A compositive architecture coupling outlier correction, EWT, nonlinear Volterra multi-model fusion with multi-objective optimization for short-term wind speed forecasting. (1st February 2022)

• Record Type: Journal Article
• Authors: Fang, Ping; Fu, Wenlong; Wang, Kai; Xiong, Dongzhen; Zhang, Kai
• Is Part Of: Applied energy. Volume 307(2022)
• Publication Date: 2022-02-01
• Subjects: Short-term wind speed forecasting -- Detection and correction of outliers -- Empirical wavelet transform -- Multi-model fusion -- Volterra fitting -- Enhanced HMOHHO; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.118191
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Data preprocessing based on box-plot eliminates outliers of original wind speed series. A simplified ConvSGRU coupling peephole structure is proposed to implement short-term wind speed forecasting. Multi-model fusion strategy is applied in Volterra fitting to improve the stability of the overall architecture. Kernel identification parameters of Volterra are optimized by enhanced HOMMO with mutation operator. Wind speed forecasting experiments as well as contrastive analysis verify the superiority of the proposed compositive architecture. Abstract: Short-term wind speed forecasting with high accuracy puts forward a positive influence on implementation of power system dispatch and wind energy utilization. Accordingly, to further promote the prediction accuracy and maintain the prediction stability, an innovation compositive architecture is developed incorporating three modules: data preprocessing, several individual predictors and Volterra multi-model fusion with enhanced multi-objective optimization algorithm. First and foremost, detection and correction of outliers are regarded as a preprocessing technique to process original wind speed sequences, after which empirical wavelet transform (EWT) is further adopted to adaptively decompose the corrected sequences into multiple subsequences. Then, individual predictors including several typical models and the proposed convolutional simplified gated recurrent unit (ConvSGRU), are applied to acquire prediction results by summating the predicted values of each component, which is constructed by PSR to obtain feature matrix as input of the above models. Subsequently, standard deviation (Std.) calculated between the actual values and predicted values of each model is severed as the basis for ensemble modeling, in which multi-model fusion based with smaller Std. on three models are adaptively selected to construct a multi-dimensional matrix as the input of Volterra. In this process, kernel identification parameters of Volterra can be acquired by multi-objective Harris hawks optimization (MOHHO) algorithm with mutation operator (HMOHHO), which can effectively guarantee accuracy and stability of the model simultaneously. Furthermore, four experiments with three datasets collected at different regions are revealed to ascertain forecasting capability of the proposed compositive architecture. The experimental results clarify that: (1) original wind speed sequences based on box-plot are employed to detect and correct, which can greatly decrease the influence of outliers on the sequences while preserving its mainstream trend; (2) the proposed deep network ConvSGRU can give assistance for the forecasting accuracy, thereby contributing to ameliorate performance of the entire compositive architecture; (3) multi-model fusion based on Volterra nonlinear fitting is conducive to further actively compensating for instability of individual models; (4) HMOHHO enables Volterra fitting to receive a series of constructive kernel identification parameters.