Modeling For Nonlinear System Based On Hybird Neural Network

A hybrid neural network had been put forward in this paper. Which consists of B-spline neural network and linear diagonal recursive neural network(DRNN). For making up the deficiency of one single neural network, such as, getting into local minimum, long training time, lower precision for testing, and more training parameters and so on; and bring the advantages of neural network into play, such as high nonlinearity, self-adaptive, self-learning and so on.CSTR model and process of PET had been applied as the simulated models for hybrid neural network, and RLS and RPE algorithm had been used for both serial and parallel hybrid networks to the simulations. The result of simulate experiments indicate that, hybrid neural network with RLS algorithm can model quickly and well for dynamic nonlinear systems of measured outputs; and hybrid neural network with RPE algorithm can model for unmeasured output systems well. Experiments proved that, hybrid neural network can reduce the difficulty of training for one single network, make up the deficiency of modeling by one network, and be propitious to resolving for control strategy.