ABSTRACT. Waiting at bus stops is often one of the most uncomfortable parts of transit use (Fan et al., 2016). Uncertain or long waits discourage ridership and diminish overall user satisfaction. While agencies usually focus resources on fleets and service frequency, targeted investments on stop-level amenities can significantly improve system comfort, safety, and equity (Kim et al., 2020; Shi et al., 2021; Watkins et al., 2011). However, deciding the placement and selection of amenities to balance social equity with system efficiency under limited budgets remains a key challenge. To address this, we used Automated Passenger Count (APC) data to measure stop-level ridership demand and combined it with two complementary data sources: Geographic Information Systems (GIS) data and the Calgary Equity Index. The GIS data describe the built environment, such as connectivity, land use, and weather exposure, while the Calgary Equity Index summarizes socio-demographic vulnerabilities, including the share of children, seniors, people with disabilities, and transit-dependent households. Together, these inputs formed the attribute set for evaluating and ranking bus stop needs. Stops were classified into three amenity tiers: basic, moderate, and maximum. The framework combines a hybrid multi-criteria decision-making (MCDM) analysis with integer programming model to determine amenity allocation across stops. The MCDM component used an entropy-weighted TOPSIS approach to rank attributes by their ability to differentiate between stops while minimizing subjective bias (Hamurcu & Eren, 2022). In parallel, empirical comparisons between stops with and without specific amenities generate utility scores that link attributes to amenities. These relationships formed a three-dimensional weighting matrices: (1) community-attribute weights (W_ac) representing the importance of each attribute for each community, (2) amenity-attribute weights (W_ja) capturing how each amenity relates to each attribute, and (3) resultant community-amenity weight (W_jc) obtained by integrating the first two to reflect the importance of each amenity level for each community. The final W_jc values are incorporated into the optimization model’s objective function. The optimization model aims to maximize the net present value (NPV) of total benefits, including reduced perceived waiting time and avoided mode-switching during delays, minus the sum of installation (fixed) and operating (variable) costs over a 20-year horizon. Constraints included (i) prevention of redundant upgrades, (ii) a fixed budget, (iii) one amenity level per stop, and (iv) equity conditions, ensuring both a fair geographic distribution across communities and proportional benefit allocation to riders. Route 32, a 30.6 km north–south corridor serving nine distinct communities, was used to demonstrate the framework. The optimization revealed significant efficiency and equity improvements. Before implementation, nearly half of boarding passengers used minimally equipped stops. After optimization, the number of passengers exposed annually to only basic amenities dropped from 22,480 to 3,203 (-86%), while the share of riders benefiting from moderate and maximum amenities increased by 63% and 52%, respectively. The required investment for these enhancements (CAD 1.25 M) and increased variable costs (CAD 839 K) yielded long-term benefits of CAD 11.05 M, producing a NPV of CAD 9.37 M. These findings show that targeted amenity upgrades can be both socially impactful and financially viable. Furthermore, the equity analysis showed that the combined equity scenario achieved the most balanced outcome, delivering the highest NPV while serving both vulnerable populations and high-demand areas fairly. and that performance remained robust under varying budget levels. Sensitivity analyses confirmed the framework’s robustness, showing consistent results across varying investment levels. They also revealed diminishing returns beyond approximately CAD 5 million, suggesting a practical threshold for cost-effective amenity expansion. This study contributes at theoretical and methodological levels. Theoretically, it reframes transit equity as a multi-dimensional objective that considers fairness both across communities and in passenger demand, rather than treating it as a secondary constraint. By incorporating NPV, the framework shifts the evaluation perspective from short-term ridership and costs to long-term passenger comfort and inclusion. Methodologically, the study introduces a hybrid framework that integrates MCDM with integer programming optimization to guide amenity allocation in a systematic way. Moreover, it accounts for equity by linking community attributes with amenity–attribute relationships to generate interpretable utility scores, which are then used to determine optimal amenity configuration. Applying the framework to Calgary Transit’s Route 32 demonstrates how it can improve comfort and equity while remaining financially feasible, offering an adaptable, data-driven tool for allocating bus stop amenities to support equitable and effective transit planning.

References Fan, Y., Guthrie, A., & Levinson, D. (2016). Waiting time perceptions at transit stops and stations: Effects of basic amenities, gender, and security. Transportation Research Part A: Policy and Practice, 88, 251–264. https://doi.org/10.1016/J.TRA.2016.04.012 Hamurcu, M., & Eren, T. (2022). Applications of the MOORA and TOPSIS methods for decision of electric vehicles in public transportation technology. Transport, 37(4), 251–263. https://doi.org/10.3846/TRANSPORT.2022.17783 Kim, J. Y., Bartholomew, K., & Ewing, R. (2020). Another one rides the bus? The connections between bus stop amenities, bus ridership, and ADA paratransit demand. Transportation Research Part A: Policy and Practice, 135, 280–288. https://doi.org/10.1016/J.TRA.2020.03.019 Shi, X., Moudon, A. V., Hurvitz, P. M., Mooney, S. J., Zhou, C., & Saelens, B. E. (2021). Does improving stop amenities help increase Bus Rapid Transit ridership? Findings based on a quasi-experiment. Transportation Research Interdisciplinary Perspectives, 10. https://doi.org/10.1016/j.trip.2021.100323 Watkins, K. E., Ferris, B., Borning, A., Rutherford, G. S., & Layton, D. (2011). Where Is My Bus? Impact of mobile real-time information on the perceived and actual wait time of transit riders. Transportation Research Part A: Policy and Practice, 45(8), 839–848. https://doi.org/10.1016/J.TRA.2011.06.010

ABSTRACT. In-person presentation (805 words)

As transit agencies work to retain existing riders and rebuild trust in a post-pandemic context, understanding how the drivers of trip satisfaction evolve as systems mature is essential to prioritizing investments and sequencing improvements. This study examines temporal shifts in the determinants of commute trip satisfaction across three of Montréal’s rapid transit services, namely the long-established Metro, the recently implemented Pie-IX bus rapid transit (BRT), and the first branch of the Réseau express métropolitain (REM) light rail, using a two-wave panel of 354 commuters surveyed in 2023 and 2024. We frame service attributes with the hierarchy of transit needs, which organizes determinants into functional (e.g., access, travel, and waiting time), security, and hedonic layers (e.g., comfort, information), and incorporate a value layer (cost-benefit) that has been highlighted in prior research. While the hierarchy posits that higher-order considerations gain salience once foundational needs are consistently met, systems may not follow a strict chronological progression as older transit services can remain at lower-order needs if foundational issues persist, while new systems may start operations meeting higher-order expectations. We estimate mode-specific, weighted Bayesian ordered probit models of trip satisfaction and evaluate the marginal predictive contribution of each layer using leave-one-out cross-validation (LOO), comparing full model specifications with versions that exclude functional, hedonic, or value attributes. This design allows us to measure not only which individual attributes matter most by mode and year but also whether the aggregate layers align with the theory’s conditional relevance. Descriptive dynamics and LOO comparisons are interpreted jointly: satisfaction scores indicate which needs are being met, while changes in each layer’s predictive salience indicate whether users are shifting the criteria they rely on when forming overall judgments. In this way, we test whether, as foundational conditions stabilize, higher-order considerations become more decisive. Three main findings emerge. First, the Metro reflects a mature system in which satisfaction is high and stable; comfort and other experience-related attributes increasingly base evaluations, while functional elements such as travel time remain universally important. In this environment, hedonic considerations carry the largest marginal contribution to predictive accuracy when contrasted with specifications that omit functional or value variables, consistent with a setting where core needs are widely perceived to be met and riders’ satisfaction relies on higher-order qualities. Second, the Pie-IX BRT illustrates how users recalibrate expectations as a newer service settles into habitual use. Waiting time initially dominated evaluations; yet between 2023 and 2024 its salience declined even as mean satisfaction with waiting time did not markedly improve. This finding suggests that adaptation and coping mechanisms can reduce the weight of unresolved functional gaps. Third, the REM findings indicate a layered but potentially transitional structure during its early operational stage. Functional needs exerted the strongest influence, yet hedonic factors such as information and comfort also played a notable role. This combination suggests that while core operational expectations like waiting and travel times are being met, unresolved challenges with station accessibility keep functional needs central. Across all three modes, travel time emerges as a universal anchor, where even as higher-order needs rise in prominence, it remains foundational to user satisfaction. Our results provide a nuanced exploration of the hierarchy of transit needs. At the attribute level, functional and hedonic drivers often operate in tandem, particularly in systems like the REM where some foundational gaps persist alongside strong operational performance. At the aggregate layer level, clearer patterns emerge: the metro aligns most closely with the theory’s expectations, the REM illustrates a potential transitional stage, and the BRT diverges as functional concerns appear to lose salience despite being unresolved. This layered perspective highlights both the value and the limits of the framework, while it provides a useful lens for interpreting satisfaction dynamics, its conditional relevance assumption does not always hold in practice. Future research could explore how factors such as service familiarity, expectation adjustment, and user coping mechanisms contribute to shifting perceptions over time, particularly in newer systems. Policy implications follow from the mode-specific results and the observed sequencing of needs. Agencies should avoid prioritizing higher-order features before basic operational features, such as reliability, speed, accessibility, are met. For the Metro, with hedonic drivers at the forefront, investments that improve the user environment (i.e., seating, cleanliness, ventilation, crowd management) are likely to yield the largest satisfaction gains, so long as foundational performance is preserved. For the Pie-IX BRT, improvements that deliver headway reliability and predictable waits should precede comfort upgrades. For the REM, a combination of operational reliability and high-quality information is crucial: real-time announcements (in loco and online) and clear wayfinding can mitigate the perception of unreliability derived from recurring service disruptions and sustain confidence during early years of service consolidation. Across all modes, travel time retains a universal role, suggesting that minimizing in-vehicle time and access burdens will continue to underlie rider judgments even as higher-order needs grow in salience.

ABSTRACT. Advances in transport technology and rapid urbanization, combined with increasing spatial separation between housing and employment, have lengthened commuting distances in many metropolitan regions. In response, high speed or regional express rail systems are being introduced as complements or alternatives to conventional urban rail and bus services. Although regional express rail shares several operational features with urban rail, it typically serves longer trips, stops only at major hubs, and therefore relies more heavily on transfers. These characteristics suggest that user satisfaction and usage behavior for regional express rail may be driven by a different structure of service attributes than for conventional urban public transport. This study investigates the causal relationships between perceived service quality and actual usage frequency in a regional express rail context, explicitly incorporating transfer related conditions and user perceived attribute importance. The study contributes to the literature in two main ways. First, whereas most satisfaction and usage studies focus on urban rail or bus and treat public transport as a single, homogeneous mode, we explicitly represent structural features of regional express rail, including its long-distance role and high transfer intensity. In particular, we introduce a latent variable for transfer service that captures the clarity of transfer routes, availability of transfer information, and transfer time. Second, we allow for heterogeneity in attribute importance. Because passengers differ in how strongly they value specific service aspects, satisfaction with an attribute that a user considers important is expected to have a stronger effect on usage than satisfaction with an attribute perceived as less important. To capture this mechanism, we jointly measure satisfaction and importance for multiple service items and construct importance weighted satisfaction indicators for each service dimension. Based on the Theory of Planned Behavior, we treat satisfaction with individual service attributes as indicators of attitude toward the regional express rail service and interpret overall satisfaction with the system as a proximal indicator of behavioral intention to use it. We then examine how this behavioral intention translates into actual behavior, measured by the self-reported frequency of regional express rail use. To quantify these relationships, we estimate a structural equation model (SEM) using 2025 survey data from 500 regional express rail passengers in Korea. Service evaluation is represented by three latent constructs: user convenience, operational service, and transfer service. User convenience is measured using indicators such as station accessibility, the number of station entrances, and the ease of access to facilities. Operational service is defined by indicators such as perceived fare reasonableness, appropriateness of in-vehicle travel time, and adequacy of headways. Transfer service is assessed using items such as the clarity of routes to connecting lines, provision of transfer information, and perceived transfer time. For each service attribute group, we construct two latent variables: an unweighted satisfaction factor based on 7-point Likert-scale satisfaction scores centered at the scale midpoint, and an importance weighted satisfaction factor based on the product of centered satisfaction and centered importance scores for the same items. Estimation results show that both unweighted and importance weighted satisfaction in all three dimensions have significant positive effects on overall satisfaction, and that overall satisfaction in turn has a significant positive effect on usage frequency. Thus, each service attribute group exerts a positive indirect effect on usage via overall satisfaction. The standardized coefficients of the importance weighted constructs are consistently larger than those of the corresponding unweighted constructs, indicating that improvements to attributes users regard as important are more strongly associated with increased usage frequency. In a model without importance weighting, operational service appears to be the dominant driver of overall satisfaction and, consequently, of usage. When importance weighting is introduced, however, the effect of transfer service increases markedly and exceeds the effects of the other latent variables, whereas the effect of operational service changes only slightly. These findings have several practical implications for the planning and management of regional express rail. First, by prioritizing improvements to service attributes that users perceive as more important, operators can design more resource-efficient enhancement strategies. The fact that importance weighted satisfaction has a stronger indirect effect on usage frequency than unweighted satisfaction suggests that focusing on high-importance attributes yields greater behavioral returns for a given investment. Second, the results emphasize the need for customized service improvement strategies that reflect user heterogeneity in attribute priorities. Because passengers vary in priorities, station-specific strategies based on user profiles may enhance service effectiveness. For instance, at stations with high transfer activity, targeted upgrades to transfer services, such as clearer signage, reduced transfer distances, or improved transfer coordination, may be more effective. In contrast, at stations serving information-sensitive users, enhancing operational transparency and real-time information provision could be more impactful. Third, the findings support the development of differentiated service level targets for each attribute dimension. In the unweighted model, operational service exerts the strongest influence on overall satisfaction and thus the highest indirect effect on usage. However, once attribute importance is considered, the contribution of transfer service increases substantially and exceeds that of the other service dimensions. Notably, the effect of operational service remains relatively unchanged after importance weighting, suggesting that passengers treat it as a basic requirement rather than a value-adding feature. In other words, while operational service does not substantially increase satisfaction when performing well, its degradation significantly harms user experience, consistent with the notion of a “hygiene factor.” Accordingly, policy efforts for operational service should focus on ensuring stability and consistency, maintaining a reliable minimum standard to avoid dissatisfaction. On the other hand, transfer service, which shows greater responsiveness to importance weighting, warrants a different policy approach. Rather than ensuring uniform quality across the network, targeted investment in transfer service, tailored to users, segments, and time periods where transfer convenience is particularly valued, can yield greater gains in satisfaction and usage. This finding is especially relevant for regional express rail systems, where transfers are frequently required due to limited station stops. To improve transfer experience, strategic policies and investments should be directed toward streamlining transfer paths, simplifying pedestrian flows, and integrating transfer-related information systems.

ABSTRACT. Preference for presentation at the conference: In-person

Background:

Public transport (PT) journeys are highly complex, comprising multiple trip legs (e.g. access, egress, waiting, transferring, in-vehicle), transport modes (e.g. walking, cycling, PT) and environments (e.g. streets, stations, vehicles). Along the door-to-door (DTD) journey, travellers may face barriers, such as an unsafe walk to the station or a difficult transfer, which can shape their trip satisfaction or render PT inaccessible. To support travellers’ well-being and sustain PT ridership, we therefore need to consider the DTD experience, and identify barriers that influence satisfaction. However, few studies account for the complexity of the DTD journey, or include predictors from each trip leg together, limiting our insights into the most important aspects and key barriers.

