Verification of neural network based power system protection schemes

Mederer C, Kordowich G, Oelhaf J, Maier A, Bayer S, Jäger J (2025)

Publication Type: Conference contribution

Publication year: 2025

Publisher: IET

Book Volume: 2025

Pages Range: 155-159

Event location: Bilbao

Issue: 5

DOI: 10.1049/icp.2025.1062

Abstract

Recent studies have highlighted the potential of neural networks (NNs) to address emerging challenges in power system protection posed by the integration of distributed energy resources. While NNs have demonstrated effectiveness in fault detection, localisation, and protection coordination, their practical adoption in real-world grid environments remains limited. This limitation is primarily attributed to their "black box" nature and vulnerability to adversarial inputs. This paper presents a methodology for validating and visualising NN decisions specifically designed for power system protection tasks. The verification process employs a Mixed-Integer Linear Programming (MILP) approach, ensuring reliable NN decisions across defined input regions rather than isolated points. The decision boundaries and regions of trust are visualised through R-X diagrams, a well-established tool among protection engineers. These visualisations enhance the transparency and interpretability of NN decision-making, paving the way for secure and dependable NN-based solutions in power system protection.

Authors with CRIS profile

How to cite

APA:

Mederer, C., Kordowich, G., Oelhaf, J., Maier, A., Bayer, S., & Jäger, J. (2025). Verification of neural network based power system protection schemes. In Proceedings of the 19th IET Conference on Developments in Power System Protection (DPSP Europe 2025) (pp. 155-159). Bilbao, ES: IET.

MLA:

Mederer, Christoph, et al. "Verification of neural network based power system protection schemes." Proceedings of the 19th IET Conference on Developments in Power System Protection (DPSP Europe 2025), Bilbao IET, 2025. 155-159.

BibTeX: Download