Research On Urban Road Traffic State Identification And Short-term Prediction Based On Multi-source Data Fusion

The identificaition and prediction of urban road traffic state are key techniques to realize the intelligence of urban traffic.On one hand,the real-time traffic state identification and prediction can help city authorities to improve the urban road system efficiency,as well as to immediately respond to traffic incidents.On the other hand,the travelers could take advantage of those real time information to make a better route planning with fewer trip time and congestion.However,limited by the quality and sampling rate of data collected in most cities,the urban road traffic state is identified at the network-level rather than the link-level.Motived by the the necessity and practical difficulties,this study proposes a data-driven modeling framework for idntification and prediction of urban road traffic state utilizing multi-source data of floating car trajectory and urban road network database.Firstly,the paper introduces the structure of the floating car trajectory data and the road network database.And the data quality inspection and preprocessing are carried out on the origin data.These works mainly include the selection of research area,cleaning of invalid data and speed correction,etc.The method for converting origin trajectory data into structured spatiotemporal data is proposed for the following research..Secondly,in order to solve the data sparseness problem in full space-time range,the paper proposes a traffic flow data completion method based on tensor decomposition.The basic theory of tensor decomposition is introduced,and the spatio-temporal correlation analysis of traffic flow based on the actual urban road traffic flow speed data is discussed.The weighted optimization method based on CP decomposition is modified by reasonablely adding factor tensor in the regular terms,which contains the characteristics of weekday,time period,road segment.The experiements prove that the above algorithm can improve the completion accuracy under different data missing degrees.Then,on the identification of urban road traffic states,the paper discusses the limitations and reliability of traditional traffic flow parameters in real situation,and several traffic state evaluation indexes extracted from GPS data including float car volume,average speed,standard deviation of speed and speed ratio.Taking the road network around Nanjing South Railway station as an example,the paper uses these parameters to describe the traffic state distribution and change in the road network to further verify the rationality of the indexes proposed.Based on the improved K-Means cluster algorithm,cluster analysis is carried out for different levels of road.Combined with the characteristics of different states and clustering results,the traffic states on urban roads can be classified into smooth,saturation and congestion.Finally,on the urban road traffic states prediction,considering the traditional convolutional neural network cannot be adopted to the urban road topology,the graph neural network theory is introduced.The road topology are described by an distance-based adjacency matrix.Based on this,the paper establishes a diffusion convolutional recurrent neural network(DCGRU)for short-term traffic state prediction to capture both spatial and temporal dependency.The traffic evaluation indexes,road geometric features and the road network adjacency matrix are the inputs of the proposed model.A numerical experiment is conducted based on one-month dataset.The results indicate that the DCGRU outperforms other traditional models in different traffic conditions.