A new kind of recurrent neural network structure-state integrated recurrent neural network (SIRNN)

which can be used in modeling nonlinear dynamic chemical process systems

is put forward here and the BP algorithm

used in training multi-layer feed-forward neural networks

is used as its learning algorithm. The special recurrent structure of SIRNN

which successfully combines state feed-back

time-series and integrated node together

is different from other recurrent neural networks that it makes the network memorize more past system states and that it keeps the network structure from being too complex. Thereby

SIRNN can be used in modeling high order nonlinear dynamic systems. Input-output models of a second order dynamic nonlinear SISO system using different recurrent neural networks are respectively built and their performance is compared. The result shows that the model using SIRNN performs better than other three recurrent neural networks and also

the SIRNN has higher ability of fault tolerance. This indicates its promising future in the application to true systems. The SIRNN is also used in modeling a MISO CSTR system and its performance is tested. The result shows that after training

the model output can well accord with the output trend of the CSTR system.