Adaptive load forecasting under regional distribution shifts: A meta-learning framework

Cross-regional short-term load forecasting represents a critical yet challenging task for modern power systems, where significant data distribution shifts between regions fundamentally hinder the accuracy and generalization of existing models. To address this core challenge, we propose STaRNet (Short-Term Adaptive Regional Network), a novel meta-learning framework designed to solve this problem through a dual-component architecture: (1) a reptile-based meta-learning strategy to learn a transferable model initialization for rapid, macro-level adaptation to unseen regions, and (2) a specifically tailored Lightweight Segment-Gated Feedforward Network (LSG-FFN) to capture fine-grained, micro-level local dynamics. Evaluated on real-world load datasets from four German Transmission System Operators (TSOs), STaRNet demonstrates superior performance. Notably, in the most challenging heterogeneous region transfer scenario, it achieves a statistically significant reduction of 10.79% in 24 h Mean Absolute Error (MAE) (p<0.05) compared to the strongest baseline. This finding empirically validates that STaRNet’s success stems from its synergistic design, which effectively integrates macro-level generalization with micro-level dynamic modeling, offering a robust and practical solution for large-scale, multi-region load forecasting.