An emergency control strategy for undervoltage load shedding of power system: A graph deep reinforcement learning method

Undervoltage load shedding (UVLS) is the last line of defense to ensure the safe and stable operation of the power system. The existing UVLS technique has difficulty adapting and generalizing to new topology variation scenarios of the power network, which greatly affects the reliability of the control strategy. This paper proposes a UVLS emergency control scheme based on a graph deep reinforcement learning method named GraphSAGE-D3QN (graph sample and aggregate-double dueling deep q network). During offline training, a GraphSAGE-based feature extraction mechanism of the power grid with topology variation is designed that can better capture the changes in system state characteristics. A novel reinforcement learning framework based on D3QN is developed for UVLS modeling, which reduces overestimations of control action and leads to a better control effect. Then, online emergency decision-making is achieved. The simulation results on the modified IEEE 39-bus system and IEEE 300-bus power system show that the proposed UVLS scheme can always provide more economical and reliable control strategies for power networks with topology variations and achieves better benefits in both adaptability and generalization performances for previously unseen topology variation scenarios.