Short-term Traffic Prediction Based On Dynamic Tensor Completion

Intelligent Transportation System (ITS) is a rapidly developing system for trafficcontrol and management in recent years. As vital component parts of ITS, both the trafficcontrol system, traffic management system and route guidance system need accurate andreal-time traffic information. And the short-term traffic prediction based on real-timeinformation is the premise of real-time traffic control and route guidance, which makes itbecome a criticle theoretical groundwork for ITS. Howerver, real-time traffic flow data isdepending on many factors and exhibites multi-mode information. It is still a chanllengeproblem to make full use of such multi-mode information on a unify framework forshort-term traffic prediction model. In this paper, the multi-mode information of traffic flowis analyzed under the umbrella of multilinear analysis, the traffic flow data is modeled asboth static and dynamic tensor pattern. Then, based on the dynamic tensor pattern, ashort-term traffic prediction method based on dynamic tensor completion is proposed. Themain content of this paper is as following:Firstly, from the view of multi-mode information mining, the traffic flow isconstructed into both static and dynamic tensor pattern and multilinear analysis is used torecognize the multi-mode low-rank property within the traffic flow data. Then, a newmethod based on mutl-mode low-rank property for solving traffic prediction problem isproposed.Then, based on the mutlilinear analysis of traffic flow data, a tensor completionmethod based on multi-mode matrix factorization is proposed for short-term traffic flowprediction. With the matrix factorization method, we can not only solve traffic flowprediction problem, but also present a new framework of tensor completion. And numerousexperiments are conducted on the proposed tensor completion framework to assess itsapplicability.Finally, the traffic flow prediction method combined with dynamic traffic flow tensorpattern and multi-mode matrix factorization based tensor completion is proposed. Theproposed method is compared with most state-of-art methods. Experiments show that tosome extent the proposed method is more accurate than state-of-arts. Moreover, theproposed method can accurately impute missing data and forecast future traffic flow on aunify framework.