To address the performance degradation of existing vehicle-following control methods under abrupt behavioral changes of leader vehicles, this paper proposes a trajectory prediction-based deep reinforcement learning framework. A Transformer architecture is introduced to construct a trajectory prediction module, which forecasts future motion states and speed profiles of leader vehicles using historical trajectories. The framework integrates a vehicle-following policy designed with the Soft Actor-Critic (SAC) algorithm. By directly mapping throttle and brake commands, this method eliminates cascaded control delays, while entropy regularization enhances policy stability by balancing exploration and exploitation. A closed-loop experimental environment built on the CARLA simulator validates the method’s performance in scenarios involving emergency braking, acceleration, and constant-speed driving. Experimental results demonstrate that, the proposed approach achieves significant improvements in efficiency and comfort while ensuring safety, exhibits rapid response to sudden speed changes, and reduces steady-state distance errors.