The trend of a “higher proportion of renewable energy and power electronics” in the new power system has changed the classical stability characteristics of the system. The stability mechanism is more complex, and the system stability modes are more diverse. Online stability control strategies based on typical operating modes face a challenge. Considering the rotor angle stability problem of the new power system, an intelligent generation stability control strategy based on safe reinforcement learning is proposed. First, a constrained Markov model for power system stability control problems is established, and the safety constraints involved in rotor angle stability control are summarized and proposed. Secondly, to improve the ability to extract spatial and temporal features of the power grid’s transient response, a feature perception network based on graph convolutional layers and long short-term memory units is constructed. Then, to improve the training efficiency of the stability control agent, a training framework of stability control strategies using proximal policy optimization algorithm based on embedded domain knowledge constraints is proposed. Finally, a case study is performed on the IEEE 39-bus system and a practical power grid. The results show that the proposed method can adaptively generate unit tripping strategies based on the system operating state and fault response, and its decision-making effectiveness and efficiency are superior to existing stability control strategies.