A Novel Hybrid Neural Network-Based Day-Ahead Wind Speed Forecasting Technique

As a dominant form of renewable energy sources with significant technical progress over the past decades, wind power is increasingly integrated into power grids. Wind is chaotic, random and irregular. For proper planning and operation of power systems with high wind power penetration, accurate wind speed forecasting is essential. In this paper, a novel hybrid Neural Network (NN)-based day-ahead (24 hour horizon) wind speed forecasting is proposed, where five hybrid neural network algorithms are evaluated. The five algorithms include Wavelet Neural Network (WNN) trained by Improved Clonal Selection Algorithm (ICSA), WNN trained by Particle Swarm Optimization (PSO), Extreme Learning Machine (ELM)-based neural network, Radial Basis Function (RBF) neural network, and Multi-Layer Perceptron (MLP) Neural Network. Single- and multi-features and their effect on the accuracy of wind speed prediction are also analyzed. The wind speed dataset used in this paper is Saskatchewan’s recorded historical wind speed data. Despite the excellent wind power potential, only 6.5% of the total electricity demand is currently supplied by wind power in Saskatchewan, Canada. This study paves the way for economical operation, planning, and optimization of Saskatchewan’s future wind power generation.


I. INTRODUCTION
Wind power plays an increasing role in the modern mixed energy landscape by producing sustainable clean energy and reducing fossil fuel-based conventional power generation. Wind turbines have been installed in windy regions, such as Saskatchewan in Canada, which benefit from ample wind resources due to their geographical locations and terrain features. In Saskatchewan, the currently installed wind power capacity only approximately provides 6.5% of the electricity demand, and the province has the plan to boost wind power up to 40% of the demand by 2030 [1]. However, one challenge associated with wind power is intermittent wind speed. The accurate wind speed prediction methods are urgently needed to improve economical, secure, and reliable operation of wind power systems [2].
The wind speed prediction can be performed in five time scales [3]- [5]: very short-term prediction (a few seconds to 30 minutes ahead); short-term prediction (30 minutes to 6 hours ahead); medium-term prediction (6 hours to 24 hours ahead); long-term prediction (24 hours to one week ahead); and very long-term prediction (one week and longer) [6]. The applications of each time-scale are provided in [3]- [5]. To realize wind speed forecasting, the following four types of methods can be used: physical or weather-based methods; statistical or time-series-based methods; Artificial Intelligence (AI)-based methods; and hybrid methods [4], [5], [7].
Physical or weather-based wind speed prediction methods use the meteorological data of wind power plants. The utilized data include essential factors, such as the land topology, atmospheric temperature, pressure, humidity, obstacles, and surface coarseness. The predicted wind speeds are used with power curves of wind turbines to derive wind power generation. However, these methods lead to high computational burden, and thus, require a long processing time [4].
Statistical or time-series wind speed prediction methods are data-driven approaches utilizing historical wind speed data, and the one-step statistical analysis [7]. Auto Regression Moving Average (ARMA) and Auto Regression Integrated Moving Average (ARIMA) are two common statistical methods being used [8], [9]. The ARIMA-based models can predict wind speed efficiently for short-term applications [9], but their performance is not acceptable for long-term applications, and thus, more improvements are needed. Fractional ARIMA (FARIMA) and Markov chain (with its robustness in sequential data modeling) methods are classified as statistical methods [10], [11]. In [10], the FARIMA method can provide effective results compared to conventional methods for long-term wind speed forecasting, especially for sites with severe intermittent wind speed. Statistical methods are suitable for short-term applications, and their implementation is very easy and straightforward [11], but the issue is they are linear-based predictors, while wind speed is a nonlinear phenomenon. In [12], the Markov chain is employed for short-and medium-term wind speed forecasting.
