F. 11850, Mexico;
[email protected]
* Correspondence: [email protected]; Tel.: +52-777-362-0090 (ext. 38010)
† These authors contributed equally to this work.
Abstract: Two on step ahead wind speed forecasting models were compared. A univariate model
was developed using a linear autoregressive integrated moving average (ARIMA). This method’s
performance is well studied for a large number of prediction problems. The other is a multivariate
model developed using a nonlinear autoregressive exogenous artificial neural network (NARX).
This uses the variables: barometric pressure, air temperature, wind direction and solar radiation or
relative humidity, as well as delayed wind speed. Both models were developed from two databases
from two sites: an hourly average measurements database from La Mata, Oaxaca, Mexico, and a ten
minute average measurements database from Metepec, Hidalgo, Mexico. The main objective was to
compare the impact of the various meteorological variables on the performance of the multivariate
model of wind speed prediction with respect to the high performance univariate linear model. The
NARX model gave better results with improvements on the ARIMA model of between 5.5% and
10.6% for the hourly database and of between 2.3% and 12.8% for the ten minute database for mean
absolute error and mean squared error, respectively.
1. Introduction
At the end of 2014, the worldwide installed wind energy generating capacity was 369,597 MW;
Europe having 134,007 MW, of which Germany and Spain stood out with 39,165 and 22,987 MW,
respectively. During 2015, 42% of electric power in Denmark was generated from wind [1]. In the
Asia-Pacific region, China had a reported capacity of 114,609 MW of a total of 141,964 MW. In North
America, the reported U.S. installed capacity was 65,879 MW with the Mexican and Canadian installed
capacities being 9694 and 2551 MW, respectively. In Latin America, Brazil was the leader, with 5939 MW
of 8526 MW total [2].
Onshore wind based power generation has reached the technological maturity of being
competitive with the lowest cost power generation options in many places. For example in Mexico in
2012 installed capacity increased by 76% with respect to the total installed wind energy generation
capacity at the end of 2011 due to increasing exploitation of the intense resource in the state of Oaxaca.
In Oaxaca in the corridor from La Venta to La Mata passing through La Ventosa, the annual
average wind speed is over 9 m/s at 30 m above ground level with a dominant wind direction
of North-Northwest/North-Northeast 70% of the time [3]. These highly favorable intense wind
conditions in Oaxaca represent an appreciable source of inexpensive renewable energy in addition to
Mexico’s large fossil fuel reserves, which makes its exploitation a priority.
In Mexico, the National Center for Energy Control (CENACE) is responsible for the dispatch
control of energy for the National Electric System. CENACE uses an information system to prepare
pre-dispatch strategies. This system takes into account: availability, derating, restrictions and other
factors that affect the dispatch capacity of generating units, as well as the electricity demand forecast.
These models are produced by CENACE. An hourly operation plan is essential for each unit [4]. Energy
producers have a responsibility to provide forecasts of wind and net energy production to CENACE a
day ahead.
Recently, a considerable number of wind speed prediction models have been developed using a
range of methods, some simple and others combining various techniques. Cadenas and Rivera [5] have
reported short-term wind speed forecasting in a region of Oaxaca using an artificial neural network
(ANN) with a representative hourly time series for the site. The model showed good accuracy for
energy supply prediction.
Salcedo-Sanz et al. [6] presented a hybrid model between a fifth generation mesoscale model
(MM5) and a neural network for short-term wind speed prediction at specific points.
Cadenas et al. [7] analyzed and forecasted wind velocity in Chetumal, Quintana Roo, Mexico,
with a single exponential smoothing method. The method was found to be good for wind forecasting
when the field data had alpha values close to one.
Li and Shi [8] compared three artificial neural networks for wind speed forecasting. These were:
adaptive linear element, back propagation and radial basis function. None of these outperformed the
others on all of the metrics evaluated.
A new short-term hybrid method based on wavelet and classical time series analysis to predict
wind speed and power was proposed by Liu et al. [9]. The mean relative error in multi-step
forecasting using this method was smaller than that from classical time series and back-propagation
network methods.
A wind speed forecasting model for three regions of Mexico was developed using a hybrid
autoregressive integrated moving average technique (ARIMA-ANN) by Cadenas and Rivera [10].
