Abstract: The model mismatch problem appears in the application of multivariable predictive control algorithms that may lead a controller not to meet control requirements. We present a model predictive control(MPC) framework that uses the frequency response error between the controlled plant and the predictive model as the criterion to determine whether model mismatch exists. If model mismatch occurs, a persistent excitation signal is added to the controlled plant first, and then the transfer function model of the plant is identified by an improved model adaptive identification algorithm. Finally, the transfer function model is transformed into a finite step response (FSR) model via inverse Laplace transform, and multivariable predictive control is reactivated. Using this new method, an offline identification test becomes unnecessary, and multivariable identification can be achieved. The dynamic characteristics of the identified transfer function model are even clearer and more convenient for analysis and modification. After inverse Laplace transform, the FSR model runs more smoothly and can eliminate the offset caused by model errors. Simulation results show the effectiveness of the proposed method.