Abstract: The artificial neural networks are now being widely studied and used in many fields because of their high capability of mapping the input-output relationship of nonlinear systems.However,for nonlinear dynamic systems,such as the fractionating equipment of refinery,the standard back-propagation(BP) neural network——which is the most often studied and used,can not perform very well when used to model these systems,especially when they are not in steady states.So,in this paper,a new structure of feed forward neural network named dynamic back propagation neural network(DBPNN),based on back propagation algorithm and time series model is presented.The DBPNN can be used to model dynamic nonlinear systems.Furthermore,with the special structure of the integrated node layer,the complexity of the DBPNN is largely reduced.By using data from laboratory model and data from the fractionating equipment of some refinery,DBPNN models are built.The result shows that the DBPNN models can identify the dynamic behavior of nonlinear dynamic systems and are more accurate and reliable,compared with the result of models of these two nonlinear dynamic systems using standard back propagation neural networks.