Research On The Traffic Flow Data Recovery And Forecasting Methods Of Expressway Based On Microwave Data

With the continuous improvement of Chinese urbanization level,the structure of the urban road network has been gradually improved.The expressway plays a key role in the urban road network system.The operation efficiency of the expressway not only affects the overall operation quality of the road network,but also affects the normal operation of the city.The intelligent transportation system(ITS)based on traffic control and dynamic route guidance technique is an effective way to raise the service level of expressways.Specially,mining massive traffic data deeply and forecasting the traffic state accurately are fundamental for improving ability of traffic control and guidance.In this paper,focusing on the missing problem and non-linear feature of traffic flow data captured by remote traffic microwave sensor(RTMS),high-efficiency models for traffic flow data recovery and forecasting are established based on tensor theory and deep learning to provide fully support for traffic management and control of expressways.Firstly,by summarizing the research status of traffic data recovery and traffic states prediction at home and abroad,it is concluded that the existing methods take the entire road sections as the research object,which is difficult to satisfy the increasing demand of the refined prediction of the traffic state.Therefore,lane sections are taken as research target and the technical route of this paper is presented.Secondly,the technique of RTMS and the processing methods of collected data are introduced.In addition,the difference and the spatio-temporal relevance of lane-level traffic flow are analyzed with help of kernel density estimation and correlation coefficient respectively,which provides a theoretical support for the researches on traffic data recovery and prediction.Thirdly,to deal with the data missing problem after collection and preprocessing,the scenarios of the data missing are divided into three scenarios including Missing Completely at Random(MCR),Missing at Random(MR),and their combination(MCR/MR).A data imputation approach(TDIM)for lane-level traffic flow based on tucker factorization of tensor is proposed.The ground-truth travel speed data captured by RTMSs installed on the 2nd Ring Road of Beijing are utilized to analyze the proposed approach under the three scenarios with different missing rate.Fourthly,this paper establishes a hybrid deep learning(HDL)model for lane-level traffic flow prediction based on convolutional long short-term memory(Conv-LSTM)neural network,which can adapt to the time-varying and nonlinear characteristics of traffic flow by capturing its spatio-temporal characteristics.The HDL model can forecast the travel speed of multiple lanes section synchronously with input tensors combining speed,volume,and occupancy.The experimental results show that the HDL model has better prediction performance than some traditional deep learning models.Finally,to further improve the stability of traffic flow forecasting and overcome the limitation of single prediction model,this paper proposes an improved bayesian fusion model(IBFM)which fuses the HDL model and other classical traffic prediction methods by means of ensemble learning theory.Validated by the ground-truth data captured by RTMSs,the IBFM model is superior to single models and traditional fusion models in terms of accuracy and stability.There are 51 figures,20 tables,and 90 references in this paper.