Exogenous Measurements from Basic Meteorological Stations for Wind Speed Forecasting

This research presents a comparative analysis of wind speed forecasting methods applied to perform 1 h-ahead forecasting. The main significant development has been the introduction of low-quality measurements as exogenous information to improve these predictions. Eight prediction models have been assessed; three of these models [persistence, autoregressive integrated moving average (ARIMA) and multiple linear regression] are used as references, and the remaining five, based on neural networks, are evaluated on the basis of two procedures. Firstly, four quality indices are assessed (the Pearson’s correlation coefficient, the index of agreement, the mean absolute error and the mean squared error). Secondly, an analysis of variance test and multiple comparison procedure are conducted. The findings indicate that a backpropagation network with five neurons in the hidden layer is the best model obtained with respect to the reference models. The pair of improvements (mean absolute-mean squared error) obtained are 29.10%–56.54%, 28.15%–53.99% and 4.93%–14.38%, for the persistence, ARIMA and multiple linear regression models, respectively. The experimental results reported in this paper show that traditional agricultural measurements enhance the predictions. Energies 2013, 6 5808


Introduction
With the deterioration of the environment and depletion of conventional resources, it has become imperative and advisable to search for alternative energy resources that are sustainable, clean and environmentally respectful.As a result, most industrialized countries have adopted policies to increase installed power with renewable energy power plants in order to comply with international environmental agreements.In particular, at the European Council in March 2007, the European Union endorsed a mandatory target of a 20% share of energy from renewable sources in overall community energy consumption by 2020.Also imposed on the Member States are individual targets in order to enable them to decide on their own, the preferred energy mix of alternative energy sources to fulfill this objective [1].
Among renewable energy sources, wind power is undoubtedly the one that has experimented the greatest growth over recent years, turning into a base pillar of the energy system in many countries and, thus, becoming the true alternative to fossil fuels.As a result, the worldwide installed wind power capacity has increased considerably in these early years of the 21st century, and it is estimated that by 2013, installed wind power worldwide would amount to 318.0 GW, compared to 120.0 GW at the close of 2008 [2].
In this regard, the principal feature of wind power concerning its integration into the grid is that it is not programmable.In contrast to conventional energy sources, wind power production cannot be specified beforehand, but depends on the incoming wind on various wind farms.Moreover, if wind power production is not known with sufficient accuracy, the power system regulators should make detailed schedule plans and set reserve capacity to prevent the possible fluctuation of this.To reduce the reserve capacity and increase the penetration of wind power, accurate forecasting of wind speed is needed [3].
Numerous studies about wind speed prediction or wind power prediction can be found in the scientific literature.The calculation of wind power production is carried out by means of the relationship between electric power production and wind speed, obtained empirically for the park itself or using the power curve provided by the manufacturer of the wind turbine.In addition to the variable of prediction, the forecasting models are classified according to the prediction horizon, the forecasting methodology and the type of data.
Short-term predictions are usually based on time series analysis, from simpler (autoregressive models, linear multiple regression and persistence) to complex structures (computational fluid dynamic, neural networks and fuzzy logic).Taylor et al. [4] compared the methods.Lei et al. [5] gave a survey on the general background and developments in wind speed and wind power forecasting.This study is completed by consulting Foley et al. [6], who gave a review of the current methods and advances.
The availability of many measuring points of wind speed might be conducive to the performance of works, like [20,21].Bilgili et al. [22] made a study to predict the monthly mean wind speed in the eastern Mediterranean region of Turkey using high-quality measurements.
This work uses exogenous data, from standard weather stations in the surroundings of the target area, to improve short-term forecasting of ANNs.The value-added novelty of the paper lies in the fact that these stations are not yet considered by the World Meteorological Organization [23].In this sense, these papers establish a scientific proposal and a method for them to be included [24,25].
The paper is organized in the following way: Section 2 presents the region and the raw data from the on-site equipment; Section 3 summarizes the theoretical framework; the experimental procedure is outlined in Section 4; and results are analyzed in Section 5; finally, conclusions are drawn in Section 6.

