AI Grid Protection – Machine Learning for Power System Protection

Coordinated power system protection using machine learning (Netzschutz-KI)

This project is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Project number 535389056.

On this page:

• Abstract
• Data and Code Availability
  • Simulation Datasets and Benchmarks
  • Code Repositories
• Publications
• Project Details

Abstract

To address these challenges, the project explores learning-based protection agents that operate directly on voltage and current measurements. Research covers fault detection, fault classification, faulted-line identification, and fault localization, together with event analysis and selectively coordinated fault clearance. Conventional methods, feature-based machine learning, and neural networks are evaluated under common protocols. A particular focus is whether these methods remain reliable across unseen fault scenarios, operating conditions, network topologies, and different levels of measurement availability.

A central focus of the project lies on coordination: moving beyond isolated relay decisions toward protection agents that support selective, system-wide fault clearance. This includes research on reinforcement learning for protection decisions and assessment through missed faults, false alarms, decision timing, and selectivity. Open resources such as PROTECT-90 and EvEMTBench support reproducible simulation-based evaluation, complemented by research on real-world disturbance recordings, fault-inception estimation, and simulation-to-real transfer. Real-time laboratory validation is a further project objective, with the long-term goal of enabling transparent and reliable AI-based protection concepts for future power systems.

Data and Code Availability

The AI Grid Protection project follows the principles of reproducible and transparent research. Publicly released datasets, evaluation frameworks, and reference implementations are linked below. Their documentation describes data access, intended use, and the protocols used to reproduce the reported results.

Simulation Datasets and Benchmarks

The project uses electromagnetic transient (EMT) simulations to provide voltage and current waveforms with documented scenario metadata. The following resources support controlled model comparisons and studies of robustness and generalization:

• PROTECT-90: A Fault Dataset for Power System Protection
  A dataset of 9,022 simulation episodes from a 90 kV double-line transmission topology. Each episode contains one second of synchronized voltage and current waveforms sampled at 6.4 kHz, with 48 channels from eight measurement locations. Metadata describe fault properties, operating conditions, and topology states.
  Dataset DOI: 10.5281/zenodo.21109169
  Paper: PROTECT-90: A Fault Dataset for Power System Protection
• EvEMTBench Simulation Dataset
  A multi-grid EMT dataset covering four grids spanning 20–345 kV, with synchronized voltage and current waveforms for faults and other operating events. Scenario metadata support research across different grids, voltage levels, and event types.
  Dataset description and access information: A simulation based dataset of faults and events for machine learning in power systems
• EvEMTBench: An Open Benchmark for Machine Learning in Power System Protection
  An executable benchmark with standardized tasks, data partitions, measurement access, metrics, and reference baselines. It supports evaluation under local, line-level, and global observability, predefined distribution shifts, and zero-shot and fine-tuned cross-grid transfer.
  Preprint: EvEMTBench benchmark paper
  Documentation and code: https://github.com/EvEMTBench/evemtbench-benchmark

The datasets provide waveforms and scenario metadata, while the evaluation frameworks specify how methods are trained, tested, and compared. Together, they support reproducible evaluation under controlled and well-documented conditions.

Real-world disturbance analysis: The project also studies field-recorded voltage and current waveforms from RTE, including fault-inception estimation and the distinction between fault-like disturbances and non-fault transients. The associated study is available as Fault Inception Detection in Real-World Disturbance Data for Power System Protection.

Code Repositories

Reference implementations, evaluation pipelines, and experiment configurations are available via the following public GitHub repositories:

• Protection Evaluation Framework
  https://github.com/julianoelhaf/protection-eval-framework
  Companion code for the IJEPES framework article, providing standardized fault classification and localization on PROTECT-90 with common preprocessing, decision windows, episode-grouped evaluation, and reproducibility artifacts.
• Deep Learning Fault Analysis Protection Benchmark
  https://github.com/julianoelhaf/dl-fault-analysis
  Reproducible benchmark accompanying the EPSR article, comparing deep-learning architectures for fault detection, classification, line identification, and localization under common preprocessing and evaluation protocols.
• PROTECT-90 Companion Repository
  https://github.com/julianoelhaf/protect90-dataset
  Dataset documentation, citation guidance, and a guided notebook for loading waveforms and inspecting scenario metadata.
• EvEMTBench Benchmark
  https://github.com/EvEMTBench/evemtbench-benchmark
  Reference implementation for protection and event-analysis tasks, including fixed partitions, baseline evaluation, observability studies, distribution shifts, cross-grid transfer, and reproducible reporting.
• Offline Reinforcement Learning for Distribution-Grid Protection
  https://github.com/julianoelhaf/offline-cql-protection
  Companion code for the PESS 2026 contribution on conservative Q-learning for line-selective protection decisions, including data preparation, training, and evaluation.

Together, these resources enable independent validation of the reported results and support further research on learning-based and coordinated protection concepts for future power systems.

Publications

Project Details

2024

• Netzschutz-KI: Coordinated grid protection based on machine learning methods

  
  (Third Party Funds Single)
  Project leader: Andreas Maier, Johann Jäger, Siming Bayer
  Term: July 1, 2024 - June 30, 2027
  Acronym: Netzschutz-KI
  Funding source: DFG-Einzelförderung / Sachbeihilfe (EIN-SBH)
  
  →More information