Regime-Aware Federated Aggregation (RAFA) for privacy-preserving energy load forecasting across heterogeneous...

Federated Learning (FL) enables collaborative model training across distributed clients without sharing raw data, making it attractive for privacy-sensitive energy forecasting. However, standard aggregation algorithms, most notably FedAvg and FedProx, assume stationary client data distributions, an assumption routinely violated in multi-national power grids where seasonal patterns, demand crises, and policy changes induce persistent regime shifts.

We present RAFA (Regime-Aware Federated Aggregation), a novel FL framework that explicitly detects and adapts to distribution shifts in client model updates without accessing raw client data. RAFA comprises three components. These are a Mahalanobis-distance detector operating on random-projected update vectors; reliability-weighted aggregation that exponentially discounts shifted clients while preserving exact FedAvg behaviour when no shift is detected; and always-on personalised head fine-tuning that adapts each client's final prediction layer to its local regime.

We validate RAFA on a new multi-national benchmark of six years (2019-2024) of hourly electrical load data for eight European bidding zones sourced from the ENTSO-E Transparency Platform. RAFA achieves a test MAE of 185.4 MW, an 11.8% improvement over FedAvg (210.2 MW), with by far the largest gain for the client exhibiting the strongest seasonal contrast (France, $-30.5\%$ ) and consistent improvements across the remaining clients (e.g., Portugal $-7.6\%$ , Poland $-7.4\%$ ). Ablation studies confirm both mechanisms contribute independently.

RAFA provides a lightweight, privacy-preserving extension of standard federated aggregation that is robust to regime shifts in energy systems. The approach generalises to any federated setting with temporally non-stationary client distributions and a separable model architecture.