Data Description
The target 20 m-high wind measurement tower, located at Universal Transverse Mercator coordinates (294,284.4175161) in Northern Andalusia (Peñaflor, Sevilla, Spain), acquires data at 10 min intervals, covering the period from September 2007 to August 2008.Figure 1 shows the wind singularities at the target station through a wind speed graph, and the frequency and speed rose (only 500 records are displayed for better visualization).

Frequency rose
To improve predictions, the ten closest stations have been selected within a radius of 40 km, and initially conceived to measure agriculture variables (Andalusian agriculture-climate information network [26]), providing hourly measurements.Wind records are not reliable enough, because most of them are located in the open air, being highly affected by obstacles (the anemometer's height is 2 m).
Figure 2 shows the location of the target and reference stations.Table 2 summarizes the geographical coordinates, average wind speed, distances and correlation coefficients of the stations with respect to the target.Note that the order shown in the Table 2 is from lowest to highest separation distance relative to the target station.

Methodology Framework
This paper describes short-term wind speed forecasting techniques based on various soft computing models.The value of the present research resides in the use of exogenous data from standard on-site measurement stations to improve prediction.This procedure is depicted in Figure 3.The best model

Prediction Models
In this article, eight forecasting methodologies are treated.Three of them [persistence, autoregressive integrated moving average (ARIMA) model and multiple linear regression model] are references, and the remaining, based on ANNs, constitute the main goal of this work.The reference models selected are as follows: • The persistence model is the most common reference forecasting method for prediction horizons up to 3-6 h [6,27,28].It states that the predicted value at t is similar to the last measurement (ŷ t = y t−1 ); it needs neither the parameters' estimation nor exogenous variables.
• ARIMA (p, d, q) constitutes a three-phase method (identification, estimation and diagnosis) for selecting an appropriate model for estimating univariate data, ranging from stationary to non-stationary and seasonal time series, and has been used extensively, e.g., [7,18,29].
• The multiple linear regression (MLR) model is a highly flexible system for examining the relationship between a collection of independent variables and a single dependent variable [30].
The considered soft computing models for wind speed forecasting are the multilayer-perceptron neural network, radial basis function neural networks, recurrent neural networks and the adaptive neuro-fuzzy inference system (ANFIS): • Multilayer ANNs, shown in Figure 4, based on error backpropagation, are the most widely used method [31,32].The network is a dynamical system that changes with the learning rule, which sequentially finds the weights that codify the knowledge.The network will then have a generalization capability that must be measured.The Levenberg-Marquardt algorithm has been chosen, minimizing the mean squared error (MSE) of the difference between the output in each step.• Radial basis function (RBF) networks have similar forms to that of a two-layer multilayer network (Figure 5) [33].The fundamental difference arises in the hidden neurons and operates on the distance between an input with respect to the synaptic vector (called a centroid).The RBF neurons comprise a localized response, because they only respond with an appreciable intensity when presented with an input vector and the centroid of the neuron belonging to a nearby area in the input space.RBF training is comprised of two stages: the first is unsupervised and accomplished by obtaining cluster centers of the training set inputs; and the second one consists of solving linear equations.• The Elman neural network is characterized by feedback from the hidden layer output to the first-layer input, which constitutes the context layer that retains information between observations [9,29,34].This recurrent connection allows the Elman network to detect and generate time-varying patterns (Figure 6).• ANFIS is a hybrid of two intelligent system models and combines the low-level computational power of a neural network with the high-level reasoning capability of a fuzzy inference system.A demonstration of the use of ANFIS can be found in [9,11,29], and Figure 7 depicts the ANFIS structure.

Comparison Procedure
Different error criteria have been proposed and used in the literature, but no single error has been proven to be the universal measure.Therefore, two strategies have been used to select the best model.The first strategy is based on quality indexes, and the second one performs a statistical analysis based on a multiple comparison procedure.
The error criteria considered in this paper are: Pearson's correlation coefficient (ρ), the index of agreement (IOA), the mean absolute error (MAE) and the MSE: (2) where σ yt ŷt is the covariance between y t (observed data) and ŷt (predicted data); σ yt and σ ŷt are their respective standard deviations; and N is the number of data used for performance evaluation and comparison.
The second strategy consists of a statistical analysis, based on multiple comparison, whose objective it is to choose which model or models are best.This analysis is composed of two tests: the parametric test [analysis of variance (ANOVA)] and the Bonferroni adjustment [35].The first one assesses the differences among more than two samples, and the second obtains the best averaged model.

