Clustering Of Traffic Congestion Patterns Based On Traffic Performance Index

In order to quantitatively evaluate the state of congestion, many domestic cities have developed traffic congestion index systems. By applying the clustering analysis model to the existing traffic data, researchers have found the regularities of traffic congestion patterns, which can be used as the basis of congestion prediction and the policy research and decision. However, most existing research were conducted by using the data of traffic flow, while very few analyzed traffic congestion patterns by using the indexes which quantitatively evaluate the whole road network.In this thesis, a traffic patterns clustering model is developed aiming to analyze the characteristics of congestion patterns based on the Traffic Performance Index (TPI) in Beijing. In order to achieve the above objective, firstly, this thesis explained the TPI in Beijing and its characteristic, as well as the regularities of traffic congestion patterns based on the TPI data. Secondly, the clustering processes are divided into four parts and the different methods in each part are explained, on the basis of which, three clustering models are developed. Then, three evaluation indexes are proposed to evaluate the results of the above three models. The data of the second season of 2014 are used to determine the optimal model, and the model is also applied to the data of the same time of 2015 to validate its stability and consistency. At last, the characteristics of traffic congestion patterns of different years in Beijing are analyzed. The proposed model are also used in other megacities like Shanghai, Guangzhou and Shenzhen to analyze the congestion patterns. The result shows that when using the indexes based on the time-varying TPI, it will generate a more comprehensive reflection of the differences among traffic congestion patterns, and it will generate a better clustering result than using the indexes based on the geometry of TPI curves. When using the maximum Silhouette measure to determine the optimal number of clusters, it will be more effective to avoid the subjectivity and get a more reasonable number of clusters. When using the CV of the indexes as the weight of clustering index, it will describe the differences among the samples better and will improve the reliability of the results of clustering. The three indexes, namely CVin,CVout, and CR have a reasonable evaluation of the effectiveness of clustering results. It is found that the days of the second season of Beijing could be clustered into six clusters by the proposed Model, which covers 81.3% of the days, and the corresponding CVin and CVout are 0.141 and 0.337. For the other typical domestic megacities like Shanghai, Guangzhou and Shenzhen, the proposed model illustrates good applicability. The clustering results of congestion patterns also represent distinguished characteristics of different traffic management policies in different cities.