Spectral Clustering of Voltage Profiles for Phase Identification in LV Grids: Accuracy Gains From Kernel Optimization

Information about the phase connectivity of lowvoltage grid customers is often missing or insufficiently accurate, and thus needs to be estimated. Various estimation methods have been proposed that utilize metering data analysis. This paper proposes modifying the classic spectral clustering methodology to estimate phase connectivity with improved accuracy, relying exclusively on the customers' voltage magnitude profiles. The modification includes the use of Gaussian Mixture Models for Kernel optimization, a technique that has been proven to improve estimation accuracy in other application contexts significantly. The paper provides the application framework necessary to implement the methodology for estimating phase-connectivity in LV grids. The proposed methodology is applied to the IEEE 116 LV grid using a week of synthetic voltage magnitude data to illustrate its advantages over other clustering implementations. It is then applied to an existing LV grid with real-world metering data for validation in practical contexts.