Deep Reinforcement Learning for Demand Response in Distribution Networks

Load aggregators can use demand response programs to motivate residential users toward reducing electricity demand during peak time periods. This article proposes a demand response algorithm for residential users, while accounting for uncertainties in the load demand and electricity price, users' privacy concerns, and power flow constraints imposed by the distribution network. To address the uncertainty issues, we develop a deep reinforcement learning (DRL) algorithm using an actor-critic method. We apply federated learning to enable users to determine the neural network parameters in a decentralized fashion without sharing private information (e.g., load demand, users' potential discomfort due to load scheduling). To tackle the nonconvex power flow constraints, we apply convex relaxation and transform the problem of updating the neural network parameters into a sequence of semidefinite programs (SDPs). Simulations on an IEEE 33-bus test feeder with 32 households show that the proposed demand response algorithm can reduce the peak load by 33% and the expected cost of each user by 13%. Also, we demonstrate the scalability of the proposed algorithm in 330-bus and 1650-bus feeders with real-time pricing scheme.