Deep Reinforcement Learning Framework for Short-Term Voltage Stability Improvement

Abstract

This paper investigates the mitigation of fault-induced delayed voltage recovery (FIDVR) using dynamic voltage support from hybrid PV plants and optimal load control using deep reinforcement learning (DRL). We characterize and quantify the delayed voltage recovery phenomenon through probability density-based metrics. We propose a DRL-based load control by optimally tripping stalled induction motor loads to recover the voltage quickly. The amount of load tripping depends on system operating conditions, so the data-driven framework gives optimal load control adaptable to the system conditions. The numerical simulations show that the dynamic reactive power injection and DRL-based load control improve the voltage recovery and decrease the amount of load tripped significantly.