Deep Learning for Fault Localization in High-Voltage Power Grids

Type: MA thesis

Status: open

Supervisors: Julian Oelhaf, Siming Bayer, Andreas Maier

 

Deep Learning for Fault Localization in High-Voltage Power Grids

Data-Driven Protection on Centralized Sampled-Value Architectures

Background

Recent developments in centralized line differential protection based on IEC 61850-9-2 sampled values enable time-synchronized waveform access across substations and vendors. Instead of exchanging phasors via proprietary interfaces, raw current signals can now be aggregated centrally using standardized communication protocols (see, e.g., Stachel & Schumacher, DPSP 2026).

While current implementations focus on analytical differential protection (87L), the availability of synchronized multi-terminal waveform data opens new research directions for data-driven fault localization.

This project investigates how deep learning models can leverage such centralized waveform architectures to estimate the continuous fault position along transmission lines.

The work is embedded in ongoing research on AI-based protection systems at LME and aligns with emerging centralized protection concepts in industry.

Project Structure – Three Levels of Complexity

The localization model will be evaluated under three structured measurement settings inspired by centralized SV-based protection architectures:

1. Local Only (Single CT/VT)

  • One relay location only
  • Three-phase voltage and current signals
  • 8 independent experiments

Focus: Classical decentralized protection view.

2. Line Only (Both Terminals of One Line)

  • CT/VT measurements from both ends of a transmission line
  • 4 independent experiments

Focus: Centralized two-terminal SV-based differential setting.

3. Full Grid (All Four Lines)

  • Measurements from all eight CT/VT locations
  • Fully time-synchronized multi-line visibility

Focus: Grid-level centralized protection and wide-area learning.

Your Tasks

  • Implement and compare modern deep learning models for fault localization
    (e.g., TCN, InceptionTime, Transformer-based models)

  • Design a structured and reproducible hyperparameter optimization workflow

  • Establish a leakage-aware evaluation protocol with strict train/validation/test separation

  • Analyze model performance under different measurement availability settings

  • Investigate the trade-off between model capacity and generalization

  • Evaluate performance using:

    • MAE / RMSE (in % of line length or km)

    • Error percentiles (P50 / P95)

  • Identify the most robust and generalizable architecture for the predefined localization task

All experiments must be reproducible and statistically sound.

Depending on progress and interest, results may contribute to a scientific publication.

Requirements

  • Computer science or related background
  • Good Python skills (NumPy/Pandas, Git, PyTorch)
  • Basic knowledge of machine learning or deep learning
  • Interest in implementing and evaluating methods from scientific papers
  • Able to attend the weekly in-person meeting in Erlangen (Tuesdays, 14:00)

Project Type

  • Master thesis (research-oriented)
  • Start date: flexible

Apply

Send one PDF to julian.oelhaf@fau.de with the subject:

Application | Thesis | DL Central Fault Localization | <Your Full Name>

Email body (max. 200 words):
Short motivation and your earliest start date.

Attach as one single PDF:

  • CV
  • Transcript (dated)
  • Optional: code links (GitHub, etc.)

📌 Incomplete applications will not be considered.

[1] Centralized line differential protection using inter-substation process bus