Initially, the ARIMA models were used to generate wind speed forecasts for the time series. The
resulting errors were used to build the ANN to account for the non-linear behavior that the ARIMA
technique could not model. This reduced the errors. The results showed that the hybrid model
produced higher accuracy wind speed predictions than those of the separate ARIMA and ANN models
for all three sites.
Kavasseri and Seetharaman [11] used the fractional-ARIMA models to predict wind speed and
power production one or two days ahead for North Dakota. Forecasting errors in wind speed and
power were compared to the persistence model. Significant improvements were obtained.
Li et al. [12] presented a robust two-step method for accurate wind speed forecasting based on a
Bayesian combination algorithm and three neural network models: an adaptive linear element network
(ADALINE), back propagation (BP) and a radial basis function (RBF). The results were that the neural
networks were not consistent for one hour ahead wind speed. However, the Bayesian combination
method could always give adaptive, reliable and comparatively accurate forecasts.
Liu et al. [13] evaluated the effectiveness of autoregressive moving average-generalized
autoregressive conditional heteroscedasticity (ARMA-GARCH) for modeling mean wind speed and
its volatility. The results showed that ARMA-GARCH could capture the trend changes of these
parameters. In this study, it was found that none of the models were consistently better than the
others over the whole range of heights considered. The authors recommended that for a given dataset,
all of the models should be evaluated to find the most appropriate (Given that the range of heights
Energies 2016, 9, 109 3 of 15
considered was from 10 to 80 m and that the swept diameter of wind turbines is of the order of 90 m
centered at a height of 70 m for 3-MW units, the heights covered by the study needed to be greater).
Guo et al. [14] developed an empirical mode decomposition (EMD) based feed-forward neural
network (FFNN) learning model, which resulted in improved accuracy over each of the two methods
individually for predicting daily and monthly mean wind speeds.
Liu et al. [15] have proposed hybrid ARIMA-ANN and ARIMA-Kalman methods for hourly wind
speed forecasting. The authors concluded that both methods gave good results and can be applied to
dynamic wind speed forecasting for wind power systems.
Kalman filtering was optimized for application to very short-term wind forecasting and applied
to wind energy for a site at Varese Ligure in Italy by Cassola and Burlando [16]. A numerical
meteorological model BOLAM (Bologna Limited Area Model) was used, and the results with the
application of Kalman filtering showed a considerable reduction in error.
On the basis of parameter selection and data decomposition, two combined strategies and four
modified models based on the first-order and second-order adaptive coefficient (FAC and SAC) were
proposed by Zhang et al. [17] for wind speed forecasting in four different sites in China. It was shown
that the approaches derived from the combined strategies obtained higher prediction accuracy than
the individual FAC and SAC models at the four sampled sites.
A hybrid model based on the EMD and ANNs named EMD-ANN for wind speed prediction
was proposed by Liu et al. [18]. The results were compared to an ANN model and an autoregressive
integrated moving average model. These showed that the performance of the model was very good
compared to the individual methods.
Two prediction methods were studied by Peng et al. [19] for short-term wind power forecasting
on a wind farm. Three key factors were used in the models: temperature, wind speed and direction.
One was an artificial neural network and the other a hybrid model based on physical and statistical
methods. The hybrid model produced higher accuracy results than the individual ANN model.
Chen and Yu [20] developed a hybrid model that integrated a support vector regression
(SVR)-based state-space model with an unscented Kalman filter (UKF). This was to predict short-term
wind speed sequences. The results gave much better performance for both one step and multi-step
ahead wind speed forecasts than support vector machine, autoregressive and ANNs.
Hocaoglu et al. [21] developed a model for the artificial prediction of wind speed data, from
atmospheric pressure measurements using the hidden Markov models (HMMs) technique. The model
accuracy was evaluated from Weibull distribution parameters. The relevance of the technique is in its
use of an additional meteorological variable (atmospheric pressure).
Hocaoglu et al. [22] used the Mycielski algorithm for wind speed forecasting. The algorithm
performs a prediction using the total exact history of the data samples. The basic idea of the algorithm
was to search for the longest suffix string at the end of the data sequence that had been repeated at
least once in the history of the sequence. It was concluded that the model was robust for different
behaviors of wind speed patterns. Experimental results also showed that the model not only provided
very consistent time variations in agreement with the actual measured data, but also provides accurate
distribution model parameters for estimating the wind power potential of a region.
