A dual model RBF(radial basis function)neural network was proposed in this paper.One is used for self-learning

which learns one time a day.The other is used for on-line correcting

which is the running model currently.Both the self-learning model and the on-line correcting model are corrected six times every day and should track the current conditions of the system quickly.At the same time

the accuracy of the two models should be compared.If the accuracy of the on-line correcting model is less than the one of the self-learning model

the latter becomes the new currently running model instead of the old one. Otherwise

the currently model is maintained.To solve the problem of neural network large prediction errors

a network algorithm analysis is given and the influence factors of the network prediction accuracy are found.At last

an improved algorithm of RBF neural network modeling is proposed

which combines K-means clustering method with the recursive descent algorithm.Simulation and practical application proved the effectiveness of the improved method.