ATT-Enhanced Inception-BiGRU-Seq2Seq Model for Short-Term User-Level Load Forecasting during Heatwaves

Wuxiao Chen^1*Zhijun Jiang¹Yuqing Cai¹Han Lin¹Nan Lin¹Zhihong Huang²

• ¹State grid Fujian electric power Co., Ltd. marketing service center, Fuzhou, China
• ²Beijing Tsingsoft Technology Co., Ltd,, Beijing, China

To address the issue of reduced prediction accuracy in user-level short-term load forecasting under extreme weather conditions, particularly during high temperature heat waves, this study introduces an Attention-Enhanced Inception-BiGRU-Seq2Seq (AIBS) model. In the encoder, an Inception module is combined with BiGRU. The Inception module leverages multiscale convolution to extract diverse local features from the input sequence, while a two layer BiGRU captures comprehensive temporal dependencies. The decoder incorporates an enhanced attention mechanism to strengthen the model's ability to focus on key historical time steps. Furthermore, the Grey Wolf Optimizer (GWO) is employed for global hyperparameter optimization, improving both accuracy and stability. Experimental evaluations using real world load data from various user categories during heatwave periods demonstrate that the proposed method consistently outperforms baseline models across multiple metrics, including RMSE and MAE, highlighting its superior predictive performance and strong generalization capability.