The time-varying characteristics of passenger flow in spatial and temporal dimensions impose significant challenges on urban rail transit operation management, while the complexity of demand-responsive train timetabling calls for more advanced algorithmic solutions. To address this, this paper introduces a deep reinforcement learning (DRL) approach to optimize subway train schedules, thereby improving the intelligence of operational management. The dynamic interaction between time-varying passenger demand and train timetable is formulated as a Markov decision process, which serves as a training environment for the learning agent. The state space includes train arrivals/departures, passenger states, and train counts, the action space is defined by dispatch intervals. A multi-dimensional reward function jointly considers passengers, trains, and stations. A multi-objective Soft Actor-Critic algorithm with adaptive dispatch intervals and train numbers is developed for efficient optimization. Hyper-parameter tuning via a small-scale case and confirms that demand-responsive timetable outperforms fixed-interval timetable. Simulations on Guangzhou Metro Line 8 show faster convergence and higher solution efficiency compared to other AI and heuristic methods. Moreover, under various passenger flow disturbance scenarios, the method can generate satisfactory operation plans in a short period of time, demonstrating its strong generalization capability. These results provide theoretical and methodological support for intelligent subway scheduling.