ABSTRACT. Author’s preference for presentation: In-person

Introduction

Tourism and transportation are characterized by strong bonds, given that the former cannot occur without the latter. In some cases, transportation is an essential element of visitors’ experiences (e.g., by providing accessibility to local destinations), while in other cases it can form the tourist activity per se (e.g., train or cruise journeys). With the term visitor, we opt for a more inclusive term, reflecting travellers taking a trip to a main destination outside their usual environment, for less than a year, and for any purpose (business, leisure, personal).

Public transportation (PT) plays a crucial role in facilitating visitors’ mobility needs once they arrive at their destination. Therefore, the correct estimation of the visitor PT demand is crucial for accurate PT supply dimensioning, accounting for visitors’ induced seasonal demand spikes. Automated Fare Collection (AFC) systems, with the use of smart cards (smart card data), have become very popular in modeling PT passenger behavior due to their rich spatiotemporal resolution. However, there is still a very limited body of research on understanding visitors’ travel behavior using smart card data.

Problem and objectives

In this study, we present a framework for identifying visitors entirely based on smart card data and defining distinct temporal and spatial visitor profiles. Our method is reproducible, given its reliance on individual mobility traces, thereby enhancing its added value. In particular, we utilize PT travel diaries across Region Stockholm, Sweden.

Methods

First, we identify visitors using smart card data with a combination of criteria, including: i) one-time users with a travel span of less than a week (activity criterion), and ii) users utilizing dedicated visitor fare products (1, 3, 7-day tickets) (fare product criterion). We opt for one-time users to avoid the risk of failing to distinguish repeat visitors from residents, and we ground the selection of the 7-day threshold in the local tourism context as well as the local visitor fare products.

Then, for the travel diaries of the identified visitors, we are able to obtain relevant attributes for their temporal and spatial activity. Prior to the clustering analysis, we explore redundancy among all variables by constructing correlation matrices, and the highly correlated variables are not jointly used in the clustering process. Last, we tested different combinations of the remaining temporal and spatial variable combinations, targeting both conceptually meaningful and distinctive temporal and spatial clusters.

For the temporal and spatial clustering process, we choose the k-means clustering algorithm given its demonstrated strong scalability in managing multidimensional data, as well as identifying natural clusters for practical applications. Finally, we explore the relationships between the resulting travel activity and spatial clustering profiles.

Results & Discussion

We account for passenger demand data from 2023 (January 1 - November 23, 05:00-22:00), accounting for 327 days in total. Specifically, we analyze 5,039,516 PT journeys from 1,162,117 identified visitors, corresponding to 21.5% of all cardholders but only 1.7% of the total 291,050,393 PT journeys recorded during the study period.

In order to assess the reliability of our method in identifying visitors using smart card data, we extract the monthly number of visitors and compare them with official statistics as reported by the Swedish Agency for Economic and Regional Growth (2025). Our results demonstrate a strong correlation with the official statistics both on the Stockholm municipality and at the Region Stockholm level (Pearson’s correlation coefficient of 0.85 for both comparison levels), confirming the effectiveness of our method.

Concerning the temporal clustering, we employ the following four variables:

number_of_days: Total number of active days in the PT system journeys_per_day: Average number of journeys/day across the active period first_journey_time: Average earliest journey start time across the active period last_journey_time: Average latest journey start time across the active period

We use internal evaluation indices such as: i) the total variance within the clusters (Inertia (Within-Cluster Sum of Squares (WCSS)), ii) the average similarity between each cluster and its most similar one (Davies-Bouldin Index), and iii) the intra-cluster cohesion – inter-cluster separation (Silhouette Score). We adopt the elbow rule, and we select to proceed with four travel activity clusters:

• A1 - Multi-travel short-stay visitors: Visitors with a short stay who make several daily journeys. They are characterized by high activity during the afternoon hours (14.7%).

• A2 - Long-stay and frequent travellers: Visitors who stay for multiple days and make more journeys/day. They are also characterized by high activity during the afternoon (17.3%).

• A3 - Day trippers: Visitors whose first and last journeys occur early in the day. They are characterized by short stays (a single day) with moderate daily activity (ca. 2 daily journeys) (38.5%).

• A4 - Evening visitors: Visitors whose first and last journeys occur late in the day. They are characterized by short stays with a moderate travel activity per day (ca. 2 daily journeys) (29.5%).

