Use this URL to cite or link to this record in EThOS:
Title: Data-driven modelling of passenger response to disruption on the London Underground
Author: Goldberg, George
ISNI:       0000 0004 7223 4770
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
In this thesis a novel approach to modelling passenger response to disruption on metro rail networks is developed, consisting of a two step process using smart-card data to produce disaggregate models expressed in terms of individual response behaviours. The work is carried out in the context of the London Underground, using a full snapshot of the Oyster smart-card travel histories of all passengers in conjunction with records of all incidents to occur on the network during the study time period. The combination of two key factors differentiates this work from existing studies in this area: the model is built on a large, passively collected dataset (smart-card data), but is expressed in a disaggregate form in terms of the actual actions taken by passengers, rather than the aggregate overall changes to the origin-destination flows. The conceptual framework underlying this work considers a three step process. First, incidents occur on the metro rail network causing disruption to train services, which consequently affects passengers using those services. Second, on experiencing disruption to their journeys, passengers may continue with them as planned, or may take a number of actions in an attempt to mitigate the effects of the disruption that they experience, such as changing to a different mode of transport or abandoning their journey altogether. Finally, using fare-collection smart-card data, it is possible to observe these responses in the individual travel histories of passengers. This thesis first presents a qualitative investigation into passenger response to disruption, with the dual objectives of confirming the assumptions of the conceptual framework and identifying the set of actions passengers take when their journeys are disrupted. The subsequent quantitative investigation is expressed in terms of this set of behaviours. The next stage consists of the development of a heuristic process to identify instances of those behaviours in a smart-card travel history dataset, in conjunction with the transport operator’s logs of the incidents which occurred during the same time period. Finally, a series of models of individual passenger response to disruption are developed from the data produced by the heuristic process. These models serve two purposes. The first is to understand the impact of a range of explanatory variables related to the passengers and the incidents on the response behaviours exhibited, for which the logistic regression family of models is used due to its easily interpretable coefficients. The second is to provide predictive modelling of the response behaviours, which imposes no requirement of model interpretability, allowing the logistic regression models to be compared with a range of black box machine learning classifiers. The initial qualitative investigation identified six behaviours exhibited by passengers in response to disruption: “carry on regardless”, “abandon journey”, “change time of travel”, “change origin or destination”, “change mode of transport” and “change route”. The first five of these can be identified from the available data sources through the heuristic behaviour extraction process. The logistic regression models identified a number of patterns, the most distinctive of which include an increase in the relative probabilities of the “change” behaviours over “carry on” as the incident length increases, an increase in the relative probabilities of “change origin/destination” and “abandon journey” as passenger average household income increases, and a corresponding substantial decrease in the relative probability of “change to bus” also as passenger household income increases. Significant evidence of individual heterogeneity in response behaviours was also identified. Overall model predictive performance was similar between the logistic regression models and a number of the investigated machine learning classifiers.
Supervisor: Polak, John Sponsor: Imperial College London
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral