A spatiotemporal wind power prediction based on wavelet decomposition, feature selection, and localized prediction

Wind power possesses a high level of non-linearity and non-stationarity which are the main barriers to developing an accurate wind power prediction (WPP). In this regard, a multiresolution wavelet decomposition (WD), based on discrete wavelet transform, is employed to decompose the wind power time series (TS) into several components. Afterward, in a feature selection (FS) stage, which benefits from the spatiotemporal relation among the wind farms, the double input symmetrical relevance (DISR) has been adopted to find the most suitable features in predicting each component. Then, to have a high-accuracy prediction with an affordable computation time, localized prediction engines have been used to predict each component. The final WPP value is obtained by superposition of all the predicted values corresponding to components. The proposed spatiotemporal WPP is evaluated using the wind power generation historical data in Saskatchewan, Canada. The performance of the proposed WPP is compared with other well-developed and widely-used WPP models. Various evaluation indices have been utilized for conducting the performance evaluation.