This paper presents an extensive dataset of real-world oscillograms that capture voltage and current signals from electrical substations. The dataset aims to advance research on power system analysis, fault detection, and machine learning-driven relay protection. It includes approximately 50,000 oscillograms recorded with sampling rates up to 8 kHz. A manually annotated subset of 480 oscillograms categorizes events into four groups: noise-dominated signals without deviations; routine equipment operations such as load switching, circuit breaker actions, motor startups, or transformer energization; non-critical deviations compliant with operational standards, including single-phase ground faults or minor voltage dips; critical faults such as short-circuits or voltage collapses requiring immediate relay protection activation. The unannotated part of the data supports self-supervised and unsupervised machine learning methods, enabling tasks like feature extraction, anomaly detection, and latent pattern identification in power networks. The dataset facilitates various applications, including the validation of synthetically trained models, the refinement of adaptive relay protection algorithms, and the development of fault detection and diagnosis systems.
Evdakov et al. (Thu,) studied this question.