A new compound wind speed forecasting structure combining multi-kernel LSSVM with two-stage decomposition technique

Abstract

The aims of this study are to develop a novel compound structure that consists of two-stage decomposition (TSD), hybrid particle swarm optimization gravitational search algorithm (HPSOGSA) and multi-kernel least square support vector machine (MLSSVM) for improving forecasting accuracy. In the most previous wind speed forecasting studies, only one wind speed signal decomposition method is considered, which is insufficient. To better deal with the wind speed time series, TSD method combining complementary ensemble empirical mode decomposition with adaptive noise with wavelet transform is firstly employed in the proposed model to preprocess the wind speed samples; then, binary-valued particle swarm optimization gravitational search algorithm is exploited as feature selection to identify and eliminate the abnormal noise signal within the input candidate matrix that is determined by partial autocorrelation function. The kernel function and the kernel parameters have great influence on the regression performance of LSSVM. To solve these problems, integrations of radial basis function, polynomial (poly) and linear kernel functions by optimal weighted coefficients are constructed as multi-kernel function for LSSVM, namely MKLSSVM, and the parameter combination is tuned by conventional PSOGSA. The feature selection and parameter optimization are realized by hybrid PSOGSA (HPSOGSA) simultaneously. Finally, comprehensive comparison and analysis are carried out using the historical wind speed data from one wind farm of China to illustrate the excellent forecasting performance of TSD–HPSOGSA–MKLSSVM.