Privacy-Preserving Federated Learning for Feeder-Level Outage Prediction in Distribution Grids

This project develops an AI system to predict feeder-level outage risk across power grids without sharing raw utility data. It leverages federated learning to enable cross-utility collaboration, combined with differential privacy and secure aggregation to ensure data confidentiality. Experiments show that federated models significantly outperform isolated local models while maintaining strong performance under strict privacy constraints (ε ≈ 1.0). The work also highlights key challenges in applying graph-based models across heterogeneous grid topologies, revealing important directions for future research in scalable, privacy-safe grid intelligence.