Reinforcement Learning Based Pricing for Demand Response

Smart pricing based demand response (DR) programs can enable the system to shape load profiles to improve system reliability and performance. Existing works on pricing based DR often assume that users' response functions are available or predictable at the load serving entity (LSE) side. Due to divergent consumption habituates of the users, it is challenging to have an accurate estimate of users' responses. Clearly, any mischaracterization of users' responses would result in higher system costs. To tackle this challenge, we leverage reinforcement learning to learn users' response functions. Specifically, we formulate the DR problem as a stochastic optimization problem, in which random responses are considered due to users' volatile behaviors. We develop a reinforcement learning based algorithm to solve a pricing strategy for DR without assuming any specific forms of users' response functions. The proposed reinforcement learning algorithm is shown to converge to an equilibrium with near optimal performance, which is corroborated via numerical simulations.