Low-Voltage Consumer Phase Identification Method Based on Randomised Domain Search with Feature Constraints

Low-voltage distribution networks (LVDNs) are crucial for maintaining user satisfaction, yet their vast and dynamic nature complicates the accurate monitoring of phase sequence information, leading to issues such as line losses and phase imbalances. Current phase identification methods, including manual inspections and signal-based techniques, are inefficient and vulnerable to external disturbances. This paper proposes a novel methodology combining Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction with an enhanced CLARAN-Medoids clustering algorithm. UMAP captures the intrinsic nonlinear relationships in voltage time-series data, while CLARAN-Medoids integrates a randomized neighborhood search and Cannot-link constraints to enhance clustering robustness against noise and phase misalignment. Experimental results from three transformer districts in Southern China demonstrate the superior performance of our approach, with accuracy improvements of up to 10.9% compared to traditional methods. This research provides a reliable solution for precise phase sequence identification in LVDNs, with potential applications in advanced metering infrastructure and dynamic grid management.