AI-based wind speed prediction methods, such as Artificial Neural Network (ANN) and Support Vector Machines (SVM), are data-driven and multi-step approaches. Unlike statistical methods (which are linear approaches), AI-based methods are suitable for nonlinear forecasting. In the literature, ANN is primarily used in the short-term wind speed forecasting. Characteristics of the wind speed forecasting problem and considerations of the analysis are two important factors in determining a proper ANN model and the number of neurons [4]. The principal of ANN techniques lies in mapping actual nonlinear input data into forecasted values. Therefore, a suitable training approach is chosen to train the ANN-based model to reach proper weight values based on the recorded data, i.e., an ANN model learns from the gained experience through the training procedure, then uses nonlinear input data to forecast wind speed [13]. A proper number of neurons should be employed depending on the nature of wind speed forecasting problem. By improving the general ANN techniques, new ANN methods are proposed for wind speed forecasting, including Radial Basis Function (RBF) neural networks, Recurrent Neural Networks (RNNs), Feed-Forward Back Propagation (FFBP) neural networks, and Multi-Layer Perceptron (MLP) neural networks [4], [5], [14], [15]. Ref. [16] reveals that RBF and FFBP neural networks are effective in improving ANN's performance by considering the nonlinearity of wind speed. In [16], SVM and MLP neural network are compared for wind speed forecasting, and SVM outperforms MLP neural network in most cases. In [17], five ANN methods are used for short-term wind speed forecasting, among them, the ANN consisting of 19 hidden layers, 4 input layers, and one output layer has the best performance, while a single ANN with a traditional training technique has the limited capability to follow the nonlinear behavior of wind speed.
Hybrid wind speed prediction methods are data-driven approaches and introduced to tackle the shortcomings of physical, statistical, and AI-based methods. Hybrid methods usually include two or more prediction methods, or they can be filter-augmented methods. Filter-augmented methods use a special filter added to a basic method to improve the forecasting performance. Kalman Filter (KF) and Wavelet Transform are two popular filters added to some statistical (ARIMA) and AI-based methods (ANN). In [18], KF and ANN methods are added to the ARIMA method to create the hybrid KF-ANN-ARIMA method. The obtained results reveal that the KF-ANN-ARIMA method is more effective, and successfully improves the efficiency and accuracy of the original ARIMA method in wind speed forecasting. Wavelet Transform is added to ANNs to create the Wavelet Neural Network (WNN), and choosing a proper Wavelet Transform affects the accuracy of wind speed prediction using WNN [4].
Statistical methods have been added to AI-based methods to improve wind speed forecasting performance. In [18], [19], three methods, ARIMA, ANN, and ANN-ARIMA, are compared for wind speed prediction in terms of accuracy, and the hybrid ANN-ARIMA method provides a better accuracy in different time scales than ARIMA and ANN, and thus, is a suitable method for datasets with nonlinear and linear patterns. The fuzzy logic concept was added to the self-learning capability of ANN to improve the performance of wind speed forecasting, which leads to the creation of Fuzzy Neural Networks (FNNs), and Adaptive Neuro-Fuzzy Inference System (ANFIS) [4]. Particle Swarm Optimization (PSO) and Whale Optimization Algorithm (WOA) can also be added to ANN. PSO is combined with ANFIS to form the PSO-ANFIS method to improve the performance of the original ANFIS method. In [20], WOA is added to the Support Vector Regression (SVR) method to improve the accuracy and simplify the implication of SVR, as SVR is basically constrained by the penalty factor (C) and kernel parameter (K). It is found that the WOA-SVR method has better performance than the PSO-SVR method. As well-known training algorithms, Imperialistic Competitive Algorithm (ICA), Improved Clonal Selection Algorithm (ICSA), and Extreme Learning Machine (ELM) are also added to hybrid methods [21]- [23].
Among wind speed prediction methods reported in the literature, hybrid methods are more advanced approaches, but their development is only at the beginning stage. Also, in most wind speed prediction work, single and multiple meteorological factors have not been considered and compared. Another issue is the lack of a comprehensive procedure for wind speed forecasting among published papers. Therefore, this paper aims to address the above mentioned technical deficiencies in wind speed forecasting area.
The major contributions of this paper include:  A novel hybrid neural network-based day-ahead (24 hour horizon) wind speed forecasting technique is proposed in this paper. Five hybrid neural network methods (WNN trained by ICSA; WNN trained by PSO; ELM-based neural network; RBF neural network; and MLP neural network) are evaluated using measured historical wind speed data in Saskatchewan, Canada.  A comprehensive procedure to implement the proposed technique is developed for wind speed forecasting. This procedure includes the training, evaluation, and comparison of the employed hybrid methods.  Two different analyses, single-feature analysis and multifeature analysis, are performed to evaluate the effect of the number of features on accuracy of wind speed forecasting.
The paper is arranged as follows: Section II proposes a novel hybrid neural network-based day-ahead wind speed forecasting technique with a comprehensive procedure. Fundamental theories of the employed hybrid neural networks are provided in Section III. In Section IV, the numerical analysis is conducted. The conclusions are drawn in Section V.

