Quantification and Prediction of Variability in Public Transport Operations

The travel time of public transport vehicles is subject to variability, which arises because of the stochastic nature of various factors influencing the operations. This variability is inconvenient for passengers and operators of public transport systems, as it introduces uncertainty and additional costs. The variability of travel times hinders passengers from knowing the precise arrival time of a planned public transport service or the expected travel time. This complicates their decisions regarding departure time, route, or even mode choice. Public transport operators are affected by the variability, as it reduces on-time performance and increases operating costs, for example, by requiring the addition of recovery times to schedules. Many strategic and operational decisions are also affected by variability and impact the cost of service. Therefore, understanding the variability of travel times can help improving current passenger service and operating efficiency. Recent advances in sensor and communication technology have significantly increased the amount and quality of data available, making it possible to model the variability of travel times better. This thesis concentrates on the modeling of variability for descriptive and predictive purposes. The three parts of this thesis shed light on different aspects of the modeling of variability. Part I focuses on the quantification (i.e., descriptive analysis) of travel time variability of urban public buses. The current state of the art on applying statistical distributions to model the variability of travel time components (i.e., dwell, running, and travel times) is reviewed. It finds that the best fitting distribution depends on the modeling context, but mixture distributions often provide a good fit to observed travel time data. To ensure the meaning of the parameters, we propose a novel approach to fit mixture distributions for bus running time distributions on a section level. The model can capture running time distributions over the course of the day and, thus, increase the explanatory power and fit compared to conventional models. Finally, we show how section-level running and station-level dwell time distributions can be combined to compute segment-level travel time distributions. The practical applicability of the models proposed in this part was demonstrated at urban bus lines in Zurich. Part II discusses the prediction of travel time variability. Whereas conventional prediction models typically limit the prediction of travel times to a single value, we propose a model to predict full probability distributions. In this way, the variability can be predicted. We introduce a probabilistic framework, basing on Bayesian Networks, to predict bus running and dwell times and the interconnected arrival times. Furthermore, we analyze in detail how the predicted variability evolves with a changing prediction horizon. Part III studies the identification of a distinct change in a public transport system. For analyzing variability, we typically assume that process generating the variability does not change over time. Hence, it is crucial to identify a change in the system. We do this at the example of a large-scale disruption in the aftermath of a tunnel collapse in Rastatt, Germany, in 2017. At this unique disruption, we could empirically identify how the disruption spreads in the Swiss railway network and how it affects the performance. The thesis suggests a meaningful way to quantify and predict variability in bus travel times and shows when the proposed methods are applicable. The results acquired in this pave the way towards probabilistic information of passengers and decision making of operators