An Effective Ensemble Algorithm for Short-Term Load Forecasting

En-Wei Zhang, Luo-Fan Wu, Chun-Wei Tsai

Abstract: Short-term load forecasting (STLF) is a critical issue for managing electricity distribution systems because its accuracy might strongly impact the performance and security of the power system. An effective load forecasting system, of course, can also be used to detect and prevent abnormal behaviors (e.g., electricity theft detection) to further provide a stable and safe power system. That is why extensive studies have been presented by using statistical, machine learning, and deep learning methods. However, each learning model has its strengths to capture specific patterns in the load profile. To integrate the strengths of different deep learning algorithms, an ensemble model, which contains three deep learning mechanisms, including convolution, self-attention, and recurrent, is proposed in this paper. Because of the high uncertainty and volatility, a single-step horizon will be focused on while predicting the load at the residential level to ensure timely and precise responses to security threats. Multiple-step forecasting will also be conducted at the aggregated level to further enhance the robustness of the power grid security. Furthermore, considering the distinct behavior in the load profile, the proposed method is designed accordingly to adapt to each level. The experimental results show the effectiveness of the proposed method compared with deep learning models on both levels across two well-known datasets.

External IDs:dblp:conf/wisa/ZhangWT24