A key challenge is the lack of large-scale data capturing both behaviour and experiences along the DTD journey. Traditional travel surveys provide length and duration of each trip leg, but lack the access/egress routes, which can be crucial for considering the built environment conditions and their effects on satisfaction. Similarly, travel card data often begin and end at stations or in vehicles, missing access/egress entirely. Satisfaction surveys, also mainly focus on stations or vehicles as these aspects fall within operators’ control. This highlights the need for data collection methods that record both travel behaviour (e.g. routes, environments and travel modes from DTD), and travellers’ experiences along the way (e.g. perceived safety, difficulty, comfort, trip satisfaction).

Aim:

In this study, we introduce a mobile-app-based data collection capturing travel behaviour and experiences along DTD PT journeys in Denmark. We present correlates of trip satisfaction in the abstract and plan to examine the effects of trip characteristics, built environment conditions, travellers’ experiences, and individual factors in a structural equation model (SEM) analysis until Transitdata 2026. This will identify strategies and design measures that improve trip satisfaction and support a more seamless DTD journey.

Data:

The data collection covers Denmark and consists of three steps: (1) an online survey that provides background information, (2) a mobile app that tracks travel behaviour, and (3) a built-in survey that measures travel experiences along DTD PT journeys.

First, we distributed a preliminary online survey to a representative sample using the Danish Central Person Register (CPR). This survey recorded respondents’ attitudes towards PT and socio-demographic features. It also recruited respondents who indicated that they are frequent PT users to the mobile app. Second, the mobile app tracked users for 14 days, and recorded their trip characteristics (e.g. route, travel mode, distance, duration) along the DTD journey, and activities during the day. Third, app users received a built-in online survey via push notifications when the app detected a PT journey. This survey measured travel experiences (perceived safety, difficulty, comfort) across different trip legs on a 5-point Likert scale: safety for all legs, difficulty for all except the vehicle, and comfort only inside the vehicle. It also recorded trip satisfaction on a 5-point Likert scale, together with experienced disruptions during the trip.

The data collection began in October 2025, and will run for a full year, recruiting new users each month to capture seasonality effects. Between 3-15 October 2025, 363 respondents completed the preliminary survey. Of these, 68 app users recorded 3,575 trips until 30 October. Most were walking trips (43.4%), followed by car (29.5%), bicycle (14.3%) and PT (11.8%). Furthermore, 31 app users completed the built-in survey for 104 PT trips. This sub-sample forms the basis of our current analysis.

Methods:

For our preliminary analysis, we examined correlations between trip satisfaction and travel experiences (perceived safety, difficulty, comfort), trip characteristics (transfers, unexpected disruptions, distance, duration) and individual factors (attitudes) as a preliminary analysis. These findings will guide the development of the SEM model to be presented at Transitdata 2026.

As trips had varying number of transfer and in-vehicle legs, we tested different ways of incorporating experiences during these legs to our analysis. We computed the minimum and average values of perceived safety and difficulty for transfers, and comfort for in-vehicle capturing both the worst and typical experiences. We also computed overall journey averages to assess whether the trip-leg specific or average experiences correlate more strongly with satisfaction.

As we were interested in testing the effects of PT attitudes, we also conducted a factor analysis with varimax rotation on eleven attitudinal items, using a larger sample from the preliminary survey (N=1,093). We identified four attitudinal factors with high internal consistency: PT autonomy, PT excitement, PT privacy, Pro-environmental intentions.

Preliminary results:

Across the 104 DTD PT journeys, 24% had a transfer, 86.5% went as expected, and the average trip duration and distance were 74 minutes and 38 kilometres. Travellers were generally highly satisfied (mean: 4.52/5), with high perceived safety and comfort and minimal difficulty at or around stations.

Our preliminary correlation analysis suggests that, among the experience variables, in-vehicle comfort has a strong correlation with satisfaction (min: 0.54, avg: 0.58). For perceived safety, in-vehicle (min: 0.26, avg: 0.32) and overall journey ratings (min: 0.32, avg: 0.27) were most strongly correlated with satisfaction. For difficulty, egress (-0.30) and entire journey ratings (min: -0.33, avg: -0.31) showed the strongest correlation. These results suggest that for each experience variable, a different trip leg might be more influential on overall trip satisfaction. We will test these differences systematically in the SEM analysis.

We also observed strong correlations between perceived safety and difficulty at and around stations, likely reflecting the built environment conditions that affect both safety perceptions and ease of access. Interestingly, correlations between trip characteristics or PT attitudes and satisfaction were weak.

Future work:

Our future work will develop an SEM model explaining trip satisfaction from a DTD perspective. By June 2026, we expect to have over 1000 app users, securing a solid data foundation for this analysis. In addition to the variables presented here, we will incorporate built environment features of observed stations and access/egress routes. To achieve this, we are creating a digital twin of travel environments in Denmark using image recognition methods. The findings will inform policymakers aiming to create safe, seamless, and inclusive DTD PT journeys.

ABSTRACT. The global rise in urban population and vehicle use has led to serious challenges such as traffic congestion, long commute times, and rising emissions. In response, many countries are turning to high-capacity, high-speed rail systems as sustainable solutions for urban mobility. These systems aim to reduce travel time, improve regional access, and shift people away from car dependence. South Korea has also been addressing these urban mobility challenges through various policy and infrastructure initiatives, one of which is the launch of the GTX (Great Train eXpress) project-a national effort to enhance transportation efficiency and equity in the Seoul metropolitan area. The first line of the network, GTX-A, began operation in 2024, connecting Seoul’s city center with major suburban hubs at speeds of up to 180 km/h. It was expected to significantly reduce commuting times and help distribute urban growth more evenly across the region. While the GTX-A line has been successfully implemented in terms of technology and operations, there remains a need to better understand how users feel about the service and how satisfied they are with various aspects of it. In particular, there is still limited understanding of how satisfied passengers are, how they perceive the service’s value, and whether GTX-A has had any effect on their daily commuting behavior. Since public transportation is increasingly evaluated based on how well it meets users’ needs and expectations, understanding user experience is essential for improving and expanding such services. This study focuses on this gap by analyzing actual user feedback from the first year of GTX-A operation. To address this objective, we used data collected in 2024 through the National Public Transportation Survey conducted by the Korea Transportation Safety Authority. The survey specifically targeted GTX-A users and collected responses across five main topics: travel characteristics, changes in public transit usage, satisfaction with service attributes, overall satisfaction with GTX-A, and socioeconomic information. To identify which aspects of service most influence overall satisfaction, a factor analysis was first applied to the 28 service-related items. This analysis grouped the items into three main factors that summarize key areas of the user experience and represent higher-level dimensions of satisfaction. These factors were then used as explanatory variables in an Ordered Logit Model, with overall satisfaction as the dependent variable. This modeling approach is suitable because the dependent variable is ordinal, meaning that the numerical ratings reflect ordered categories rather than continuous values. The first factor, Service Reliability and Safety, included user perceptions about how regularly the trains run, the absence of unexpected delays, and the adequacy of safety features and emergency facilities. It also reflected evaluations of whether actual travel time was consistent with expectations and whether the system minimized issues such as noise or vibration. The second factor, Transfer and Accessibility Convenience, reflected how easy it is to navigate stations, the clarity of transfer routes, the availability of bus routes around stations, the adequacy of real-time transfer information, and whether transfer time felt appropriate relative to the entire trip. The third factor, In-Station and In-Train Service Quality, captured aspects such as the clarity of travel information, the comfort and cleanliness of facilities, the onboard environment, and whether users perceived the fare to be fair and reasonable. The ordered logit analysis showed that all three factors significantly affected users’ overall satisfaction. Users who rated any of these service areas more positively were more likely to give higher overall satisfaction scores. The findings clearly indicate that reliability, accessibility, and service quality each play meaningful roles in shaping users’ evaluations of GTX-A. Beyond service-related factors, the analysis also found that individual characteristics played an important role in shaping satisfaction. These characteristics included trip purpose, frequency of use, transfer experience, car ownership, and income. Users commuting for work or school tended to report lower satisfaction, likely because they were more exposed to peak-hour crowding. Those who used the service more frequently reported higher satisfaction, possibly because they became more familiar and comfortable with the system over time. Users who did not need to transfer were more satisfied than those who did, which may reflect the deep underground design of GTX stations that often makes transferring more difficult and time-consuming. People without private vehicles showed higher satisfaction, possibly because they rely more heavily on public transportation and therefore perceive greater benefit from GTX. Users with higher incomes reported greater satisfaction, perhaps because premium rail services like GTX align better with their expectations for speed, comfort, and reliability. These findings demonstrate that both service quality and user context shape how people experience GTX-A. Evaluating large-scale transportation projects should therefore consider more than just operational metrics such as speed or punctuality. It must also include how the service feels and works for different groups of users with different needs and travel behaviors. By identifying the three core areas of service that matter most to users—reliability and safety, transfer and accessibility, and overall service quality—this study offers a useful framework for setting future priorities in GTX operations and investment. The strong effect of the transfer and accessibility factor highlights a key weakness in the current design: even if the main line is fast and modern, users can still be discouraged by poor last-mile connections, long walking distances, or complex station layouts. Improving these aspects is essential to achieving the full potential of the GTX network. The role of individual characteristics also suggests that a one-size-fits-all approach may not be effective. For example, commuters who transfer frequently may benefit from clearer guidance, shorter walking distances, improved signage, or station-level design changes. Providing tailored strategies could enhance satisfaction for specific user groups. As South Korea continues to build out the GTX system, the early lessons learned from GTX-A provide valuable guidance. Addressing pain points in accessibility, transfer convenience, and service reliability should be prioritized. Ultimately, the success of the GTX network will depend not just on its speed or scale, but on whether it consistently delivers a reliable, convenient, and satisfying daily commute for all who use it.

ABSTRACT. Transit agencies have increasingly recognized the value of customer feedback as a rich, near-real-time source of operational intelligence. However, this unstructured data is historically laborious to analyze and glean insights from. This study demonstrates how a Large Language Model (LLM) transforms GO Transit Customer Satisfaction (CSAT) survey comments into clear, actionable insights. Using Databricks’ built in AI functions, the developed pipeline tags over 89 categories of comments, identifies mode of transport, conducts aspect based sentiment analysis (ABSA) including recognizing sarcasm, and the location the comment refers to (station, stop, corridor etc.). The CSAT AI tool operates within the Metrolinx’s governance boundary, ensuring auditability, reliability and security. At current scale of approximately 1,000 respondents per month, end to end processing completes in two hours, replacing 20 hours of manual work, while eliminating sampling bias, and adding granular route/stop insight. Accuracy is reported against historical human-labelled CSAT survey comments and cross-checked with other business metrics such as on time performance and incident logs. Aggregated outputs feed station and route level dashboards enable tracking of hotspots and trends for planning. The results demonstrate a practical, reproducible path to turn qualitative CSAT feedback into operational intelligence, while complying with stringent public sector privacy and AI policies.

ABSTRACT. Transit overloads are caused by excessive boardings, which fill transit vehicles to, or past, capacity, and prevent would-be riders from boarding public transportation. Minimizing overloads is a key concern for transit agencies as they strive to provide a positive customer experience and reliable service. This concern is particularly acute in Halifax, NS, where the population has grown rapidly since the COVID-19 pandemic (the city now boasts a population of over 500,000), and the transit system has struggled to keep up with an increased demand for its services. This has led to a far larger number of overloads than the city experienced pre-pandemic and called for urgent action from Halifax Transit staff. AVL systems are installed on all Halifax Transit vehicles, automatically returning information on position several times per minute. Traditionally, this data has been used to track buses in real time, allowing transit operations staff to reroute vehicles around possible obstacles on-the-fly. This data has also allowed for dynamic decisions to be made on when and where to activate replacement buses, by giving insight on when routes fall far behind schedule or have been stuck at a standstill for some time. We used transit vehicle AVL data to identify the start and end locations of overloads throughout Halifax to enhance service planning and performance management. This was done by accessing the automatically collected GPS coordinates from all buses and cross-referencing this data with reported overloads and route information. The processing was performed by an automated script, which also cleaned the data and created GIS assets to demonstrate which routes were overloading and pinpoint where these overloads were occurring and ending. The script’s creation highlighted some of the methodological challenges in using automated data, including verifying data points and high computational costs. However, these challenges were managed with efficient coding, and specific queries to process overloads at times known to be the most troublesome. Halifax Transit staff used the GIS assets to visualize overload data at several different time intervals and recommend changes to resource distribution. The resulting maps highlighted major travel patterns, and trip attractors which were being underserved by existing service. Changes made as a result of this study reduced the number of overloads in the transit system and provides an example which other transit agencies can use to reduce overloads and evaluate overload patterns quickly.

ABSTRACT. Planned special events, such as sporting events, can cause sudden demand surges and disruptions, thereby placing additional strain on urban mobility networks. Assessing their impact on urban transportation systems is essential for improving operations during such events and for informing future transportation planning, ranging from public transportation service adjustments to synergies between different transportation modes. This study proposes a framework for assessing the impact of planned special events that is based exclusively on using automated data sources. In particular, recurring sporting events (baseball, hockey, and basketball) in Washington, DC, US, are considered in the analysis using smart card data, bike-sharing, and ride-hailing trip data. Results indicate that public transportation (metro) accounts for the largest modal share (25%-35%) across all sporting events during the pre- and post-game periods, whereas the share of shared bikes is negligible (<1%). Ride-hailing is more prevalent in areas with lower public transportation accessibility, and varies across sporting events held at the same venue. Overall, this study provides a reproducible, data-driven framework for evaluating the system-wide impacts of planned special events, serving as a solid foundation for future event transportation planning.

