Online learning for demand response

摘要

Demand response is a cheap source of flexibility for power systems with renewables. Loads are uncertain due to unavailable physical models, random factors such as weather and human behavior, and privacy. Load aggregators can use online learning to characterize loads as they use them, reducing uncertainty without relying on costly pilot studies. In this talk we discuss multi-armed bandit and online convex optimization-based approaches to demand response. In both cases demand response motivates novel extensions to standard frameworks, for which we give low-regret algorithms that perform well in numerical studies.