The relationship between input and output of a dynamic system is analysed and a new - type structure of neural networks is proposed in this paper. There is a specific layer added in the network which is the first hidden layer. Each node in the first hidden layer only receives the input having the same property and produces an integrated signal. The network

as compared with the conventional network

is greatly reduces the number of weights. So it is of great benefit to speed up the training network and to improve the stability and convergence of the network. Then the inverse dynamic neural network controller has been discussed. It is suggested to add a control deviation in the input layer to improve the model accuracy of the controller. Finally

the effects of the structures above mentioned are demonstrated by using CSTR dynamic simulation and control.