Demand-responsive Disruption Management in Mass Transit Systems: An Agent-Based Simulation Model to Assess Disruptions and Resilience-Enabling Rescheduling Measures

Transit systems are indispensable to the growing urban population. However, severe unforeseen and unprecedented disruptions place planners and operators in a position where typical risk mitigation and reliability improvements cease to provide manageable solutions. Instead, resilience assessment may lead the way towards an alternative perspective on handling disruptions. Firstly, it does not necessarily predispose that incidents are known or predictable. Moreover, it conceptualises that a system will undergo a number of phases during the draw-down and draw-up cycle following an adverse event, that each can be assessed and exploited for system improvements. This resilience perspective has been the premise of this thesis and drove development of computational models for the assessment of possible disruption management and decision-making strategies that remedy the impact of adverse events in transit systems. By taking resilience assessment as a key requirement, computational models need to capture the complete and dynamic draw-down and draw-up cycle during a disruption. At the same time, this thesis emphasizes that system performance ought to be, at least partly, expressed in terms of the inherent level-of-service that a transit system provides to passengers. Any evaluation or optimisation of system planning and operational rescheduling measures thus requires an explicit treatment of passengers and their interaction with the transit system supply. In addition, uncertainties related to variability of demand and supply as well as limited or lack of knowledge about current and future system conditions are prevalent. On the one hand, this means that models need to be able to capture these uncertainties. On the other hand, planning strategies and operational measures need to account and exploit these uncertainties to better cope whenever disruptions impinge on the network. In order to tackle the aforementioned challenges, this thesis starts by analyzing and reproducing the user demand patterns via the development of a static Origin-Destination (OD)-estimation problem for networked transit systems. Time series measurements of the number of passengers entering and exiting at stations are the only observable data. These data are used to infer the missing OD-coefficient matrix, where each coefficient denotes the share of passengers that enter at a specific origin station and are bound to travel to a designated destination station. The proposed solution method is a Bayesian model formulation and Markov Chain Monte Carlo (MCMC) sampling approach to find estimates of the parameter posterior distributions. The method is scalable to instances with hundreds of network nodes and superior over alternative Quadratic Programming (QP) formulations and solutions. In addition, the Bayesian model approach enables the explicit treatment and quantification of uncertainty, which not only provides a means of assessing the reliability of parameter estimates, but also has advantages in model testing and verification. The inference model is demonstrated on the New York City (NYC) subway network to showcase its advantages regarding the robustness to high-dimensional and underdetermined OD-estimation problems. Solving the OD-estimation problem efficiently is a prerequisite for many of the tasks that follow; a main objective of this thesis is to develop a predictive model of the effects of transit system disruptions that can be validated against a real-world system and further be leveraged for alternative design and operational strategy optimisation. Deriving OD-demand estimates for the NYC subway network, which is the reference network of this thesis, is thus a necessary step. This thesis then continues to delineate the development of the agent-based discrete-event simulation model to capture dynamic processes that occur during transit system disruptions. The model is a dynamic transit assignment model that emulates passengers as individual entities (i.e., agents) that interact with the transit system supply network. The agent-based simulation approach enables the generation of heterogeneous passenger agent populations and asynchronous sharing of system and timetable information with these passenger agents. Although targeted towards assessing disruption scenarios, the model is equally suitable for normal, un-disrupted (i.e., nominal) conditions. By expressing each individual passengers' en-route behavior, passenger flows emerge bottom-up, as opposed to the top-down representation of typical transit assignment approaches. The novelty of the proposed model lies with the re-routing module of passenger agents, which accounts for multiple choice options when faced with service disruptions, including the option to terminate a journey early. Moreover, the simulation programme is able to handle multiple local faults and dispatch strategies simultaneously, during which passenger agents can choose to freely re-route their journeys. Furthermore, this thesis applies the developed agent-based simulation model to the NYC subway network. The results are validated against real-world observations of passenger access and egress rates at stations during nominal and disrupted conditions. This validation step establishes is that the simulation model is able to closely capture the real-world observations of passenger egress rates. In addition, the simulation enables high-resolution predictions of passenger movement throughout the network and uncertainty estimates of these predictions. This approach therefore facilitates reliable estimates of crowding and congestion levels on-board trains and at stations. Moreover, the simulation model approximately resolves the changes in passenger access and egress rates at stations following a specific fault event and ensuing service disruption. This fault event occurred on May 9th 2017 and is due to a power outage and resulting signalling defect, halting the operations at a station in the network for one hour. Despite being an isolated local fault, the simulation model reveals that this contingency causes system-wide and long-lasting ripple effects on service level losses and timetable deviations. Therefore, this thesis underlines that demand-supply interaction during disruptions is not limited to local delays and re-routing decisions along a single line, for instance, but instead propagates across the entire network. Seemingly intact and undisrupted network