II. THE PROPOSED APPROACH WITH A COMPREHENSIVE PROCEDURE TO PREDICT WIND SPEED
In this section, the main idea of the proposed approach for wind speed forecasting can be implemented in a comprehensive procedure as shown in Fig. 1. It consists of four stages: Stage 1-Data Processing; Stage 2 -Employed Models; Stage 3-Training Approaches; and Stage 4 -Testing and results analysis. The detailed explanations of each stage are given as follows:

1) Stage 1:
This stage is dedicated to processing/preparing wind speed data as training and testing datasets. It consists of three steps as discussed below:  Step 1: The wind speed data is entered into the forecasting system.  Step 2: The meteorological data are split into training and testing datasets. Based on the day-ahead application in this paper (a medium-term prediction), 1440 samples (about 98.4% of the whole data) are used for training, and 24 samples (one day's data, which is about 1.6% of the whole dataset) are used for testing.  Step 3: In this step, the autocorrelation function is used to select the most relevant input set for training. The large set of input elaborates the extraction of input/output mapping function of the forecasting process for the prediction engine, and thus, reduce its performance. Further explanations regarding using the autocorrelation function are provided in Section IV.

2) Stage 2:
In this stage, one model from the five chosen hybrid models is selected to be used for wind speed forecasting. The function of the models and their related activation functions are briefly mentioned in Fig. 1. Further details of each model are discussed in Section III.

3) Stage 3:
In this stage, according to the selected model, an appropriate algorithm must be used to train the forecasting system. Because various models are used in this study, different appropriate algorithms have been designed to train the proposed approach. As shown in Fig. 1, the designed algorithms for WNN, RBF, and MLP models are similar, but the designed algorithm for the ELM model is completely different. Therefore, this stage consists of several steps considering the employed algorithm/model. The details of each step/algorithm are evaluated below. As the employed algorithms for WNN, RBF, and MLP models are similar, the related steps for their algorithms are provided as follows:  Step 1: The parameters of the WNN, RBF, and MLP models are initialized.  Step 2: The training method for each model is chosen.
The two methods, ICSA and PSO, are considered to train the WNN model. The ICSA method is also considered to train the RBF and MLP models, i.e., ICSA is the default training method for the RBF and MLP models.  Step 3: After training the forecasting system, parameters of the models are updated.  Step 4: In this step, the termination criterion is checked.
If the required condition is not met, the algorithms follow the procedure back to Step 2, the system will be re-trained, and parameters will be updated again until meeting the required condition for termination. If the required condition is met, the algorithms move on to Step 5.  Step 5: After reaching the termination criterion, the obtained parameters are considered as the final results for the model parameters, which will be used to form the prediction models. In the ELM-based model, which uses a single-layer FFNN, the ELM training technique is used to train the neural network and determine its parameters. The related steps of this training method are provided as follows:  Step 1: The ELM method is used to train the neural network.  Step 2: The obtained parameters in Step 1 are considered as the final parameters of the ELM model to form the prediction model.

4) Stage 4:
This stage is to test the prediction systems and choose the best approach for wind speed forecasting. It consists of four steps as follows.  Step 1: The testing set is applied to the forecasting system.  Step 2: Wind speed is forecasted based on input data.

Multi-feature Analysis
Temperature Direction Humidit y Speed

III. FUNDAMENTAL THEORIES OF THE EMPLOYED HYBRID NEURAL NETWORK METHODS AND TRAINING APPROACHES
In this paper, five hybrid NN methods based on four basic models, WNN, ELM, RBF, and MLP, are evaluated for dayahead wind speed forecasting using the proposed technique introduced in Section II. Two training approaches, ICSA and PSO, are chosen in this study for these hybrid models. In this section, fundamental theories of the basic hybrid forecasting methods along with ICSA are briefly explained. The details of the WNN method trained by PSO can be found in [24].

