Safety Deep Reinforcement Learning Approach to Voltage Control in Flexible Network Topologies

Recently, Deep Reinforcement Learning (DRL) methods have become increasingly common in voltage control problems of power distribution systems. However, existing DRL methods either lack a theoretical guarantee of voltage stability or cannot maintain their optimality when the network topology changes. This paper proposes a DRL algorithm based on Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) framework, Flexible-MATD3, which can maintain the system stability and optimality even when the topology and impedance parameters change. We proposed a partial monotonic neural network to constrain the policy search space of our DRL algorithm so that our policy can always guarantee safety. In addition, the experimental results show that Flexible-MATD3 achieves better performance than the baseline controllers for different network topologies and line impedance parameters without retraining the neural network. © 2024 IEEE.