A Privacy-Preserving Federated Learning Method for Probabilistic Community-Level Behind-the-Meter Solar Generation Disaggregation
Jun Lin, Jin Ma, Jianguo Zhu
Abstract
Accurate estimation of residential solar photovoltaic (PV) generation is crucial for the power distribution and demand response program implementation. Currently, most distributed PVs are installed behind-the-meters (BTMs), and are thus invisible to the utilities. The existing methods separate the BTM solar generation from the available net load in a centralized manner assuming that all data are accessible to utilities. However, this can cause privacy issues, since the data are owned by different utilities and they may be unwilling to share their data. To this end, a novel method is proposed for disaggregating community-level BTM solar generation using a federated learning-based Bayesian neural network (FL-BNN), which can preserve the privacy of utilities. Specifically, a Bayesian neural network (BNN) is designed as the probabilistic energy disaggregation model with the ability to capture uncertainties. The BNN training process is extended into a decentralized manner based on the federated learning framework. To enable the model customized for each community, the layers of BNN are categorized into shallow and deep layers, and a layerwise parameter aggregation strategy is proposed to update the model. Both community-specific features and community-invariant features can be learned. The effectiveness of the proposed method is validated on a publicly available dataset.