ABSTRACT. Managing and scheduling bus movements during the Hajj season presents one of the most complex logistical challenges in the management of large-scale public transport globally. The complexity comes from the massive number of pilgrims, the diverse and time-sensitive nature of their movements between the holy sites, and the geographically constrained environment of Makkah and the surrounding sacred areas. These challenges are exacerbated by the concurrent operations of multiple transportation service providers, the stringent requirements for schedule compliance, and the imperative necessity to sustain traffic flow while minimizing waiting times for both buses and pilgrims. Over the years, the Ministry of Hajj and Umrah, in close coordination with the General Transport Centre and key operational stakeholders, has developed pre-season scheduling and transport strategies to better manage crowd flows, ensure schedule accuracy, reduce network congestion, and improve overall fleet utilisation. The core movement patterns revolve around three primary operations Hajj transportation phases: 1. Pilgrims are carried from their accommodation in Makkah to the ritual sites of Mina and Arafat, requiring dense scheduling and large-scale convoy movements. 2. After the day in Arafat, pilgrims move from Arafat to Muzdalifah within a few nighttime hours, using more than 20,000 buses to ensure timely arrival for overnight rituals. 3. From Muzdalifah, pilgrims are transported to the Grand Mosque in Makkah for key rites, before returning to Mina, with each phase conducted on constrained infrastructure under extreme fleet demand that can reach thousands of buses per hour. Once the assignment of service centres to pilgrim residences is finalised, all relevant data is automatically gathered and centralised using a unified digital platform. This platform integrates multiple data layers, including the number of pilgrims per residence, the operational company and its service centre, bus details data, the designated entry gates to holy sites, and the allocation of camps in Mina and Arafat. This rich dataset serves as the foundation for advanced bus scheduling, allowing precise, real-time, and optimised route generation throughout the transport network. The objective of this study is to introduce an integrated and data-driven framework to improve the distribution and operational management of bus fleets during the Hajj season. The approach combines advanced traffic modelling, spatial analytics, and multi-objective optimisation to enhance both pre-planned scheduling and adaptive real-time operations. It supports both strategic planning and dynamic reallocation in the field of transport resources in response to changing demand or operational interruptions. The proposed framework addresses the transport challenge from three principal technical perspectives. 1. Spatial Clustering of Pilgrim Accommodations: The clustering of residential accommodations is conducted using spatial algorithms, e.g., DBSCAN, OPTICS. The clustering framework incorporates multiple operational and user-centric constraints, such as pilgrims’ preferences, demographic attributes, infrastructure capacity limits, and accessibility to designated entry points. This multifactor approach enables planners to form demand groups and assign operational zones that best reflect the diverse requirements and restrictions embedded in the urban transport network. By considering these heterogeneous constraints, the system not only reduces route overlaps and improves the logical segmentation of transport responsibilities, but also adapts routing and resource allocation to the unique needs and backgrounds of each cluster, fostering a more responsive and efficient operational plan. 2. Automated Route Generation and Scheduling Engine: Building on the spatial clustering output, the system then generates optimized bus routes connecting accommodations to specific destinations for all the movements types. The route generation module is governed by a set of constraints, including maximum trip durations, allowed departure intervals, staging area capacity limits, and predefined access rules for specific routes or zones. This module applies multi-objective optimization algorithms to simultaneously minimize total vehicle hours, reduce wait times, and improve schedule adherence. Aternatives routes are also envisaged in the generation for emergency situations. 3. Real-Time Monitoring and Decision-Support The system provides a real-time monitoring and decision-support module based on continuous bus tracking. Live AVL/GPS data from the bus fleet is streamed into an operational dashboard that visualizes current bus locations, trip progress, schedule adherence, and operational alerts across the Hajj transport network. This enables control-center staff to monitor service levels, detect delays or bottlenecks as they occur, and trigger corrective actions such as rerouting buses, adjusting departure times, or reallocating vehicles between routes. Historical tracking data can also be analyzed to identify recurrent congestion areas and operational patterns, supporting continuous improvement of plans and resource allocation for future Hajj seasons. Key Performance Indicators (KPIs) have been defined to quantitatively measure the performance of the system and the impact of automation on service quality. These include:

- Average trip duration per route - Average waiting time per pilgrim at key stops points (residence, camp, parking) - Vehicle utilization rate - On-time departure and arrival compliance - Schedule deviation metrics - Congestion Monitoring Indicators, These KPIs not only support real-time operations but also serve as post-event assessment tools to guide improvements for future seasons. In conclusion, this research presents a comprehensive, intelligent transportation framework that transforms raw operational data into actionable strategies for large-scale transit coordination. It bridges disciplines including spatial data science, operations research, and intelligent transport systems to create a unified, adaptive platform for fleet scheduling, route optimization, and performance management. The framework has demonstrated measurable benefits in terms of reducing congestion, improving schedule adherence, and enhancing service quality, making it a viable model for future implementations in similarly complex mobility environments.

ABSTRACT. Presenters: Anna Victoria Polski, Ian Thistle, Jay Westreich Preference for presentation: Undecided WMATA (Metro) has a practice of running additional “event trains” before and after events (usually sports games) in the Washington, DC region. These event trains run special patterns and provide additional service in the gaps between regularly scheduled headways to alleviate extreme crowding on platforms and in trains at key stations. With improving scheduled service on Metro’s Green Line as Metro’s service recovered after the COVID-19 pandemic, Operations leadership asked teams at Metro to examine whether additional service was still needed after Washington Nationals’ baseball games, and to understand whether cost savings could be found by running fewer additional trains while recognizing any impacts to customer experience. Using data processed by Metro’s vendor Korbato into our “Trace” model combined with Major League Baseball data, Metro staff estimated Metrorail ridership to and from each baseball game and evaluated the impacts on crowding from running a fixed number of additional trains compared with different options of running fewer trains when ridership was expected to be lower. Staff also evaluated game attendance projections from the Nationals compared with actual attendance to develop methodologies for more accurate projections. Working with Metro Rail Transportation as well as Rail Planning and Scheduling, staff developed a plan to issue recommendations before each Nationals series on how many additional trains were needed to avoid in-train crowding. Staff also recommended the optimal time to dispatch trains given travel times from the yard to the stadium. Staff developed dashboards that could be updated after each game which displayed whether event trains were used, how crowded each train was, and whether trains were dispatched at the optimal time to address the peak level of crowding. Analysis teams worked with Operations to understand considerations made in deviating from these recommendations. Iterative work continues to maintain a path of improvement, including working on recommendations for other large events in the Washington, DC area such as music concerts. This presentation will cover the initial work evaluating potential optimal service levels, examine the real-life considerations that guided Operations’ decisions made during the 2025 baseball season, and provide an evaluation of the value added by this exercise and lessons learned.

ABSTRACT. [Introduction] Increasing presence of automated passenger counting (APC) systems provides new possibilities for (pro)active travel management in public transport (PT) networks. In particular, the APC data can be processed to generate real-time crowding information (RTCI) on-board the PT vehicles and disseminate it to passengers. State-of-the-art studies already provide certain evidence and toolset to demonstrate the RTCI benefits for travel experience and operational efficiency. Methods for generating and disseminating the RTCI from APC data have been also developed, showing the potential to achieve high precision quality. However, these findings often stem from limited-scale, specific case studies. More research is needed to understand the attainable RTCI accuracy on the network scale and for different prediction time horizons.

[Objectives] Our study addresses this gap and leverages on a real-world APC dataset to explore the RTCI prediction characteristics. Specifically, we address the 2 principal research questions: (1.) What is the attainable RTCI prediction accuracy in the PT network, depending on input data availability levels – i.e. historical APC (from previous days) or current-day APC (from previous departures or stops)? (2.) What RTCI accuracy evolution patterns can be distinguished with respect to the prediction time horizon – i.e. emerging risks of crowding under- or overestimation?

[Methodology] Research methodology consists of the following stages: (1.) The RTCI generation (prediction) framework. We formulate the core analytical problem as predicting the load factor (LF), i.e. passenger volume-to-capacity ratio, for a given segment of trip (departure) r from stop s to stop s+1 on day d,, as a function of observed LF values (explanatory features) from previous departures, stops or days. We utilise here the light gradient boosting (LGB) methodology to train and test the RTCI generation models, aiming to minimise the gap between the predicted and actual LF values. (2.) Calculating the RTCI predictions for different time horizons. The RTCI prediction models are evaluated for 5 availability levels of APC data, corresponding to different possible prediction horizons. First, the RTCI based on historical data, i.e. LF values of target departure from previous days (d-1, d-2, d-3). Second, we supplement these with LF values for previous departures of the same day (r-1, r-2, r-3). In the final 3 models, we incorporate also the LF values of approaching target departure from various upstream stops (s-1, s-2, …, s-7). (3.) Classification of RTCI accuracy evolution patterns. We perform the K-means clustering of all the data points (prediction instances) based on differences between predicted and actual LF values, as obtained for the 5 RTCI prediction models. This allows to distinguish main patterns of emerging RTCI prediction (in)accuracy, with respect to (both) prediction error signs and discrepancy ranges.

[Case study] We utilise an extensive APC dataset for the urban PT network of Geneva (Switzerland), provided by TPG (Transports Publics Genevois). Daytime PT supply consists of 5 tram, 73 bus and 6 trolleybus lines. The whole PT vehicle fleet is equipped with APC data collection systems, providing a high-quality and continuous dataset for the RTCI prediction modelling. After discarding the erroneous or incomplete records, the processed input dataset contains ca. 3,044,000 records from 19 working days for the month of October 2023 (excluding first 3 days which are used as “warm-up” period), with complete historical and current-day APC data for testing and evaluating the RTCI prediction models.

[Results] Key findings from our investigation are as follows: (1.) Historical (previous-days) APC data already enables moderate RTCI precision (R2 = 0.68), albeit with significant crowding underestimation risk. Inclusion of current-day APC data for previous departures has very little impact, in contrast to the APC of target departure from upstream stops. The latter allows to substantially improve the RTCI precision in short time horizon of 5 mins (R2 = 0.84) to 10 mins (R2 = 0.77) before the departure of trip r from stop s. (2.) Analysis of feature importance for RTCI predictions: for historical data, the most relevant features are LF values from previous days, followed by hour (time) interval, stop index position and total line demand. For current-day data, the LF value from the nearest upstream stop dominates the remaining features. (3.) Clustering analysis reveals 4 evolution patterns of RTCI accuracy. Approx. 60% of data instances belong to the “accurate” cluster, with closely aligned LF predictions vs. realisations already for the historical APC model. About 17% comprise the “overestimation” cluster, where predicted LF is on average 10 p.% higher than actual realisation but this discrepancy decreases with current-day APC data. The remaining 2 clusters correspond to “minor -” and “major underestimation” patterns. They comprise ca. 20% and 4% of data points respectively, but over 34% and 10% if weighed by passenger numbers. In the latter cluster, mean predicted LF is 25 p.% lower than realisation (for historical data) and improves only in the very short time horizon before target departure. (4.) The most dominant predictor for cluster membership is the actual LF of target departure. First 2 clusters (accurate and overestimated) correspond to no or little crowding instances, i.e. LF < 40%. Clusters with underestimation risks contain mostly higher LF values, up to 60% (minor underestimation) and beyond (major underestimation).

[Conclusions] Our findings demonstrate the feasibility of RTCI predictions, even if relying exclusively on limited historical APC data, especially for lower crowding instances. Prediction quality can be much enhanced with current-day APC data for the target departure from upstream stops, which is especially crucial to mitigate underestimation risks for high-demand PT lines and stops. Study findings show that APC systems can be an instrumental dataset for network-wide RTCI provision, with overall moderate prediction accuracy. This baseline RTCI framework can be extended with further modelling developments, e.g. customised methods for more challenging instances, to ensure more robust and high-precision passenger crowding predictions.

ABSTRACT. Reduction of denied boarding and alleviation of crowding on bus services is one of the major challenges for large cities without high-capacity metros. It is often seen that crowding levels of bus services on the same route are not proportionate to headways because 1) some services pick up many passengers at transferring stops while others do not, 2) service origins are different, etc. Whilst it has been widely known that holding is an effective strategy for bus service regulation, if crowding levels are not proportionate to headways, this strategy may not work. In this work, we have developed a strategy which regulate bus services based on crowding levels of bus services, not headway. The method uses actor critic offline reinforcement learning techniques where the algorithms are trained using real bus GPS and crowding level data of Sofia (Bulgaria) and Bucharest (Romania), rather than being coupled with simulation model (and hence no need to have a simulation model). Evaluation are conducted against the baseline scenarios where the traditional headway-based control is used (and where no control takes place). The algorithms have been developed, and we are now conducting tests, which are scheduled to be complete by the end of 2025 (when the project completes). The results will be presented in the Transit Data 2026.

ABSTRACT. Modern metropolitan areas rely heavily on urban rail networks to facilitate the movement of millions of commuters every day. While these systems provide efficient and sustainable mobility, they face persistent challenges, particularly overcrowding during peak hours and underutilization during off-peak periods. Overcrowding on transit vehicles is not only perceived as longer travel time but also imposes safety and security concerns for commuters. Past studies have explored strategies to minimizing overcrowding during peak hours and maximizing the utilization of the transit system during off-peak hours using strategies, such as collective incentive strategy, dynamic incentive pricing, peak-hour surcharge, or off-peak discounts, hybrid fare scheme, and flexible work start time. Studies have also reported that commuters are willing to change their departure time rather than switch their transportation mode to avoid congestion. The departure time flexibility in literature is modeled using the mixture of uniform and beta distribution for regular commuters. However, studies have not jointly considered departure time flexibility along with congestion pricing in a comprehensive behavioral framework. This study develops a data driven simulation-optimization framework that integrates passenger departure time flexibility and train scheduling to reduce onboard crowding while minimizing congestion-related surcharges in a commuter rail system. An agent-based simulation model is used to model individual passenger choices, where each agent selects departure times and trains based on preferred arrival times within departure time flexibility ranges, congestion-related surcharges, and perceived crowding. Smart card data are utilized to characterize passenger behavior and departure time flexibility distributions, and GTFS data is used to characterize the operational service environment. A multi-objective genetic algorithm (GA) generates alternative combinations of train departure times, service frequencies, and station and time specific congestion-related surcharge levels. For each candidate solution, the agent-based simulation is executed to evaluate both objectives, systemwide peak-hour onboard crowding and total congestion pricing paid by passengers, simultaneously. These paired objective values guide Pareto-based selection, crossover, and mutation in the GA, enabling the search for solutions that reflect realistic trade-offs rather than sequential or single-objective optimization. The resulting Pareto front highlights operational and pricing strategies that jointly balance crowding mitigation and passenger cost burdens. This integrated simulation–optimization approach provides a robust and behaviorally grounded framework for improving demand management and service efficiency in urban rail systems. The framework considers the Beijing Metro network as a case study, beginning with the calibration of the agent-based model and validation against observed smart card passenger flows to ensure the simulation accurately reproduces real-world travel patterns. Once validated, the optimization module identifies operational and pricing strategies that minimize crowding and congestion-related surcharge costs across different peak periods. The resulting set of optimized policies provides transit planners with a flexible tool to evaluate station- and time-specific pricing scenarios and to design targeted interventions for managing peak-hour demand more effectively.

ABSTRACT. Submitted for in-person presentation at Transit Data 2026, Toronto, ON Canada.

Light rail transit (LRT) investments are often promoted as catalysts for shifting travel behavior, improving accessibility, and reducing automobile dependence. Yet, the success of such investments depends not only on the quality of service delivered but on how different segments of the population respond to the new infrastructure. Understanding who adopts new systems, who does not, and how expectations translate into actual behavior is vital for planning effective transit networks. Most studies investigating the impacts of new LRT infrastructures on the transit market rely on cross-sectional snapshots taken before or after implementation; consequently, little is known about how market composition and individual travel patterns evolve once new infrastructure becomes operational. This study fills that gap by examining the introduction of Montréal’s Réseau express métropolitain (REM), a new automated light rail system, and analyzing how transit market segments shifted before and after its launch through a combination of cross-sectional and panel data. Specifically, the study addresses three research questions:

• To what extent do user profiles remain stable over time, and how do their defining characteristics evolve as the new infrastructure becomes operational? • How do individuals transition between user segments over time, particularly with the opening of the REM? • How do respondents stated pre-launch intentions to use the REM align with their actual ridership behaviors once the system becomes operational?