Concerning the spatial clustering, we employ the following four variables:

distance: Average road network distance traveled/journey total_destination_stops: Total number of unique destination PT stops during the active period main_OD_%journeys: Share of PT journeys concentrated in the most visited Origin-Destination (OD) intra_stockholm_%journeys: Share of PT journeys within Stockholm’s inner city

Similarly, adopting the elbow rule, we select to proceed with three spatial clusters:

• S1 - City center visitors: Visitors mainly traveling within the city center. They make short-distance trips to a limited number of destinations (31.3%).

• S2 - Outside the city center visitors: Visitors mainly traveling outside the city center for longer distances to a limited number of destinations (37.3%).

• S3 - Mixed visitors: Visitors traveling both to central and non-central areas. They demonstrate exploratory behavior due to their tendency to visit multiple destinations over longer distances (31.4%).

Concerning the relations between the clusters, we highlight that City center visitors (S1) are primarily associated with Day trippers (A3) and Evening visitors (A4), highlighting the importance of Stockholm’s urban core for diverse activities throughout the day.

Last, the identification of distinct spatiotemporal visitor profiles can inform public transportation planning in relation to tourist destinations and contribute to the strategic assessment and refinement of policies such as fare structures.

ABSTRACT. The widespread availability of real-time transit information (RTI) via mobile applications presents a paradigm shift in how transit services are delivered and experienced, as many commuters are now routinely relying on such data for travel decision-making. Despite this, current passenger route choice models still frequently treat commuters as static decision-makers who rely on fixed or perceived information, and assume stable day-to-day travel conditions. This limited static framework fails to capture the transformative impact of RTI, which:(1) provides commuters with dynamic awareness of the transit network's status, enabling adaptive decision-making and en-route adjustments; and (2) expands the bounded rationality of travellers by reducing the cognitive load required to evaluate options, thereby lowering a critical barrier that causes travellers to exclude services perceived as inconvenient. Therefore, this research investigates how real-time transit information (RTI) influences the temporal dynamics of commuters’ route choices.

Past research aimed to understand the influence of RTI on commuters’ travels through traditional data sources, such as surveys. However, surveys have significant limitations to capture longitudinal travel data (over months or years) due to high acquisition costs, participant burden, and recall bias. Thus, this study utilizes a novel type of tracking data from a transit mobile application, as it becomes a viable cost-effective tool to collect long-term crowd-sourced data firsthand from app users who are willing to participate as it requires little effort. This study used long-term data collected through “OneBusAway”, an open-source mobile transit application that provides real time transit information to users in multiple cities across the U.S.. From 2019 to 2023, OBA began to conduct an experiment to obtain long-term travel information. OBA users were asked on the app if they are willing to be tracked throughout daily activities for research purpose. By this, OBA managed to collect both trajectory data (Dataset 1) and OBA app history usage (Dataset 2) from 445 Puget Sound (Washington State, US) residents who contributed data for more than 300 days. Dataset 2 is their transit information inquiries, consisting of information on the specific bus stops, bus lines, and bus arrival times they viewed. Thus, these tracking data provide a perfect foundation for the analysis between RTI and commuter route choice patterns over time. Google Activity Recognition Transition API within Android smartphone was used by OBA to record “activity transitions points”. These points represent when users changed their activities among {Still/Walking/Bicycle/InVehicle} based on the mobile device’s movement. Although these transition points are sparse and do not capture full trajectories, the low-frequency collection provides a trade-off as it minimizes battery drain to be nearly unnoticeable. Hence, the strength of this data is the ability to obtain travel records continuously and that participating users can provide data over long periods of time.

The first task of this study is to address the sparseness of the data (recorded transition points). While Google API provided activity-labels based on device’s speed, preliminary analysis revealed their accuracy to be questionable. Therefore, the study instead focuses on the location and timestamp recorded within each point, since past studies have confirmed its positional accuracy to be mostly within 50 meters. First, spatio-temporal clustering is employed to extract individual trips from each user's daily data, converting these point series into an Origin-Destination (O-D) form. For each trip, the study generates a potential choice set based on the O-D pair and its associated timestamp. Using the Google Directions API, supplemented by open data from OpenStreetMap (OSM) and the General Transit Feed Specification (GTFS) for Puget Sound, the study generates a “reasonable” choice set based on labelling criteria (least travel time, least walking distance, fewest transfers). This set also includes driving/walking routes, along with transit alternatives constrained to those taking less than three times the minimum travel time, three times the minimum walking distances, and two or fewer transfers. Finally, a simple form of Inverse Distance Weighting (IDW) is used to map-match the user's recorded points to these generated routes, identifying the most likely route taken for each trip. This methodology enables the longitudinal reconstruction of each user's route choices over months or years.