Development of the Proposed Models
After the experimental procedure has been explained in Section 3, the reference models (persistence and ARIMA model) are assessed using the wind speed time series acquired at the target station (the wind speed in Peñaflor).
The implementation of the persistence model does not present any problems.However, the ARIMA model needs preliminary actions to establish the order p, d and q which best fit the target time series.The Box-Jenkins methodology is adopted to identify these parameters.In the Box-Jenkins methodology, a differencing approach is used to stabilize the original data, and both an autocorrelation function (ACF) and a partial autocorrelation function (PACF) are utilized to decide the autoregressive or moving average component, which should be included in the ARIMA model.The choice of the appropriate values of the model is based on the ACF and PACF characteristics shown in Table 3.With this goal, the autocorrelation coefficients and the partial autocorrelation coefficients are evaluated and depicted in Figure 8. Finite: cut off at lag q Infinite, but convergent ARMA (p,q) Infinite: exponential and/or sine-cosine wave decay Infinite: exponential and/or sine-cosine wave decay The autocorrelation coefficients shown in Figure 8 decay as the time-lag increases, but they have positive correlations for many time-lags.This conveys the idea that one order of differentiating is needed [36].• PACF cuts beyond the second lag.According to the above explanations, ARIMA (2,1,0) should be selected; • ACF decays in the second lag.Thus, ARIMA (0,1,2) is selected; • ACF and PACF have a decreasing and oscillating phenomenon that begins in the second lag.Consequently, ARIMA (2,1,2) has been selected.In the estimation process, the errors of each model [ARIMA (2,1,0), ARIMA (0,1,2), and ARIMA (2,1,2)] have been calculated along with their parameters attending to the maximum likelihood paradigm.The best model is selected with the help of Akaike and Bayesian information criteria [37]: ARIMA (2,1,0) with coefficients φ 1 = −0.2352and φ 2 = −0.0922.
The last step of the ARIMA process is the diagnosis testing.In this step, the assumptions on the residuals are validated, so that the model can be used to forecast.
For the predictor, MLR, which is based on a statistical linear model, no tuning parameters are needed, and the inputs used for training and validation are the same as those used in the ANNs.
The following models are tested in this paper: backpropagation network with one and two hidden layers (BP1 and BP2), RBF network, Elman neural network (ELM) and ANFIS.The following premises are considered: • the reference stations in Section 2 have been used as exogenous variables to improve the prediction; • data are normalized, so that they are in the interval [−1, 1] for a faster computation; • the dataset was divided into three subsets: training, evaluation and test sets.The training and validation sets, with 70% and 15% of the data, respectively, were used for ANN model building; and the third set, with the last 15%, was used to test the predictive power of a model on the out-of-sample set.The building model is performed in two phases: the first one uses the training set to obtain the parameters for non-linear predictors; and, the second phase uses the validation set to choose the optimum; • the training of the tested networks is carried out until the validation error starts increasing.At this point, we stop the training, and the performance of the network is measured in the test.
One hundred experiments have been launched for each model in order to achieve statistically meaningful results which rule out the random factors influencing the ANN, and we keep the best results.
Hereinafter, we detail the particularities of the models.Table 4 collects parameters, corresponding to the network architecture and activation functions of neural networks.The rule to select the range of neurons in the hidden layers for BP1, BP2 and ELM models is as follows.The number of neurons in the first hidden layer is the mean of the neurons between the input and output layers, and the number of neurons in the second hidden layer is one half of the neurons in the first hidden layer [38].As the maximum number of inputs is eleven and the number of outputs is one, the ranges for the first and second hidden layers are [4,10] and [2,5], respectively, as shown in Table 4.
For the RBF model, we must specify the appropriate value of the Gaussian Kernel width or spread.The higher the value assigned to this parameter, the smoother the approximation function.Too large a spread means a lot of neurons are required to fit a fast-changing function.Conversely, too small a spread means many neurons are required to fit a smooth function, and the network might not generalize well.Different training architectures were analyzed depending on the spread and the number of hidden neurons.We specify the first variable in the range of [1,20] and the second variable between 1 and 150 neurons.
A total of 500 ANFIS models have been designed considering different combinations of inputs and rules.Similar to ANNs, the input space is divided into two groups: one with 70% of the data for the training and the other with the remaining 30% for validation of the model.Table 5 resumes the parameters used to specify the assessed ANFIS models.In all tested models (ANNs and MLR), the wind speed and direction data of the reference stations have been applied as exogenous variables.The prediction of the objective variable at an instant, t, is done using the exogenous data of a previous time (t−1).Then, data are presented in an autoregressive matrix.For example, the autoregressive matrix shown in Table 6 is associated with wind speed prediction supported by four exogenous variables.
Figure 10 shows the evolution of the performance of the models based on the MSE criterion and the type of variable used.As can be seen, first, the simulations begin without exogenous variables, and then, they are progressively inserted from one to ten stations (from highest to lowest correlation).In principle, simulations only consider the wind speed time series at each station as exogenous variables; then, both wind speed and direction are applied.On the basis of a visual assessment, it is evident that the insertion of wind direction data does not lead to the improvement of the performance of the models; so henceforth, only the wind speed data for each station will be taken into account as exogenous variables.Each model's performance is summarized by the former four quality indexes.

