Deep Reinforcement Learning Based Coordinated Voltage Control in Smart Distribution Network
Daner Hu, Yonggang Peng, Jinxiang Yang, Qingtang Deng, Tiantian Cai
Abstract
This paper designs a reactive power optimization strategy based on multi-agent deep reinforcement learning soft actor-critic (MASAC) algorithm. Compared with the traditional method, our proposed method does not depend on accurate power flow modeling based on the data prediction of day-ahead load and distributed generation data. The proposed strategy uses deep neural network to train the action functions of solar photovoltaic (PV) and wind turbine (WT) inverters as multiple agents and completes the training of deep neural networks in the process of interaction with the distribution network environment. Additionally, centralized training and decentralized execution framework are adopted to solve the Volt-VAR control problem. In our MASAC algorithm, each agent is regarded as an actor trained with a critic during the training process. Finally, a numerical study is given to verify the effectiveness of the MASAC algorithm.