Aim at solving the strong nonlinear control problem, a multi-step predictive control method is proposed, which uses a RBF neural network as a model. A multi-step predictive model is constructed, a Jacobian matrix computing method for predictive error about control sequence is given, a receding horizon optimization policy is designed by using L-M algorithm, feedback correction is achieved by modifying reference input according the error, and the stability of the system is proved. Simulation results of the control method validate desirable performances.