1) Wavelet Neural Network:
In the literature, Wavelet Transform is usually used as a preprocessor to decompose time series wind speed to a set of subseries, and to forecast future values of these subseries using statistical or machine learning-based methods; the inverse wavelet transform is then used to construct the original wind speed time series [24]- [26]. In this paper, the wavelet is used to form WNN by considering the wavelet function as an activation function in hidden neurons of the neural network. Mexican hat and Morlet wavelets are two options that can be used as an activation function to predict wind speed [22], [27]. We choose the Morlet wavelet function as shown in Fig. 2. The structure of the WNN is illustrated in Fig. 3. This type of structure is known as a three-layer feed-forward structure, where   1 2 , , , m X x x x   is the input data, and y is the model output, known as a target variable.
The activation function in the hidden layer, which is the multi-dimensional Morlet wavelet function, is as follows: where i a and i b are scale and shift parameters, respectively.
  x  is the Morlet wavelet function. n is the number of hidden neurons of the wavelet neural network. Eventually, the output of the WNN forecasting engine is obtained as follows: In Error! Reference source not found., i w indicates weights between the th i neuron and the output, while the th j input and the output are connected by j V , respectively. The parameters of WNNs, which should be specified by the improved clonal section as a training method, are shown as follows: , , , , , ,.. , ,

2) Extreme Learning Machine-based Neural Network:
ELM is a type of training strategy that is used in a single hidden-layer Feed-Forward Neural Network (FFNN) as shown in Fig. 4 [23], [25], [28].   In ELM, the input weights and biases are randomly selected, and only the output weights are determined mathematically using a simple matrix. Considering various  , the ELM having Y hidden neurons and the activation function (.) q can estimate M samples without errors and can be formulated as follows: In (5), i w is the weight between input data and hidden neurons, and i  is the weight between hidden neurons and output nodes. i b indicates the bias value of the hidden neurons. For the simplicity purpose, Eq. Error! Reference source not found. can be expressed as follows: where H indicates the matrix of the hidden nodes output.  is a matrix connecting hidden nodes to output nodes. T represents the target matrix. As in this strategy, the input weight and hidden neuron biases are produced randomly, the matrix of the hidden neuron output can be specified.
Accordingly, the approximated output weights   can be obtained by minimizing the following objective: The estimated output weight can be computed using the following simple inverse: † H T    (9) In (9), † H indicates the Moore-Penrose generalized inverse of the matrix associated with the output of hidden neurons H . Thanks to this simple matrix, the training speed is very high [28].

3) Multi-Layer Perceptron Neural Network:
The MLP neural network is a FFNN model, which is a specific form of supervised NNs. MLP is capable of creating a mapping function between input datasets and the related output data. The MLP structure consists of multiple stacked layers of nodes, so-called neurons in neural networks. Each layer is connected to the next layer through the neurons in a onedirectional manner. The general structure of the MLP neural network is shown in Fig. 5 [22], [29].
As shown in Fig. 5, the MLP structure consists of three layers: 1) Input layer, which feeds the network with input variables; 2) Output layer, which produces the final outputs; 3) Hidden layers, which are the stacked layers of neurons between input and output layers. The general process of the network is done through neurons, which consist of the activation function. Various activation functions can be used in the MLP structure, such as linear, logarithmic sigmoid, and hyperbolic tangent sigmoid functions. The MLP parameters that are determined through the training process include: weights connecting the input layer to a hidden layer; weights connecting a hidden layer's output to the next hidden layer or the output layer; and hidden biases [22], [29].
Each layer can be represented mathematically as (10). w represents the bias of the neuron i in the layer l . The first layer 0 l  is the input layer of the network, whose output length is 0 n and is characterized as (0) O x  . Also, the last layer l L  is the output layer of the network, whose output length is L n and is characterized as ( ) L O y  . In (10), ()  is the activation function of the network, which is hyperbolic tangent sigmoid function in (11).

4) Radial Basis Function Neural Network:
The RBF neural network is another FFNN model. Its general performance is similar to the MLP neural network discussed previously. The general structure of RBF neural networks is shown in Fig. 6 [22], [30], which differs from its MLP counterpart in terms of the number of hidden neurons, only one hidden layer is used In RBF neural networks. Also, activation functions of this structure are Radial Basis Functions. The mathematical equations of the RBF neural network are expressed by 1 1 1 ( ) , 1, 2,..., / ( ) where j v is the weight between the output and the hidden layer, ij w represents the weight between the input data and the hidden layer. () f is the activation function for this neural network in this paper [30].

B. TRAINING STRATEGY FOR WNN, RBF, AND MLP
The ICSA is used to train WNN, RBF, and MLP models. The original Clonal Selection Algorithm (CSA) is firstly explained; the ICSA is then introduced [31].

