Short Term Wind Power Prediction Based on Improved Kriging Interpolation, Empirical Mode Decomposition, and Closed-Loop Forecasting Engine

Nima Amjady, Oveis Abedinia
2017 Sustainability  
The growing trend of wind generation in power systems and its uncertain nature have recently highlighted the importance of wind power prediction. In this paper a new wind power prediction approach is proposed which includes an improved version of Kriging Interpolation Method (KIM), Empirical Mode Decomposition (EMD), an information-theoretic feature selection method, and a closed-loop forecasting engine. In the proposed approach, EMD decomposes volatile wind power time series into more smooth
more » ... into more smooth and well-behaved components. To enhance the performance of EMD, Improved KIM (IKIM) is used instead of Cubic Spline (CS) fitting in it. The proposed IKIM includes the von Karman covariance model whose settings are optimized based on error variance minimization using an evolutionary algorithm. Each component obtained by this EMD decomposition is separately predicted by a closed-loop neural network-based forecasting engine whose inputs are determined by an information-theoretic feature selection method. Wind power prediction results are obtained by combining all individual forecasts of these components. The proposed wind power forecast approach is tested on the real-world wind farms in Spain and Alberta, Canada. The results obtained from the proposed approach are extensively compared with the results of many other wind power prediction methods. Sustainability 2017, 9, 2104 2 of 18 relevant and least redundant input variables in [8] . In [9], wind power forecast by Ridgelet Neural Network (RNN) has been proposed in which Ridgelet is used as the activation function of the hidden nodes. Moreover, a new Differential Evolution (DE) method has been presented in [9] for training the RNN. A hybrid forecasting method based on Enhanced Particle Swarm Optimization (EPSO) has been introduced in [10] for wind power prediction. This hybrid method is composed of persistence technique, back propagation neural network, and Radial Basis Function (RBF) neural network. Also, EPSO further optimizes the weight coefficients in this hybrid technique. In [11] , another version of EPSO with a hybrid NN has been proposed for wind power prediction. The training mechanism of the hybrid NN is composed of the EPSO as well as Levenberg-Marquardt (LM), Broyden, Fletcher, Goldfarb, Shannon (BFGS), and Bayesian Regularization (BR) learning algorithms. Wind power forecast methods have been reviewed in [12] [13] [14] [15] . Various research works have focused on maximum power point tracking (MPPT) in photovoltaic (PV) generation systems, wind generation systems and their hybrid wind-PV systems, such as [16] [17] [18] [19] . A neuro-fuzzy wavelet-based adaptive MPPT control for PV Systems [16], dynamic operation and control for a hybrid wind-PV-fuel cell microgrid [17], intelligent MPPT control for a grid-connected hybrid solar power, wind power, and diesel-engine system [18] , and fuzzy MPPT controller for a hybrid solar power and battery system [19] have been presented in the literature. In recent years, different transformation techniques have been presented by researchers to enhance the accuracy of forecast methods. Fourier Transform (FT) is one of the earlier techniques which gives the frequency spectral content of the signal [20] . However, FT application is limited to stationary signals. It cannot give information about where in time the frequency spectral components appear. Short-time FT (STFT), which provides the time information by calculating many FTs for sequential time windows and putting them together [21] , has been proposed to deal with non-stationary signals. However, STFT provides a fixed resolution at all times, while low/high frequency behaviors require long/short analysis windows. Wavelet Transform (WT) alleviates the limitations of FT and STFT methods by using functions that retain an appropriate compromise between time location and frequency information. The basic concept of this method begins with the selection of a proper wavelet function, called mother wavelet, and then combining its shifted and scaled versions. Several wind power forecast methods using WT have been presented so far, such as [22] [23] [24] [25] . However, WT is a linear signal-processing tool which may not be able to thoroughly analyze nonlinear signal variations. Additionally, the effectiveness of WT is dependent on the proper selection of mother wavelet, while it has to be given before the analysis. Mother wavelets are usually chosen by trial-and-error and heuristic methods. The problems of WT have been extensively discussed in [26] . EMD is a well-organized decomposition approach, which can mitigate the problems of WT regarding estimating instantaneous frequency [27] . EMD has been used for wind power and wind speed prediction in [28] . However, EMD has a disadvantage to process non-stationary signals. This disadvantage is generating fake extremes [26] . To alleviate this problem a new version of EMD is proposed in this research work. The new contributions of this research work are as follows:
doi:10.3390/su9112104 fatcat:ug77jmxm4rgfbkc7k5quy3lbc4