This study draws on the Montréal Mobility Survey (MMS), a multi-wave, bilingual, online longitudinal survey administered by the Transportation Research at McGill group across the Greater Montréal region. This study focuses on respondents residing in the catchments area of the REM’s first operational branch, including the South Shore, Nun’s Island, and a 2 km buffer around Gare Centrale station in downtown Montréal. Our analysis draws on two key MMS waves: Wave 3 (2022), the most recent pre-launch survey, and Wave 5 (2024), the first post-launch survey. Wave 3 captures baseline travel behavior, intended REM use, and project perceptions prior to opening, while Wave 5 reflects actual ridership and travel patterns one year after the South Shore branch became operational. Within the study area, 623 valid responses were retained in 2022 and 1,645 in 2024. A longitudinal subsample of 175 respondents completed both waves, allowing for direct comparisons between pre-launch intentions and post-launch behavior. All survey samples were weighted to reflect age, gender, income, and regional mode share distributions from the 2021 census. To identify the composition of the transit market along the South Shore branch, we employed a two-stage approach to market segmentation. First, exploratory factor analysis identified latent dimensions related to REM perceptions, travel behavior and mobility preferences, gentrification concerns, income and telecommuting frequency. Second, weighted k-means clustering grouped respondents based on the identified factor structure. Four clusters consistently emerged in both waves, allowing direct comparison. To assess behavioral dynamics at the individual-level, we constructed survey-weighted transition flows from respondents’ original 2022 cluster assignments to their 2024 counterparts. Finally, we cross-tabulated pre-launch intentions to use the REM with post-launch revealed ridership to examine intention–behavior alignment. A cluster solution of four profiles was found to provide the best qualitative description of the market at both points in time. In the pre-implementation period (2022), the clusters were identified as potential REM adopters, potential REM telecommuters, car-oriented individuals, and low-income individuals. Following implementation (2024), a similar segmentation structure emerged. The clusters were characterized as frequent REM riders, telecommuter REM riders, car-oriented individuals, and low-income individuals. While modal behaviors and demographic compositions shifted slightly within certain groups, the overall structure of the market remained consistent, with no new clusters emerging in the post-launch period. Nonetheless, the longitudinal analysis reveals more fluidity at the individual level than aggregate stability would suggest. Potential REM adopters displayed the strongest alignment between intention and behavior, with over 70% transitioning into frequent REM riders after launch. In contrast, potential REM telecommuters were far more fluid, with fewer than 40% becoming telecommuter REM riders, while nearly half reverted toward car-oriented profiles. This pattern highlights how post-pandemic adjustments in work routines translated into divergent travel outcomes, sustaining telecommuting for some while reinforcing auto reliance for others. Car-oriented individuals remained the most resistant to change, with a majority staying in the same cluster and a minority consistently adopting the REM, reflecting the difficulty of shifting entrenched driving habits. Low-income individuals also showed stability, with small increases in transit use but continued barriers to REM adoption. Overall, these transitions illustrate a duality of stable markets but shifting individuals: market profiles persist as meaningful categories, yet individuals within them adjust in response to new opportunities and constraints. REM adoption has been constrained by two intertwined forces: shifting work patterns and spatial misalignment. Many residents who intended to ride before launch telecommute several days a week and use the line only occasionally, while others appear to have returned to full-time on-site jobs in places the REM does not serve well shrinking the pool of daily commute trips the REM can capture. These labor dynamics meet a structural barrier as the branch primarily targets downtown; for suburban travelers whose jobs, schools, and errands are elsewhere, distance to stations and limited geographic coverage reduced uptake despite mostly favorable attitudes. Early recurring technical and weather-related disruptions likely further decreased demand among riders with travel alternatives. The segmentation results highlight that maximizing the benefits of the REM requires a combination of targeted and overlapping policy strategies. Some interventions, such as strengthening feeder bus networks, improving last-mile access, and ensuring reliable operations, would benefit multiple groups simultaneously, from frequent riders who need dependable service to low-income and car-oriented residents who face distance barriers. At the same time, other strategies must be tailored to specific market segments, such as flexible fare options for telecommuters. By grounding policy action in the distinct needs and barriers of each group, while recognizing the overlaps across them, planners can move beyond one-size-fits-all approaches.

ABSTRACT. Urban rail transit systems exhibit inherently complex passenger flow patterns shaped not only by large-scale origin-destination (OD) trips but also by the cumulative effects of numerous OD pairs with statistically small-scale trips (hereafter termed SmallODs). While existing passenger flow assignment models have predominantly focused on major passenger movements, the role of SmallODs, which can account for 25-50% of all OD pairs and 10-30% of total passenger volume, remains poorly understood. This research addresses this critical gap by developing a comprehensive framework to analyze the impact mechanisms of SmallODs on network-wide passenger flow distribution using the Nanchang Metro in China as a case study. We define SmallODs using a statistically-grounded threshold based on confidence interval theory, considering route choice probability distributions rather than arbitrary passenger count cutoffs. Our methodology employs a three-scenario comparative analysis framework to examine flow variations under different behavioral assumptions. Scenario 1 establishes a baseline using field-adopted proportions currently employed for fare clearing operations. Scenario 2 assigns all SmallODs to shortest paths while maintaining field-adopted proportions for remaining flows. Scenario 3 explores the opposite extreme, routing SmallODs through longest reasonable paths. This approach creates a comprehensive envelope of potential impacts, revealing network sensitivities that would otherwise remain hidden. To quantify spatial and temporal variations across scenarios, we develop four complementary indicators. The Mean Relative Difference (MRD) captures normalized spatial flow variations, enabling fair comparisons across sections with different baseline volumes. The Hourly Absolute Difference (HAD) measures temporal flow magnitude changes throughout operational hours. The Peak Misalignment Rate (PMR) identifies temporal inconsistencies in peak-hour definitions across network sections. The Misalignment Impact Flow (MIF) quantifies the passenger volumes affected by these temporal misalignments. Our spatial analysis reveals three categories of vulnerable network sections particularly sensitive to SmallOD behaviors. Core hub areas consistently exhibit high flow variations (MRD > 0.15), indicating their susceptibility to cumulative effects of SmallOD routing decisions. Sections adjacent to transfer stations show elevated sensitivity due to the complex interaction of multiple lines and passenger flow convergence. Corridors connecting suburban areas to city centers demonstrate significant fluctuations, reflecting the aggregation of diverse trip patterns. Network vulnerability maps generated through our analysis provide actionable insights for targeted infrastructure investments and operational improvements. Temporal analysis uncovers substantial peak-hour misalignments challenging conventional network management. Approximately 25% of network sections experience peak-hour misalignment, affecting roughly 20% of total passenger flow. Downstream direction exhibits more severe misalignments than upstream, with both the number of affected sections and network-wide misalignment rates showing significant directional asymmetry. Line-specific analysis reveals heterogeneous impacts, while Line 2 experiencing misalignment rates approximately double those of other lines in both directions, suggesting line-specific operational challenges. Surprisingly, off-peak periods demonstrate the highest variability in both mean values and variance, surpassing peak-hour measurements. Off-peak MRD values show 30% higher average impacts compared to peak hours, while maintaining comparable fluctuation levels. This counterintuitive finding suggests that the absence of dominant flow patterns during off-peak hours amplifies the relative impact of SmallOD routing decisions. The sectional time window analysis, conducted at 10-minute intervals, demonstrates that sections experiencing misalignment cluster temporally around traditional peak periods (07:30-07:45 and 17:25-17:35), yet specific timing varies significantly across sections, indicating that network-wide peak definitions may inadequately capture section-specific demand patterns. The integration of spatial and temporal analyses reveals strong correlations between network topology and temporal stability. Transfer stations and their adjacent sections show 40% higher temporal variability compared to linear sections. While individual SmallODs contribute minimal passenger volumes, their aggregated effect at critical network locations can form substantial flows affecting system performance, particularly pronounced during peak-off-peak transitions. Our findings have significant implications for transit planning and operations. The identification of vulnerable network sections enables targeted capacity management strategies that account for flow uncertainties introduced by SmallODs. The temporal misalignment patterns suggest the need for section-specific operational schedules rather than network-wide uniform timetables. The unexpected instability during off-peak periods indicates opportunities for dynamic pricing or service adjustments to better manage SmallOD-induced variations. This research contributes to transit data analytics by providing a reproducible framework for quantifying the previously overlooked impacts of SmallODs. The multi-scenario approach offers a robust methodology for bracketing potential flow variations, essential for risk-aware network planning. The developed indicators provide transit agencies with practical tools for monitoring and managing flow uncertainties. By revealing how SmallODs collectively influence network performance, this study challenges the current paradigm of focusing primarily on high-volume flows and demonstrates the necessity of considering the full spectrum of passenger movements in urban rail transit systems. Future research directions include developing predictive models that incorporate SmallOD volatility, designing adaptive routing algorithms that respond to real-time flow variations, and extending this framework to multi-modal transit networks where SmallODs may have even more complex aggregation patterns.

ABSTRACT. 1. Introduction The introduction of large-scale transportation infrastructure such as the Great Train eXpress-A (GTX-A) corridor triggers deep and system-wide modifications in metropolitan mobility. While conventional policy evaluations emphasize observable outcomes—ridership gains, average travel-time reductions, and modal shifts—these metrics capture only behavioral responses, not the structural reconfiguration of the underlying connectivity network. A district may display minimal change in observed performance while experiencing substantial alterations in how it participates in the broader mobility system: becoming more central, more peripheral, or more strategically connected. These latent shifts in connectivity roles determine long-term impacts on urban equity, resilience, travel demand redistribution, and land-use dynamics. Yet they remain unmeasured in most before–after analyses because dynamic OD networks are high-dimensional, dense, and temporally unstable. Traditional spatio-temporal GNNs are unsuitable due to over-smoothing and spectral instability, which obscure or distort structural comparisons across time. This study addresses the gap by developing a stable and interpretable structural-embedding framework designed to detect and quantify latent structural changes in OD networks surrounding the opening of GTX-A. The approach explicitly separates intrinsic node identity from evolving flow dynamics, enabling rigorous evidence-based evaluation of how districts reposition within the metropolitan system.

2. Data and Problem Definition The analysis draws on high-resolution OD movement data aggregated at the legal district (Dong) level across more than two years. Each day consists of multiple 120-minute bins, with OD flows enriched by travel-time metrics, transfer counts, and modal-share breakdowns. Across the entire period, the union of all origin/destination occurrences forms a unified node set, ensuring stable temporal alignment even when districts appear or disappear. Each time bin is encoded as a directed, weighted OD graph: · Nodes: legal districts (fixed index) · Edges: OD flows with multiple attributes · Node features: inflow/outflow intensity, district-level aggregated travel times, modal-share diversity The objective is not route optimization nor prediction of a single variable, but rather: learning a time-indexed embedding that encodes each district’s structural role in the metropolitan mobility system and quantifying how that role changes around the GTX-A opening. We frame this as a dynamic structural-embedding problem, where embedding trajectories capture gradual or abrupt reconfigurations of the city's connectivity patterns.

3. Methodology To identify structural shifts reliably—without smoothing biases or spectral instabilities common in spatio-temporal GNNs—this study designs a modular pipeline with four key components. 3.1 Node Alignment and Graph Construction · All nodes are assigned stable indices using the union of all OD occurrences across time. · Each txt file is decomposed into multiple time-bin graphs, enabling a uniform temporal grid. · Edges consist of OD flows, and node features are aggregated from outgoing and incoming statistics within each time bin. This produces a consistent, dense sequence of graphs suitable for temporal modeling. 3.2 Structural Encoding (GNN-Free) Dynamic graphs complicate spectral methods: Laplacian eigenvectors shift, flip signs, or reorder under minor structural perturbations. To avoid these well-known instabilities, we employ Random Walk Structural Encoding (RWSE) computed from a pre-opening baseline graph. RWSE produces walk-return probabilities across multiple hop lengths, encoding: · local redundancy, · hubness and structural centrality, · multi-scale connectivity. Because RWSE is computed once and reused across all time steps, embeddings are anchored to a stable structural coordinate system, enabling meaningful pre–post comparison. 3.3 Edge-Aware Graph Transformer Instead of GCN-based smoothing—which obscures role differentiation—the spatial encoder employs a lightweight Graph Transformer. Edge attributes such as travel-time, intermodal transfer costs, and modal-share asymmetries modulate attention weights, ensuring that: · strengthened corridors receive higher propagation weights, · weakened or bypassed connections diminish in structural importance, · spillover effects (e.g., secondary accessibility gains) are learned organically. This selective propagation is essential for detecting the nuanced and uneven spatial effects characteristic of major rail openings. 3.4 Temporal Transformer Embedding sequences for each district are passed through a Temporal Transformer that captures long-range temporal dependencies, seasonal variations, and transitions around the policy intervention. Temporal attention allows the model to learn: · persistence or volatility in district-level structural roles, · structural breakpoints corresponding to the opening event, · lagged responses such as gradually intensifying shift patterns. The resulting embeddings provide a continuous structural trajectory for every district.

4. Results The GTX-A opening triggers substantial structural reorganization across the OD network, far beyond what is apparent from direct travel-time indicators: Key structural findings include: · Influence-zone amplification: Districts surrounding GTX stations exhibit embedding shifts several times larger than placebo regions, reflecting expanded reach and more central positions within the OD system. · Redistribution of structural centrality: Certain southern and eastern suburban districts experience marked increases in structural prominence despite modest changes in observed trip flows. · Reconfiguration of structural neighborhoods: Many districts change 30–50% of their top structural neighbors, signaling new connectivity affinities and shifting travel corridors. · Emergence of secondary hubs: Several intermediate districts gain structural importance due to increased through-flows and new potential for multi-modal integration. · Stability checks: Placebo regions remote from the GTX corridor show negligible change, validating model specificity and ruling out confounding seasonal variation. These findings reveal latent and spatially heterogeneous impacts that conventional before-and-after measures fail to capture. These findings reveal latent and spatially heterogeneous impacts that conventional before-and-after measures fail to capture.

5. Discussion This study demonstrates that structural-embedding analysis offers a distinct and increasingly indispensable perspective for transport policy evaluation. Three implications stand out: (1) Beyond Performance Metrics: Capturing System-Level Change Traditional evaluation metrics cannot represent how the network itself reorganizes. As cities adopt higher-capacity and high-speed systems, structural effects—changes in strategic reach, corridor dominance, and network hierarchy—will increasingly determine the long-term success of projects. (2) Detecting Emerging Spatial Inequalities Structural role shifts often precede observable travel demand changes. Districts that gain or lose connectivity relevance can be identified early, providing planners with forward-looking indicators of equity and development pressure. (3) Informing Next-Generation Network Design Structural embeddings expose how new lines interact with existing bus and rail subsystems, revealing: · unexpected bottleneck transfers, · corridors with emergent strategic value, · areas where complementary investments would maximize system efficiency. Such evidence supports integrated decision-making across modes and scales.

6. Conclusion The proposed framework introduces a robust approach to uncovering latent structural transitions in dynamic OD networks. By integrating RWSE-based structural anchoring, edge-aware spatial attention, and temporal sequence modeling, it provides a fine-grained and interpretable assessment of how large-scale transit investments reshape urban connectivity. Applied to the GTX-A opening, the method uncovers substantial, spatially uneven structural reorganization—patterns that conventional evaluation metrics overlook. As transportation networks grow more multimodal, interconnected, and data-rich, structural-embedding approaches will become crucial tools for diagnosing system-wide impacts, anticipating future demand, and guiding equitable and resilient transport planning.

