Classification and characterization of intra-day load curves of PV and non-PV households using interpretable feature extraction and feature-based clustering

Abstract Load pattern categorization plays a significant role in enhancing the understanding of demand characteristics of different cohorts of energy customers on a distribution network. For a distribution network with embedded photovoltaic (PV) systems, it is also desirable to enumerate these premises since unreported PV installations can have a detrimental impact on voltage control, frequency regulation, and incur reverse power flows. In this paper, a holistic smart meter data analytics approach is proposed for classifying and characterizing the intra-day load curves of PV and non-PV households. Unlike existing studies based on raw time-series data, a series of interpretable and discriminating global and peak-period features are first developed to extract the physical information of load patterns. A two-step feature-based clustering is then applied to classify the load profiles of PV and non-PV households. A post-clustering quantitative compositional analysis is developed to characterize the energy use variability and compositional changes for each household. The proposed PV household identification method allows network owners to identify the existence of PV installations in a particular area for voltage control and ancillary service provision. In the longer term, effective load pattern categorization and post-clustering characterization can help better understand demand characteristics of different cohorts of customers, network utilization, and network-level changes in load growth.