A Few-Shot Multi-Energy Demand Forecasting Method Based on Meta-Learning Graph Neural Network

Forecasting user-level multi-energy demand under data scarcity is challenging due to the strong spatial interdependencies among users and the complex coupling across heterogeneous energy carriers. To address the challenges, this paper proposes a few-shot forecasting approach for integrated electricity-gas-heat systems, enabling reliable prediction with limited historical data. First, a meta-learning graph neural network is developed to extract generalizable energy conversion features adaptively across user patterns, effectively alleviating data insufficiency. Second, to accurately capture system dynamics, a multi-relational spatiotemporal graph module is designed to hierarchically model intra-energy dynamics and inter-energy coupling, achieving adaptive fusion of heterogeneous spatiotemporal features. Furthermore, a knowledge-enhanced training mechanism is introduced by embedding physical energy conversion laws into meta-training via pseudo-tasks, enhancing physical consistency and preventing overfitting in few-shot adaptation. Extensive experiments on multi-energy datasets ultimately demonstrate that compared to state-of-the-art forecasting models, the proposed approach achieves significant improvements in average metrics 3.80% for MAE, 7.89% for RMSE, and 13.12% for MAPE).