Deep Reinforcement Learning-Based Fast Prediction of Strategies for Security Control

The hybrid connection of regional power grid via UHV AC / DC and increasing penetration of renewable energy brings a rapid growth of uncertainties. The operation modes of power grid are more changeable and the transient stability problem becomes more serious. In consequence, ensuring the stability of power systems is gaining more attention today than ever before. So, in this paper, aiming at the angle stability problem of large-scale system after UHVDC bipolar block, from the perspective of data-driven, a deep reinforcement learning-based method is adopted to quickly predict strategies of the security control for different power grid operation modes corresponding to the above-mentioned uncertainties. In detail, a Deep Q Network (DQN)-based model is established to formulate the security control prediction problem with the system information before disturbance as input and security control strategies as output. The reasonable design and tuning for the reward function of the agent is the most important for the settings of DQN and by which the effectiveness of security control strategies under different operation modes is guaranteed. Detailed analysis and validation on a simplified regional power grid in China are presented to prove feasibility and performance of the security control strategy prediction method.