Behind-the-Meter PV Disaggregation With Copula-Constrained Bayesian Variational Autoencoder and Feature Clustering

A quantitative assessment of the uncertainty of behind-the-meter (BTM) PVs is critical for accurate forecasts, voltage control, and distribution system security. This paper proposes a novel distribution-constrained Bayesian variational autoencoder approach with feature-based clustering (DC-BVAE-FBC), to disaggregate BTM PV generation from net load data probabilistically. By incorporating Bayesian inference into the variational autoencoder framework, the approach accurately captures uncertainty in the disaggregation process. A significant advance lies in constraining the latent space of the BVAE using a Gaussian C-Vine copula, which models the target component’s distribution—either native load or PV generation, ensuring that latent samples align with real-world patterns, thereby enhancing disaggregation accuracy. A feature-based clustering (FBC) strategy is also proposed for group days based on meteorological factors influencing power consumption and PV output, enabling tailored DC-BVAE models for each cluster to optimize predictions. Validated with real smart meter data with BTM PVs from London Advanced Metering Infrastructure (LCL-AMI) and Pecan Street Inc. Dataport (PS-ND/PG), this methodology demonstrates superior performance over existing techniques, offering a robust solution for modern grid management challenges.