IIWFedAvg: Physics-Informed Federated Learning Boosts Grid Resilience 75%

A new paper from Ibrahim Shahbaz, Omar Al-Refai, and Eman Hammad introduces Inertia-Informed Weighted FedAvg (IIWFedAvg), a physics-informed federated learning framework designed to enhance the resilience of distributed smart grids. Unlike standard federated averaging that treats all generators equally, IIWFedAvg incorporates each synchronous generator's inertia constant into the global model aggregation, allowing the control policy to account for the physical heterogeneity of the grid. The framework also leverages interpretable Chebyshev Kolmogorov-Arnold Network (ChebyKAN) controllers augmented with Rate-of-Change-of-Frequency (RoCoF) features, giving the system dynamic awareness of frequency disturbances.

Tested on the IEEE 39-bus benchmark under fully decentralized deployment, IIWFedAvg achieved a 75% generalization success rate across unseen fault contingencies, significantly outperforming individually trained models that fail to generalize. It also surpassed a centralized training baseline in two out of three stabilized fault scenarios, while delivering a 3x improvement in stabilization speed—all without requiring any centralized coordination overhead. The approach is accepted for IEEE SmartGridComm 2026, signaling a practical step toward resilient-by-design grid control that can maintain stability under physical disturbances and communication failures.