Control and Coordination of Multiple PV Inverters in Power Distribution Network using Multi Agent Deep Reinforcement Learning

The growing power demand can be realized with the increased saturation of PVs in the distribution network (DN) of the power system. Moreover, low-cost energy with less emission of polluted gases can be achieved. Along with these advantages, it has some disadvantages as well. The integration of a high number of PVs in DN causes voltage deviation, which is undesirable. To reduce voltage deviation and keep the voltage within a particular range, the agents (PVs) must be controlled and coordinated in real-time. This real-time control and coordination are achieved through a multi-agent scheme of deep reinforcement learning (DRL). An Integrated PV inverter is considered an agent and its action can be divided into actor networks (AN) and critic networks (CN). PV inverter has an Actor-network, having the capability of producing or absorbing the reactive power accordingly. The CN evaluates the performance of the AN and produces a Q-value according to the action. Each agent tries to maximize its Q-value. Moreover, all the agents are arranged in a distributed and decentralized scheme to achieve real-time coordination among them. The proposed framework is analyzed on the PV-integrated IEEE-33 power buses. Reactive power control of all the PVs collectively retains the voltage within a safe range of 5%.