Connecting the Patches: Exploring Multi-Scale Relations in Patches via Graph Neural Networks for Short-Term Electricity Load Forecasting

Precise short-term forecasting of electricity load is essential to ensure the consistent and dependable functioning of power systems. Yet, the intricate and evolving relationships among load sequences and various external variables (e.g., temperature, electricity cost) present major difficulties in multivariate modeling. Existing graph-based forecasting models typically construct graphs based on variables or time points, but often neglect local spatio-temporal patterns, making it difficult to capture rapid feature variations within short time windows. Moreover, due to the multi-periodic nature of load sequences, their correlations with external features vary across time scales, further reducing forecasting accuracy. To address these challenges, we propose a novel Multi-scale Dynamic Patch Graph Neural Network (MDPGNN). In the local channel, the model constructs patch-level graphs and integrates dynamic graph learning with adaptive graph convolution to jointly model the internal structure of load sequences and their multi-scale dependencies with external features. In the global channel, a global correlation graph is designed to preserve long-range temporal patterns and inter-variable relationships. Furthermore, a Mixture-of-Experts (MoE) mechanism is employed to effectively fuse multi-scale dependencies and the outputs of both local and global channels. Experimental results on two real-world load datasets show that MDPGNN consistently outperforms state-of-the-art GNN-based models in terms of prediction accuracy and stability.