Neural Network Structure And Hybrid Modeling

In this paper, single objective genetic algorithm (SOGA) and multiobjective genetic algorithm (MOGA) are developed to optimize the architecture of radial basis function neural network (RBFNN). The training error and the number of hidden layer units are combined as the optimal objective-of SOGA. In MOGA, optimal objective consists of the number of hidden layer units, training and generalization error of RBFNN. By minimizing these objective functions, optimal configuration of RBFNN is obtained. Through experiments, it is shown that the RBFNN obtained by SOGA has less hidden layer units, better training and generalization performance than the one derived by orthogonal least square algorithm (OLS). The RBFNN designed by MOGA has much better generalization than that of SOGA, and also keeps simple architecture.In this paper, the hybrid modeling method is studied and proposes parallel and serial hybrid modeling plans. RBFNN is the main part of the parallel hybrid model, prior knowledge of the modeled system is added to it, which endow some physical mechanism to RBFNN and help to avoid the disadvantage of statistical model. The principle model is the main part of the serial hybrid model, RBFNN models the mathematical relation between two parameters of system, which can not be expressed exactly by principles. Take the typical object in petrochemical industry and final condensation polymerizationprocess in Polyethylene Terephthalate (PET) production as the modeled system, the two kinds of hybrid model are established. The result verifies that the hybrid modeling. incorporating the merits of RBFNN and principle modeling, has great advantage, and can be a feasible way to improve the modeling accuracy.