PROTECT-90: A Fault Dataset for Power System Protection

PROTECT-90 is an open high-voltage fault waveform dataset for reproducible research in power system protection, transient waveform analysis, and learning-based fault analysis. The dataset contains 9,022 electromagnetic transient (EMT) simulation episodes generated on a standardized 90 kV double-line transmission topology. Each episode provides 1 second of synchronized three-phase voltage and current waveforms sampled at 6.4 kHz, corresponding to 128 samples per 50 Hz cycle. Measurements are recorded at eight protection-relevant locations, resulting in 48 waveform channels per episode. All episodes are released as raw instantaneous waveforms without feature engineering. Each waveform file is accompanied by structured, machine-readable metadata describing the fault type, faulted line segment, normalized fault location, fault resistance, fault inception time, randomized grid parameters, load-flow conditions, and topology switching states. Start here Dataset download and DOI: this Zenodo record Companion repository, documentation, citation guidance, and guided notebook: https://github.com/julianoelhaf/protect90-dataset Paper preprint: https://arxiv.org/abs/2606.24298 Project context: AI Grid Protection project page Dataset contents 9,022 EMT simulation episodes 90 kV double-line transmission topology 1.0 s duration per episode 6.4 kHz sampling rate 48 synchronized voltage and current waveform channels Eight protection-relevant measurement locations Raw instantaneous waveforms without feature engineering Machine-readable metadata for fault, grid, load, and topology parameters Intended use PROTECT-90 is intended as a physically grounded reference benchmark for protection-oriented signal-processing and machine-learning research. It supports studies on: Fault detection Fault classification Fault localization Faulted-line identification Transient waveform analysis Reduced-observability studies Robustness and domain-shift evaluation Reproducible benchmarking of learning-based and signal-processing methods Scope and limitations PROTECT-90 is a simulated EMT dataset. It is designed as a transparent and reproducible research benchmark, not as a replacement for utility-specific protection studies, field validation, or certified protection functions. Users are encouraged to report task definitions, train/test partitioning, observability assumptions, preprocessing steps, and evaluation metrics explicitly when using the dataset. Citation If you use PROTECT-90, please cite both the dataset DOI and the accompanying paper where appropriate. Citation information is provided in the companion repository and in the metadata of this Zenodo record.