Accelerated DRL Agent for Autonomous Voltage Control Using Asynchronous Advantage Actor-critic

This paper presents a novel data-driven parallel framework for autonomous voltage control (AVC) of the power grid. The proposed framework employs a distributed Deep Reinforcement Learning algorithm named Asynchronous Advantage Actor-Critic (A3C) to regulate voltage profiles in a power grid. A well-trained accelerated agent is obtained in the proposed framework by employing multiple workers simultaneously and interacting with a power grid simulator repeatedly. With the proposed framework, multiple threads can run in parallel. A well-trained agent, which utilizes the parameters acquired by the joint training of multiple workers, is obtained and tested through a realistic Illinois 200-bus system with consideration of N-1 contingencies. The training and testing results show the significant speedup capability and excellent numerical stability of the proposed framework.