The study then conducts the analysis in two parts. First, it identifies factors influencing the demand for RTI inquiries, aiming to identify what information interest commuters when searching within their journey. A Fixed-Effect Negative Binomial panel regression is employed to model the number of inquiries per user-month (the dependent variable) as a function of service related factors. These include number of bus trips taken, intended service headways, number of available transit alternatives, and service delay time (extracted from recorded RTI). Results show all factors are positively correlated with RTI usage. Notably, the number of available transit choices provided the strongest coefficient, highlighting the critical role of RTI in commuter’s route choice decision under complex transit network. Second, ongoing work is set to assess travel behavioural impacts between user groups with varying levels of RTI adoption, focusing on route diversification and chosen route performance among their "regular" trips. A regular trip is first identified by spatially clustering all O-D pair locations into 500-meter zones using DBSCAN; the most frequent O-D pair is then selected as the individual’s regular trip. Route diversification is quantified using the normalized Shannon entropy, calculated from the observed variety and frequency of routes taken within regular O-D pairs for more than 30 days. Route performance is then evaluated based on adaptive adjustments (both pre-trip and en-route), based on the assumption that a user, when presented with a superior alternative, will choose it. The results are expected to show that higher RTI usage correlates with greater route diversification over time, even for regular trips. Overall, the study provides valuable insights on the ability to link RTI inquiries with actual travel patterns, offering transit agencies and app developers evidence-based recommendations for improving information systems and optimizing services. In addition, questionnaire surveys are also administered to these volunteer users to validate the results.

ABSTRACT. In-person presentation

Reliability is one of the key indicators of public transport efficiency, reflecting the service certainty as experienced by passengers in relation to the planned service (van Oort, 2014). Unreliable services lead to delayed arrival times, longer waiting times, increased crowding, bunched vehicles and reduced passenger satisfaction (Rezazada et al., 2024). Service reliability can be measured in terms of regularity, the consistency of actual travel times, and punctuality, the deviation from the scheduled timetables (Dixit et al., 2019). While most transport authorities traditionally rely on schedule-based indicators such as on-time performance, reliability can also be measured using regularity-based ones, including the headway coefficient of variation (CV), average excess waiting time (EWT), and headway adherence (Cats, 2014).

Regularity is particularly relevant for services operating with short headways, where passengers arrive more randomly without consulting the timetables (Chen et al., 2025; Ingvardson et al., 2018). In contrast, punctuality could be more meaningful for longer headways. Passenger-experienced reliability has been studied under various travel conditions, such as crowding and seating (Jenelius, 2018), as well as across different modes and routes for the same origin-destination (OD) pair (Dixit et al., 2019), serving as a tool for service planning and policy evaluation. Nevertheless, it remains unexplored which dimensions of reliability passengers prioritise along their journeys, given the substantial heterogeneity in preferences and behaviours.

This study addresses the existing literature gap by investigating the extent to which passengers consider punctuality and/or regularity when choosing their public transport trips. We estimate revealed preferences using large-scale route choice data while incorporating both headway- and schedule-based reliability indicators. The case study focuses on the multimodal network of Eastern Denmark, which includes regional and local trains, S-trains (suburban trains) and metro services. Our analysis draws on approximately 140,000 trips recorded via the Danish smart travel card (Rejsekort) between September 1 and 15, 2023, during weekday morning peak hours (07:00–09:00). In addition, we utilise automatic vehicle location (AVL) data, which provide planned and observed arrivals and departures at the stop level for the same period.

The modelling framework is based on the Random Utility Maximisation (RUM) paradigm, where each individual, n, is assumed to choose the route i that maximises their utility. The route utility captures observed trip attributes, including in-vehicle time for every public transport mode, transfer-type penalties for specific transfer combinations (e.g. train to train, bus to train, etc.), transfer walking time, waiting time at the transfer, and planned waiting time at the origin. To capture route overlap, we employ the Path Size Logit (PSL) model that accounts for correlations among alternatives (Bovy et al., 2009). Route choice sets are based on the observed smart card routes within the study period, with alternative routes generated by including stops within 500 m of the chosen origin and destination.

