A time series model based on hybrid-kernel least-squares support vector machine for short-term wind power forecasting. (February 2021)

• Record Type: Journal Article
• Authors: Ding, Min; Zhou, Hao; Xie, Hua; Wu, Min; Liu, Kang-Zhi; Nakanishi, Yosuke; Yokoyama, Ryuichi
• Is Part Of: ISA transactions. Volume 108(2021)
• Page Start: 58
• Page End: 68
• Publication Date: 2021-02
• Subjects: Short-term wind power forecasting -- Time series forecasting model -- Least-squares support vector machines -- Wavelet decomposition; Engineering instruments -- Periodicals; Engineering instruments; Periodicals; Electronic journals; 629.805
• DOI: 10.1016/j.isatra.2020.09.002
• Languages: English
• ISSNs: 0019-0578

Abstract

Abstract: In this paper, a time series model based on hybrid-kernel least-squares support vector machine (HKLSSVM) with three processes of decomposition, classification, and reconstruction is proposed to predict short-term wind power. Firstly, on the basis of the maximal wavelet decomposition (MWD) and fuzzy C-means algorithm, a decomposition method decomposes wind power time series and classifies the decomposition time series components into three classes according to amplitude–frequency characteristics. Then, time series models on the basis of least-squares support vector machine (LSSVM) with three different kernels are established for these three classes. Non-dominated sorting genetic algorithm II optimizes the parameters of each forecasting model. Finally, outputs of forecasting models are reconstructed to obtain the forecasting power. The proposed model is compared with the empirical-mode-decomposition least-squares support vector machine (EMD-LSSVM) model and wavelet-decomposition least-squares support vector machine (WDLSSVM) model. The results of the comparison show that proposed model performs better than these benchmark models. Highlights: The forecasting framework with decomposition, classification and modeling is proposed based on the analysis of amplitude–frequency characteristics of wind power. The maximal wavelet decomposition decomposes wind power time series for obtaining stationary time series components, and the fuzzy C-means classifies these decomposed components into 3 classes based on amplitude–frequency characteristics. LSSVM models with three different kernels are built for these 3 classes to improve the accuracy of wind power forecasting.