Conditional diffusion modeling for probabilistic behind-the-meter PV disaggregation

Distributed behind-the-meter (BTM) PV systems are expanding rapidly, yet their generation is often metered together with demand, resulting in operational challenges for grid operators. This paper introduces a novel generative diffusion-based methodology for probabilistic BTM PV disaggregation that directly learns the distribution of PV generation from low-resolution smart meter data with 30-minute granularity. The proposed conditional diffusion model combines a domain-specific conditioning strategy with a sequential U-Net architecture to generate multiple plausible daily PV profiles, enabling both accurate point estimates and well-calibrated prediction intervals. On a real-world Australian benchmark, the model outperforms traditional deterministic baselines and representative probabilistic alternatives, achieving a mean absolute scaled error (MASE) of 0.954 and a coverage probability of 61.5% for the 50% prediction interval. Sensitivity studies highlight the effect of conditioning signals and training data volume, while cross-dataset experiments confirm generalization to Dutch residential data. By disaggregating BTM PV and quantifying uncertainty, this study demonstrates the potential of generative models to enhance grid visibility and support applications such as load forecasting, planning, and flexibility assessment.