ASA-Net: Adaptive Sparse Attention Network for Robust Electric Load Forecasting
Yinqiang Deng, Xu Wang, Yong Liao
Abstract
Electric load forecasting (ELF) is always employed to perform power systems management. However, it is difficult to predict electric load due to the following issues: 1) electric load prediction is prone to external interference, e.g., temperature and weather; 2) the user behaviors are random, such as family gatherings and business rush orders; and 3) electric load consumption varies significantly in different time periods. To solve such problems, an adaptive sparse attention network (ASA-Net) is proposed for ELF, where the adaptive sparse spatial attention (ASSA) module is first designed to increase the anti-interference ability by capturing the detail change caused by external interference; next, the adaptive sparse channel attention (ASCA) module is developed to enhance the tolerance to local outliers by learning their feature information; and finally, the adaptive sparse batch attention (ASBA) module is devised to model the dependencies of the timestamp to reduce the time impact on ELF. Experiments conducted on the benchmarks show the excellent performance of ASA-Net for ELF, and it can further provide valuable assistance for the smart grid.