Generator tripping scheme(GTS)is the most commonly used scheme to prevent power systems from losing safety and stability.Usually,GTS is composed of offline predetermination and real-time scenario match.However,it is e...Generator tripping scheme(GTS)is the most commonly used scheme to prevent power systems from losing safety and stability.Usually,GTS is composed of offline predetermination and real-time scenario match.However,it is extremely time-consuming and labor-intensive for manual predetermination for a large-scale modern power system.To improve efficiency of predetermination,this paper proposes a framework of knowledge fusion-based deep reinforcement learning(KF-DRL)for intelligent predetermination of GTS.First,the Markov Decision Process(MDP)for GTS problem is formulated based on transient instability events.Then,linear action space is developed to reduce dimensionality of action space for multiple controllable generators.Especially,KF-DRL leverages domain knowledge about GTS to mask invalid actions during the decision-making process.This can enhance the efficiency and learning process.Moreover,the graph convolutional network(GCN)is introduced to the policy network for enhanced learning ability.Numerical simulation results obtained on New England power system demonstrate superiority of the proposed KF-DRL framework for GTS over the purely data-driven DRL method.展开更多
基金supported by National Natural Science Foundation of China(No.U22B20111,No.U1866602)。
文摘Generator tripping scheme(GTS)is the most commonly used scheme to prevent power systems from losing safety and stability.Usually,GTS is composed of offline predetermination and real-time scenario match.However,it is extremely time-consuming and labor-intensive for manual predetermination for a large-scale modern power system.To improve efficiency of predetermination,this paper proposes a framework of knowledge fusion-based deep reinforcement learning(KF-DRL)for intelligent predetermination of GTS.First,the Markov Decision Process(MDP)for GTS problem is formulated based on transient instability events.Then,linear action space is developed to reduce dimensionality of action space for multiple controllable generators.Especially,KF-DRL leverages domain knowledge about GTS to mask invalid actions during the decision-making process.This can enhance the efficiency and learning process.Moreover,the graph convolutional network(GCN)is introduced to the policy network for enhanced learning ability.Numerical simulation results obtained on New England power system demonstrate superiority of the proposed KF-DRL framework for GTS over the purely data-driven DRL method.
