A Real Controlled-Fault Benchmark for Power-Anomaly Detection and Identification on IoT Edge Nodes (three-station 1 Hz dataset)

Three-station, 1 Hz power-monitoring benchmark with real, physically induced faults, recorded by ESP32 + PZEM-004T nodes over seven consecutive days per station (3 × 604,800 samples = 1,814,400 labelled samples, ≈99.8% directly measured). Sixteen fault types in eight categories (load, voltage, frequency, power quality, sensor, outage, contextual, compound) with five severity grades and per-sample labels (anomaly, anomaly_type, fault_category, severity, event_id). Each CSV has 29 columns including raw electrical quantities (V, I, P, energy, frequency, power factor), data-quality flags, and precomputed features. Grid-side categories (frequency deviation, sag/swell, outage) were emulated at the supply side under controlled protocols (programmable AC source / variac / MCCB switching); see DATASHEET.md for full collection details. Includes feature-extraction and cross-validation code reproducing all tables and figures of the companion paper (submitted to Electrical Engineering, Springer). License: CC BY 4.0.