GARL: A Generative Adversarial Reinforcement Learning for Transferable Power Grid Dispatch

This paper proposes a novel deep reinforcement learning approach, named GARL (Generative Adversarial Reinforcement Learning), for intra-day optimization dispatch in high-proportion new energy grid. The proposed method embeds a Generative Adversarial Network (GAN) within the Deep Deterministic Policy Gradient (DDPG) framework to learn the reward distribution of the power grid in the future operation situation, so as to realize the cumulative reward prediction in the dispatching period and ensure the optimization of the dispatching decision. Comprehensive experiments are conducted on the SG-126 power grid simulator under three representative scenarios to evaluate the performance and generalization capability of GARL. The results show that GARL can learn the distribution characteristics of stateaction-reward sequence from historical experience, establish the mapping relationship between current state and future cumulative reward, and improve the accuracy of future power grid operation trend prediction, which is always better than the independent DDPG algorithm. Furthermore, the study explores the potential of transfer learning between scenarios and finds that GARL exhibits a promising generalization capability.