Results
The number of models tested in this work amounts to more than 3200 configurations, attending to their typology (BP, RBF, ELM or ANFIS), tuning parameters and exogenous variables used in the simulations: 98.72% of them are better than persistence, 98.60% are better than the ARIMA (2,1,0) model and 41.30% are better than the MLR model.For the sake of comparison, Table 7 shows the results of the quality indexes on the test set obtained by the ten best models that generated the smallest MSE and also the values associated with the reference models [persistence, ARIMA (2,1,0) and MLR].Note that these results are ordered in terms of the MSE criterion, because this is the objective function adopted in the learning process of the models.As shown in Table 7, the results of the models supported by low-quality stations significantly improve the reference models: the persistence, ARIMA (2,1,0) and MLR.The best of all the models tested is a backpropagation network with one hidden layer (BP1) and five neurons within the hidden layer.There are seven inputs for this model: one of them is the wind speed of the target station, and the rest are exogenous speed values registered at the six closest stations.This model reduces the MAE and MSE with respect to the persistence model by 29.10% and 56.54%, respectively.The percentages of improvement over the ARIMA (2,1,0) model are 28.15% and 53.99%, and the percentages of improvement over the MLR model are 4.93% and 14.38%.None of the ten best models uses directions as exogenous variables, indicating that this information is not significant.
Once the analysis based on quality indexes has been completed, a multiple comparison procedure is carried out with the best models selected in the Table 7.As was discussed in Subsection 3.2, the analysis is composed of the parametric test (ANOVA) and the Bonferroni adjustment.
Table 8 shows the results, in terms of MSE, obtained for each of the ten best models selected in Table 7.One hundred experiments were performed for statistical analysis.As illustrated in Table 8, the model, BP 1(5) 6 ExogVar, proves the robustness and capacity for enhancing the prediction errors of the reference models.This is confirmed by means of the statistical tests (ANOVA and Bonferroni's test).As mentioned before, results show that the best model using the two selection processes (quality indexes and statistical analysis), called BP 1(5) 6 ExogVar, enhances the effectiveness of the forecast through exogenous information received from agricultural measurement stations.
Figure 11 depicts the performance of the best model, i.e., the one shown in Table 7 with the best quality indexes.The same graph also shows the three reference models (persistence, ARIMA and MLR).The data belong to the test set, and only 50 h are shown for a better visualization of the results.As expected, the time series by the best model based on the ANN is clearly more accurate than the reference ones.The overall improvements obtained by the selected neural network models on the reference models are shown in Figure 12.In addition, the improvements of the models are also depicted, ranked by percentage.Once the best model has been selected through the two procedures, it is assessed in terms of the number of exogenous variables used.Hence, we started using the time series registered in the target as the sole input (without exogenous variables).Then, one by one, the exogenous time series were used to evaluate their benefits.As is possible to see in Figure 13, the MAE and MSE indexes are improved by the inclusion of exogenous measurements.The achievement is optimum using the speed registers from the six nearest stations.

