An improved generalized predictive control algorithm is presented in this paper by incorporating offline identification into online identification. Unlike the existing generalized predictive control algorithms, the proposed approach divides parameters of a predictive model into the time invariant and time-varying ones, which are treated respectively by offline and online identification algorithms. Therefore, both the reliability and accuracy of the predictive model are improved. Two simulation examples of control of a fixed bed reactor show that this new algorithm is not only reliable and stable in the case of uncertainties and abnormal disturbances, but also adaptable to slow time varying processes.