Travel time prediction for the planning of mass transit companies: a machine learning approach

Abstract

In this thesis we undertook a study in order to know how travel time prediction can be used in mass transit companies for planning purposes. Two different problems were identified: the definition of travel times (1) for timetables and (2) for bus and driver duties. All these studies assume the existence of data on the actual trips, typically obtained from Automatic Vehicle Location (AVL) systems. The first problem is a well-known problem with several related studies in the literature. Our approach is not analytical. Instead, we have designed and developed a decision support system that uses past data from the same line and representative of the period the timetable will cover. This problem was the least studied. With respect to the second problem, travel time prediction three days ahead, we focused on how much we can increase in accuracy if we predict travel times for the definition of bus and driver duties as near the date as possible, instead of using the scheduled travel times (STT). The reason for doing this is that, if the increment is important, it is expected to reduce operational costs and/or increase passengers' satisfaction. In this second problem we used machine learning approaches. However, we started by defining a baseline method (in order to evaluate comparatively the results obtained with more sophisticated methods) and an expert based method using the knowledge we had at the time together with the tra±c experts from the STCP company. Then, we tried three different algorithms with reported good results in different problems. They were: support vector machines, random forests and projection pursuit regression. For each of these algorithms, exhaustive tests were done in order to tune parameters. Other tests were done using the three focusing tasks: example selection, domain values selection and feature selection. Accuracy was improved using these approaches. The next step was to experiment heterogeneous ensemble approaches in order to ameliorate further the results by comparison with the use of just one model. An extensive survey on ensemble methods for regression was undertaken. Several experiments using the dynamic selection approach were executed. Approaches using ensembles have improved results consistently when compared to the use of just one model. Experiments on the second problem finished by comparing the baseline, the expert based, the best single algorithm (with the respective tuned parameters and focusing tasks), and the ensemble approach, against the use of STT, on various routes. Results gave a small advantage in terms of accuracy to the ensemble approach when compared to the expert based method. However, the expert based approach needs less data and is much faster to tune. The actual method used by STCP (the use of STT) was competitive for circular routes. However, this result can be explained, at least partially, by how these routes are controlled. On the rest of the routes tested, it was clearly beaten. We also try to give practical answers in using travel time predictions for the planning of mass transit companies, using the STCP company as study case. The impact of travel time prediction in the business goals, namely clients' satisfaction and operational costs, is not addressed despite it is the natural step forward of this research.