Efficient Reactive Power Control Using Reinforcement Learning under Inaccurate Power Network Model

This paper utilizes distributed multi-agent reinforcement learning (MARL) based optimization technique to arrive at the efficient reactive power control (ERPC) in electric power systems by considering the inaccurate power network. Every agent will incorporate with the adjacent agent only if it is electrically combined for data sharing utilizing global information sharing by using the consensus-based algorithm. A Q-learning-based distributed RL-based algorithm is used and incorporated with the consensus algorithm for the global reward search to limit the power losses and operational constraints. It is demonstrated to be applicable for online learning, and a model-free methodology has been considered and provides the optimal solution. It has been observed that supervised learning can speed up the learning process and reduce undesired results.