 Step 1:
The primary population of the clonal selection is randomly created in an acceptable range. This population, known as an antibody in the clonal selection, is a candidate solution for solving optimization problems. Each member of this population contains a parameter of WNNs, known as a decision variable in the optimization problem, which can be determined from Error! Reference source not found.. Accordingly, these four types of decision variables are randomly produced in the range of 1 1

 Step 2:
The objective function is used as a criterion to specify the affinity of antibodies. In this problem, the training error, Mean Square Error (MSE), is considered as an objective function to be minimized.  Step 3: According to the calculated objective function, the antibodies are sorted, and the antibody with the lowest value of the objective function is considered the best solution and ranked first.  Step 4: The following equations can be used to replicate antibodies based on the position that they gain in Step 3.

1, 2, ,
where p na is the number of antibodies replicated from th p antibodies. Round() is a function that rounds the obtained value to the nearest integer value.  is the rate of replication. NA is the total number of replicated antibodies.  Step 5: The total number of antibodies is mutated by  Step 7: In this step, N NT  antibodies are created for the next generation.  Step 8: While the number of generations is increased, this process ends when the termination criterion is met and reaches the maximum iteration of the optimization problem. Finally, the antibody with the lowest MSE is chosen as a solution. There is another modification added to the original CSA as shown in (17) based on which optimal antibodies are mutated less compared to others.
where M p N indicates the number of gen associated with th p antibodies that are going to be mutated [22].

IV. NUMERICAL ANALYSIS USING THE PROPOSED APPROACH
In this section, a numerical analysis using the proposed technique for day-ahead wind speed forecasting using historical wind speed data measured in Saskatchewan is demonstrated. Two error evaluation indices are used to evaluate the performance of the proposed approach.

A. DATA DESCRIPTION
In this study, we use historical hourly wind speed data measured in Saskatoon, SK, Canada, which is obtained from Historical Climate Data on the Government of Canada website [32]. The rated power of a wind turbine of 1.8 MW is considered in this paper with the cut-in, rated, and cut-out wind speeds at 15 km/h, 50 km/h, and 90 km/h, respectively [33]. The measured wind speed, wind direction, humidity, and temperature are demonstrated in Fig. 7. The 60 days of these observed data before the prediction day are taken as the training data set. Accordingly, we have 1,440 hours of training data (24 hours × 60 days). March 2, 2021, December 1, 2020, September 1, 2020, and June 1, 2020 are chosen as the test days. Furthermore, the autocorrelation function is used to determine the most effective candidate input for the forecasting engine. More specifically, 400 hourly lagged values of wind speed, wind direction, humidity, and temperature are served as the candidate input data in this study, which are processed by the autocorrelation function to identify a minimum subset of the most informative features to be incorporated into the proposed model. The autocorrelation function results of wind speed for the test day on March 2, 2021 is shown in Fig. 8, which demonstrates the relationship between the input data and the target value. The data with a high autocorrelation function value is chosen as an input of the forecasting engine.

B. WIND SPEED FORECASTING EVALUATION
Each method should be evaluated in terms of performance and efficacy. For the performance evaluation, two factors arise: 1) data size, and 2) statistical criteria. The required data size depends on the employed method [34]. Statistical criteria are determined considering the nature of the methods. Therefore, it is necessary to review existing approaches for the performance evaluation in the context of wind speed forecasting. Since wind speed forecasting belongs to numeric methods, their performance evaluation is mostly based on numerical-error evaluations. In this regard, various criteria are used to evaluate the performance of wind speed forecasting, including MSE, Mean Bias Error (MBE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Skill Score [3], [4], [35], [36].
In this study, two error indices are used: the normalized Root Mean Square Error (nRMSE) and the normalized Mean Absolute Error (nMAE), which are formulated as follows:

