DCPower-ICS: A Labeled ICS Anomaly Detection Dataset for Data Center Power Infrastructure

DCPower-ICS is a labeled anomaly detection benchmark dataset for data center power infrastructure, generated from a physics-faithful digital twin simulator. The dataset covers a 480V data center power plant including utility feed, point-of-common-coupling (PCC) breaker, automatic transfer switch (ATS), backup generators (GEN1, GEN2), battery energy storage system (BESS), UPS, critical IT load, cooling load, and non-critical shed-able load. The dataset contains 172,800 time-series observations at 1-second resolution across 40 sensors, divided into a 24-hour training split (normal operation with planned maintenance events) and a 24-hour test split (normal operation interleaved with 254 labeled attack windows). The test split covers all 22 attack and anomaly scenarios across four categories: maintenance errors, commissioning faults, cyber-physical attacks, and operational disturbances. All five unsupervised baseline detectors (Isolation Forest, One-Class SVM, LOF, PCA reconstruction, AutoEncoder) achieve segment recall of 1.0 across all 22 scenarios, with global F1 ranging from 0.675 to 0.752. An interactive live demo is available at https://dcpower-ics.onrender.com. Generation code and evaluation scripts are available at https://github.com/vheydari/dcpower-ics.