Reactive Power Optimization of Distribution Network Based on Deep Reinforcement Learning and Multi Agent System

This paper presents a reactive power optimization method of distribution network based on deep reinforcement learning (DRL) and multi-agent system. Firstly, a method of extracting distribution network fusion features is proposed, which extracts statistical features from the operation data of distribution network. Then, the statistical features and historical control strategy are taken as inputs, and the network loss and voltage deviation are taken as outputs respectively. The network loss agent and voltage deviation agent are trained. The reactive power optimization problem is transformed into a multi-step Markov decision-making process, and the Double DQN algorithm is used to construct DRL agent. A multi-agent system is composed of network loss and voltage deviation agents and DRL agents. Network loss and voltage deviation agents provide rewards for DRL agents, while DRL agents provide strategies for network loss and voltage deviation agents. They cooperate with each other to obtain the optimal strategy. The reactive power optimization control experiment of IEEE-37 bus distribution network is carried out. The results show that the proposed method can effectively reduce the node voltage offset and network loss. It has nothing to do with the model and parameters of the distribution network system, and has fast online decision-making speed. It can improve the operation economy of the distribution network.