Federated Learning for Local Fault Analysis in Power Systems

Abstract

This project investigates federated learning for local fault analysis in power systems using the PROTECT-90 EMT waveform dataset. Each relay location is treated as an independent client that trains collaboratively without exchanging raw voltage and current measurements. The study compares local-only models, centralized training, federated global models, and locally fine-tuned federated models for fault detection, fault classification, and fault localization. By evaluating performance separately at each relay location, the project aims to determine whether shared fault representations can improve local protection models under heterogeneous relay perspectives.