A Robust and Heterogeneity-Aware Federated Learning Framework with Knowledge Distillation for Cross-Regional Load Forecasting

Accurate load forecasting is fundamental for power system operation and planning. While traditional single-region approaches are constrained by limited local data, cross-regional forecasting leverages larger datasets to achieve higher accuracy. Federated learning (FL) emerges as a promising solution, enabling cross-regional collaboration. However, existing FL-based approaches struggle with model, system, and statistical heterogeneity, along with security vulnerabilities. To address these issues, this paper proposes a robustness-enhanced personalized federated learning framework that integrates knowledge distillation for cross-regional load forecasting. Proxy models are utilized to enable secure knowledge transfer while preserving local model adaptability, thereby resolving model heterogeneity. Perturbed Gradient Descent (PGD) mitigates statistical heterogeneity, and a dynamic exit mechanism reduces computational costs by allowing clients to exit early upon meeting accuracy thresholds, addressing system heterogeneity. Case studies on open-access energy dataset from six European countries show that the proposed method outperforms conventional FL models, personalized FL methods, and traditional robust aggregation schemes in terms of accuracy, robustness, and model resilience.