Short-term Traffic Prediction Based On Tensor Model

With the development of the transportation industry,more and more people begin to pay attention to the problems caused by traffic.Traffic problems not only affect the efficiency and safety of travel,but also restrict the development of society.Based on the continuous growth of traffic data,people are eager to find effective ways to solve some traffic problems by using the hidden information in traffic data.Intelligent Transportation System(ITS)Development provides us with an effective direction.Short-term traffic flow forecasting,as a research hotspot in the field of intelligent transportation,is one of the key technologies to realize intelligent traffic control and traffic guidance.According to the multi-dimensional characteristics of traffic data,a dynamic tensor model is constructed to study the short-term traffic flow predction method under the tensor model framework.Meanwhile,in view of the disadvantageous effect of data missing in traffic data on short-term traffic flow forecasting,combining with the low rank characteristic of tensor decomposition,the recovery strategy of traffic data is studied,further analysis of short-term traffic flow forecasting.The main research work and contributions of this thesis are summrised as follows:1.According to the multi-dimensional and non-linear characteristics of traffic flow data,the tensor model is introduced to analyze traffic flow data and proposes a short-term traffic flow prediction method.Firstly,clustering algorithm is used to cluster the traffic flow with similar properties in space,extraction intersections with powerful internal correlation for traffic flow prediction.Secondly,combined with multi-level temporal dimension(X as its variable parameter),the short-term traffic flow prediction model based on the tensor decomposition of “Location-X-Time” is constructed,mining traffic flow trends from multiple perspectives.Finally,in view of the dynamic characteristics of traffic flow,the sliding window is introduced alleviate the problem of sparsity effectively and improve the accuracy of the short-term traffic flow prediction.The experiment shows that the method combines the dynamic structure and the multi-dimensional characteristics of traffic data,then improve the problem of low prediction accuracy in the case of sparse data,improve the prediction accuracy.2.The lack of traffic data has negative impact on the analysis of traffic data and deep mining.Aiming at the problem of data missing,based on the construction of traffic data tensor model,the alternating direction multiplier method is introduced to transform the tensor loss recovery problem into a low rank recovery model and estimate the lost data.The model is not only effectively maintains the multi-mode spatial-temporal correlation characteristics of traffic data,but also conducive to tensor expansion according to each mode and fully mining the multi-mode information of traffic data.This work achieves partial recovery of missing data and promotes the research of short-term traffic flow forecasting technology.Finally,the real data set collected at the traffic junction is used for experimental verification.The experimental results show that the algorithm which combines traffic flow similarity measure and tensor model can improve the accuracy and relieve the problem of low prediction accuracy in the case of sparse data.At the same time,under certain conditions,using tensor reconstruction method to restore traffic data achieves good results.The recovered traffic data also show better performance in short-term traffic flow forecasting.