Distributed Multi-Community-Collaborative Method for Multi-Site Behind-the-Meter Solar Disaggregation for Heterogeneous Communities

To effectively manage distribution network (DN) operations, it is imperative for DN operators to accurately disaggregate the behind-the-meter (BTM) solar generation for various communities at concerned DN nodes. In numerous regions, most communities may have very few or even no submeters for measuring solar generation, making it very difficult to achieve a satisfactory result using common solar disaggregation methods. This paper proposes a distributed multi-community-collaborative BTM solar disaggregation method, by tactically utilizing the data from various communities at different DN nodes. A two-stage collaborative framework is designed to tackle the notorious data heterogeneity issue among various communities. In the first stage, we train general models and a set of personalized models that are tailored for partly observable communities (POCs). In the second stage, a suitable general model is selected and then adapted through adversarial training to enable reasonable disaggregation for an unobservable community (UOC). The entire framework is executed in a distributed manner to avoid heavy computational burdens on the DN control center. To address communication dropout issue, a modified aggregation scheme is further proposed. Case studies demonstrate the efficacy of the proposed method. In particular, its solar disaggregation error can be reduced by 30% or more for both POCs and UOCs in comparison with baseline methods.