Federated Learning for Resilient Smart Grid Monitoring: A Physics-Based Simulation Framework for Distributed Energy Networks

Federated learning (FL) can support collaborative smart-grid monitoring without centralizing raw telemetry, but its value relative to local and centralized learning depends on both the operating regime and the learning model. This study evaluates that hypothesis in a physics-based AC power-flow simulation framework built on IEEE 14-, 39-, and 118-bus benchmark systems over 30-day horizons at 1 h resolution. The framework is a controlled simulation environment rather than a field-coupled or real-time digital twin. Twenty independent physical seeds were used to compare centralized XGBoost, local-only learning, FedAvg, SCAFFOLD, and a study-specific validation-blended personalized FedAvg, with pooled same-optimizer centralized linear/logistic controls used to isolate federation from model-family effects. Under nominal operation, centralized or personalized models were strongest; on IEEE 118, XGBoost achieved the lowest line-loading and power-loss MAEs (1.331 percentage points and 2.880 MW), while local-only and personalized learning achieved voltage-deviation MAEs of 0.000327 and 0.000330 p.u. Under IEEE 118 N-1 voltage-instability detection, centralized logistic gradient descent, FedAvg, and SCAFFOLD all achieved AUPRC values of approximately 0.993, demonstrating that much of the apparent XGBoost-to-FL gap was model-family-dependent. Under cross-attack transfer, SCAFFOLD retained small but statistically significant advantages over the matched centralized control. A targeted operational-threshold sensitivity analysis added 1500 refits across 15 threshold settings and 20 seeds; the qualitative separation between the federated and non-federated approaches remained stable, although the exact ordering among federated variants was metric- and threshold-dependent. Moderate telemetry noise, missingness, and bounded reporting delay caused at most 1.7% relative MAE degradation, whereas physical-regime changes produced substantially larger effects. The results support a regime-dependent and model-controlled interpretation of FL for supervisory smart-grid monitoring.