Deep Learning Architectures for Fault Analysis in Power System Protection: A Reproducible Evaluation on Public Waveform Data

Oelhaf J, Kordowich G, Bergler C, Maier A, Jäger J, Bayer S (2027)

Publication Language: English

Publication Type: Journal article, Original article

Publication year: 2027

Journal

Original Authors: Julian Oelhaf, Georg Kordowich, Christian Bergler, Andreas Maier, Johann Jäger, Siming Bayer

Book Volume: Volume 265, Part C

Article Number: 114169

URI: https://www.sciencedirect.com/science/article/pii/S0378779626014574

DOI: 10.1016/j.epsr.2026.114169

Open Access Link: https://www.sciencedirect.com/science/article/pii/S0378779626014574

Abstract

The increasing operational variability of power systems motivates data-driven methods for fault analysis in protection applications. This paper presents a systematic evaluation of deep learning models for fault detection, classification, line identification, and localization. Models are trained on EMT-simulated voltage and current waveforms from the public PROTECT-90 dataset, a 90 kV double-line benchmark. Recurrent, convolutional, hybrid, and transformer-based architectures are compared under shared data splits, temporal decision horizons, and leakage-aware validation protocols to assess predictive performance, inference latency, and model complexity. The results reveal distinct task characteristics. Detection, classification, and line identification reach near-saturated performance once short post-fault transients are available, whereas localization depends strongly on temporal context and improves with longer observation windows. Additional sensitivity analyses contextualize localization under reduced relay observability, test-time communication perturbations, and impedance-based reference methods. Across tasks, compact recurrent and hybrid architectures nearly match the predictive performance of larger architectures while maintaining shorter inference times, indicating diminishing returns from increased capacity under centralized EMT-based measurements. The findings provide guidance for architectural selection in learning-based protection and post-fault analysis under controlled, reproducible EMT sensing assumptions.

Authors with CRIS profile

Related research project(s)

Involved external institutions

Ostbayerische Technische Hochschule (OTH) Amberg-Weiden Germany (DE)

How to cite

APA:

Oelhaf, J., Kordowich, G., Bergler, C., Maier, A., Jäger, J., & Bayer, S. (2027). Deep Learning Architectures for Fault Analysis in Power System Protection: A Reproducible Evaluation on Public Waveform Data. Electric Power Systems Research, Volume 265, Part C. https://doi.org/10.1016/j.epsr.2026.114169

MLA:

Oelhaf, Julian, et al. "Deep Learning Architectures for Fault Analysis in Power System Protection: A Reproducible Evaluation on Public Waveform Data." Electric Power Systems Research Volume 265, Part C (2027).

BibTeX: Download