An Approach Of Temperature Modeling With Echo State Network For Tubular Polymerization Reaction

How to establish a dynamic neural network model of the temperature distribution in tubular polymerization reaction with echo state network (ESN) is the main target in this paper. The temperature distribution is the most equal parameter of the molecular weight distribution. It makes a practical and great significance to establish the temperature distribution model. This paper includes the following aspects:Firstly, it describes the basic principles of echo state network and introduces the reasons that ESN is chosen as modeling network. The black-box model is built after setting the parameters of ESN. The simulation results show that the generalization error of black-box modeling can be fit with temperature distribution well, which provides the foundation for the gray-box model.Secondly, based on the known information and mechanism knowledge two kinds of structure-approaching gray-box modeling method are proposed. The system feature groups (SFG) from prior knowledge are selected as the node of reserve pool in the first gray-box model, which would better use of the reserve pool’s ability that its internal linear or nonlinear elements can be flexibly combined. Considering the deficiency of SFG selection process, we deeply extract the important functions again and choose the optimal SFG with stepwise regression analysis in the second gray-box model. The structure of ESN in the second gray-box model is optimized. Compared with the previous models, it shows that the accuracy of generalization error and the speed of training have been significantly improved.Finally, during the research it found the temperature distribution along the tubular polymerization reaction has different characteristics in each state. Based on this fact, the segmented series connection modeling method is proposes. Its main idea is to transform distributed parameter systems into the series connection model of lumped parameter system. The simulation results demonstrate the effectiveness of the method.This paper’s research supplies a new thinking to establish the temperature model based on the prior information. The method would be a reference to be applied to other chemical reactions.