Multi-agent deep reinforcement learning for mitigation of unbalanced active powers using distributed batteries in low voltage residential distribution system

High penetration and uneven distribution of single-phase rooftop PVs and load demands in power systems can cause unbalanced active powers, which in turn can adversely affect power quality and system reliability. This paper proposes a multi-agent deep reinforcement learning-based strategy to compensate for the unbalanced active powers by employing single-phase battery systems distributed in the LV residential distribution system and subsidized by the utility. First, the unbalanced active powers are formulated as a Markov game. Then, the Markov game can be solved by a multi-agent deep deterministic policy gradient algorithm. The proposed strategy uses only local measurements, and the experiences of the agents are shared in a centralized manner during training to achieve cooperative task. Information about phase connections of the battery systems is no longer required. The proposed strategy can learn from historical data and gradually become mastered. The four-wire LV residential distribution system uses real data from rooftop PVs and demands for verification. As adaptive agents, the battery systems are able to cooperatively operate by charging/discharging active powers so that neutral current at the point of common connection can be minimized.