On Privacy Preservation of Online Learning for Optimal Participant Selection in Demand Response

Online learning offers significant potential for load aggregators to optimize participant selection, thereby enhancing the efficiency of residential demand response programs. However, such learning processes rely on sensitive residential data—such as load adjustment and indoor temperature—to inform decisions and update parameters. The exposure of this data could lead to the revelation of electricity usage patterns and personal living habits, raising serious privacy concerns. To address this issue, this letter proposes a privacy-preserved online learning algorithm, for the first time, seamlessly integrating a data encryption scheme within a contextual multi-armed bandit framework. The proposed algorithm is theoretically validated to ensure computational accuracy, robust privacy protection, and effective learning outcomes. Numerical simulations further demonstrate the algorithm’s ability to efficiently learn residential demand response behavior while safeguarding privacy.