The Research And Application Of Radial Basis Function Neural Network In Process Modeling

Radial Basis Function Neural Network (RBFN) is one kind of forward neural network, which has three layers with only one hidden layer. For RBFN can approximate any continuous function with any precision, it has been used to many fields broadly, such as function approximation, pattern classify, forecasting and so on, and it has acquired good application impression. The training algorithms of RBFN and its practice application in process modeling are the main researches in this paper. The research of this paper is divided into two phases: algorithm research and algorithm improvement and application.In the phase of algorithm research, according to the demand of the subject, after reading and researching lots of international forward literatures, a kind of algorithm which has a reasonable structure, a fast training speed and a mezzo precision is selected from comparing four algorithms in common use by two simulation examples that are one-dimension function approximation and time serial forecast. Then the MATLAB program is changed into JAVA language which is required. The JAVA language is tested by the former two simulation examples at last. In the phase of algorithm improvement and application, the application of RBFN in electric power load forecasting is researched in this paper. The tradition k-means clustering algorithm in common use anciently has two limitations: the optimal result depends on the initial value badly and the number of the cluster can not be solved. To overcome the two inherent limitations, a method of adding clustering number step by step is proposed, and a self-adapt clustering algorithm is advanced, furthermore, the type of date and surrounding factors are synthetically considered when forecasting. The results of practical applications show that not only the training speed of RBFN can be improved, but also higher forecasting precision can be obtained, the proposed algorithm has better practicability.