With advancement of intelligent urban rail transit, automatic train operation needs to achieve an effective balance between energy efficiency and punctuality. However, existing dispatching regulations are mostly expressed in natural language, making them difficult to directly incorporate into control modeling. To address this problem, this paper proposes an automated train operation control method integrating large language models (LLMs) and proximal policy optimization (PPO). The proposed method employs an LLMs to perform semantic parsing of dispatching regulations and operational experience, and automatically constructs a Markov decision process (MDP) model under the constraints of a knowledge graph (KG). The generated MDP includes the state space, action set, reward function, and safety constraints. On this basis, an action masking mechanism and an adaptive reward-weight adjustment strategy are introduced, and PPO is used to train a safe and controllable automated train operation policy in a simulation environment. Simulation results show that, compared with conventional rule-based control, Model Predictive Control (MPC), and Deep Q-Network (DQN), the proposed method reduces the average traction energy consumption by approximately 2.5%, 1.5%, and 1.0%, respectively, and reduces the average delay time by approximately 13.7%, 19.6%, and 16.3%, respectively. In addition, the average jerk is reduced to 0.59 m/s³. These results verify the effectiveness of combining LLM-based semantic modeling with reinforcement learning for automated train operation control. Further hardware-in-the-loop tests indicate that KG-based consistency checking can reduce rule omissions and physical inconsistencies in semantic parsing. The action masking mechanism embeds hard constraints, including speed limits and safe headway distances, into policy execution, while the adaptive reward-weight mechanism adjusts the trade-off among energy consumption, punctuality, and ride comfort according to operational feedback. The proposed method therefore establishes a closed-loop framework of semantic modeling, policy optimization, and feedback adjustment, providing an interpretable approach for incorporating natural-language dispatching knowledge into automated train operation control.