ABSTRACT. Background and Objectives

In large metro systems, the internal spatial complexity of transfer stations, together with signage quality, visibility constraints, and multi-level layouts, has a substantial impact on passenger navigation efficiency, especially for visitors and infrequent riders. Unfamiliar passengers often struggle to find their way, causing delays for themselves and contributing to crowding for others. Even regular riders may unintentionally choose suboptimal routes that concentrate flows and worsen bottlenecks. However, transit agencies currently lack scalable and data-driven methods to evaluate wayfinding conditions and to assess the effectiveness of design measures intended to reduce confusion, delay, chokepoints, and peak-period capacity inefficiencies. The Washington Metropolitan Area Transit Authority (WMATA) is actively implementing improvements and systemwide design updates. The deployment of these interventions offers a natural experiment to assess how design changes influence passenger circulation and wayfinding outcomes, thereby providing valuable support for future operational and planning efforts.

This study aims to develop a data-driven framework for the WMATA to understand how station design, passenger behavior, and wayfinding interventions jointly shape navigation efficiency in metro stations, with a particular focus on the more complex transfer stations. Our research focuses on three key dimensions: (1) identifying when and where existing passenger movement patterns generate inefficiencies within multi-level transfer stations, (2) examining whether spatial and visual characteristics of stations contribute to passengers' vertical circulation time and overall path choice; and (3) assessing whether recent wayfinding interventions have improved navigation success and reduced unnecessary circulation. A deeper understanding of these mechanisms is essential for transit agencies aiming to enhance wayfinding performance at complex transfer stations and for planners seeking to implement design interventions that reduce congestion, improve circulation efficiency, and create a more seamless passenger experience.

Methods

Our analysis integrates a multi-source dataset that combines structural, behavioral, and visual information from the WMATA system. First, GTFS Pathways provides detailed representations of station topology, including corridors, walkways, stairs, elevators, and vertical connections. Second, Automated Fare Collection (AFC) records and Origin, Destination, and Transfers (ODX) data offer detailed patterns of passenger use of entrances, mezzanines, faregates, platforms, inferred route and rail vehicle assignment, and vertical circulation elements across the WMATA network. Third, vision-based data to enrich the structural and behavioral layers: including WMATA Station View (an interactive virtual tour tool provided by the agency) and Google Street View imagery supply fine-grained visual cues such as signage, orientation markers, and visibility constraints, while CCTV feeds processed with computer vision provide observations on the distribution and movement of passengers across elevators, escalators, stairs, and platforms. Together, these data sources enable a more comprehensive and spatially resolved understanding of both station environments and passenger navigation behavior.

We construct passenger wayfinding behavior measures by modeling vertical circulation and path choices using ODX data and validating these patterns against observed flows to ensure that the inferred movements provide realistic and representative approximations of rider experience. CCTV feeds processed with computer vision provide direct observations of passenger movements and travel times within transfer stations, offering ground-truth benchmarks for validating and calibrating the vertical circulation flows and times inferred from ODX. To characterize the visual environment within stations, we extract visual features from WMATA’s in-station image inventory and Street View imagery using computer vision tools, including measures of visual complexity, semantic segmentation–based visibility metrics, and signage quality and density. We then align each interior viewing position with its corresponding GTFS Pathways graph node, enabling every node in the multi-level station graph to be assigned a visual feature vector. Finally, we will analyze how these spatial and visual characteristics relate to passenger behavior and evaluate these relationships through a pre–post assessment of recent WMATA wayfinding interventions, treating the staged rollout as a natural experiment.

Anticipated Results

Reconstructing WMATA station topology from GTFS Pathways and conducting exploratory analysis of ODX data will help identify recurring patterns of how passengers use vertical connections across stations and time periods. The availability of interior Station View imagery further enables the incorporation of visibility conditions, visual complexity, and signage information into station-level analysis once processed through computer vision tools. Integrating these structural, behavioral, and visual components within a unified multi-level station graph is expected to identify where complex layouts, constrained sightlines, or limited signage contribute to inefficient circulation and suboptimal path choices. Leveraging the staged rollout of WMATA’s wayfinding interventions as a natural experiment, we also anticipate detecting measurable changes in vertical circulation time and the distribution of flows across elevators, escalators, and stairs. These results are expected to highlight which spatial or visual characteristics most strongly influence navigation efficiency and where targeted guidance or infrastructure adjustments could most effectively improve station navigation and passenger experience.

Potential Contributions

This study will contribute a multi-source analytical framework that links station geometry, visual features, and passenger behavior to in-station navigation outcomes. The results are expected to clarify the mechanisms that produce inefficient circulation, provide empirical evidence on the effectiveness of wayfinding interventions, and offer generalizable methods for evaluating guidance and infrastructure strategies in complex public transit stations.

ABSTRACT. Presenter Preference: In-Person

Canadian transit systems are evolving, with the opening of new lines, overhauls to networks, and renovation of terminals. However, these improvements often necessitate changes to the flow of passengers as they navigate their commute. This transition from familiar to unfamiliar routing can increase journey times and stress as passengers re-learn their routes through their terminals and transit systems. These aspects can have operational, safety, and rider perception impacts, but are challenging to capture and incorporate into crowd models. The YorkU Fire Research Team’s Human Factors Lab is quantifying these effects using LiDAR and real-world field studies.

A portable LiDAR-Based Crowdflow monitoring system has been developed for automated, long-term collection and analysis of pedestrian behavior and crowdflow in large structures such as transit terminals. The LiDAR methodology allows for micro-scale quantifications across hours, days, and weeks, which may also have additional applications in evaluating and validating changes in passenger demand, operational modifications, wayfinding updates, or emergency situations.

This system was deployed at a transit station’s concourse to evaluate the crowdflow impacts following the partial relocation of buses to a new terminal. Sensors were placed to cover multiple entry points from a rail platform and multiple exits to the old and new bus terminals. Following an approved ethics protocol, data collection occurred across 11 weeks including, 2 changes, and 1 holiday period. The system was able to observe changes to origin-destination patterns, movement speeds, crowd densities, flow rates, and other behaviors including pausing or running. Of particular interest is the increase in passenger volumes to and from the older terminal, which peaked following the change in station design before subsiding over subsequent weeks. Aside from the first day of the change, the greatest impacts were seen amongst the commuter demographic. Lessons learned from initial system deployment and additional results from currently ongoing studies may also be discussed.

The results can be attributed to the theories of familiarity and stickiness, indicating that there are temporal and demographic elements to consider for planning, modelling, and implementing terminal changes. The level of familiarity with a station’s old layout can result in observable deviations in behavior from assumed crowd modelling parameters. By understanding the impact of familiarity, engineers, planners, and operators can plan interventions that allow passengers to more easily adapt to new and modified station layouts for a smoother, safer transition experience.

ABSTRACT. The dynamic and interdependent nature of urban rail transit systems poses significant operational challenges. Passenger demand, vehicle capacity, and service schedules continuously evolve over time and interact with one another, leading to phenomena such as crowding and denied boarding. These internal dynamics are further influenced by external processes, including maintenance closures and emergency disruptions, which affect passenger movement and service performance. Capturing such complex and interdependent dynamics requires a modeling framework capable of representing system-wide interactions at fine temporal and spatial resolutions. Dynamic simulation with explicit representation of individual passengers, service runs, and infrastructure therefore provides a powerful tool for analyzing these phenomena.

In this study, we present an open-source integrAted dAta-driven Agent-based siMulation for the urban rail transit systems, referred to as the AAAM. The model operates at a mesoscopic spatio-temporal granularity, enabling dynamic updates of occupancy levels on platforms and in vehicles. Individual passengers are explicitly modeled as agents, allowing traveler-level decision-making and behavioral heterogeneity to be incorporated. The main characteristics of the AAAM are described next and a benchmark case study is presented of overcrowding simulation on a busy section of the Beijing Subway.

Input: The AAAM utilizes four categories of inputs, all of which can be sourced from standard transit operation records. Demand inputs capture passenger-level travel information, including trip origin, destination, and departure time. These data can be obtained from smart card tap-in and tap-out records or from travel surveys. Supply inputs describe transit operations, such as vehicle schedules and train capacities. Infrastructure inputs characterize the physical components of the transit system, including platform size and capacity, while the network inputs define the connectivity among stations, platforms, and lines. The supply, infrastructure, and network inputs can be collectively derived from the General Transit Feed Specification (GTFS), which is widely available for many urban transit systems.

System construction and initialization: The transit system is represented as a network graph, where nodes correspond to stations and platforms, and edges represent either train service links or pedestrian walking links between stations and platforms. These network elements can be constructed automatically from GTFS inputs. The simulation includes three types of agents: passenger agents, service run (vehicle) agents, and platform agents, all embedded within the network environment. During initialization, passenger agents are assigned a “pre-trip” status, while passenger counts on all platforms and service runs are set to zero, representing an uncongested initial system state.

Temporal updates: The simulation advances in discrete time steps, with the temporal resolution specified by the user. A time interval of 5–20 seconds is generally recommended to capture key operational events such as vehicle dwell times, passenger boarding, and alighting. Service run agents update their states based on scheduled operations, alternating between dwelling at platforms and traveling between stops. Passenger agents progress along their assigned routes (e.g., dynamically updated fastest paths), transitioning through a series of states including “pre-trip”, “walking”, “waiting”, “aboard”, “transfer”, and “arrival”. When a vehicle reaches its capacity limit, additional passengers experience denied boarding and must wait for subsequent trains.

Outputs: The model produces time-stepped outputs capturing the dynamic states of train, passenger, and platform agents at each simulation step. These outputs enable the computation of passenger volumes as well as crowding status at individual platforms, along line segments, and at the system-wide level.

Benchmark case: Beijing Subway network A benchmark case study is conducted to demonstrate the usage of the AAAM by estimating the effective operational capacity of Beijing Subway Line 6 segment using smart card tap-in and tap-out data. Given the heavy reliance on the subway for commuting and the subsequent overcrowding, actual occupancy levels are believed to exceed designed capacity during the peak hours while the walking speeds to reduce. The AAAM is used to demonstrate the quantification of these operating phenomena. Three passenger routing strategies are evaluated for the full network: shortest dynamic travel time, shortest fixed (scheduled) travel time, and minimum number of stops. Validation results indicate that routing based on shortest dynamic travel time provides the most accurate representation of observed passenger flows. To benchmark peak-hour capacity and passenger walking times, a series of simulation scenarios are tested on the eastern section of Beijing Subway Line 6, comprising 14 stations. Five train capacity scenarios and four walking-time scenarios are evaluated. The capacity scenarios include the design capacity of 1,960 passengers per service run (p/s), three higher capacities (2,100, 2,300, and 2,500 p/s), and a scenario with effectively unlimited capacity. Walking times between stations and platforms are set to 30, 50, 70, and 90 seconds. For each scenario combination, simulated passenger journey times are compared with observed journey times derived from smart card tap-in and tap-out data. The results indicate that a maximum occupancy of 2,300 p/s, combined with an average walking time of 70 seconds during peak periods and 50 seconds during off-peak periods, provides the closest match to observed travel times. This implies that peak-period train occupancy exceeds the design capacity by approximately 17.34%, while walking times between stations and platforms increase by about 40%. Together, these findings highlight the impact of overcrowding on both train loading and passenger movement in the subway system, and the ability to capture these variations by the AAAM.

Summary: This study presents one of the first integrated, data-driven agent-based modeling frameworks (AAAM) for simulating individual-level behavior and dynamic operations in urban rail transit networks. By combining multi-source operational data with agent-based mesoscopic simulation, the AAAM captures time-dependent vehicle loading, platform crowding, and passenger interactions with transit services. Apart from simulation-based performance evaluation, the AAAM enables the estimation of real-world operational characteristics such as peak-hour train occupancy and station walking times. The flexible structure can be leveraged to support actionable decisions for short-term operations and long-term capacity planning. The framework is extensible to the analysis of passenger behavioral dynamics, policy interventions, and system resilience, and can be integrated with broader modeling paradigms such as dynamic traffic assignment and rail operations models. These capabilities position the AAAM as a flexible foundation for future research on overcrowding management, disruption response, and data-driven optimization of urban rail transit systems.

Author’s note: This abstract summarizes ongoing work from a journal manuscript under review. The preprint is available at: Zhao, Bingyu and Tang, Kelly and Soga, Kenichi and Zhou, Xuesong (Simon) and Yang, Hai, Dynamic Passenger Crowding and Operations in Urban Rail Transit Systems: A Validated Framework of Data-Driven Agent-Based Simulation. Available at SSRN: http://dx.doi.org/10.2139/ssrn.5052651

ABSTRACT. **Preferred Presentation Format: In-person (virtual is also acceptable)**

As the most widely used transit data standard in the world, the General Transit Feed Specification (GTFS) allows transit agencies to share service information to riders across common trip planning applications such as Google and Apple Maps. But, for other apps, researchers, or transit data enthusiasts, this data can be hard to find.

That’s where the Mobility Database comes in. The Mobility Database is an open repository of over 3300+ feeds from across the world, complete with search functionality, historical data, and importantly, visualizations of these GTFS feeds. Anyone can access the database to gather data for a number of purposes.

When an agency makes their data open and available in the Mobility Database, they’re contributing to a global pool of knowledge. Analysts, researchers, urban planners, governments, trip planning applications, and even other agencies can access this data knowing its relevance, history, and quality.

Our presentation will focus on the interaction of data visualization and data quality. We’ll showcase real-life transit data from Canadian cities and demonstrate how you can leverage the tools within the Mobility Database to ensure high-quality GTFS data. Specifically, we will: - Demonstrate how to use the Mobility Database's validation reports to quickly detect potential data quality issues within GTFS feeds. - Illustrate how the database's built-in visualization tools (e.g., mapped routes, stop locations, color schemes) can make errors easier to spot, whether it's a misplaced stop, a deviated route line, or a non-legible color scheme. - Present a statistical visualization analysis of the current quality of public GTFS feeds data from Canadian transit agencies, highlighting common trends and areas for improvement.

This session will provide practical insights for researchers, planners, and agencies on utilizing open data platforms for the continued improvement of, in turn, raising the quality, reliability, and coverage of information available to riders.

MobilityData is the steward of GTFS, maintaining the specification, coordinating its evolution with the global community, and supporting its adoption by transit agencies and app developers worldwide.

