Proximal policy optimization through a deep reinforcement learning framework for remedial action schemes of VSC-HVDC

A proximal policy optimization (PPO)-based back-to-back VSC-HVDC emergency control strategy based on multi-agent deep reinforcement learning (DRL) approach is proposed for use in an energy management system (EMS). In this scheme, an advanced DRL algorithm is proposed by implementing both PPO and a communication neural network for large power systems. The PPO modeled as intelligent agents with objective functions have shown a higher convergence performance than have existing DRL algorithms. Further, the model was demonstrated to effectively address voltage variances caused by the high penetration of renewable energy sources. By implementing PPO, the learning procedure is stabilized and made robust to continuous changes in network topology. To escalate the effectiveness of the proposed algorithm, a comprehensive case studies were conducted on an standard test systems and Korean power system considering variations in load and PV generation and a weak centralized communication environment. The results indicate that outstanding control performance and autonomously regulated bus voltage and line flows, thereby validating the effectiveness of the method.