Bridging Transient and Steady-State Performance in Voltage Control: A Reinforcement Learning Approach With Safe Gradient Flow

Deep reinforcement learning approaches are becoming appealing for the design of nonlinear controllers voltage control problems, but lack stability guarantees hinders their deployment in real-world scenarios. This paper constructs a decentralized RL-based controller inverter-based real-time distribution systems. It features two components: transient policy and steady-state performance optimizer. The is parameterized as neural network, optimizer represents gradient long-term operating cost function. parts synthesized through safe flow framework, which prevents violation reactive power capacity constraints. We prove that if output bounded monotonically decreasing with respect to its input, then closed-loop system asymptotically stable converges optimal solution. demonstrate effectiveness our method by conducting experiments IEEE 13-bus 123-bus test feeders.