Research Of Multivariable Time Series Prediction Algorithm Based On BBN-SVM And PCV

The time series prediction plays more important role in industrial production and agricultural production and our lives. Its application spreads many fields. Support vector machine is a statistic learning method. Because of its better performance and better generalization ability, it is widely applied to many fields.At present, a majority of prediction methods using support vector machines are single-variable time series prediction algorithms. The paper raises a multivariable prediction method for improving prediction accuracy. In the issue of selecting multiple variables, we introduce Bayesian network in order to analyze relationship among the variables and then select variables. At first, we determine the predicted variable. After the Bayesian network structure learning, we can get a model. Then, the correlative variables set are selected according to this model, and the training set is prepared. This method can effectively consider the interaction among the variables, so that it can improve the prediction accuracy.The generalization ability of support vector machine depends on parameters selection in great degree. How to select optimal parameters for improving generalization performance of SVM is a hotspot at present. In this paper, we design a parallel K-fold cross validation algorithm(PCV algorithm) based on the MPI programming model. The algorithm overcome the shortcoming that the traditional K-fold cross validation is inefficient in large database. The algorithm help us select optimal parameters of SVM, meanwhile, it declines runtime greatly.We discuss and compare some kinds of basic kernel, especially the Polynomial kernel and Gaussian radial basis kernel. Moreover, the method of combination kernel is raised and the comparing experiments have been made.At last, the raised methods are used in weather time series and stock market time series prediction. The experiment results show that multivariable time series prediction can make the average error within ten percent, and have better generalization ability. The parallel K-fold cross validation method helps us select parameters and make the run time be 1/p times of traditional method.