Federated Learning for Local Fault Analysis in Power Systems

Motivation

Protection systems make decisions based solely on local current and voltage measurements. When machine learning models are trained only on data from a single relay location, they may learn patterns that are highly specific to that particular electrical perspective.

Federated learning enables multiple relay locations to collaboratively train a shared model without exchanging raw waveform data. This project investigates whether federated training can improve the performance of models that are ultimately deployed locally at individual relays.

Objective

The goal of this project is to evaluate whether federated learning across multiple relay locations improves local performance for:

• Fault detection
• Fault classification
• Fault localization

The central research question is:

Can federated learning improve the performance and generalization of locally deployed fault analysis models under heterogeneous relay perspectives?

Tasks

1. Local Baselines
   Implement a deep learning model for fault analysis and train separate models using only data from individual relay locations.
2. Centralized Reference
   Train a model on pooled data from all relay locations to establish a performance reference and upper bound.
3. Federated Training
   Implement a federated learning strategy such as FedAvg.
   Treat each relay location as an independent client and train a shared global model without exchanging raw data.
4. Personalization
   Fine-tune the federated model locally and compare the global model with locally adapted versions.
5. Evaluation
   Compare:
   • Local-only training
   • Centralized training
   • Federated global model
   • Federated plus local fine-tuning

   Analyze results separately for each relay location and for each task.

Methodology

• Dataset: PROTECT-90
• Multi-relay EMT waveform data
• Each relay location is treated as one federated client
• Evaluation on relay-specific test sets

Metrics

• Detection and classification: Accuracy and F1-score
• Localization: MAE and RMSE in percent of line length

Expected Outcomes

• A structured comparison between local, centralized, and federated training
• Insight into whether shared fault representations exist across relay perspectives
• Understanding of when federated learning improves local performance and when it does not
• Practical recommendations for collaborative learning in protection systems

Prerequisites

• Python and PyTorch
• Basic knowledge of deep learning
• Interest in power systems and applied machine learning

Organization

• Weekly meeting Tuesday 14:00 in person
• Independent implementation and experimentation
• Final report and presentation