Smart-meter phase clustering in low-voltage distribution network

Low-voltage (LV) distribution networks have evolved over decades with limited prioritization of detailed documentation and digitization. Unlike high-voltage (HV) networks, which are well-mapped and monitored, LV networks often lack critical data, such as labelled feeders and phase identification in GIS systems. This data gap combined with the growing integration of distributed energy resources (DER) and the adoption of time-of-use tariffs presents challenges for effective planning and operations. This paper introduces a method for phase identification based on voltage measurements from smart meters (SMs). The analysis utilizes a real-world dataset comprising 15-minute sampled voltage time series from 89 and 109 single-phase SMs connected to two transformers. The developed machine learning pipeline thoroughly addresses data pre-processing and applies spectral clustering and unsupervised clustering evaluation. The pipeline was designed to ensure robustness against distinct voltage profiles often observed for different transformers. By utilizing existing SM data, the proposed approach eliminates the need for costly on-site data collection. It enables distribution system operators to automatically map phase connectivity, delivering immediate benefits such as improved load balancing, loss reduction, and enhanced network planning. This method provides a scalable, cost-effective solution for addressing critical data gaps in LV distribution networks.