Metropolitan Traffic Sensing And Forecasting Based On Big Data Analysis

With the advancement of urbanization in China,the scale and structure of the metropolitan road network become larger and more complex.Along with the increment of registered vehicles,all these phenomena have caused tremendous pressure on the urban transportation systems and made the congestion problem remarkable.Also,the development of traffic management system is relatively slow,difficult to meet the needs of traffic management,and has further intensified the congestion problem.Specifically,it is difficult to effectively estimate the road conditions using traditional sensing schemes and algorithms,not mention to predict the future conditions precisely and reliably.However,the rapid development of Internet of Vehicle and Big Data has brought new ideas for solving these issues.In the scenario of Internet of Vehicle,vehicles which play the roles of mobile sensors can provide data far more than before.Besides,Big Data technologies provide efficient tools for traffic data analysis and mining.Thus,the application of these technologies has already highly concerned in academia.Nevertheless,existing work still has shortcomings.Traditional data recovery algorithms are not sufficient to overcome the sparsity and systematic errors of single-source traffic data.Meanwhile,traditional prediction algorithms can hardly deal with the strong spatiotemporal correlations in the large-scale road networks with complex structure.Therefore,this article takes Shanghai as the research object and provides solutions for urban traffic monitoring and prediction through the processing of traffic big data,the modeling of the road network,and the research of global prediction algorithm.The multi-sources heterogeneous traffic data this article used gathered from Shanghai Transportation Information Center.Firstly,this research eliminates the error and incomplete records with data cleansing and preprocessing technics and extracts valuable information from different datasets through the slide-window based filtering algorithms.Secondly,this article structures and implements a heterogeneous data fusion system aiming to alleviate the sparsity problem mentioned before.This system consists of two coarse-fine grained map-matching algorithms and a data fusion method.The former algorithms are designed to meet the different scale of data.The latter method is built on the theoretical evidence of the linear relationships among different sources.This theoretical evidence is provided by a modified Granger causality test which suppresses the influence of samples' noise using a robust regression method.Real-world data-driven experiments indicate the accuracy of the proposed solution and show that the sampling rate of the traffic monitoring system can be improved notably by 38.85% relatively.Thirdly,this article further studies the traffic short-term prediction problem based on the historical monitoring information given by last part of work.The analysis in this research indicates that the strong spatiotemporal coupling existed in the large and complex road network greatly increases the difficulty of traffic prediction,which is different from the commonly discussed time-series prediction scenario.Furthermore,the propagation pattern of traffic flow is the key to reveal the law of traffic conditions' variation.To excavate such a pattern,this article remodels the topological structure of the road network and defines the "Linkage Network".Linkage Network releases the redundancy in the traditional model,decouples the propagation patterns through the introduction of the element "Linkage",and enhances the expressive ability of the graph model.Based on the Linkage Network,Graph Recurrent Neural Network(GRNN)is proposed to predict the future traffic conditions.GRNN is well designed to mine the spatiotemporal relationship through its highly compressed propagation module.The mathematical derivation of GRNN's learning algorithm indicates that GRNN can learn the propagation patterns dynamically.The theoretical analysis proves the extreme efficiency of GRNN.In detail,compared with the traditional approaches,GRNN reduces the space complexity from O(nm)to O(n + m)and the time complexity from O(nm)to approximately O(n).Finally,comparative experiments based on historical data show that GRNN has higher accuracy on traffic prediction task.This study solves the problem of low efficiency and low accuracy of road condition monitoring in the traffic management system,improves the ability of supervisory departments to sense and control the overall traffic condition of the road network,and provides technical support for the development of urban traffic management system.Finally,this article discusses further research work.