Of all of the models reviewed, only two use additional meteorological parameters other than
wind speed, such as pressure and temperature.
In this work, wind speed forecasting for La Mata, Oaxaca, and Metepec, Hidalgo, was carried
out using univariate and multivariate models. To achieve higher accuracy forecasts, wind speed
models using non-linear auto-regressive exogenous (NARX) modeling were developed. This technique
uses additional exogenous variables (i.e., other than wind speed) to generate more accurate forecasts
with respect to ARIMA models solely based on wind speed time series. The meteorological variables
used in this study were: wind speed and direction, solar radiation, temperature and pressure. In the
generation of the NARX model, only solar radiation or relative humidity was used due to the results
from a correlation study.
Energies 2016, 9, 109 4 of 15
2. Experimental Data
Two weather databases were used as sources of information to allow the wind speed prediction
models to be tested under a range of conditions. The time series of the variables used in this analysis
are shown in Figures 1–6.
Figure 1. Wind speed time series from La Mata, Oaxaca (hourly averages), and Metepec, Hidalgo
(ten-minute averages).
(a) (b)
Figure 2. Wind rose of the studied sites. (a) La Mata, Oaxaca; (b) Metepec, Hidalgo.
One data set was from observations in the town of La Mata in the state of Oaxaca, Mexico. This
was provided by the Mexican Federal Electricity Commission (Comision Federal de Electricidad (CFE))
and has 8759 data points corresponding to one year of hourly averaged measurements taken at 40 m
above ground level from 1 May 2006 to 30 April 2007. The measurements were: wind speed, WS (m/s);
wind direction, WD (◦ ); barometric pressure, P (mbar), air temperature, T (◦ C), and solar radiation,
SR (W/m2 ).
The other data set was from observations in the town of Metepec in the state of Hidalgo, Mexico.
This was provided also by the CFE and has 68,550 data points which correspond to just over a year
and three months of ten minutely averaged measurements. The measurements were made at a height
Energies 2016, 9, 109 5 of 15
of 50 m above ground level from 22 November 2007 to 12 March 2009. The measurements were: wind
speed, WS (m/s); wind direction, WD (◦ ); barometric pressure, P (mbar), air temperature, T (◦ C), and
relative humidity, RH (%).
Figure 1 shows the measured wind speed for both stations. It can be appreciated that there are
no tendencies for neither periodic nor seasonal wind speed variations in the time series. The average
wind speeds are 10.9 and 5.2 m/s for La Mata and Metepec, respectively.
Figure 2 shows the wind rose for both sites, where 0◦ , 90◦ , 180◦ and 270◦ denote North, East,
South and West directions, respectively. In the case of La Mata, Oaxaca, the dominant wind direction
is from the South (S). It should be noted that the wind direction is in the range from 335◦ to 15◦ for
59.4%. Periods of calm (<1 m/s) represent 1.56% of the total sample. In the case of Metepec, Hidalgo,
the dominant wind direction is from West–Northwest (W–NW). It should be noted that 53.4% of the
time the wind directions is in the range from 85◦ to 135◦ . Periods of calm (<1 m/s) represent 5.31% of
the total sample.
Figures 3 and 4 show hourly average air temperature and barometric pressure, respectively, for
both sites. The solar radiation series shown in Figure 5 for the La Mata site of course shows a daily
cycle. The relative humidity for the Metepec site is shown in Figure 6.
Figure 3. Air temperature time series from La Mata, Oaxaca (hourly averages) and Metepec, Hidalgo
(ten-minute averages).
Figure 4. Barometric pressure time series from La Mata, Oaxaca (hourly averages), and Metepec,
Hidalgo (ten-minute averages).
Energies 2016, 9, 109 6 of 15
Figure 5. Solar radiation time series from La Mata, Oaxaca (hourly averages).
Figure 6. Relative humidity time series from Metepec, Hidalgo (ten-minute averages).