C. CALCULATING WIND POWER FROM WIND SPEED
To obtain wind power using wind speed, the following equation is used:

D. INFLUENCE OF THE NUMBER OF FEATURES
In this study, five different hybrid NN techniques (WNN trained by ICSA, WNN trained by PSO, ELM, RBF, and MLP) are utilized to forecast Saskatchewan's day-ahead wind speeds. To evaluate the impact of the number of features on the accuracy of wind speed prediction, two analyses have been performed: 1) Single-feature Analysis, only wind speed is considered as the input data, i.e., it only uses one feature; 2) Multi-feature Analysis, four input data are used including wind speed, wind direction, humidity, and temperature, i.e., it uses four features.
Single-feature Analysis: Table I  On the other hand, ELM is a forecasting engine with highspeed and the least prediction time. In Table I, the ELM model forecasts wind speed in 4.76 seconds, the MLP, RBF, WNN trained by ICSA, and WNN trained by PSO models use 38.68 second, 523.83 seconds, 3,021 seconds, and 5,186 second, respectively. Such difference on computing time among models is due to different training strategies and different structures of neural networks. For instance, ICSA and PSO are iteration-based optimization methods, and they can discover the optimal solution after at least 100 iterations, which are time consuming training strategies. However, the ELM training approach is exceedingly fast due to a simple matrix computation using (7). Multi-feature Analysis: The performance evaluation of forecasting engines considering four meteorological variables (wind speed, wind direction, humidity, and temperature) as features are shown in Table II. It is observed that WNN with the ICSA and ELM-based training strategies outperform other methods from the accuracy point of view. The ELM model also has the least computational time, which is significantly less than other models.
By comparing Tables I and II, we can see the multi-feature analysis for the day-ahead wind speed forecasting in this study does not improve accuracy of wind speed prediction for all models; in some cases, the accuracy is even reduced. It is mainly because when the number of input increases, it becomes difficult for the forecasting engine to find the best input/output mapping. Moreover, using multi-features lead to the increase of computational time for all models. Therefore, in short-term wind speed prediction, the generated wind speed forecasts are merely based on the past wind speed data, and it is better not to include wind direction, temperature, and humidity in the forecasting model. However, these meteorological variables may be useful for a longer forecasting horizon.

E. DAY-AHEAD WIND SPEED AND WIND POWER FORECASTING RESULTS
Figs. 9-12 illustrate wind speed and wind power prediction outcomes using the five hybrid NN models for all test days (June 1, 2020, September 1, 2020, December 1, 2020, and March 2, 2021) considering only wind speed as the input data. According to these figures, the WNN trained by ICSA and ELM models can provide satisfactory trends and ramps following the actual wind speeds and wind power; while the other three models, WNN trained by PSO, MLP, and RBF, may not be able to follow the measured time series accurately. In wind power prediction figures, the wind turbine cannot generate power when the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed.
Scatter plots for wind speed prediction on September 1, 2020 are shown in Fig. 13 for all models. The scatter plots of the WNN trained by ICSA and ELM models lie closely along the diagonal line, which confirms the high capability of these forecasting techniques to map input-output data. In contrast, the scatter plots of WNN trained by PSO, MLP, and RBF models are scattered and not distributed closely along the diagonal line. In this study, among the five evaluated hybrid NN methods, it is found that the WNN trained by ICSA model and the ELM model have better performance than other three methods. This is proved by both prediction accuracy and computation time aspects shown in Tables I and II. Figs. 9-12 graphically confirm that the WNN trained by ICSA and ELM models can accurately follow ramps and trends of the measured wind speed data. Finally, scatter plots affirm the superiority of the WNN trained by ICSA and ELM models over other methods.
Between the two best performing models, the ELM model is much more computationally efficient than the WNN trained by ICSA model, which can be valuable for conditions, in which the computation time plays a critical role. The WNN trained by ICSA model can be utilized in the planning stage since it has the highest forecasting accuracy.

V. CONCLUSION
Wind power is a promising clean energy source, especially in windy regions like Saskatchewan, Canada, but the intermittent nature of wind is one of the major challenges for wind power integration. In this paper, a novel hybrid neural network-based day-ahead wind speed forecasting technique is proposed with a comprehensive procedure. Five hybrid methods are considered using historical wind speed data recorded in Saskatchewan, Canada: 1) WNN trained by ICSA, 2) WNN trained by PSO, 3) ELM-based Neural Networks, 4) Radial Basis Function Neural Networks; and 5) Multi-Layer Perceptron Neural Networks. To evaluate the effect of the number of features, both single-feature analysis using only wind speed and multi-feature analysis using wind speed and other meteorological variables are conducted. It is found that for the day-ahead wind speed forecasting, only wind speed needs to serve as the feature, multi-features do not necessarily improve the accuracy, but increase the computation time.
Among the five hybrid NN models, the best performing models are the WNN trained by ICSA and ELM-based NN models. This is proven by the error indices calculation, graphically comparison between measured and forecasted wind speed curves, and scattered plots. Between the two best performing models, the WNN trained by ICSA model has the highest accuracy, while the ELM-based NN model has the least computation time. Therefore, which model to choose should be based on the particular applications.
The results in this study show that the proposed novel technique with a comprehensive procedure is very effective for day-ahead wind speed forecasting. The proposed method is very valuable to improve the economical, secure, and reliable operation of future wind power plants in Saskatchewan and beyond.