Fault Detection and Localization in LV Distribution Networks Using Anomaly Detection in Smart Meter Data and Distribution Network Digital Twin

This paper presents a novel data-driven technique for fault detection and localization in low voltage distribution networks (LVDNs). The proposed method leverages smart meter (SM) time series voltage magnitude data and a Distribution Network Digital Twin (DNDT) to address key limitations of traditional approaches, including their reliance on customer trouble calls, the inaccuracy of impedance-based methods and the need for large, complex training datasets for machine learning (ML)-based fault localization. Voltage anomalies in half-hourly SM data are first detected using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, enabling automatic identification of both permanent and temporary faults. The DNDT then maps affected customers to narrow the fault search area to specific network branches. The Branch-Current Branch-Voltage (BCBV) matrix, combined with last-gasp SM voltage data identifies the exact faulted branch. Finally, a Random Forest (RF) model, trained on synthetic fault scenarios generated by DNDT, precisely locates the fault and classifies its type. Validation on both IEEE 123-bus test system and a large-scale practical Midwest U.S. 240-bus test system, demonstrates minimum F1-score of 99.6% for fault type classification and R-Squared score 99.8% for fault-pinpointing, with shorter training time compared to benchmark methods. Key advantages include eliminating the need for customer trouble calls, compliance with customer data privacy regulations, and reduced computational complexity, enhancing fault management efficiency in modern LVDNs.