Universal and Relay-Generalizable Machine Learning for Protection in Power Grids

📋 Type Project
⚡ Status finished
📅 Duration Apr 1, 2026 – Aug 31, 2026
👤 Primary supervisor Julian Oelhaf
👥 Co-supervisor Andreas Maier
🎓 Student Nithin Pradeep Nadayil Kizhakkethil Electromobility-ACES (M.Sc.)

Abstract

Machine learning models for power system fault detection are typically trained and evaluated at fixed relay locations, implicitly assuming that training and deployment data share the same measurement distribution. In practical transmission networks, however, protective relays are installed at multiple locations whose electrical measurements differ systematically due to network topology, fault proximity, and local operating conditions. This work investigates cross-relay generalization failure using exhaustive pairwise transfer experiments across eight transmission-line protective relays in a 110 kV electromagnetic transient simulation dataset. We show that standard aggregate metrics such as accuracy and F1-score show only modest changes under cross-relay transfer, while the FPR increases from 0.4506 under within-relay evaluation to 0.6299 under cross-relay transfer, with individual relay pairs reaching an FPR of 0.993. Representation analysis further reveals that the learned latent representations retain coarse measurement-location information structured along bus-group boundaries, consistent with the topology-linked transfer difficulty observed across the pairwise transfer matrix, and that latent distributional shift quantified via squared maximum mean discrepancy (MMD²) shows a moderate positive association with cross-relay FPR. These findings provide a systematic characterization of cross-relay generalization failure and highlight the limitations of relying solely on aggregate performance metrics and the need for evaluation protocols that explicitly account for relay-location distribution shift when assessing model readiness for multi-relay deployment.

Universal and Relay-Generalizable Machine Learning for Protection in Power Grids

Motivation

Modern power system protection relies on local relay measurements obtained via current and voltage transformers at circuit breaker locations. Machine learning approaches have demonstrated strong performance for fault detection, classification, and localization tasks.

However, existing work typically trains and evaluates models on the same relay locations, implicitly assuming fixed spatial measurement points. In real-world deployment, protection algorithms must generalize across:

  • Different relay locations
  • Changing grid configurations
  • Varying short-circuit levels
  • Diverse operating conditions

A key open research question is therefore:

“Can machine learning models learn spatially invariant fault representations that generalize across relay locations while still allowing local adaptation?”

This thesis addresses this question through a structured two-layer learning approach.

Research Objectives

The thesis investigates the design of a universal, relay-generalizable protection model that:

  1. Learns generic fault representations independent of spatial location
  2. Quantifies cross-relay generalization behavior
  3. Enables efficient fine-tuning for specific installation points
  4. Research Questions

The work will address the following scientific questions:

  1. How large is the cross-relay generalization gap in fault detection and localization tasks?
  2. Which relay pairs exhibit “easy” vs. “hard” transfer and why?
  3. Can domain generalization strategies improve spatial invariance?
  4. Does universal pretraining reduce the amount of local data required for adaptation?

Methodology

Step 0 – Baseline: Cross-Relay Generalization Analysis

  • Train a model on measurements from a single relay.
  • Evaluate on all other relays using a leave-one-relay-out protocol.
  • Construct a full transfer matrix across all relay pairs.
  • Analyze:
    • Same-line vs. cross-line transfer
    • Influence of topology and infeed
    • Spatial symmetry effects

This establishes the relay generalization gap.

Step 1 – Universal Pretraining

  • Train a neural network using domain-randomized grid simulations.
  • Learn grid-invariant fault representations.
  • Incorporate robustness against:

Step 2 – Location-Specific Fine-Tuning

  • Adapt the pretrained model to a specific relay.
  • Extend from fault detection to fault localization.

Expected Contributions

The thesis is expected to provide:

  • A formal evaluation framework for relay generalization
  • Quantitative analysis of spatial transfer behavior
  • Identification of invariance-inducing training strategies
  • Practical recommendations for ML-based protection deployment