components may thus become unusually stressed and face passenger demand levels that the standard (nominal) timetable and operational schedule are not designed for. By taking as reference the same aforementioned initiating fault that caused the real-world disruption event in the NYC subway network on May 9th 2017, this thesis assesses the potential of intervening rescheduling strategies to partly restore the system service level during and after the disruption. Without any rescheduling interventions, the simulation assumes that services scheduled to run along the blocked segment will stack up until the blockage is cleared. The proposed intervening rescheduling measure short-turns trains nearest to the fault location, thus maintaining train movement upstream of the fault location and covering for the missing opposite direction services. By studying the dynamic response of system-wide link flow levels over a several-hour long time horizon, this thesis points out that this strategy, on average, reduces deviations from the undisrupted link flows. However, it also identifies abnormal and unfavourably high demand levels along certain line corridors that are not directly affected by the fault or rescheduling strategy. It therefore shows that the conceptual abstraction of the resilience draw-down and draw-up cycle of the full network cannot be generalised to individual components of the network. In return, the developed simulation model and decision-making framework can support the identification of balanced intervening rescheduling measures which minimize the potential for creating abnormal conditions in the remaining operational sections of the transit system. Having established that the simulation model is suitable for predicting the immediate and hour-long aftermath of disruption events, a next step is to explore possible remedial strategies for timetable adjustment and rolling stock rescheduling through a structured simulation-based optimisation approach. This approach is deemed suitable for the medium-to-long-term tactical planning phase in railway and metro networks to preemptively assess rescheduling measures, inform contingency plans, identify potential bottlenecks, or prepare component and process redundancies to be swiftly engaged and dispatched when needed during disruptions. The developed agent-based simulation model is therefore coupled to a multi-objective simulation-based optimisation tool known as SOCEMO (Müller, 2017), and applied to a subcomponent of the NYC subway network. A benchmark standard timetable for nominal operations is determined that adjusts the dispatch headways every two hours throughout the day in order to adapt to the daily passenger peak and off-peak demand profiles. The standard timetable optimisation objectives account for level-of-service and rolling stock requirements, where level-of-service is measured in terms of the average travel time of passengers and rolling stock requirements are measured in terms of the number of allocated trains needed to fulfill the standard timetable. Once in place, the benchmark timetable is tested on a scenario that considers multiple track blockages at separate locations in the network. An optimal short-turning strategy is determined, where the set of standard timetable objective functions is extended to also account for passenger inconvenience measured in terms of the number of agents that need re-route to a new destination. The decision variables are the times and locations to short-turn trains along the affected line. The chosen Pareto-optimal short-turning strategy ensures that passengers who reach their intended destinations experience a similar level-of-service as for the standard timetable, and limits the inconvenience to passengers who need to re-route to a new destination or cut their journey short while travelling, all while reducing the number of required trains to uphold the dispatch schedule. This strategy is compared to an ad-hoc strategy of short-turning trains nearest to the fault location, which is also termed on-location short-turning. For the specific disruption scenario under study, on-location short-turning slightly increases level-of-service and reduces passenger inconvenience versus the Pareto-optimal strategy. However, it requires more trains to be dispatched. The exploratory analysis of optimal and on-location short-turning strategies shows that considerable improvements are possible to retain level-of-service and reduce passenger inconvenience, while mitigating severe deviations from the planned rolling stock circulation during a disruption. However, the computationally costly simulation-based optimisation approach is unsuitable for real-time applications. At the same time, it is conjectured that the on-location short-turning strategy is not guaranteed to always utilize available resources most efficiently. Therefore, a control methodology is proposed that is embedded into a real-time flexible rescheduling strategy, denoted as fuzzy short-turning. Under this strategy, trains are variably and flexibly short-turned upstream of a track blockage. The short-turning times and locations are determined from a fuzzy control algorithm that accounts for the incertitude regarding passenger congestion and crowding levels as well as the deviations from the planned rolling stock schedule. The assessment of multiple disruption scenarios with variable durations shows that the fuzzy short-turn control strategy generally abates passenger inconvenience more efficiently, when compared to the on-location short-turning strategy; only in few scenarios does on-location short-turning outperform fuzzy short-turning. The proposed fuzzy control approach is therefore a promising decision support system for disruption management in high-frequency urban transit networks, especially given that its underlying control rules can be verbosely formulated by professional operators. Moreover, it is poised to be compatible with mathematical optimisation formulations for timetable and rolling stock rescheduling. Overall, this thesis achieves to address the three key objectives originally set out: (1) Gaining an understanding of the (re-)distribution of transit system passenger flows during normal and disrupted conditions; (2) Facilitating experimental, close-to-reality testing of various disruptions and operational rescheduling interventions in a controlled and observable environment; and (3) Identifying suitable and tuned vehicle operation and timetable rescheduling strategies that can control the disruption draw-down and draw-up cycle under uncertainty.