Traffic Flow Forecasting Application Research Of Wavelet SVM

Intelligent transportation system is a system of time-varying, random and complexity. Its significant feature is provided with tremendous uncertainty which brings big difficult on traffic flow forecasting. Recent years, Support Vector Machine has been created by Vapnik based on statistical learning theory. As the key point, kernel function is very important at learning accuracy and generalization ability of regression model. For the choice of kernel function, many deep researches have been proposed, but there is not reasonable and common guideline and method.The aim of this paper is to construct a new kernel function that is suitable for traffic flow forecasting and can improve the forecasting performance. Firstly, by analyzing theory and experiment result, the shortage at traffic flow forecasting of RBF kernel function which is common used as SVM kernel function includes information lost, long training and prediction time, poor forecasting accuracy for bound signals and multi-scale signals. Next as for these bottlenecks, wavelet theory is pulled in to optimize SVM performance. Wavelet function can approximate square integrable space and is provided with the advantages of localization, multi-level and multi-resolution so as to suit for multi-scale analysis, which can make up RBF kernel function's shortage at application of traffic flow forecasting.Secondly, according to Mercer dot product wavelet kernel lemma and translation invariance lemma, wavelet kernel function is constructed and Bubble wavelet is chosen to be mother wavelet. Because wavelet is suitable for multi-resolution, a new multi-scale kernel function based on Bubble wavelet is constructed and precision-improvement distinguish method is designed to optimize scale selection.At last, the generalization performance is evaluated by real traffic flow data of highway in Los Angeles, USA and a variety of experiments are carried out. Compared to RBF kernel function, Bubble wavelet kernel function and multi-scale wavelet kernel function have much more forecasting accuracy and higher efficiency, can better satisfy real-time demand of traffic flow forecasting.