There are many loops in the subway station air-conditioning system, and it has strong coupling and nonlinear characteristics, resulting in difficulties to adjust the parameters of the PID controller. And it is impossible for PID to consider both passenger comfort and energy saving. In addition, due to the difficulty of system modeling and the huge computational effort of nonlinear optimization algorithms, it is difficult to realize the engineering application of intelligent control methods. Therefore, a model predictive control strategy is established for the subway platform air conditioning system to improve passengers comfort and achieve energy efficiency. Firstly, based on the heat and humidity load balance and the law of energy conservation, a prediction model for the thermal dynamic characteristics of the subway platform is developed. And then, an artificial neural network(ANN)-based predictive controller is designed, in which the weights and thresholds of the ANN controller are optimized online using the Lagrange-variational-based gradient descent training algorithm, to minimize the system cost function with small computational and storage needs. Simulation results show that the proposed predictive control strategy can shorten the system response time by 39.6% and reduce the energy consumption of the end fan by 73.39% compared with the PID control on the premise of meeting passenger comfort requirements.