Multi-Task Learning for Integrated Fault Analysis in Power System Protection

📋 Type Project
⚡ Status finished
📅 Duration Oct 1, 2025 – Apr 1, 2026
👤 Primary supervisor Julian Oelhaf
👥 Co-supervisors Siming Bayer Andreas Maier
🎓 Student Rahul Bhagwandas Motwani Artificial Intelligence

Abstract

This project evaluates whether a single neural network can detect a fault, classify its type, identify the affected line and estimate the fault position along that line. A shared multilayer perceptron with four task-specific outputs was trained on simulated three-phase voltage and current measurements from the PROTECT-90 dataset. Compared with four separately trained models, joint training used about one quarter as many parameters and largely preserved fault detection, line identification and localization performance. Fault classification performance declined sharply when all classes were weighted equally. An additional test using measurements from one selected relay showed that line identification and localization were particularly sensitive to losing measurements from the other relays.

Conference Paper · IEEE PESS 2026

The paper based on this project was accepted for presentation at the IEEE Power & Energy Student Summit 2026 in Karlsruhe.

Evaluating Multi-Task Learning for Integrated Fault Analysis in Power System Protection
Rahul Bhagwandas Motwani, Julian Oelhaf, Christian Bergler, Andreas Maier and Siming Bayer

Across five training runs with measurements from eight relays, the shared model reduced total training time per run from 104 to 40 minutes compared with training four separate models. Overall fault classification accuracy was 98.7%, but macro F1, which gives each class equal weight, fell from 0.998 to 0.531. The largest errors occurred for two-phase-to-ground faults.

At the selected single-relay location, fault-detection F1 was essentially unchanged, while mean localisation error rose from 8.88% to 18.45% of line length. This shows how strongly the value of measurements from multiple relays can vary by task.

Conference programme