Modeling of Chemical Dynamic Systems Based on Modified Recursive Neural Network Structures

A modified recursive neural network structure—dynamic-hide-output recurrent neural network (DHORNN), based on the partial recursive neural network Elman, Jordan, SIRNN,is put forward. Structures of state feedback from hide layer, output feedback, and time-delayed nodes are well combined in the structure. The Levenberg-Marquardt algorithm is successfully used to train the network and to improve its nonlinear-dynamic-modeling capability. The modified recursive neural network structure is used to build models for a continuously stirred tank reactor (CSTR), and then the models are compared with other recursive neural network models. The results showed that the models based on the DHORNN are more effective in chemical dynamic system modeling.