A model of neural network based on system architecture is presented, and the technique of using this type of neural networks in modeling and parameter optimization is introduced in detail. This network model is structurized in essence. The architecture of network is corresponding to that of the system. Hence, the number of neurons in network is definitive, and the neurons are connected according to the relationship among its subsystems. Parts of the weights in network, which are corresponding to the structure parameters of system, have obvious physical meanings. So further optimization of system structure parameters can be made after modeling.