Tables 1 and 2 give the basic statistical characteristics (Central tendency and dispersion) for each
of the measured variables for La Mata and Metepec respectively. The calculation of the mean wind
direction and standard deviation requires special attention because the wind direction is a circular
function resulting in a discontinuity (0◦ –360◦ ), so that the arithmetic mean cannot be used. Therefore,
the mean wind direction was calculated using the arctangent function of the averages of the sine and
cosine of the wind directions data.
Table 1. Descriptive statistics of the involved variables of La Mata. WS, wind speed; WD, wind
direction; T, temperature; P, pressure; SR, solar radiation.
Table 2. Descriptive statistics of the involved variables of Metepec. RH, relative humidity.
the wind speed as a variable. The multivariate method uses a non-linear autoregressive exogenous
(NARX) model using wind direction, air temperature, barometric pressure, solar radiation and relative
humidity, in addition to wind speed.
Multivariate analysis allows simultaneous consideration of diverse datasets allowing optimal
decisions to be made considering all of the information.
where φi for the purpose of stabilizing the variance, i is the i-th autoregressive parameter, θ j is the j-th
moving average parameter and et is the error term at time t.
ARIMA models are used in a wide range of applications from engineering to economics. In cases
such as the prediction of power demand, wind speed and stock market value behavior, that is
things that can be represented as a time series with sufficient measurements, these can be modeled by
this technique.
The Box-Jenkins method was followed to model the time series from La Mata and Metepec. This
is basically a three-step iterative process: model identification, parameter estimation and diagnostic
checking [24]:
should be transformed to stabilize the variance. In the following Estimation step, the parameters in
potential models were calculated, and suitable criteria were used to select the best model (Figure 8).
Finally, ACF and PACF were used to test the residuals as the Diagnostic Checking stage. Normality
and “t” tests were applied to the residuals to find their closeness to white noise (Figures 9 and 10).
For La Mata and Metepec, the ARIMA models were ARIMA(1,1,0) and ARIMA(1,1,1), respectively.
Table 3 shows the details of each of the two models. In the case of La Mata, the coefficient corresponding
to the first term of equation: 1.1 yt−1 indicates the importance of the wind speed one hour earlier. This
shows that the wind speed is persistent in this region. In the case of Metepec, the second term of the
model apparently has a higher relevance: 0.6258 yt−2 ; however, the third term, the error coefficient
0.660 et−1 , which only involves a delay, has a similar order of magnitude. These two terms are thus of
similar importance.
Energies 2016, 9, 109 9 of 15
Histograms of the residuals between the models and the measured time series for La Mata and
Metepec are shown in Figures 9 and 10, respectively. The histograms show the form of the residuals’
distribution. It can be seen qualitatively that both cases are sharp normal distributions. The Metepec
database has a lower scatter and higher frequencies because of the larger dataset. Both histograms are
symmetrical, and this is borne out by the measures of central tendency, such as the mean, median and
mode coinciding. This shows that there are no other patterns present in the data.
These two lines make up the input neural layer of a multilayer perceptron [26]. The following
expression describes the model’s dynamic behavior:
n
1
MSE =
n ∑ e2t (7)
t =1
Hindcasts are a way of evaluating the difference between model results and measurements.
The MAE is a measure of the average of the absolute error whose advantage is that it is easier for
non-specialists to understand. MSE is similar, but the values are all positive due to the squaring; this
makes it easier to use in an optimization technique [28]. To achieve a higher degree of certainty when
comparing models, the MAPE was calculated by means of the percentage error (PEt ) and the mean
percentage error (MPE) using the following expressions:
Yt − Ft
PEt = · 100 (8)
Yt
n
1
MPE =
n ∑ PEt (9)
t =1
n
1
MAPE =
n ∑ | PEt | (10)
t =1
The MAPE makes the comparison of results between from the two models easier because it is
percentage based. This gives an indication of the size of the prediction errors in comparison to the
measured values in the series.
To see how the prediction models fit to the real data, Figures 13 and 14 show the last 50 h of
data allowing a qualitative comparison. These correspond to 50 data points for La Mata and 300 for
Metepec. In these figures, the solid curve is the measured data, the long dashed curve is the ARIMA
modeling and the shorter dashed curve is the NARX modeling. From these figures, the importance of
the meteorological values of the previous time step (previous hour for La Mata and previous 10 min
for Metepec) is obvious for one step ahead wind speed prediction, indicating the persistence of the
wind speed on the short-term for these sites.
