The reaction mechanism of chemical process is complex. There is a modeling error between the mechanism model and the actual reaction system. At the same time

there are complex slow time-varying features

such as catalyst deactivation

fuel coking

etc. Thus there will be a mismatch between the process model and the actual process system. A self-adaptive iterative hybrid model (SAIHM) is established to reflect the dynamic characteristics of the process accurately over a long period. The mechanism model and the data-driven model are effectively combined to improve the prediction accuracy of the model; the data-driven model uses the deep recurrent neural network (DRNN) to fully exploit the timing relationship between adjacent conditions; the data-driven model is automatically updated based on the evaluation indicators to resolve the contradiction between accuracy and efficiency. The simulation and comparison results of the self-adaptive iterative hybrid model and the existing mechanism model established based on the historical operation data of an acetylene hydrogenation adiabatic reactor show that the self-adaptive iterative hybrid model can more effectively track the actual system.