By analyzing the traction -braking system of EMU, this paper establishes the linear model of each power unit of EMU. Considering the existing nonlinear problems such as air resistance and unknown interference during EMU operation, the error between actual output and linear model output is described as unmodeled dynamics. As a nonlinear problem, EMU operation process is described as an integrated model composed of linear model and nonlinear unmodeled dynamics. BP neural network is used to estimate the unmodeled dynamic items online, and the recursive identification algorithm is used to update the model parameters online. In the controller design, the generalized predictive controller and unmodeled dynamic compensator are combined to design the multivariable nonlinear generalized predictive controller. Based on the simulation of CRH380A EMU, the highprecision tracking control of the given speed is realized, and the on -time, comfortable and safe operation requirements of the EMU are realized.