Based on the relevance vector machine(RVM), a nonlinear model predictive control algorithm is developed to deal with the difficulties in the modeling and real-time optimization problem for the controlled plant with nonlinearity and the optimization objective function with non-convexity. The nonlinear predictive model is established by RVM, the nonlinear constrained programming is on-line performed by employing differential evolution(DE) algorithm which possesses the advantages such as global optimum, robust and fast convergence. The proposed algorithm is applied in the polypropylene process, and the simulation results show that the performance of the grade transition control can be greatly improved.