The novelty of this study lies in incorporating and comparing multiple reliability indicators into behavioural models to examine how passengers account for reliability in their route choices. Specifically, we include the average excess waiting time due to irregular services, the standard deviation of headways and schedules, and both the headway- and schedule-adherence (on-time performance). Reliability is assessed at the corridor level, defined as the set of services connecting the same OD pair for each leg of the journey. Assuming that passengers board the first arriving vehicle at stops, the indicators are derived from the combined service pattern of all lines operating within each corridor. This approach captures the effective OD-level reliability experienced by passengers, rather than the reliability of individual lines.

Preliminary results indicate that buses are perceived far more negatively than metro services, all else equal. Travelling one minute on the metro is equivalent to approximately 5 minutes in a bus, highlighting a strong preference for using metro. Similarly, trains and S-trains are perceived more attractive compared to buses. Transfer penalties are substantial, while transfers between trains are more acceptable than those including bus. Introducing reliability indicators, particularly headway adherence, reduces substitution rates, suggesting that the high reliability of metro services is a key driver of passenger preference and a substantial part of what might otherwise be considered an alternative-specific constant in favour of the metro. Schedule-based indicators, such as the maximum standard deviation across journey legs, better capture choices for other trains and buses. The metro operates at a very high frequency, without a public timetable, publishing only an approximate headway of 2 minutes. Therefore, passengers’ perceptions are best explained by the service regularity.

On-going work includes estimating and analysing a latent class model to capture heterogeneity among passengers with respect to reliability considerations. This method will allow us to identify passenger segments with different sensitivities to regularity and punctuality and assess whether a latent class approach improves model specification. By examining how reliability factors jointly or independently influence passenger preferences, this research aims to provide a deeper understanding of passenger route choice in multi-modal public transport systems.

References Bovy, P. H. L., Bekhor, S., & Prato, C. G. (2009). The factor of revisited path size: Alternative derivation. Transportation Research Record, 2076, 132–140. https://doi.org/https://doi.org/10.3141/2076-16 Cats, O. (2014). Regularity-driven bus operation: Principles, implementation and business models. Transport Policy, 36, 223–230. https://doi.org/10.1016/j.tranpol.2014.09.002 Chen, D., He, J., Lin, S., & Yang, Z. (2025). Passenger arrival patterns and its implications for bus operation: The impact of schedule reading behavior on average waiting times at bus stops. Transport Policy, 163, 310–322. https://doi.org/10.1016/j.tranpol.2025.01.021 Dixit, M., Brands, T., Van Oort, N., Cats, O., & Hoogendoorn, S. (2019). Passenger travel time reliability for multimodal public transport journeys. Transportation Research Record: Journal of the Transportation Research Board, 2673(2), 149–160. https://doi.org/10.1177/0361198118825459 Ingvardson, J. B., Nielsen, O. A., Raveau, S., & Nielsen, B. F. (2018). Passenger arrival and waiting time distributions dependent on train service frequency and station characteristics: A smart card data analysis. Transportation Research Part C: Emerging Technologies, 90, 292–306. https://doi.org/10.1016/j.trc.2018.03.006 Jenelius, E. (2018). Public transport experienced service reliability: Integrating travel time and travel conditions. Transportation Research Part A: Policy and Practice, 117, 275–291. https://doi.org/10.1016/j.tra.2018.08.026 Rezazada, M., Nassir, N., Tanin, E., & Ceder, A. ( (2024). Bus bunching: A comprehensive review from demand, supply, and decision-making perspectives. Transport Reviews, 44(4), 766–790. https://doi.org/10.1080/ 01441647.2024.2313969 van Oort, N. (2014). Incorporating service reliability in public transport design and performance requirements: International survey results and recommendations. Research in Transportation Economics, 48, 92–100. https: //doi.org/10.1016/j.retrec.2014.09.036

