A novel ensemble support vector regression for load forecasting under data attacks

Load forecasting is significantly crucial for optimal operational management of smart grid within the energy industry. As the incidents of cybersecurity and data attacks happen more and more frequently, it is very necessary to timely produce accurate load forecasts under data attacks. In this study, we propose a novel ensemble iteratively re-weighted least squares (IRLS) support vector regression (En-IRLS-SVR) method for electric load forecasting without or with data attacks as follows. First, an optimization-based method is introduced to select optimal features for effective load forecasting. Then a fuzzy membership function based on IRLS is introduced for efficiently and iteratively calculating the relative importance of each observation in the load history, and incorporated into the SVR model to develop a novel IRLS-SVR model. Finally, we incorporate the bootstrap method and developed IRLS-SVR model to propose the En-IRLS-SVR method by incorporating multi-model forecasting outputs into original data features for final load forecasting. Comprehensive numerical results on real-life load datasets demonstrate the superior performance of proposed En-IRLS-SVR method over well-known robust load forecasting models and ensemble learning methods without or with three types of data integrity attacks.