Behind-the-Meter PV Disaggregation Under Limited Sample Budget: A User-Level Active Learning Strategy

Accurate estimation of Behind-the-meter Photovoltaic (BTM PV) generation is essential for load forecasting and grid planning. Most distributed PV systems are installed behind customer meters, making their output unobservable. Disaggregating PV output from net load is therefore critical for improving distribution network observability. However, existing deep-learning-based disaggregation methods require large labeled datasets, and obtaining such data is costly. Under limited budgets, only a few users can be labeled, which constrains model performance. This paper proposes a user-level BTM PV disaggregation method based on active learning with adaptive weight updates, aiming to maximize model performance with minimal labeling cost. We design a multi-dimensional user value assessment system incorporating epistemic uncertainty, aleatoric uncertainty, and representativeness. In each iteration, the most informative users are selected for sub-meter installation. To dynamically optimize the selection strategy, we propose an adaptive weight update mechanism that adjusts the weights for the next round based on performance improvement gradients across dimensions. This closed-loop feedback enables the strategy to capture evolving model needs and prioritize users that yield the greatest performance gains. The proposed method is validated on the public Ausgrid dataset, and experimental results demonstrate its effectiveness under limited budgets.