ABSTRACT. Transit investment has profound effects on people’s daily lives. Beyond changes in travel time and accessibility, new transit infrastructure can reshape travel behaviour, particularly through shifts in mode choice. When individuals move from driving to public transit, their travel patterns often incorporate more walking to and from stations, more predictable travel time, and lower travel-related stress. These changes can translate into improvements in physical activity, mental well-being, and overall working performance. As a result, perceived impacts of commute on quality of life may improve in ways that are not immediately apparent through standard transport metrics. Public transit investment may advance public health as well as they incorporate walking and reduce stress associated with commute. In this context, the Réseau Express Métropolitain (REM) in Montréal offers a timely opportunity to examine how a $9.4B CAD new light-rail system influences everyday life of its users. This study investigates whether and how the REM affects self assessed work performance, physical and mental health, and perceived quality of life. This study uses wave 5 of the Montréal Mobility Survey (MMS) collected in 2024 (N=3226), one year after the opening of the first branch of the REM. The MMS is a bilingual online survey providing detailed information on respondents’ travel behaviour, including commute mode, trip duration, trip purpose, and access and egress patterns. It includes residential and workplace locations, as well as socio-demographic characteristics such as age, gender, and income. The survey contains self-reported assessments of how commuting affects performance at work, physical and mental health, and overall quality of life. Respondents rated six outcome dimensions: energy, productivity, and punctuality at work, and physical and mental health, and quality of life, on a five-point scale ranging from Very negatively to Very positively. The sample used for this study has an average respondent age of approximately 43 years and is balanced across gender and income groups. On average, respondents report about one and a half days of teleworking per week and spend roughly seven hours per day sitting. The dominant commute modes are driving and metro, while REM serves as the primary mode for about 6% of the sample. Commute durations typically range from about 20 minutes to one hour. A substantial share of respondents self-identified as cyclists or pedestrians. Overall, the sample is well suited for investigating the impacts of the REM on work outcomes, health, and well-being. To analyze these relationships, we estimated an ordered probit model for each of the six outcome dimensions: energy, productivity, punctuality at work, physical health, mental health, and quality of life. A broad set of explanatory variables was considered, including trip characteristics (trip purpose, weekday versus weekend travel, parking time, and REM access and egress modes), travel experience (satisfaction with REM), and built environment indicators (job accessibility and the CanALE index). These variables were ultimately excluded from the final specifications due to limited statistical significance. Marginal effects were calculated to illustrate how changes in key explanatory variables influence the probability of reporting positive outcomes in work performance, health, and quality of life. One of the most important findings is that REM commuters report significantly better physical health outcomes compared to drivers. The positive effect of REM travel exceeds that of both the metro and bus. Relative to driving, REM commuting increases the probability of reporting positive physical health outcomes by 22.16 percent, compared with 20.14 percent for metro users and 11.29 percent for bus users all else equal. This effect is not statistically significant for working performance or mental health. The model shows that cyclists and pedestrian commuters experience the most positive outcomes across all six dimensions. Compared with drivers, active commuters report higher probabilities of positive impacts on energy at work (by 16.57 percent), productivity (by 9.34 percent), physical health (by 47.10 percent), mental health (by 25.51 percent), and quality of life (by 20.05 percent) ceteris paribus. These results reinforce the substantial benefits associated with active travel. With respect to working performance, teleworking slightly reduces perceived energy and productivity, while longer commute times significantly decrease both outcomes. Relative to drivers, respondents travelling primarily by walking or cycling report substantially higher energy, productivity, and punctuality. In contrast, commuters using public transit report similar or slightly lower levels of working performance compared to drivers. For health outcomes and quality of life, teleworking and long sitting times are associated with lower physical health. Transit commuters and active commuters report significantly better physical health than drivers, with active commuters showing the strongest effects. REM users report more positive physical health impacts than metro and bus users, which suggests that the REM may provide additional benefits beyond those of conventional transit. Active commuters experience significantly better mental health outcomes than drivers. Bus users report more negative commuting impacts on mental health than drivers, while REM and metro users do not differ significantly from drivers. A similar pattern appears for the overall quality of life measure. Longer commute durations are consistently associated with declines in perceived mental health and quality of life. Overall, the findings suggest that transit investments such as the REM generate meaningful health and well-being benefits that are often overlooked in traditional transport evaluation frameworks. The results indicate that improvements in physical health and quality of life may be more effectively supported through enhanced transit systems, which reshape daily travel behaviour in ways that advance broader public health objectives.

ABSTRACT. In-Person Presentation Preferred.

Motivation. Routine and scheduled events are central to public transport operations and often generate a significant amount of ridership. However, traditional event-related analysis based on automated fare collection (AFC) has key shortcomings. In many bus systems, passengers are not required to tap out when alighting, which means there are no direct records of whether riders get off near the event venue. After events, agencies often relax fare control to speed up passenger clearance, allowing people to board without tapping at all. In such settings, AFC data will structurally fail to capture the full event ridership. At the same time, automatic passenger counting (APC) systems provide rich stop-level records of boardings and alightings, but it is rarely used to model specific rider subgroups or to reconstruct origin–destination (OD) matrices. Being a one-dimensional count of “how many got on and off where and when”, it cannot distinguish between event and non-event riders who board and alight at the same location. This leads to a methodological blind spot: APC is only used as a tool for crowdedness and total loads rather than describing the travel patterns of specific events and rider groups, while most travel behavior analysis and OD matrix estimation frameworks rely on AFC trip-chaining. When fare card data are incomplete or biased, it becomes difficult for agencies to accurately estimate event-specific demand.