ABSTRACT. Redesigning a bus network is a way to make it more attractive to customers and more efficient given limited resources. This process involves changes to routes, frequencies, and service coverage to better meet customer needs and travel habits while reallocating resources where they will have the greatest impact. Data plays a major role in a network redesign. Scenario analysis requires measuring indicators related to operations, customer impacts, competitiveness, accessibility, equity and so on. The quality of these indicators directly influences the quality of the selected scenario and how it is communicated to stakeholders—an essential element for its acceptability. The potential use of passive data from various systems—particularly smartcard fare collection systems—in transit analysis and planning has been extensively demonstrated in the literature, including Trépanier (2011) and Monteiro-Fialho et al. (2025). The latter summarizes the contribution of smart card data to several topics relevant to network redesign, including the identification and measurement of destinations, routes, travel times, activities, accessibility, multi-day usage patterns, and transfers. It also highlights that half of the studies combine other data sources such as land use and demographics. In practice, however, the adoption of data and analytics by transit agencies has face historically many challenges, such has constrained ressources and a primarly focus on operation (Hemily, 2024). He emphasizes the importance of disseminating best practices and raising awareness among policymakers about the value of this data. The objective of this presentation is therefore to share how passive data (Smartcard, CAD-AVL, APC) is being used in practice as part of a complete bus network redesign for a mid-sized transit system. The Société de transport de Laval (STL) operates a network of approximately 350 buses serving the city of Laval, a northern suburb of Montreal and the third-largest city in Quebec. Since the implementation of the OPUS fare collection system in 2008, STL has introduced several methodological advancements, such as determining boarding and alighting locations and calibrating simulation models. These steps in transforming and structuring raw data into actionable planning information have improved the understanding of travel behavior, and this experience now enables the development of robust indicators to objectively compare redesign scenarios. The redesign project has the following strategic objectives: • Increase frequency on key corridors; • Provide more opportunities for Laval residents; • Ensure greater equity for vulnerable populations; • Make the network simpler and more intuitive; • Improve efficiency. Characterizing the different scenarios relies on defining and using relevant indicators to translate these strategic objectives into concrete, comparable measures. A “scorecard” composed of key indicators was developed to assess the real impact of proposed scenarios on accessibility, equity, and service attractiveness. Indicators are organized into seven main categories, each aligned with a strategic objective: • Accessibility to transit service; • Accessibility to major destinations; • Accessibility to employment; • Local accessibility; • Service attractiveness; • Customer journey; • Operational resources requirement. Accessibility emphasizes the ability for all Laval residents to use transit to reach their destinations. The main accessibility indicator concerns travel time to major regional destinations. This makes it possible to evaluate the proportion of residents who can access a given generator within a specified time and compare this across scenarios. Attractiveness focuses on the level of transit service and its competitiveness relative to other modes. A key indicator in this category relates to the competitiveness of transit travel times compared to driving. Measured for all Laval residents to each major destination, this indicator helps identify the most attractive scenarios and those likely to generate modal shift (ridership gains). The customer journey addresses the trips of current transit users. It includes indicators for travel time components (walking, waiting, in-vehicle) as well as time gains and losses between scenarios. Producing these indicators required a combination of tools and data processing techniques. In addition to leveraging passive data and its enrichments, transit itinerary calculators and transit modeling tools were used to estimate and break down travel times and to assess vehicle-kilometer traveled (VKT) resource requirements. Road network simulators under congestion conditions were also employed to measure competitiveness against driving. GIS tools were used to measure service abundance and level of service near residents. Furthermore, the development and automation of tools for producing inputs and result files facilitated analysis and iterations. To enhance analysis and communication of results, particularly to visualize differences between scenarios, maps and interactive dashboards were also produced. In summary, this presentation will provide an overview of the key indicators used in the redesign, the methodologies applied, and the challenges encountered during implementation. It will also illustrate and describe how these indicators are visualized and presented using maps, charts, and dashboards. Visualization sample

References Pelletier M., Trépanier M., Morency C. (2011), Smart card data use in public transit: A literature review, Transportation Research Part C: Emerging Technologies, Volume 19, Issue 4 Monteiro-Fialho L., Cuto-Rubio E., Granell C., Trilles S. (2025), Transportation Analyticss Using Smart Card Data : A Systematic Review, IEE Transactions on Intelligent Transportation Systems, Vol.26, no.7, July 2025 Hemily B. (2024), Applying Transit Data and Analytics Research into Practice; Challenges and perspective, ITS Canada-APTS Committee, January 25th 2024, https://www.youtube.com/watch?v=O5VD77jYOSY

ABSTRACT. Urban bus network design aims to select routes, frequency, and stop patterns based on rider demand, the built environment, and objectives such as reliability, efficiency, and environmental impact. Transit agencies now have access to wide swaths of automated data, including Automatic Vehicle Location (AVL), Automatic Fare Collection (AFC) derived ridership patterns, and census-based demographics. However, few existing tools coherently combine these datasets into a computationally manageable design framework due to the complexity of balancing equity requirements, coverage targets, fleet and labor constraints, and institutional priorities. Thus, redesign efforts still largely focus on heuristic adjustments and ad-hoc methods, which limits the ability of planners and stakeholders to reason openly about tradeoffs.

This study develops a general-purpose framework for representing and redesigning bus networks. Through the monotone theory of co-design, which offers a formal process for decomposing a bus redesign problem into components that maintain relationships, this framework is able to represent complex interplay between different datasets and stakeholders involved in transportation planning. Adaptable to the varying requirements of planners, this framework aims at incorporating insights from diverse data sources into a unified platform, including OpenStreetMap road network information, General Transit Feed Specification (GTFS) schedules, Origin-Destination (OD) travel pairs, AVL-informed travel times, and U.S. American Community Survey (ACS) demographic information. The co-design formalism generates explicit Pareto frontiers across various functionalities and requirements of the underlying system, revealing how competing objectives relate to one another. These fronts allow planners and stakeholders to jointly explore and weigh alternative optimal configurations while maintaining transparency about constraints such as fleet composition, crew assignments, and prioritization of different customer segments.

A key feature of the framework is its ability to encode existing bus route designs in the same format used to generate new alternatives. This makes it possible to evaluate current networks directly and to compare them with designs produced through a Mixed-Integer Linear Programming (MILP) formulation. Stakeholders can therefore identify whether the current network is efficient, equitable, or dominated by other feasible configurations, then determine where improvements are structurally achievable or constrained by the design space.

We demonstrate the suitability and relevancy of the general framework through a case study of school bus transit in Framingham, Massachusetts, where we assess the current routing system and generate alternative designs based on cost, coverage, and equity priorities. We assess where different implementations lie on the generated Pareto front, illustrating how the framework can be used to explore real-world design tradeoffs in a network restructuring process transparently, while also offering a reproducible basis for planning discussions.

By combining a formal representational structure with integrated data pipelines and linear optimization methods, our approach aims to highlight how our proposed approach can help planners and operators design transit systems that are both cost-effective and socially equitable, ultimately improving access across diverse urban communities. Eventually, the goal of this framework is to promote tractability, transparency, and stakeholder-aligned tradeoff analysis. We hope this offers a scalable and reproducible method for bus network redesign.

ABSTRACT. Public transit data is widely accessible through the standardized General Transit Feed Specification (GTFS). Data accessibility, with advancements in data analytics and automation, presents a great opportunity to develop advanced tools for transit system performance analysis. This study develops a Python framework providing a high-level analysis of GTFS data. The study focuses on extracting timetable information, operational key performance indicators (KPIs), and developing geo-based visualizations, as examples of GTFS data automation. The framework is applied to the Hamilton Street Railway (HSR) bus transit system. HSR operates a set of 34 routes in Hamilton, a mid-size city in Canada. The framework utilizes routes, trips, stops, stop times, and calendar data frames from the GTFS feed. After data extraction, a four-step process is developed: splitting trips based on route and direction, storing service IDs for days of the week, linking journey information to stop times and locations, and arranging the results into a timetable for outbound and return trips. Framework 1 represents a section of the process. Framework 1. An example of automated GTFS timetable generation

Input: GTFS tables (routes, trips, stop times, calendar, stops) Output: Timetables (CSV) and KPI-ready datasets for dashboard integration

Phase 1: GTFS data loading and provisioning Dataset routes ← Import routes Dataset trips ← Import trips Dataset stop times ← Import stop times Dataset calendar ← Import calendar Dataset stops ← Import stops Phase 2: Route-level filtering and weekday service selection For each route ID in the Dataset routes, do:   Route metadata ← Extract route information (route ID, short name, long name)   Route trips ← Filter the Dataset trips by route ID and direction ID   Weekday services ← Filter Dataset calendar for weekday service IDs (Monday to Friday)   Weekday trips ← Join route trips with weekday services on the service ID Phase 3: Stop time operation and creation of timetables Trip stop times ← Join weekday trips with Dataset stop times on the trip ID   Trip stops ← Join the trip stop times with the Dataset stops on the stop ID   Stop name ID ← Combine the stop name and stop ID for each record   Timetable ← Pivot Trip stops:    Rows ← trip ID    Columns ← stop name ID    Cells ← arrival time   Timetable ← Normalize all time values to a valid 24-hour format Phase 4: Export and integration for analytics Add route short name, route long name, and direction ID to the timetable   File name ← Create file name using route ID and direction ID   Save the timetable to CSV using the File name End For Return: Full set of route-level timetables for KPI extraction

The study automatically extracts KPIs beyond timetables, such as service frequency on weekdays, number of stops, and stop-per-trip ratio, providing an indication of service coverage. These KPIs are complemented by OpenStreetMap (OSM), enabling the visualization of service performance at the network-level. The open-source framework produces ready-to-use maps, system-wide KPI summaries, and individual CSV timetables for each route. The developed framework is scalable, transferable, and could be readily implemented across the board. As such, it offers a glimpse of the potential of data analytics and automation for transit operations and planning.

ABSTRACT. 1 Introduction Public Transportation (PT) stands at a pivotal crossroads in its digital evolution. This is not a prediction but a reflection of an ongoing technological revolution transforming industries worldwide, powered by advances in Artificial Intelligence (AI). The fuel of this transformation is data—collected, integrated, and analyzed at unprecedented scale—while its engine is built on algorithms capable of generating images, text, and predictions with remarkable precision (UITP AI Workgroup, 2025). Yet, for Transit Agencies (TAs), the challenge lies not in adopting AI tools but in orchestrating the data-intensive ecosystems that sustain them. This challenge also represents an opportunity: to rethink the architectural foundations that will support the next generation of urban mobility.

As shown in Figure 1, the two dominant AI workflows—traditional machine learning and foundational models (FMs)—place distinct and often demanding requirements on the underlying data ecosystem. Traditional ML depends on well-engineered features, consistent temporal granularity, and highly structured inputs, whereas FM workflows introduce new constraints related to reliability, latency, and cost, especially for multimodal models fine-tuning. Integrating these tools into the pipelines and data flows of PT’s existing subsystems and architectural domains is particularly challenging. Another significant shift introduced by foundational models concerns data modeling practices. Whereas traditional approaches depend heavily on well-curated datasets at the outset of solution development, foundational models allow early-stage prototyping to precede large-scale data collection, enabling proof-of-concept development and iterative exploration before full data pipelines are established.

Our central argument is that PT differs fundamentally from other industry sectors in how it generates, consumes, and depends on data—making a domain-specific architecture essential for effective AI deployment. Yet discussions of architectural paradigms remain largely absent from transportation research and practice, creating a persistent gap between analytical advances and the infrastructures needed to support them. Most studies emphasize application-level analytics while overlooking the architectural foundations required for scalability, transit interoperability, and AI readiness. Consequently, innovation remains fragmented, constrained by proprietary systems, vendor lock-in, and architectural designs that do not reflect the technical and operational realities.

This work addresses these needs by proposing a structured path for AI-ready PT ecosystems. Our contribution is threefold: 1-Awareness and discussion: Recognizes that the strategy for a PT service and its operations requires short, medium, and long-term planning of the systems architecture to ensure AI-readiness. 2-Conceptual synthesis: Integrates modern data architecture paradigms into the context of transit ecosystems. 3-Illustrative application: Demonstrates how different strategic choices support interoperability, governance, and AI-readiness.

2 Data Architecture Strategies for PT For decades, traditional database systems have relied on monolithic, tightly coupled designs in which storage, computation, and memory operate as a single, inseparable unit. This architecture works well when data is highly structured and stable, as in classic relational tables with fixed schemas, but it becomes restrictive when TAs must manage the complex, fast-changing, and often unstructured data typical of modern PT operations. While these systems excel at ensuring reliability and transactional integrity, they are built on proprietary storage formats and vertically scaled hardware, making them difficult to modify, integrate, or extend. For TAs, this rigidity translates into practical limitations: integrating new data sources requires expensive custom interfaces; deploying modern AI models demands data preprocessing pipelines that legacy systems cannot support; and scaling to meet regional interoperability goals—such as combining data across agencies for multimodal coordination—becomes nearly impossible. In effect, the very design principles that made traditional data management stable and robust now constrain innovation, reducing the flexibility needed to adopt cloud-native tools, distributed processing frameworks, or foundational models capable of learning from diverse, high-volume mobility data.

Figure 2, illustrates not only the range of architectural configurations and associated transit interoperability objectives, but also how these strategic choices progressively enable advanced technical capabilities in PT. In particular, it highlights the transition from the centralized, monolithic architectures that still predominate in most TAs toward more decentralized, modular, and flexible designs capable of supporting Mobility as a Service (MaaS), cross-agency integration, and large scale end-to-end AI-enabled data pipelines.

3 Cybersecurity

With the increasing integration of AI into data management platforms and operational systems within PT, data security has become a foundational architectural concern rather than a peripheral technical issue. While AI-enabled analytics enhance forecasting, optimization, and real-time decision-making, they also expand the system’s attack surface by introducing new dependencies on data pipelines, automated inference processes, and model-driven control logic.

These developments expose PT systems to novel cybersecurity risks that extend beyond traditional network intrusions. Data poisoning, model manipulation, and adversarial attacks can compromise AI-driven predictions and recommendations, potentially leading to degraded service reliability, unsafe operational decisions, or cascading disruptions across interconnected services. Furthermore, the increasing use of federated, cloud-native, and API-driven data architectures—often spanning multiple organizations—raises challenges related to access control, data provenance, and trust management. Cybersecurity must be addressed holistically at the architectural level, encompassing secure data governance, resilient system design, continuous monitoring, and robust fail-safe mechanisms. Reframing cybersecurity as an integral component of transit data architecture is therefore critical to ensuring the reliability, safety, and public trust of AI-enabled public transportation systems.

As illustrated in the Figure 3, cybersecurity is positioned as a transversal layer that spans the entire PT software architecture and data lifecycle, rather than being confined to a single component or stage. From the initial generation of data at diverse sources, through integration, storage within a centralized data lake, and eventual consumption by applications and business processes, security and governance mechanisms continuously regulate how data is accessed, transformed, and used. This cross-cutting layer explicitly encompasses data governance, security, and metadata management, ensuring consistency, traceability, and protection across technology and infrastructure, application, and business layers. Importantly, the figure highlights that these controls apply equally to AI technologies developed in-house and to externally sourced solutions embedded within the PT ecosystem.

4 Conclusions While the proposed framework establishes a foundation for AI-ready data ecosystems in public transportation, it also opens several promising avenues for further research. Advancing this agenda requires not only technical innovation but also institutional and governance transformations that shape how transit agencies collect, manage, and deploy data. Five interrelated domains merit particular attention: federated and collaborative AI, ontology-driven governance, benchmarking and reproducibility, socio-technical adoption, and cybersecurity.

