Accurate short-term load forecasting becomes particularly challenging when models are deployed across heterogeneous regions with distinct consumption structures, climatic conditions, and socio-economic characteristics. While existing deep learning approaches achieve strong in-region performance, they often experience severe degradation when transferred to unseen regions, limiting practical scalability in interconnected power systems. This study investigates cross-regional generalization in short-term load forecasting and proposes a structured spatiotemporal representation decomposition framework that explicitly separates region-invariant load generation mechanisms from region-specific modulation effects. The framework integrates an invariant backbone, orthogonality-constrained local adaptation, and cross-regional similarity alignment within a unified learning objective that directly embeds robustness and generalization considerations. Extensive experiments on multi-regional datasets demonstrate that the proposed approach reduces cross-regional forecasting error by more than 40 percent under zero-shot deployment compared with conventional fine-tuning strategies, while achieving up to 35 percent improvement over training-from-scratch baselines in limited-data scenarios. At the same time, competitive in-region accuracy is maintained. These findings indicate that structural disentanglement of transferable and localized dynamics offers a principled and scalable pathway toward robust load forecasting across heterogeneous power systems. • A transferable spatiotemporal load forecasting framework is proposed. • Load dynamics are decomposed into invariant and region-specific components. • Orthogonality constraints prevent leakage of local features into invariant space. • The framework enables zero-shot and few-shot cross-regional forecasting. • Superior generalization is achieved under strong regional heterogeneity.
Zhang et al. (Sun,) studied this question.