Short-term load forecasting by a hybrid attention scheme: Multi-feature attention and context awareness

Accurate short-term load forecasting (STLF) is critical to optimize power generation scheduling and ensure grid stability with increasing penetration of renewable energy. Short-term power load variations typically exhibit strong locality and regularity. Despite significant progress, existing forecasting models still face two major challenges in handling short-term forecasting tasks: (1) Conventional static feature weighting strategies (feature selection, uniform attention, etc.) fail to adaptively capture dynamic interdependencies among heterogeneous features; (2) Transformer-based models suffer from high computational costs and inadequate local pattern extraction, limiting their effectiveness in modeling short-term dependencies. To address these challenges, we propose a hybrid attention scheme with multi-feature attention and context awareness (MFACA). First, the multi-feature attention (MFA) layer dynamically adjusts feature weights in each time step through output dependencies, enabling feature-sensitive prioritization of critical features. Second, the context awareness (CA) layer dynamically weights contextual information based on its correlation with encoder output, thus enhancing the model’s ability to simultaneously decode local fluctuations and global periodic trends. Finally, MFACA jointly optimizes multiscale feature interactions and temporal dependencies within an encoder–decoder architecture. Extensive evaluations using three real-world power load datasets confirm the effectiveness of the proposed model, demonstrating superior performance on multiple metrics, with the MFA and CA components contributing significantly to the improvement.