ABSTRACT. Introduction While bus networks undergo detailed planning that includes built-in slack and contingency measures, various sources of uncertainty can still cause operations to deviate from scheduled timetables. It is therefore essential to monitor operations in real time, as the longer a deviation persists, the more difficult it becomes to recover the schedule. These uncertainties may arise from fluctuating traffic conditions, driver behavior, passenger demand, and travel time variability. In low-frequency bus services, where headways typically exceed 10 minutes, maintaining on-time performance is especially critical, as even minor early departures or delays can lead to significant passenger costs. A common approach to maintaining schedule adherence and improving operational resilience is to use control strategies such as holding, stop-skipping, and speed adjustment, though stop-skipping is typically avoided in low-frequency services due to high passenger costs. Speed adjustment offers a less disruptive alternative that can systematically guide buses to recover their schedules, though its effectiveness in non-autonomous fleets relies heavily on driver compliance and accuracy. Methodology In this study, we propose and evaluate three Reinforcement Learning (RL)-based models, each equipped with different sets of control strategies. These models are compared against two baselines: a ground truth model (NC) in which no corrective actions are taken, and a benchmark rule-based model (HO) that prevents early departures through holding. The three RL-based models are each equipped with a set of control actions from holding, stop-skipping, and speed adjustment. The first RL model (RL-HO) applies only the holding strategy for early buses but, unlike the benchmark model, where the holding time equals the earliness, it allows the agent to determine the optimal holding duration, providing flexibility to learn the best decision. The second model (RL-HS) adds the option of stop-skipping for late buses, while the third (RL-HSS) incorporates speed adjustment into the action set. Case Study and Scenarios The models are tested on a low-frequency bus line from the Auckland Transit network in New Zealand, which includes 36 stops and operates with a 30-minute headway . The dataset includes APC and AVL data for one year from July 2023 to July 2024 and includes scheduled and actual arrival and departure times at each stop. The APC data includes passenger boarding and alighting at the stop level. Using the data from 7:30 AM to 9:30 AM (the morning peak) along with observed passenger arrival rates, we simulate passenger arrivals per stop by drawing from a truncated normal distribution for low-frequency services. Moreover, in low-frequency routes, where passengers tend to coordinate their arrival with scheduled bus departures, based on historical observations, a log-normal distribution is the best fit for modeling their arrival times. Additionally, AVL data is utilized to estimate the parameters of link travel times, which are modeled using a truncated normal distribution. Dwell times are generated through a linear regression model that predicts stop time duration based on boarding and alighting passenger counts. To evaluate the total passenger cost, we need to use the value of the waiting and in-vehicle time. According to the New Zealand Transport Agency (NZTA) Research Report 565 (2015)1, the value of in-vehicle time for Auckland bus passengers is NZD $9.70 per hour per passenger, and waiting time carries a weight of 1.46 relative to in-vehicle time, implying a waiting time value of approximately NZD $14 per hour per passenger. Two scenarios are considered in this study: in the first, during regular operations, minor delays and early arrivals occur at various stops, while in the second, in addition to the regular deviations, a significant delay affects one bus at a particular stop which simulates a breakdown. The performance of each model in recovering system performance from passengers’ perspective is evaluated and compared across these scenarios. Results and Discussion Results show that the RL-HS and RL-HSS models, which can apply stop-skipping, and speed adjustment, achieve the best overall performance in minimizing total passenger journey times . It can be concluded that although in low-frequency operations, stop-skipping is not the best option during regular operations with small delays due to long headways, in case of significant delays, stop-skipping could be considered as the best strategy to recover the on-time performance. The same trend can be observed for the total passenger waiting time and journey cost, where the value of time for the waiting and riding time is considered. Looking at the distribution of passenger waiting times with different models, it can be observed that the RL-HSS model has the highest percentage of waiting times less than 5 minutes. Moreover, in this model, only 15.2% of passengers experience waiting times longer than 10 minutes, while this number for the other models is more than 27.9%. Furthermore, since stop-skipping is the only control strategy in the RL-HS model for delays, the waiting time does not have significant improvements in this model. The last model also has the highest percentage of passenger-weighted on-time departures, which considers the mean number of passengers at each stop. Moreover, in cases of substantial delay, the RL-HS and RL-HSS models can reduce total journey times and total waiting times more efficiently in case of a breakdown. In the scenario with a significant delay, the percentage of passengers that experience waiting times shorter than 5 minutes is 50% and the percentage of passengers that have to wait longer than 10 minutes is 24.2% while this value for the no control and benchmark models is 42.3% and 33.3%, respectively. Finally, the average load profile of different scenarios shows a lower maximum passenger load in the RL-HS and RL-HSS model, which indicates a lower chance of overcrowding and uneven loads. As future work, the proposed RL-based models will be extended to high-frequency services to assess their effectiveness under conditions that may lead to bus bunching. Moreover, the efficiency of the models in reducing the transfer time in case of intersecting lines will be studied.

1 Douglas, N. J. (2015). Pricing strategies for public transport. NZ Transport Agency.

ABSTRACT. Incidents disrupt public transport systems, but their empirical study is often limited by a reliance on simulations. While real-time data like GTFS-RT feeds are increasingly available, the unstructured and noisy text of ServiceAlerts remains an underutilized resource for incident analysis. This study develops an unsupervised machine learning framework to automatically reconstruct discrete incident events from historical, multimodal GTFS-RT ServiceAlerts. We compare three clustering approaches: Two-Stage Constrained Clustering (TSCC), Single-Stage Fusion Clustering (SSFC), and Semantic-temporal Clustering (STC), with SSFC demonstrating superior performance in aligning with manually labeled ground truth. The reconstructed incidents are then systematically linked to operational and smart-card data to quantify supply-side (service degradation) and demand-side (ridership deviation) impacts. Applied to a Stockholm case study, this methodology transforms raw alert data into actionable insights on incident dynamics and their multimodal consequences, enabling data-driven analysis for improved resilience and service planning.

ABSTRACT. To be presented in-person.

When it comes to creating a transportation network that will attract riders and effectively move people around in an urban area, one of the most important metrics that agencies strive to improve is reliability, and one of the hurdles they face in achieving a high degree of reliability is incident management. If incidents are not managed properly and efficiently, small delays can compound through entire systems and lead to large consequences for reliability. Despite this, because the timeline is so tight in such an emergency situation, many transit agencies rely almost entirely on the experience and institutional knowledge of dispatchers and supervisors, who have to execute time consuming manual processes during incident management. This makes it difficult to implement solutions that are data-driven. To this end, the ADOP (Automated Dispatching and Operation Planning for Incident Management research) project is meant to develop a software tool that uses data to quickly identify these solutions in a semi-automated way, while also simplifying and automating the implementation of dispatching measures in the process, which would ease the burden on dispatchers, especially during staffing shortages which are a common problem. ADOP, as an extension to current CAD/AVL systems, receives the details of an incident as an input, including which specific roads/lanes/intersections/stations are blocked or closed, as well as the timing and anticipated duration, and then use historical (and, where available, real-time) data to propose possible detour routes for review to the dispatchers. Based on existing data about traffic, travel times, dynamic passenger volumes, route connections, the presence of infrastructure such as traffic signal priority (TSP), and more, the algorithm performs a weighted shortest path analysis to build potential detours. These detours are then presented to the dispatcher as options based on different priorities, such as the route with the least additional time, fewest additional kilometers, fewest missed boardings on average, etc., or as an amalgamation of these priorities based on their relative importance. The tool displays all of the relevant metrics for each of the proposed routes, so that the dispatchers are able to quickly compare the impacts of different options and make a more informed decision. In addition, the tool allows for dispatchers to create their own routes manually by adding constraints to the path, while still providing details about the impact of the resulting detours, and allow for more granular details to be specified; different routings can be applied to different directions and different trips, all within a single workflow. This allows the system to be a useful tool and point of reference for the dispatchers, without inhibiting them from still being able to incorporate their knowledge and experience into the solution. The system is also designed to learn from previous incidents and the matching responses. Data about how successful the previous detours were is used to ascertain the appropriateness of the proposed measures, suggesting better solutions when similar incidents reoccur in the future; for example, if a certain intersection being blocked resulted in much longer travel times on a specific adjacent road in the past, the system is able to take this into account and suggest a different route instead. In order to make the transition to and from the temporary routing as seamless as possible, the system is able to draw on existing schedule information to allow the dispatcher to create variations in the detour routing down to the trip level, and make suggestions for how the transition to and from the contingency schedule can be managed. This allows the system to aid the dispatcher in managing the impact of the detours, while remaining flexible enough to accommodate important variables like the timing and location of shift changes or range limitations on electric vehicles. The other very important aspect of the project is automating the process of responding to an incident by providing a simple and singular user interface that can handle the process from start to finish. This includes displaying incident details and detour routings to dispatchers in both text and GIS formats, automatically identifying and selecting impacted routes and stops, allowing for different routes to be specified for individual trips, automatically updating transit data feeds, automatically sending detour instructions directly to operators, and automatically triggering public-facing service alerts, all based on the detour that is created in the system. In summary, the ADOP project seeks to create a software tool, embedded in the CAD/AVL system which will help dispatchers and supervisors create semi-automated incident response solutions that are data-driven, and implement them as quickly as possible to minimize the impact of incidents on transit systems and to improve overall reliability.

ABSTRACT. Bus holding control is a challenging sequential decision-making problem because stochastic disturbances propagate through interacting vehicles and destabilize service headways. While reinforcement learning (RL) has shown promise for this task, most existing methods rely on fixed-length state vectors that do not adequately capture the variable-size and relational structure of upstream fleet observations. This paper proposes ProxSetFormer (PSF), a proximity-aware dynamic state encoder for RL-based bus holding control. The framework represents the system state as an ego bus together with a variable-length set of leading buses, and uses a set-transformer-based architecture to model inter-leader dependencies and ego-conditioned upstream relevance. The method is evaluated in a stochastic, capacity-constrained simulator derived from a real-world Chengdu transit dataset. The results show that the proposed method provides the strongest overall balance between service regularity and passenger impact. In particular, the PSF-LookBack (PSF-LB) variant achieves the lowest headway standard deviation and best gap reduction while keeping passenger waiting time, generalized perceived travel time, and denied boarding close to baseline levels. The proposed approach also shows superior passenger-cost efficiency per unit of regularity gain and remains effective under restricted control coverage. Finally, action-sensitivity analysis and SHAP-based feature attribution indicate that the learned policy captures both local instability and upstream propagation effects. These findings highlight the value of relation-aware dynamic state representation for RL-based transit control.

ABSTRACT. 1. INTRODUCTION

Public transit agencies commonly manage high-frequency urban bus corridors using timetable adherence. However, as real-time Automatic Vehicle Location (AVL) and Automatic Passenger Counting (APC) data become more ubiquitous, an increasing number of agencies are focusing on headway regularity instead of fixed schedules. This paradigm shift acknowledges that for high-frequency services, waiting time consistency and extreme gap avoidance are more important than punctuality [1], especially in situations where punctuality cannot be achieved. This headway regularity can be applied at the dispatch level, reducing start headway variability, or during a trip, by applying strategies such as bus holding and stop skipping. However, buses are still assigned only to a single route based on a pre-computed static schedule, which can be severely disrupted in adverse conditions.

A possible direction to address this issue is the concept of Dynamic Interlining (DI) [2], where a set of shared buses can be dynamically assigned to trips in different routes, using optimization models or heuristics to determine the next assignments. Zahedi et al. [2] show that DI can improve the efficiency of a fleet of shared buses under independent, random travel-time noise. However, real-world transit networks are dominated by systemic, spatiotemporally correlated shocks, such as heavy snow, car accidents, or subway shutdowns.

Dynamic Interlining (DI) can compensate for these delays; however, since it focuses on schedule adherence, after a departure miss, the scheduler attempts to “catch up" with the published timetable, resulting in a large variance in headways. A possible better solution is to use schedule-free operations, such as Multiline Vehicle Dispatching Problem (MLVDP) [3], in which arriving buses at a terminal are assigned to lines and positions in the upcoming sequence of departures.

However, both DI and MLVDP were evaluated only on benign scenarios of independent travel-time noise, which allows the optimizers to easily compensate for delayed trips using buses arriving earlier from other routes. However, in real-world scenarios, travel-time noise is often correlated, and a corridor can have persistent delays for long periods, affecting multiple routes. Or, in city-wide events, such as heavy rain or snow, the entire network can be affected. Finally, existing formulations do not consider passenger demand and can leave high-demand routes unattended for long periods.

In this work, we present a Headway-Based Dynamic Interlining for Improved Service Regularity, utilizing a Mixed-Integer Linear Programming (MILP) formulation with a demand-weighted objective function to handle vehicle departures during severe disruptions.

2. METHODOLOGY

We focus on the problem of a single hub with multiple routes, where a fraction of buses are shared between routes, while the others are dedicated to a single route. On each departure time, we solve an optimization problem to determine which bus should be assigned to each of the next k departures of each route and when each bus should depart, using a demand-weighted headway objective function. The penalty is the sum of the squared value of the headway deviations, normalized by the target headway, for each departure slot, with additional large penalties for missed departure slots. It also includes constraints to ensure that each bus is assigned to at most one slot, departure times respect bus arrival times, and departures within each route follow a non-decreasing sequence of times.

We model trip travel times using a lognormal factor model driven by corridor-level continuous-time autoregressive AR(1) processes. Each corridor maintains a latent congestion state that evolves over time, allowing it to capture temporal delay persistence. Additionally, multiple routes may share a corridor, enabling the model to capture spatial correlations.

To prevent vehicle depletion in scenarios with persistent shocks, we use a state-dependent optimization framework. For this, we perform predictions for the departure times of the next k = 3 trips per route. If there are delays in arriving buses, the MILP will increase the headways of all trips to minimize headway deviations and reduce the quadratic loss.

3. EXPERIMENTAL EVALUATION

We performed preliminary evaluations of this framework using an event-driven simulator, considering a hub with 4 routes. We simulated a scenario with a four-route configuration, comprising two high-frequency trunk routes (5- and 6-minute headways, and 30- and 36-minute trip times) and two lower-frequency routes (10- and 12-minute headways, and 60- and 72-minute trip times). Each route has 6 buses. We evaluate our framework in scenarios with disruption in a single route (route 1), caused by a transit accident, and with disruption on all routes, caused by a weather event.

We compared the headway coefficient of variation (CoV) for the schedule-based scenario with dedicated buses, headway-based MILP with dedicated buses, and headway-based MILP with a full shared fleet. With headway-based dispatching, we see a strong reduction in the CoV, from 0.67 to 0.30, even when using dedicated buses. Using a shared fleet resulted in additional decreases on the higher-frequency routes (1 and 2), from 0.3 and 0.35 to 0.18 and 0.19, respectively, and a small increase on route 4, from 0.25 to 0.37, demonstrating that the optimizer can successfully decrease headway CoV in high-frequency routes while maintaining quality service on low-frequency routes.

4. CONCLUSION

Our findings suggest that shifting from schedule adherence to demand-weighted headway regularity and utilizing longer planning horizons can enable transit agencies to effectively reduce the impact of correlated disruptions in the bus network at the hub level. We still need to perform additional analysis on other scenarios, including multi-hub ones, and with different fractions of shared fleets.

Acknowledgments: This work was supported by FAPESP grants 23/17501-4 and 21/11959-3

References: [1] Carlos F. Daganzo. A headway-based approach to eliminate bus bunching: Systematic analysis and comparisons. Transportation Research Part B: Methodological, 43(10):913–921, 2009. [2] S. Zahedi, Haris N. Koutsopoulos, and Z. Ma. Dynamic interlining in bus operations. Transportation, 52(3):827–850, 2025. [3] Felipe Delgado. Dynamic multiline vehicle dispatching strategy in transit operations. IEEE Transactions on Intelligent Transportation Systems, 23(12):24918–24928, 2022.

