Fast valving is one of the classic control methods to improve transient stability in power systems. However, its control variables are high-dimensional and discrete, and improper parameter tuning may trigger subsequent power angle oscillations and instability. The complexity of the control strategy development makes it difficult to apply fast valving closure in online and real-time decision-making. To address this challenge, a fast valving control decision method based on deep reinforcement learning is proposed. First, a deep reinforcement learning-based emergency fast valving decision-making framework is constructed. Then, the fast valving control problem is transformed into a Markov decision process (MDP). A reward function is designed to balance optimal stability control performance and minimized control cost, and the proximal policy optimization (PPO) algorithm is used to solve it, yielding a rational configuration of the fast valving strategy. Finally, the effectiveness of the proposed method is verified using the improved SG-77 system developed by CEPRI. Simulation results show that the proposed method ensures both the effectiveness and timeliness of the fast valving strategy, enabling correct decision-making under mismatched contingency scenarios, and improves transient stability and dynamic response capability of power systems.