MAE MSE
Finally, and to complete this study, the best architecture (BP1: five hidden neurons) is analyzed in the function of the forecasting horizon.The assessment is based on the forecasting horizons of 1, 2, 4 and 6 h.As Figure 14 shows, the MSE and MAE indexes get worse as the forecasting horizon progresses, although the best model still exceeds the three reference models.

Conclusions
The main contribution of this work is that wind speed predictions in a target station can be significantly improved by the use of low-quality wind data acquired at surrounding agriculture measurement stations.The originality of this study concerns the application of stations which are generally excluded from wind assessment.These low-quality data are excluded because of their low reliability, and the measurements locations are not optimal.
In the present case, the best model to perform 1 h-ahead prediction is a backpropagation network with one hidden layer and five neurons.This model uses six exogenous speed values and meaningfully improves the three witness models according to four quality indexes: ρ, IOA, MAE and MSE.The main enhancements are more visible in MAE, which achieves a value of 0.65 m/s, and MSE with 0.67 m 2 /s 2 .The percentages of improvement with respect to the reference models are 29.10% and 56.54% for the persistence, 28.15% and 53.99% for the ARIMA (2,1,0) and 4.93% and 14.38% for the MLR model.
In summary, it might be concluded that data from low-quality stations, which are ruled out because of low reliability and do not achieve the World Meteorological Organization requirements, provide useful information for the control of a wind farm.This information optimizes the control in two aspects: short-term predictions and wind farm maintenance.This fact could benefit the insertion of this renewable energy source into areas where the wind energy represents a high percentage of the total electric power.

Figure 1 .
Figure 1.Wind characteristics at the target station.

Figure 2 .
Figure 2. Map of the studied zone.T = target station.

Figure 3 .
Figure 3. Graphical abstract of the paper and the research.

Figure 4 .
Figure 4. Example of a single hidden layer neural network.

Figure 9
Figure 9 shows the prospective results of the model from the first derivative time series.The selection of three potential models has been established based on the inspection of this graph:

Figure 9 .
Figure 9. Differentiated original time series wind speed, ACF and PACF plots at Peñaflor.

Figure 10 .
Figure 10.Evolution of the performance of the models.

Figure 11 .
Figure 11.Original time series and forecasted data with the best obtained model.

Figure 12 .
Figure 12.Overall improvements of the models.

Figure 13 .
Figure 13.Evolutions of the mean absolute error (NAE) and mean squared error (MSE) errors obtained by the best model in the function of the exogenous variables used.

Figure 14 .
Figure 14.Evolutions of the MAE and MSE errors obtained by the best model in the function of the forecasting horizon.

Table 1 .
Table 1 lists the main characteristics of the measured variables.Main characteristics of the target station.

Table 2 .
Information of the used stations in this study.

Table 4 .
Parameters of the network models used.BP: backpropagation; RBF: radial basis function; ELM: Elman neural network.

Table 6 .
Example of the autoregressive matrix.

Table 7 .
Parameters and results achieved on the test set by the reference models vs. the ten best models.IOA: index of agreement; MAE: mean absolute error; MAPE: mean absolute percentage error; and MSE: mean squared error.

Table 8 .
Evaluation of MSE for the 10 best models.The results are averaged over 100 runs.

Table 9
shows the models, the corresponding mean errors and the Bonferroni's test results.The order shown in this table is from lowest to highest MSE mean.As we can see, Model 1 belongs to the group of models that do not have significantly different means.Therefore, this model should be selected coinciding with the results obtained in Table7.