Research On Time Series Prediction Based On Kernel Learning Methods

In recent years,as an important method of nonlinear signal processing,kernel learning methods based on Reproducing Kernel Hilbert Space have attracted extensive attention.Kernel learning methods use nonlinear mapping to embed the input samples into the high-dimensional feature space,that is,the nonlinear data is processed linearly in the high-dimensional feature space,mapping the nonlinear problem to the linear problem in the high-dimensional feature space,which are widely used in pattern recognition and other nonlinear regression problems.There is great important significance for how to improve the prediction accuracy and meet the requirements of real-time of time series prediction.This thesis mainly studies the traffic flow and network flow time series prediction based on the kernel adaptive filtering algorithm.Online adaptive characteristic of kernel adaptive filtering algorithms satisfies the actual demand for traffic flow and network traffic prediction and improve the prediction accuracy.The research contents of this thesis are:(1)Study Support Vector Machine(SVM),Kernel Extreme Learning Machine(KELM)and other modeling methods,analysis the basic theory and algorithm basis.(2)To improve the prediction accuracy of Henon and Mackey-Glass chaotic time series prediction,study Kernel Principal Component Analysis-Kernel Extreme Learning Machine(KPCA-KELM)and Kernel Partial Least Square(KPLS)methods based on feature extraction.To measure its effectiveness,compared with SVM,LSSVM,KPCA-SVM,KPCA-LSSVM,KELM methods under the same conditions.The experimental results show that KPCA-KELM and KPLS methods can achieve good prediction results.(3)In order to meet the requirements of prediction accuracy and real-time performance in practical applications,a kind of online kernel adaptive filtering prediction models are proposed,which include Kernel Recursive Least Square(KRLS)and Kernel Least Mean Square(KLMS).(4)KRLS and KLMS methods are applied to the measurement of short-term traffic flow and network flow time series prediction based on BC-pAug89.TL,BC-pOct89.TL and BC-Oct89 Ext.TL data sets recorded by Bell Computing Labs,and compared with KPLS and KPCA-KELM methods under the same conditions.The experimental results show that the proposed KRLS method and KLMS methods can effectively improve the prediction accuracy and accelerate the convergence speed.Further comparison with the existing related literature,under the same conditions,the prediction results of the KRLS method is slightly better than the results which using improved T-S fuzzy neural network and the variational Gaussian process in short-term traffic flow prediction.