1School of Electric Power Engineering, South China University of Technology, Guangzhou 510641, China;2Energy Development Research Institute, China Southern Power Grid, Guangzhou 510663, China

Renewable energy can rapidly regulate the power in the new power system, demonstrating the potential of participating in overload emergency control for lines. However, when it is adopted, generation methods for the emergency control strategy based on deep reinforcement learning face the challenges of excessively large decision space and high solution complexity. To address this issue, a hybrid learning method for emergency control with dynamic iteration of decision space and strategy model is proposed. First, a dual-network model comprising a control location network and a control value network is constructed, and an iterative learning framework for both networks is designed. Second, the control location network and its learning objectives are introduced, and a sensitivity-based sample generation method is designed to learn the control location network. Then, a deep reinforcement learning method for the control value network is proposed, and a segmented exploration strategy is designed for efficient learning of the control value network. Next, a dynamic iteration implementation process between the control value network and control location network is designed. Finally, the effectiveness of the proposed method is validated in the IEEE 39-bus system, IEEE 300-bus system, and a provincial power grid of China.