Load Forecasting for Integrated Energy Systems Based on Causality and Dual Layer MMoE Multi-Task Learning

Short-Term load forecasting with high precision is critical for the economic dispatch and efficient operation of the integrated energy system (IES), which involves multiple energy carriers with strong randomness and coupling relationships. To address the issues, this paper proposes a multi-energy load forecasting method based on causality analysis and a dual-layer Multi-Gate Mixture-of-Experts (MMoE) network. Firstly, a comprehensive index selection method based on correlation analysis and reverse information entropy causal inference is developed to select weather factors in the MMoE information sharing layer to learn shared information. The extracted shared features as well as individual task features are combined and fed into the MMoE subtask layer. Specifically, considering the causality among different loads, depth information-sharing mechanism based on causality is designed. To further explore hidden information from historical data and alleviate the load fluctuations, similar day loads and load participation factors as contextual features are utilized in the subtask layer and sharing layer, respectively. Experimental results on a campus IES dataset demonstrate that the proposed model overall outperforms other popular approaches, confirming the effectiveness of integrating causality and dual-layer MMoE in enhancing multi-energy load forecasting accuracy.