Framework. To address the research gap, we develop a systematic Event–APC–OD framework that identifies event-related ridership and rider origins, aiming to provide a higher-resolution understanding of event ridership and rider characteristics. First, we jointly leverage GTFS and APC to characterize the volume of the target subpopulation. Given an event with location and time, GTFS identifies the plausible bus stops, routes, and services. We then conduct large-scale comparisons of APC data between event days and matched non-event days, controlling for hours, weekday–weekend patterns, seasonality, and typical baseline demand. Second, we move beyond the APC-derived subgroup vectors into a probabilistic OD inference process. There is well-developed literature on OD inference, with iterative proportional fitting as a classical statistical method and more recent approaches in neural-network-based models such as variational autoencoders. In contrast to existing OD estimation methods that use AFC as the primary signal and APC only as a supplementary or validation source, our framework uses APC as the main demand signal and treats AFC as a partial, noisy observation of the same underlying flows.

Case Study. We have conducted a case study in Washington, D.C., with the Washington Metropolitan Area Transit Authority (WMATA) in a project to understand and improve student travel: School trips are conceptualized as a recurring “routine event” with fixed event locations and time. District of Columbia Public School (DCPS) K–12 students do not get a dedicated school bus system; instead, students receive free-travel smartcards on WMATA through the Kids Ride Free program. However, non-tapping behavior on buses is common, either students just ignore or operators waive tapping to speed up boarding for free riders. As a result, AFC data substantially underestimate student bus ridership. By contrast, Metrorail requires mandatory faregate taps, which inflates the relative share of rail trips in AFC-based OD reconstruction. Empirically, schools adjacent to Metrorail stations exhibit significantly higher AFC-recorded ridership than schools primarily served by Metrobus, even though field observations reveal comparable levels of crowding on school-related bus trips. Routing engine analysis similarly confirms the systematic bias and suggests that the majority of students’ optimal routes should be Metrobus-based.

Data Processing. Specifically, our analysis integrates WMATA’s GTFS, APC, and AFC datasets for August–September 2023, alongside DCPS calendar and bell-time schedules. We identify school-relevant stops and routes to build APC-based time series of trip-level loads. By comparing school days to baseline non-school workdays, we isolate the portion of ridership attributable to school commute events and then infer these patterns to OD flows. For after-school periods, we link boardings to plausible downstream alightings along each trip; for before-school periods, we identify feasible upstream origins into the observed alightings. APC serves as the primary constraint on demand, while both the underestimated AFC records and Metrorail faregate data are incorporated as prior information for model calibration.

Preliminary Results. The inferred student ridership by route aligns well with prior qualitative knowledge about school commutes in DC. We observe distinct ridership spikes around almost all major schools served by WMATA Metrobus, with magnitudes consistent with our field observations: for example, approximately 20–30 student riders on the first bus departing immediately after dismissal at a school with roughly 300 students. The inferred spatial OD patterns also match expectations: student origins are dispersed across residential neighborhoods but show strong concentrations at key Metrorail transfer stations. In busy commercial areas such as Columbia Heights, high volumes of non-school riders further contribute to bus loads. Overall, these patterns show corridor-level details that conventional APC analyses fail to capture.

Applications. The Event–APC–OD framework opens several practical opportunities for transit agencies. For student travel, WMATA can gain a more accurate understanding of actual student travel volume, travel segments, and distances, enabling better route planning and timetable optimization, in the absence of accurate AFC data. WMATA can also use the framework to improve operational practices around fare media and tap compliance. By estimating tap rates for different schools and comparing them with indicators such as mobile card usage or student card distribution, the agency can more effectively evaluate and pilot targeted measures to encourage more consistent tapping among free student riders.

Conclusions. Conceptually, this study reframes APC from a system-level counting tool into an instrument for analyzing specific rider groups (DCPS students) in ways typically attempted only with smartcard data. Methodologically, it extends APC beyond ridership reporting to OD reconstruction for a defined subpopulation, showing that under appropriate event and network constraints, APC can produce realistic OD estimates even when AFC data are incomplete. This provides agencies with a new approach for understanding event-driven ridership and rider origins, supporting more informed service and planning decisions.

