Data center cooling system has nonlinear, strong coupling and large hysteresis characteristics. The commonly used PID method cannot achieve overall system energy efficiency improvement, and the existing nonlinear system optimization algorithms are computationally intensive and not easy to implement. This paper proposes a model predictive control strategy for the data center cooling system. The upper optimization layer designs a predictive control strategy to reduce energy consumption on the premise of meeting the cooling load of the IT servers. A neural network is used as the feedback controller, and the optimization cost function of the system is used as the performance index of the neural network controller, and combining the variational method and stochastic gradient descent method to perform online receding horizon optimization to obtain the optimal set values of the controlled variables for each loop in the lower layer. The optimization algorithm occupies moderate storage space and small computation. The lower field control layer makes the controlled variables track the optimal set value through real-time control, which can realize optimization without destroying the original field control system. A Trnsys-Matlab simulation platform is constructed and simulation experiments are conducted for summer, transition season and winter conditions. Experimental results show that the proposed control strategy can achieve energy efficiency improvement with good robustness while meeting the premise of safe operation of the data center.