Seasonal similarity for behind-the-meter PV disaggregation in LV networks

Behind-the-meter photovoltaic (PV) generation distorts low-voltage (LV) net-load measurements, reducing visibility of true demand. This paper tests whether seasonal labelling changes measured day-to-day demand similarity in Northern Ireland LV monitoring data, and whether that behavioural context can improve unsupervised PV disaggregation. Using 2024 net-load data (resampled to 15-minute resolution) from NIE Networks with CAMS irradiance and ERA5-Land temperature, PV-inactive intervals are identified from low irradiance and nighttime periods and used to compare daily profile similarity under calendar, DSO, irradiance-derived, and temperature-derived seasons. The key novelty is a similarity-conditioned centroid selection strategy that replaces night-only matching with irradiance-informed PV-inactive interval comparison. Temperature-based seasons provide the most consistent within-season behaviour and the clearest weekday–weekend separation, so they are used to condition demand-prototype selection in a centroid-library nonnegative least squares disaggregation. In a May 2024 case study (five sites, 122 site-days), similarity-conditioned selection reduces mean RMSE from 14.9 to 14.1 kW (5.1%) and mean MAE from 10.9 to 10.5 kW (4.7%).