ABSTRACT. Presentation preference: In-person

Introduction In public transit research, the concept of Level of Service (LOS) is the most widely recognized framework for assessing service quality. One main measure for assessing LOS in terms of transit availability in time and space is the service frequency. Increasing service frequency reduces the average waiting time for passengers, making the service more attractive. However, higher frequency also increases operational costs, as more buses must be dispatched per unit time, and in some cases may require a larger fleet. A bus route should operate under the condition where the tradeoff between the operations cost and average passenger wait time cost is minimized, a principle known as the “square root dispatching policy” proposed by Newell (1971). Passenger cost depends on the number of passengers arriving within a headway and their average waiting time. Operator cost is the dispatching cost per round trip multiplied by the dispatch rate, or equivalently the cost per round trip divided by the headway. The total cost per unit time is the sum of passenger and operator costs, and the optimal dispatching rate is obtained by minimizing this total. In practice, transit agencies often face resource constraints that prevent them from operating at an optimal fleet size, resulting in longer waiting times for passengers. In such scenarios, the agency implicitly defines a value of waiting time (VoWT) for passengers, which may differ from how passengers perceive waiting time, as this perception is shaped by factors such as income, trip purpose, and demographic characteristics. When the implied VoWT is lower than the perceived value, it indicates that the service falls short of expectations and points to potential areas for service improvement. By accessing network-wide Automatic Passenger Count (APC) equipped with Automatic Vehicle Location (AVL) features, this study aims to evaluate service quality for Calgary Transit through a network-level analysis of implied VoWT. Calgary Transit operates the fourth-largest transit network in Canada, serving over 100 million trips in 2024. The network covers 206 communities with diverse population densities, making it essential to evaluate LOS from both passenger and operator perspectives. To this end, the study proposes a framework that quantifies LOS based on the gap between the implied and perceived VoWT. Furthermore, the stochastic effects of passenger demand on the implied VoWT are incorporated through a novel formulation. The results can be used to optimize dispatching policy by highlighting areas where buses can be reallocated from overserved to underserved routes while maintaining the total fleet size.

Methodology The methodology begins by estimating the passenger demand per unit time for each bus route. In real-world operations, most bus routes operate as many-to-many services, with multiple stops serving as both boarding and alighting passengers. To unify these cases, stop-level demands are shifted backward by their corresponding travel times from the route origin. This adjustment allows the calculation of route-level demand per unit time. Using APC data, which provide stop-level boarding counts for each trip, the cumulative passenger demand at the origin is obtained by summing all boardings up to a given time. Since passengers board only at scheduled departures, the cumulative demand curve takes the form of a step function. To allow differentiation, this step function is smoothed into a continuous curve through a continuum approximation, from which the passenger arrival rate over time can be derived. This continuous demand can then be used to inform analytical models for optimal headways and the value of passenger waiting time. Since the operational headway in service may differ from the optimum, using it in the “square root headway” formulation while treating VoWT as an estimable parameter reveals the implied VoWT. By applying Lindley’s theorem (Lindley, 1965), the stochasticity of passenger demand is also incorporated in calculating implied VoWT using APC data collected over time. By distinguishing between day of the week and different periods within a day, demand variability is quantified through the mean and standard deviation of passenger demand rates. The final formulation shows that the expected waiting time increases slightly with variation in passenger demand, while its variance scales with the square of demand variability.

Results The analysis so far is based on ten weeks of APC data from Route 1, which spans east–west through downtown, and BRT Route 307, which provides express service between eastern Calgary and downtown. The results show that Route 1 performs well, with implied VoWT consistently exceeding the perceived value across all periods, and with a larger gap during off-peak hours. In contrast, Route 307 underperforms during peak periods, particularly the morning peak. This is noteworthy because Routes 1 and 307 operate with similar headways, and Route 1 also carries higher demand. However, Route 307 exhibits consistently lower demand variability per unit time, leading to lower mean and variance in implied VoWT compared with Route 1. This highlights the importance of considering demand stochasticity, as routes with consistent passenger demand per unit time may require improved service to maintain performance. Further analysis will extend to six BRT routes and twenty high-ridership bus routes. To ensure network representativeness, the selected routes include between-regional services, such as those feeding downtown across all quadrants and crosstown services connecting the quadrants, as well as within-regional services operating locally in each quadrant. The expected results are as follows: 1- The distribution of the implied VoWT for the routes, capturing the stochastic effect of demand variability. 2- LOS measures across weekdays and weekends, as well as across different time periods within a day. 3- A unified and aggregated LOS index to evaluate service quality in providing mobility both between and within city regions. 4- Optimization of the dispatching policy to reallocate fleets between and within regions, throughout weekdays and weekends, and across different time periods, while maintaining the total fleet size.