The mean absolute error and the mean squared error were used to quantitatively evaluate which
model best predicts the wind speed behavior. Table 5 shows these measures of forecasting error.
The improvement of the NARX model over the univariate ARIMA model based on the MAE and MSE
results was calculated using the following equations:
MAE ARI MA − MAEN ARX
PMAE = · 100 (11)
MAE ARI MA
MSE ARI MA − MSEN ARX
PMSE = · 100 (12)
MSE ARI MA
Energies 2016, 9, 109 13 of 15
La Mata Metepec
Model
MAE MSE MAE MSE
ARIMA 0.91 1.51 0.44 0.39
NARX 0.86 1.35 0.43 0.34
Improved Percentages 5.5% 10.6% 2.3% 12.8%
From Figures 13 and 14, it can be seen that both models are effective, but it is not obvious which is
the best. The error measures in Table 5 confirm the satisfactory performance of both models; however,
it is clear that the NARX model is significantly better than the ARIMA model. The MAE of the NARX
model for La Mata is 5.5% better than the ARIMA model and 2.3% better for Metepec. The MSE
Energies 2016, 9, 109 14 of 15
percentage improvements are better than for MAE with values of 10.6% and 12.8% for La Mata and
Metepec, respectively.
The additional information provided by more meteorological variables and the non-linear nature
of the multivariate NARX model explain the superiority of this model.
5. Conclusions
A procedure to analyze and predict wind speed using standard meteorological variables
was developed. Firstly, using traditional statistical techniques, such as the ARIMA model, and,
secondly, by using a multivariate artificial neural network technique: the NARX model. Wind speed
predictions given by both models were analyzed and compared qualitatively and quantitatively with
measured data.
The results obtained show reasonable one step ahead wind speed prediction can be made with
the univariate ARIMA model. However, by using a multivariate NARX model, more accurate results
were obtained. The inclusion of additional meteorological variables is thus recommended in wind
speed forecasting models if they are available.
As well as being a multivariate model, the NARX neural network is a class of discrete-time
non-linear techniques that can represent a variety of non-linear dynamic systems, as in the case of
wind speed time series.
Acknowledgments: We are thankful to Comision Federal de Electricidad (CFE) for allowing the use of the
databases of the measurements of the variables used in this study.
Author Contributions: Erasmo Cadenas: Proposed the original idea for the paper, compiled the data bases,
carried out the calculations for wind speed and wrote the original manuscript. Replied to referees’ comments.
Wilfrido Rivera: Proposed the original paper and contributed to the results analysis and discussion. Replied
to referees’ comments. Rafael Campos-Amezcua: Contributed to the compilation of the data bases, results
analysis and discussion, the writing of the manuscript and replies to the referees’ comments. Christopher Heard:
Contributed to the writing of the manuscript and the improvement of the grammar and English style, the
bibliographic review and replies to the referees’ comments.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Denmark Broke World Record for Wind Power in 2015. Available online: http://www.theguardian.com/
environment/2016/jan/18/denmark-broke-world-record-for-wind-power-in-2015 (accessed on 16 February
2016).
2. Global Wind Report. Annual Market Update, 2014; Global Wind Energy Council: Brussels, Belgium, 2015.
3. Cadenas, R.; Saldivar, G. Wind power plant La Venta II. Available online:
http://www.revista.unam.mx/vol.8/num12/art90/int90.htm (accessed on 16 February 2016).
4. Reglas de Despacho y Operacion del Sistema Electrico Nacional; Centro Nacional de Control de Energía:
Mexico City, Mexico, 2001. (In Spanish)
5. Cadenas, E.; Rivera, W. Short term wind speed forecasting in La Venta, Oaxaca, México, using artificial
neural networks. Renew. Energy 2009, 34, 274–278.
6. Salcedo-Sanz, S.; Perez-Bellidoa, A.; Ortiz-Garcia, E.; Portilla-Figueras, A.; Prieto, L.; Paredes, D. Hybridizing
the fifth generation mesoscale model with artificial neural networks for short-term wind speed prediction.
Renew. Energy 2009, 34, 1451–1457.
