Dynamic Distribution Network Reconfiguration Using Reinforcement Learning

• Published in: 2019 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm) pp. 1 - 7
• Main Authors: Gao, Yuanqi, Shi, Jie, Wang, Wei, Yu, Nanpeng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2019
• Subjects: Conferences; Dynamic distribution network reconfiguration; Heuristic algorithms; Network topology; Q-learning; Reinforcement learning; Smart grids; Substations; Switches; training data augmentation
• DOI: 10.1109/SmartGridComm.2019.8909777

Dynamic distribution network reconfiguration (DNR) algorithms perform hourly dynamic status changes of sectionalizing and tie switches to reduce network line losses, minimize loss of load, or increase hosting capacity for distributed energy resources. Existing algorithms in this field have demonstrated good results when network parameters are assumed to be known. However, in practice inaccurate distribution network parameter estimates are prevalent. This paper solves the minimum loss dynamic DNR problem without the network parameter information. We formulate the DNR problem as a Markov decision process problem and train an off-policy reinforcement learning (RL) algorithm based on historical operation data set. In the online execution phase, the trained RL agent determines the best network configuration at any time step to minimize the expected total operational cost over the planning horizon, which includes the switching costs. To improve the RL algorithm's performance, we propose a novel data augmentation method to create additional synthetic training data based on the existing data set. We validate the proposed framework on a 16-bus distribution test feeder with synthetic data. The learned control policy not only reduces the network loss but also improves the voltage profile.