A Multi-Agent Deep Reinforcement Learning based Voltage Control on Power Distribution Networks

Due to the high penetration of renewable energy in the distribution network, the exponential increase in the amount of data collected and variables makes it difficult for centralized control methods to achieve real-time voltage regulation. Besides, hardware conditions (e.g., communication equipment) limit its application in practice. Therefore, a model-free multi-agent deep reinforcement learning (MADRL) voltage control strategy is developed in this paper. The proposed MADRL control strategy is carried out with a framework of centralized training and distributed execution. We apply deep deterministic policy gradient algorithm to help each agent control its corresponding PV inverter in a distributed manner. However, during the training process, the agent could observe other agents' information to improve training. The simulation on a 33-bus distribution network is carried out to illustrate the effectiveness of the proposed method, and its superiority is also validated by comparing with traditional methods.