7. Cadenas, E.; Jaramillo, O.; Rivera, W. Analysis and forecasting of wind velocity in Chetumal, Quintana Roo,
using the single exponential smoothing method. Renew. Energy 2010, 35, 925–930.
8. Li, G.; Shi, J. On comparing three artificial neural networks for wind speed forecasting. Appl. Energy 2010,
87, 2313–2320.
9. Liu, H.; Tian, H.; Chen, C.; Li, Y. A hybrid statistical method to predict wind speed and wind power.
Renew. Energy 2010, 35, 1857–1861.
10. Cadenas, E.; Rivera, W. Wind speed forecasting in three different regions of Mexico, using a hybrid
ARIMA-ANN model. Renew. Energy 2010, 35, 2732–2738.
Energies 2016, 9, 109 15 of 15
11. Kavasseri, R.; Seetharaman, K. Day-ahead wind speed forecasting using f-ARIMA models. Renew. Energy
2009, 34, 1388–1393.
12. Li, G.; Shi, J.; Zhou, J. Bayesian adaptive combination of short-term wind speed forecasts from neural
network models. Renew. Energy 2011, 36, 352–359.
13. Liu, H.; Erdem, E.; Shi, J. Comprehensive evaluation of ARMA-GARCH(-M) approaches for modeling the
mean and volatility of wind speed. Appl. Energy 2011, 88, 724–732.
14. Guo, Z.; Zhao, W.; Lu, H.; Wang, J. Multi–step forecasting for wind speed using a modified EMD-based
artificial neural network model. Renew. Energy 2012, 37, 241–249.
15. Liu, H.; Tian, H.; Li, Y. Comparison of two new ARIMA-ANN and ARIMA-Kalman hybrid methods for
wind speed prediction. Appl. Energy 2012, 98, 415–424.
16. Cassola, F.; Burlando, M. Wind speed and wind energy forecast through Kalman filtering of numerical
weather prediction model output. Appl. Energy 2012, 99, 154–166.
17. Zhang, W.; Wu, J.; Wang, J.; Zhao, W.; Shen, L. Performance analysis of four modified approaches for wind
speed forecasting. Appl. Energy 2012, 99, 324–333.
18. Liu, H.; Chen, C.; Tian, H.; Li, Y. A hybrid model for wind speed prediction using empirical mode
decomposition and artificial neural networks. Renew. Energy 2012, 48, 545–556.
19. Peng, H.; Liu, F.; Yang, X. A hybrid strategy of short term wind power prediction. Renew. Energy 2013, 50,
590–595.
20. Chen, K.; Yu, J. Short-term wind speed prediction using an unscented Kalman filter based state-space support
vector regression approach. Appl. Energy 2014, 113, 690–705.
21. Hocaoglu, F.; Gerek, O.; Kurban, M. A novel wind speed modeling approach using atmospheric pressure
observations and hidden Markov models. J. Wind Eng. Ind. Aerodyn. 2010, 98, 472–481.
22. Hocaoglu, F.; Fidan, M.; Gerek, O. Approach for wind speed prediction. Energy Convers. Manag. 2009, 50,
1436–1443.
23. Burton, T.; Jenkins, N.; Sharpe, D.; Bossanyi, E. Wind Energy Handbook, 2nd ed.; John Wiley & Sons, Inc.:
Hoboken, NJ, USA, 2011.
24. Box, G.; Jenkins, G. Time Series Analysis: Forecasting and Control; Holden Day: San Francisco, CA, USA, 1970.
25. Gao, Y.; Er, M. NARMAX time series model prediction: Feedforward and recurrent fuzzy neural network
approaches. Fuzzy Sets Syst. 2005, 150, 331–350.
26. Gardner, M.; Dorling, S. Artificial neural networks (the multilayer perceptron)—A review of applications in
the atmospheric sciences. Atmos. Environ. 1998, 32, 2627–2636.
27. Diaconescu, E. Use of NARX neural networks to predict chaotic time series. WSEAS Trans. Comput. Res.
2008, 3, 182–191.
28. Makridakis, H.; Wheelwright, S. Forecasting Methods and Applications, 3rd ed.; John Wiley & Sons, Inc.:
Hoboken, NJ, USA, 1998.
c 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons by Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).