Knowledge-enhanced deep reinforcement learning for intelligent event-based load shedding

Abstract

Event-based load shedding (ELS) is an important emergency countermeasure against transient voltage instability in power systems. At present, the formulation of ELS measures is usually determined offline by the experience of experts, which is inefficient, time-consuming, and labor-intensive. This paper proposes a knowledge-enhanced deep reinforcement learning (DRL) method for intelligent ELS. Firstly, the Markov decision process (MDP) of the knowledge-enhanced DRL model for ELS is established based on transient stability simulation. Different from traditional response-based MDP, this MDP is event-based. Then, compared to conventional exponential decision space, a linear decision space of the DRL agent is established to reduce the decision space and training difficulty. Furthermore, the knowledge of removing repeated and negative actions is fused into DRL to improve training efficiency and decision quality. Finally, the simulation results of the CEPRI 36-bus system show that the proposed method can accurately give effective ELS measures. Compared with the pure data-driven DRL method, the knowledge-enhanced DRL method is more efficient.