An intelligent identification method based on self-adaptive mechanism regulated neural network for chemical process

Background: The complex characteristics of chemical process, such as multivariable, nonlinear, time-varying and strong coupling often lead to the poor effect of traditional identification theory in practical application. The development of deep learning in recent years has brought a breakthrough for nonlinear system identification, but more progress is still needed. Methods: This paper proposes a nonlinear identification method based on self-adaptive mechanism regulated Long Short-Term Memory (LSTM) network for chemical process dynamic simulation. First, to increase the reliability of the application of neural network (NN) for chemical process identification and improve the generalization ability, the known differential equation describing the mechanism is taken as a regularization term to constrain the output of the NN. Then, a specific training method is proposed, which introduces trainable self-adaptive weights to force the neural network to focus on the regions with large training error. In addition, a semi-supervised network training method is proposed for the case that some parameters in the mechanism equation are unknown. Finally, a dynamic virtual device (VD) model is established, which can simulate the dynamic response of controlled objects. Significant findings: To evaluate the efficiency of the developed identification method, various comparative ex-periments are conducted on pH neutralization and continuous stirred tank reactor (CSTR) processes. The experimental results show that the proposed identification method can obtain a nonlinear dynamic model with robustness, high accuracy and strong generalization ability.