Research And Application On Advanced Public Transportation System Data Mining

It is the focus areas of advanced public transportation systems (APTS) research to discover the patterns of travel times and passenger flows and assess the public transit network performance. Along with the rapid development of APTS, mass data have been collected and stored in smart card fare payment subsystem, scheduling subsystem and automatic passenger counting subsystem. And so, data mining, as one powerful data analysis technique, has been applied by APTS researchers to discover significant patterns and rules and retrieve potential high-level knowledge underlying the data.In view of the requirements and characteristics of APTS, an extensible architecture for APTS data mining is first proposed. This architecture tries to incorporate the basic data mining algorithms and facilitate further developments to satisfy the needs of practical applications.Travel time mining algortihm and dynamic OD matrix estimation algorithm are then presented. Finally, accessibility assessing method based on travel time and dynamic OD matrice is discussed.Research to address these issues is significant to public transit planning, management and assessment. In general, the main contents and achievements of this dissertation consist of the following aspects:1) Mining travel times from smart card fare payment data.The recent adoption of smart card technology as a fare payment system by many transit operators provides a new way to infer travel times between adjacent stops in public transit network. One method is proposed to infer travel times from smart card fare payment data and bus scheduling data. An experiment is designed to test this algorithm with real-world data and the outcomes prove that the error of this method is small and the convergence is fast.The method first classifies two sequential swipes to decide whether they occurred at the same stop with Naive Bayes Classifier (NBC). Travel times are then estimated from the NBC results using Maximum Likelihood Estimation (MLE), Dynamic Programming (DP) and Quadratic Programming (QP) methods.To solve the problem with imprecise initial parameters, alternative MLE method is proposed, which updates parameters and estimates values alternatively until convergence.2) Dynamic OD matrix estimation of public transit network.In the context of APTS, boarding and alighting passenger counts at each stop and travel times between adjacent stops are available. Dynamic OD matrix estimation in congested transit network from these data is examined. A dynamic OD matrix estimation model based on path flows estimation is developed. To solve the problem by path flows estimation, it is shown that, under dynamic user equilibrium constraints, if the capacity constraints are ignored, transit path flows satisfying boarding and alighting count constraints are minimum cost path flows. And then it is shown that capacity constraints are implied by boarding and alighting count constraints. Finally, the dynamic user equilibrium constraints are added to the objective function as one penalty term. And so, the dynamic transit OD matrix estimation problem is transformed into a nonnegative convex quadratic programming problem. Sherali algorithm is used to reduce the dimension of path flows. The experiment in a small size transit network shows that the error of prior OD matrix and the error of boarding and alighting count have an influence on the error of estimating results, and the proposed model and algorithm are feasible.3) Accessibility assessment of public transit network. Accessibility is a fundamental indicator of location advantages. This dissertation presents one method that seeks to measure time accessibility of public transit network with travel time and dyamic OD matrix information. It first bisects the service area of each bus stop by Voronoi diagram, and then calculates the expectation commute time of each grid in the service area. The public transit network of Kunming is assessed using this methodology. The service area is first divided into 10m*10m grids, and then the commute time of each grid is calculated, which forms time contours.