Self-Healing Medium-Voltage Distribution Networks for Hyperscale Data Centres Through Digital Twin-Based Fault Prediction

Abstract

Medium-voltage (MV) distribution feeders are becoming a critical part of the infrastructure at hyperscale data centres, which are required to experience very low fault rates, but these fault rates are not compatible with overhead or underground MV lines. This review combined seventeen studies related to self-healing distribution automation, the Digital Twins engineering and fault prediction using artificial intelligence to assess the use of a Digital Twins layer for self-healing logic to ensure critical data-centre loads. It chronicles the growth from manual switching and centralized supervisory control and data acquisition (SCADA) restoration to decentralized multi-agent coordination and reconfiguration using deep reinforcement learning; and compares fault detection and location techniques from impedance-based estimation to wavelet-transform machine learning and traveling-wave techniques. The evidence from the literature reviewed is compiled into seven comparative tables and six figures, all of which relate to restoration-time, reliability-index, detection-accuracy, and cybersecurity-mitigation evidence. The synthesis suggests that the mean restoration time decreases from about 185 minutes (when switching is manual) to less than seven minutes if the mean restoration time is shortened using digital-twin-assisted predictive self-healing; the System Average Interruption Duration Index (SAIDI) decreases from a baseline of 142 minutes per customer per year to around 17 minutes when the mean restoration time is shortened using digital-twin-assisted predictive self-healing; the accuracy of fault detection increases above 98 percent when the hybrid machine-learning classifiers are fused with digital twin state estimates. Digital-twin-based cybersecurity monitoring is also demonstrated to increase anomaly-detection and anomaly-mitigation rates by 30-45 percentage points, compared to conventional monitoring, in five threat categories studied. These results validate the self-healing MV networks as a technically viable solution for achieving near-continuous supply reliability needed by hyperscale computing facilities, and highlight several areas of interoperability, model fidelity, and security that would benefit from further research.