Interpretable feature-temporal transformer for short-term wind power forecasting with multivariate time series

Abstract

The inherent randomness and volatility of wind power generation present significant challenges to the reliable and secure operation of the power system. Therefore, it is crucial to have interpretable wind power forecasting (WPF) to ensure seamless grid integration and effective risk assessment. Existing forecasting models often focus on improving WPF performance and ignore the interpretability of the model, resulting in ambiguous forecasting. In this paper, the interpretable feature-temporal transformer (IFTT) for short-term wind power forecasting with multivariate time series is presented. The model uses an encoder-decoder architecture to effectively integrate historical information and future prior information from multiple variables. The designed decoupled feature-temporal self-attention (DFTA) module and variable attention network (VAN) effectively realize the interpretability of temporal information and multi-variable inputs while extracting important features. The Auxiliary Forecasting Network (AFN) plays a key role in providing pseudo-future wind speed predictions, which serve as an essential input for the model's decoder, and enhancing forecasting accuracy through multi-task learning. Experimental results on multiple datasets in different geographical locations show that the proposed algorithm is superior to various advanced methods. Besides, the interpretability of the IFTT model offers valuable insights for ensuring the safety of wind power utilization and supporting informed risk decision-making.