Key points are not available for this paper at this time.
Accurate short-term Multi-Energy Load Forecasting (MELF) is critical for the efficient dispatch of Integrated Energy Systems (IES). However, stochastic fluctuations driven by meteorological conditions and consumer behavior pose significant predictive challenges. Existing methods often struggle to effectively integrate multi-scale temporal dependencies, which undermines forecasting precision. Furthermore, the limited interpretability of deep learning models hinders their practical engineering deployment. To address these issues, this study proposes a novel short-term load forecasting framework that combines adaptive multi-scale feature fusion with model interpretability. The proposed approach utilizes a Mixture of Experts (MoE) architecture and a Patch Attention mechanism to extract and fuse hierarchical temporal features. Subsequently, a Long Short-Term Memory (LSTM) network is employed to capture both transient and long-range dependencies within historical sequences. Finally, SHAP (SHapley Additive exPlanations) analysis is incorporated to enhance the model’s transparency and provide a robust interpretive layer. Experimental results demonstrate that the proposed model significantly outperforms benchmark models. Specifically, forecasting accuracy for electric, thermal and cooling loads improve by 19%-45%, 9%-45% and 12%-58% respectively. These findings indicate that the proposed framework possesses superior predictive precision, strong generalization capabilities, and high practical utility for energy management and optimization.
Zhang et al. (Fri,) studied this question.