The subway station air conditioning system consumes a lot of energy, and traditional control methods cannot take into account the comfort and energy saving issues together, resulting in poor control effect. Moreover, the current subway station air conditioning control systems control the air system and water system separately, which cannot guarantee the energy saving effect of the whole system. Therefore, this paper proposes an energy-saving control strategy for the system based on reinforcement learning. Firstly, this paper uses a neural network to establish an air conditioning system model as a simulation environment for offline training of the agent to solve the problem of long convergence time of model-free reinforcement learning methods for online training. Then, in order to improve the efficiency of the algorithm and also to address the characteristics of the multidimensional continuous action space of the air conditioning systems, this paper proposes a deep deterministic policy gradient algorithm based on multi-step prediction and designs an agent framework that will be used to interact with the environment model for training. In addition, in order to determine the optimal number of training times, the agent training termination condition is also set, which further improves the algorithm efficiency. Finally, simulation experiments are conducted based on the measured operational data of a subway station in Wuhan, and the results show that the proposed control strategy has better temperature tracking performance and can ensure the comfort of the platform, and the energy saving is about 17.908% compared with the current actual system.