Dynamic Graph Neural Networks for Uncertainty-Aware Residential Photovoltaic Power and Load Disaggregation

As residential rooftop photovoltaic (PV) adoption increases, more local solar output becomes hidden behind-the-meter (BTM). Moreover, the stochastic and volatile nature of PV generation, combined with its invisibility, increases the variability and unpredictability of overall energy demand. Disaggregating BTM PV generation from actual net load is essential for reliable distribution grids operation. This paper presents an uncertainty-aware dynamic graph variational framework that disaggregates BTM PV generation and load using only smart-meter net load measurements. Using a dynamic graph, we represent 100 PV-equipped Dutch households as nodes, where edges encode correlations among their net demand time series. The Graph Attention Recurrent Neural Network encoder processes this graph with graph convolutions, self-attention, and recurrent layers to capture spatial relationships and temporal dynamics. A multi-quantile gated recurrent unit decoder then infers each home's separate solar and load time series, outputting probabilistic estimates that quantify uncertainty in each output. Designed for real-time inference, it enables online distribution network level monitoring. We validate the method on real data demonstrating accurate disaggregation and reliable uncertainty quantification.