ABSTRACT. Not Decided

Assessing transit system vulnerability requires capturing not only how disruptions affect network performance, but also the stochastic nature of incident occurrence and the network-wide propagation of their impacts. This paper proposes a comprehensive probabilistic, data-driven framework for evaluating the vulnerability of urban rail networks under multiple incidents. Using high-resolution observed journey times derived form Wi-Fi data, we first construct time-of-day and day-of-week specific distributions of usual travel times for all origin-destination (OD) paths. These distributions are paired with path-level travel time distributions during disruptions and integrated through a limit-state reliability formulation to estimate the likelihood that each path is negatively affected by a given incident. To characterize the probability of disruption across the network, we also develop a segment-level incident occurrence probability model, capturing the heterogeneous and over-dispersed nature of daily failure patterns. The two probabilistic component, including path-level disruption likelihoods and segment-level failure probabilities, are then embedded into a new vulnerability index based on a ridership-weighted global efficiency (GE) metric. This metric quantifies the expected loss of network efficiency during incidents, accounts for temporal variation in OD demand, and aggregates event-level functional losses into segment-level consequences. The resulting vulnerability index provides a system-wide, comparable measure of expected efficiency degradation attributable to the failure of each segment. By jointly modeling the probability of incident occurrence, the stochastic impact of disruptions on journey times, and the propagation of efficiency loss through real-world ridership patterns, the proposed framework offers a generalizable and operationally meaningful tool for identifying critical segments in transit networks. The methodology enables transit agencies to prioritize mitigation strategies, allocate resources effectively, and enhance the resilience of the system under recurrent and simultaneous disruptions.

ABSTRACT. (Planning to present in person)

Introduction: When planning and operating rail traffic, capacity calculations are important to manage the system effectively. Being able to quantify the capacity with a measurable indicator is helpful in decision-making and planning for improvements and future investments, such as upgrading and dimensioning new lines and stations. It also supports tactical planning, assessing the system's current status, and predicting future scenarios. For the assessment of a timetable, capacity utilisation can be used to indicate how much of an infrastructure a timetable utilises within a predefined time period. In our previous work, a third-party software was used to extract timetable data, including data for signal block occupation, to calculate the capacity utilisation of that timetable. The timetable information was processed to construct blocks of occupation for each train and infrastructure data was used to decide which blocks can or cannot be used simultaneously. Here, a block can sometimes be the whole signal block section, but more often blocks will be sub-parts of a block section, which can be released in sequences as a train clears them. To estimate the capacity utilisation, the compression method based on UIC 406 was used. While the data extracted is on a microscopic level, being detailed and reflecting well the characteristics of real operation, the process is rather time-consuming and dependent on access to this software. Therefore, we want to explore the potential to use other data sources as input for capacity analyses. For this purpose, this study aims to examine data automatically collected by the signalling system, in the network of Sweden. The data is managed by the Swedish Transport Administration (Trafikverket), which is responsible for the state-owned railway network, including the infrastructure and traffic allocation. By analysing this operationally collected signalling log data, we investigate whether it can serve as a feasible and efficient alternative input for capacity assessments.

Background: In rail operation, each train along a line will move from one signal block section to the next. Signals indicate whether the section(s) ahead are free or occupied. The block section will be considered occupied as soon as the train enters it until it exits; during this time, no other train may enter that same block. The time of occupation depends on the speed of the train. To make sure that the infrastructure ahead is available, the train can reserve (preoccupy) one or more block sections before the actual occupation. Headway is the time between two consecutive trains on the same track section. The data used comes from a system called LUPP, provided by Trafikverket, which is used to track realised traffic and derive statistics from the system, such as punctuality. Every train run is registered and reported to the system, along with any deviations from the timetable. Data from this system doesn’t provide any information on the specific track(s) trains use at stations, only signals. The system provides different exports of the data; one possibility is to access registrations from the signals, with reservations and occupation by trains utilising the signal, in this abstract referred to as signalling log data. In detail, the data contains the signal number, ID, and the related station of the signal. Furthermore, the exact timestamp (yyyy-mm-dd, HH:mm:ss) when each (block) signal is either reserved, entered or released, together with train ID information and deviations from the planned timetable, thus offering a detailed chronological record of train–infrastructure interactions.

Method: The signalling log data is derived per train and date of operation , and Norrköping C (Nr) station is used for the case study, likewise in our previous work, to enable comparability between the old data and newly processed data. Nr is an intermediate station along the Southern mainline in Sweden, served by different types of trains, such as commuter trains, regional, long-distance, high-speed passenger trains, and passing freight trains. For the commuter trains, Nr is the start/end station, thus using the station for turnarounds. For capacity calculations using the compression method to estimate the capacity utilisation of a line or station, information on block occupation is required. As the signalling log data can be used to extract all block occupations, the application of the compression method becomes feasible using this data source.

Expected results and contribution: Preliminary outputs from the current stage of the study from a small data set show promising results for the possibility of using the extracted signalling log data for analyses on railway capacity. So far, the data has been extracted and processed to show occupation on the signal block level in order to reconstruct the realised timetable. The advantage of this dataset is that it shows the realised traffic, which makes it possible to estimate the capacity utilisation under both normal and disrupted operating conditions. On the other hand, this can also be a challenge, depending on the aim of the analysis. For instance, if a train is cancelled, thus not registered, and the operation is compared to the planned timetable. Future work will continue to explore the possibility of using signalling log data as input for capacity calculations and station analyses by engaging a larger data set and comparing with results from previous work, where we expect the data to be useful for validating and potentially strengthening the conclusions from our earlier work.

Conclusion: We believe that this study contributes to capacity research by demonstrating the potential of using automatically collected signalling log data as an alternative input for timetable analyses in order to assess station capacity.

ABSTRACT. Urban railway services often constitute the backbone of the public transport offer of the territories in which they are deployed and shape the mobility patterns of travellers thanks to their high capacity. Although the attractiveness of a railway infrastructure can be largely attributed to its structure and frequency of trains, it is also dependent on the quality of feeder services that allow users to reach and leave the rail stations from and to further areas. Transportation models allow to estimate the impact that new services or policies can have on travellers. Traditional four step models consider mobility pattern on the level of large flows in a territory, making them suitable for study of rail services[6]. However, they lack precision when it comes to understanding the much more disaggregated needs for local feeder services. The increasing popularity and recent developments of agent- based models, where each traveller and each vehicle is considered individually[5], allows to build simulations with a level of details that is sufficient to study local feeder services in relationship to region-wide rail infrastructure. In this work, we propose an approach to assess the impact of future planned rail infrastruc- ture in conjunction with the specification of feeder services. As a case study, we consider the Île-de-France region, which contains the city of Paris, where various new subway lines are being developed under the Grand Paris Express (GPE) project. The new lines will be delivered between 2026 and 2030 and will extend over a length of XX kilometers and include XX new stations. They are primarily intended to connect the suburban areas of the region without having to go through Paris. Whereas the future of operation of these future lines has been known since a few years ago, the feeder services, especially those linking new subway stations, remain to be specified. This can be explained by the fact that feeder services, whether they consist of fixed-route bus lines or other more dynamic services, are much more easily deployed than heavy rail infrastructures. Two studies on the impact of GPE on mobility in the Île-de-France region using agent-based models have been conducted in the past, relying on the MATSim simulation framework[4] and the open Île-de-France simulation[3]. In first study, the assessment was conducted by assuming an on-demand service that can be used exclusively as a feeder[2]. In the second, feeder bus lines have been optimized using Mixed Integer Linear Programming[1]. In both cases, the investigation only focused on a reduced area concerned by the GPE project due to the computational complexity of the approaches.

n this research, the integration of future GPE lines is systematized by a process that takes in the GTFS specification of the existing transportation offer in the Île-de-France region (available as open-data) and a specification of new railway services (station locations, lines and their on-peak and off-peak frequencies) and generates an updated GTFS that integrates the future offer. As for feeders, to go further in the impact assessment and in 1 the scale of the analysis in comparison to the two previous approaches, we consider them here exclusively from the perspective of their operational implications on the users: the area around the rail station covered by the service, the average speed of the service and its frequency. Thus they are note integrated in the GTFS dataset but rather directly in the simulation model which is able to integrate them with the rest of the PT offer and propose alternatives that use feeder services if they allow to obtain better travel routes. Building on these functionalities, we perform two types of analysis comparing between three sets of simulation scenarios. As for the scenarios, we consider the baseline simulation for the Île-de-France region, the GPE-NoFeeder scenario where GPE lines are added without any new feeder service, and GPE-Feeder scenarios where feeder services are added around all new stations resulting from the GPE project. As for the analyses, the first one consists in, without performing simulations with the considered public transport offers, a routing- based analysis where the best public transport route for the whole population’s trips is computed under each scenario and compared. This allows to assess, on the population level, the share of public transport trips that are improved by the GPE lines and then by feeder services. The set of improved trips will then constitute potential levers for change in mobility behaviour. The second analysis directly looks at the modal shift by running full MATSim simulations, enabling mode choice for travellers and propagating the effect of better public transport alternatives on the overall trip chain. This ongoing study is meant to be fully replicable by relying on open-source software, open-data, and by sharing all the specific software components and simulation workflows in open-source

References: [1] T. Chouaki, S. Declerq, and S. Hörl. Towards an integrated approach for the automatic design of feeder bus lines using agent-based simulation and combinatorial optimization. In The 12th Symposium of the European Association for Research in Transportation (hEART), 2024. [2] T. Chouaki, S. Hörl, and J. Puchinger. Towards reproducible simulations of the grand paris express and on-demand feeder services. In 102nd Annual Meeting of the Transportation Research Board (TRB 2023), 2023. [3] S. Hörl and M. Balac. Synthetic population and travel demand for paris and île-de-france based on open and publicly available data. Transportation Research Part C: Emerging Technologies, 130:103291, 2021. [4] A. Horni, K. Nagel, and K. W. Axhausen, editors. The Multi-Agent Transport Simulation MAT- Sim. Ubiquity Press, Aug. 2016. [5] E. Maggi and E. Vallino. Understanding urban mobility and the impact of public policies: The role of the agent-based models. Research in Transportation Economics, 55:50–59, 2016. [6] M. G. McNally. The four-step model. In Handbook of transport modelling, volume 1, pages 35–53. Emerald Group Publishing Limited, 2007

ABSTRACT. The adoption of Artificial Intelligence (AI) in railroads and transit agencies has the potential to revolutionize operations by enabling faster data analytics and informed decision-making. However, a critical challenge lies in the collection and structuring of data from diverse systems, such as rolling stock, track infrastructure, signaling, and cameras, in a format that is usable for AI applications. Without clear guidelines, agencies often struggle to specify data collection requirements, standardize data formats, and identify appropriate AI training models, limiting the effectiveness of AI-driven solutions. This paper provides a comprehensive roadmap for transit agencies to address these challenges. It outlines best practices for specifying data collection requirements in technical specifications, including the structure, granularity, and frequency of data collection. The paper emphasizes the importance of adopting standardized data formats and protocols to ensure compatibility with AI models and interoperability across systems and vendors. Additionally, it explores the types of data—such as sensor readings, diagnostic logs, and video feeds—that are critical for AI applications, and how to preprocess and label this data for effective model training. The paper also reviews open-source AI training models and frameworks that can be adapted for transit-specific use cases, such as predictive maintenance, anomaly detection, and real-time operational insights. It highlights the importance of integrating edge computing, cloud-based analytics, and Internet of Things (IoT) technologies to streamline data collection and processing. By providing actionable recommendations on what data needs to be collected, how it should be collected, and how it can be utilized, this paper seeks to empower transit agencies to overcome data collection barriers and fully harness the transformative power of AI.

ABSTRACT. Preference for Presentation: Undecided

1. Introduction

Public transit serves as the backbone for sustainable and equitable transportation in major metropolitan areas. However, the reliability and attractiveness of these systems are frequently compromised by operational challenges. Among these challenges is overcrowding, particularly during peak hours, which significantly degrades the passenger experience. This can result in denied boardings and negatively impact overall transit performance. When comfort levels drop, the perceived quality of service diminishes, often leading passengers to abandon transit in favor of other modes.

In addition to that, transit systems are vulnerable to acute disruptions. These disruption events range from localized operational failures, such as track fires, power outages, or medical emergencies, to large-scale external shocks like extreme weather (flooding, snowstorms, heatwaves) or major sporting events. Unlike minor delays, these disruptions introduce non-linear effects, where a single failure can cascade through the network, creating overcrowding that significantly degrade the passenger experience. Given the mass population’s dependence on these transit services, a disruption does not merely delay a train; it potentially affects the daily commute of thousands who need access to jobs, healthcare, and education.

Therefore, understanding public transit not merely as a static infrastructure but as a dynamic operational system is critical. While improving infrastructure is a long-term goal, a key priority for operators is enhancing resilience, which simply means the ability of the system to absorb shocks, mitigate impacts through backup planning, and recover swiftly.

To achieve this, planners require tools that can accurately simulate the dynamics of a disrupted network. While traditional macroscopic models exist, many rely on static assignment and frequency-based assumptions which often fail to capture the temporal dynamics of overcrowding, specifically 'denied boarding.' This is a non-linear phenomenon that can disproportionately affect transit performance and overlooking it often leads to underestimating congestion effects under high-demand or disrupted conditions. Towards that, this study will utilize a developed agent-based transit simulation, named as integrAted dAta-driven Agent-based simulation (AAAM) that integrates General Transit Feed Specification (GTFS) with Smart Card data (Zhao et al., 2025). The AAAM dynamically models the interactions between platform, passenger, and vehicle agents, capturing granular transit operational details, such as real-time capacity limits and denied boarding events, to accurately simulate how local failures propagate through the system. The purpose of this work is to utilize the AAAM to:

a) Simulate both business-as-usual operations and disruption scenarios to analyze how local failures cascade through the network and affect passenger journeys b) Evaluate resilience strategies, such as bus bridging and re-routing, to measure their effectiveness in mitigating these disruptions.

2. Case Study: Edmonton Transit Service (ETS) The proposed scenario framework will be applied to the City of Edmonton, Alberta, Canada. The Edmonton Transit Service (ETS) serves as a testbed for this analysis as it represents a typology common among mid-sized North American cities: an LRT-based system supported by a comprehensive bus backbone. The network consists of three Light Rail Transit (LRT) lines serving approximately 123,000 daily commuters, integrated with a bus network serving 115,000 daily commuters. The presented case study here would enable other rail-based systems to apply similar scenarios for their transit network, such as Calgary, San Diego, Portland, and Charlotte.

By subjecting the ETS network to a series of disruption scenarios (such as flooding, power outage, debris) we aim to generate actionable insights and in-depth analyses to enhance the system’s resilience during emergencies and operational disruptions.

Reference

Zhao, Bingyu and Tang, Kelly and Soga, Kenichi and Zhou, Xuesong (Simon) and Yang, Hai, Dynamic Passenger Crowding and Operations in Urban Rail Transit Systems: A Validated Framework of Data-Driven Agent-Based Simulation. (under review). Available at http://dx.doi.org/10.2139/ssrn.5056234