Machine Learning on Real-World Power Grid Oscillograms

📋 Type Project
⚡ Status open
👤 Primary supervisor Julian Oelhaf

Abstract

This project investigates machine learning for event analysis in real-world power-grid oscillograms using the OscGrid dataset. The focus is on reproducing existing baselines, building recording-independent and leakage-aware evaluation protocols, and assessing how standard ML and deep-learning models perform under causal, transition-aware conditions. The project emphasizes reproducible research, robust evaluation, and realistic power-system event classification.

Machine-learning methods for power-system event analysis are often developed and evaluated on simulated data. Real substation recordings are considerably more challenging: switching operations, abnormal operating conditions, faults, measurement imperfections, and event transitions can produce similar waveform signatures.

In this research project, you will work with OscGrid, a recently published dataset containing more than
50,000 real-world COMTRADE recordings from electrical substations. Of these, 480 carry per-timestamp expert annotations; this labelled subset is the basis of the project. The recordings originate from 0.4 – 35 kV distribution networks and are dominated by bus-transfer operations, motor starts, and single-phase earth faults rather than transmission-level short circuits. You will reproduce the published machine-learning baseline and investigate how performance changes when the evaluation is made more rigorous and closer to realistic operation.

What you’ll do

  • Reproduce a published ML pipeline for classification of real power-grid oscillograms [1]
  • Document discrepancies between the paper, the released code, and the dataset documentation. Exact reproduction is not expected; characterising the mismatch is part of the result.
  • Audit the dataset and evaluation pipeline, including labels, channels, preprocessing, and data splits
  • Implement recording-independent evaluation so that data from the same physical recording cannot occur in both training and test sets
  • Train and compare standard ML and deep-learning models, such as Random Forest, MLP, CNN, and GRU
  • Evaluate causal event recognition around real event transitions using only waveform information available up to the decision time
  • Analyze where models fail, particularly for switching, abnormal operating conditions, and protection-relevant events

Depending on progress, additional experiments may investigate observation windows, voltage vs. current measurements, or alternative input representations.

What you’ll learn

  • Applied machine learning on real-world power-system waveform data
  • Time-series classification and signal-processing fundamentals
  • How data leakage and experimental design affect ML results
  • Reproducible research workflows using Python and Git
  • Evaluation of ML for safety-critical engineering applications
  • Basics of power-system protection and substation event analysis

Requirements

Required:

  • Solid Python programming skills
  • Good understanding of basic machine-learning concepts
  • Experience with scikit-learn and/or PyTorch
  • Ability to work independently with an existing research codebase
  • Experience using Git
  • Careful and reproducible working style

Helpful but not required:

  • Courses in Machine Learning, Pattern Recognition, Deep Learning, Signal Processing, or Time-Series Analysis
  • Experience with scientific Python, Linux, or HPC systems
  • Basic knowledge of electrical engineering, power systems, or protection

Project Type

  • Bachelor / Master Project (research-oriented, typically 10 ECTS)
  • Duration: approximately 10–12 weeks
  • Start date: flexible, by arrangement
  • Location: FAU Erlangen-Nürnberg

Attendance at the weekly research meeting in Erlangen is mandatory.
The meeting currently takes place on Tuesdays at 14:00. Students are expected to work independently between meetings and present concrete results, code, and open questions each week.

Apply

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

Application | Project (10 ECTS) | Real-World Power Grid ML | <Your Full Name>

Email body (max. 200 words):

  • Short motivation
  • Your earliest possible start date
  • Confirmation that you can attend the weekly Tuesday 14:00 meeting in Erlangen
  • Relevant ML / signal-processing courses you have completed

Attach as one PDF:

  • CV
  • Current transcript of records
  • Link to GitHub/GitLab or one representative programming project

References

[1] A. Evdakov et al., A Dataset of Real-World Oscillograms from Electrical Power Grids, Scientific Data, 2026.
Paper |
Dataset |
Code

[2] J. Oelhaf et al., A Scoping Review of Machine Learning Applications in Power System Protection and Disturbance Management, International Journal of Electrical Power & Energy Systems, 2025.
Paper

[3] S. Kapoor and A. Narayanan, Leakage and the Reproducibility Crisis in Machine-Learning-Based Science, Patterns, 2023.
Paper