Traffic Congestion Propagation Pattern Recognition And Visualization For Heterogeneous Urban Networks

Studying the laws and patterns of traffic congestion propagation plays an important role in traffic management and control for large cities.Currently,there are two branches of studies associated with traffic congestion propagation.The first is traffic bottleneck identification and the second is traffic congestion propagation modeling.Although numerous researchers have explored the two research topics and achieved many significant results,two major issues still remain:?1?For traffic bottleneck identification,most current studies are built on small-scale example road networks with regular topologies or small-scale regional networks.However,the urban road networks in reality tend to be large-scale and heterogeneous,implying the occurred traffic congestion is unevenly distributed in space and inconsistently distributed in time.The two heterogeneous characteristics bring great challenges in congestion propagation modeling with existing methods;?2?For traffic congestion propagation modeling,most existing studies aim to provide the decisionmakings for local traffic control on a single segment or at an individual intersection,lacking the consideration on network-wide traffic control.Meanwhile,the current studies are mainly based on traffic simulation,which depend on artificial experience and require extensive calibration work,leading to the unreliability and difficulty in practical applications.To solve the two issues,this study takes a large-scale and heterogeneous urban road network as the research object and adopts data-driven and artificial intelligence methodology to analyze and extract the patterns of congestion propagation using traffic big data,including four major topics,i.e.traffic state definition and quantification,traffic bottleneck identification,traffic congestion propagation modeling,and traffic congestion spatiotemporal correlation analysis.The main research results and conclusions achieved are summarized as follows:?1?To obtain sufficient traffic data for modeling congestion propagation in large-scale road networks,a traffic state characterization approach based on road conditions of navigation maps is explored.Taking the road condition images as the inputs,a simple yet effective road object segmentation method is proposed by using image processing techniques.Considering the input images do not contain specific road location details,the network is divided into multiple cells.Subsequently,two indicators defining the traffic state in the network are presented.One is a congestion index which characterizes continuous traffic states,and the other is a state classification index which characterizes discrete traffic states.Accordingly,two categories of visualization components,i.e.congestion index heat map and congestion state classification heat map are constructed.Based on the road condition images of Los Angeles urban network from Google Maps,the study data set was established.By comparing the generated traffic heatmaps with the original images,it is shown the proposed can reasonably reflect the spatiotemporal evolution characteristics of traffic congestion in the network,laying a good foundation for the follow-up research topics.?2?Based on graph theory and machine learning,an approach to identifying traffic bottlenecks in heterogeneous urban road networks is put forward.By comprehensively reviewing the concepts of traffic bottleneck,the definition of the traffic bottleneck in this study is proposed.Further,a bottleneck identification method based on graph theory is developed.Traffic bottlenecks in one week in Los Angeles urban network were identified.The experimental results show block 8₁3 is a very significant bottleneck in the top 10 bottlenecks.It ranks in the top 2 in 6 days,and has been congested during 6:00 AM to 20:00 PM,lasting more than 13 hours during Monday to Thursday and 9 hours during Saturday and Sunday.Additionally,block 9₁1 is another significant bottleneck,which ranks in the top 10 for 5 days.By inspecting the activation time of each bottleneck,it is found that the traffic congestion in Los Angeles road network has significant heterogeneous characteristics in space and time.?3?A data-driven traffic congestion modeling approach is presented by constructing congestion propagation directed graph.Considering the heterogeneous road network traffic could experience multiple congestion spread and dissipation periods in a day,a congestion propagation period definition method is developed.Considering there could be multiple congestion bottlenecks in heterogeneous urban road network,and eventually multiple independent congestion propagation branches could be shaped,a method to identify independent congestion propagation branches based on spatiotemporal clustering is proposed.Next,a datadriven traffic congestion propagation modeling method based on congestion propagation directed graph is established.By using the proposed approach,12 propagation periods and the associated congestion propagation branches were extracted.Further,the congestion propagation patterns of the typical congestion propagation branches with node number greater than 25 were analyzed.Results show the spatiotemporal scale of congestion spread on Tuesday,Wednesday,and Thursday is larger than the other days.The propagation process of most of the propagation branches follows the pattern of “diffusion?peak?stable state?dissipation”,and a small number of branches follows the pattern of “diffusion?peak?dissipation?diffusion?peak?stable state?dissipation”.The larger the time and space of congestion propagation involved,the more the congestion consolidation and differentiation times lasted.For the congested branch of weekdays,the key congestion consolidation behavior mainly occurs over different time intervals from 7:00 AM to 8:30 AM and from 14:00 PM to 17:30 PM.For the congestion propagation branches on Saturday and Sunday,the congestion propagation time scale is small,and the key congestion consolidation behavior rarely occurs.The congestion type of most of the selected congestion propagation branches is plane-congested,and the other is line-congested.?4?Based on the extracted traffic indicators,a spatiotemporal congestion correlation analysis model based on an attention-based recurrent neural network is developed,where the congestion propagation pattern is analyzed from a quantitative perspective.Taking the time series of the historical congestion index of the typical bottleneck blocks and their surrounding blocks as input and the current congestion index of the bottleneck block as the output,a spatial attention mechanism is introduced in the encoder stage to capture the correlation of congestion in space,and a temporal attention mechanism is introduced in the decoder stage to capture the correlation of congestion in time.Two categories of visual components are proposed,i.e.congestion spatial correlation heat map and congestion temporal correlation heat map.Based on the proposed approach,the spatiotemporal correlations between the typical bottleneck blocks and their surrounding blocks were analyzed.Results show the developed model can reasonably reflect and capture the correlation of traffic congestion in space and time by minimizing the prediction errors and analyzing the extracted spatial and temporal attention weight matrices.