References Lindley, D. V. (1965). Introduction to Probability and Statistics from a Bayesian Viewpoint (Vol. I: Probability; Vol. II: Inference). Cambridge: Cambridge University Press. Newell, G. F. (1971). Dispatching Policies for a Transportation Route. Transportation Science, 5(1), 91–105. https://doi.org/10.1287/trsc.5.1.91.

ABSTRACT. We are undecided to attend the conference in person or virtually.

Transit service reliability is a cornerstone of public transportation performance, shaping passenger satisfaction, ridership retention, and overall confidence in transit systems. When buses arrive significantly earlier or later than expected, riders encounter uncertainty, missed transfers, disrupted schedules, and diminished trust in the service. These reliability issues not only affect individual travel experiences but also influence broader perceptions of transit as a dependable mode of transportation. With the growing availability of app-based reporting and survey tools, it has become increasingly feasible to examine how passengers perceive reliability and how these perceptions correspond to operational performance measures commonly used by transit agencies. This study investigates the relationship between rider-reported reliability perceptions and a diverse set of operational indicators, including on-time performance (OTP) and schedule deviation metrics. OTP remains one of the most widely adopted measures for evaluating reliability. Because OTP is widely measured by transit agencies at route level, stop-level and route-level OTP measures were incorporated to evaluate their explanatory power regarding user experience. These indicators were analyzed alongside rider perceptions to identify which operational metrics best align with passengers’ perceptions. The study focuses on stop–route-level conditions in Winnipeg, using rider-reported assessments and operational data collected between September and December 2022. This level of granularity was chosen because a single stop often serves multiple routes that differ substantially in reliability, frequency, headway consistency, and operational context. Aggregating data solely at the stop level can obscure these differences and weaken the correspondence between measured performance and actual rider experience. By linking operational and perception data at the stop–route level, this study provides a more accurate representation of the conditions riders face when boarding buses serving specific routes at specific stops . Passenger perceptions were collected through the a popular transit mobile application (Transit App), which allows riders to report whether a bus arrived later than expected, more or less on time, or sooner than expected. These responses were aggregated at the stop–route level to capture service-specific reliability rather than generalizing across all routes serving a stop. Operational performance data were derived from scheduled and observed transit operations using Winnipeg Transit’s vehicle location system records and General Transit Feed Specification (GTFS) datasets. Reliability indicators captured the magnitude of schedule deviation, the distribution of early, on-time, and late trips, and variability in arrival times. Additional route-level and trip-level metrics were incorporated to determine whether broader operational patterns contributed to rider perceptions. Two analytical approaches were employed. First, correlation analysis was conducted to identify pairwise associations between operational reliability indicators and rider perception. Second, ordinary least squares (OLS) linear regression models were used to evaluate the explanatory influence of operational indicators on each perception outcome. The analysis focused on three core perception categories (i.e., dependent variables) reflecting rider assessments of service punctuality. These categories correspond to the proportion of trips perceived as later than expected, more or less on time, and sooner than expected. The correlation analysis revealed consistent relationships between operational delay patterns and rider perceptions. Rider responses indicating that buses arrived later than expected showed moderate positive correlations with indicators representing late service and delay magnitude. For instance, both the proportion of late trips and the average absolute schedule deviation were positively correlated with perceptions of late arrivals. Conversely, these responses showed weaker negative correlation with on-time performance measures. These findings suggest that riders’ perceptions are associated by late service conditions at the aggregate stop-level of analysis, even when their assessments are subjective and based on personal expectations rather than precise time measurements. Perceptions of more or less on-time arrivals displayed the opposite pattern, showing positive associations with OTP and negative associations with delay-related indicators. This indicates that passengers tend to classify service as on time when schedule deviations are relatively small and consistent across trips. The weakest associations were observed for perceptions of sooner than expected arrivals, which showed only small positive correlations with indicators representing early service. This may reflect the lower frequency of early departures during winter operations or reduced passenger ability to detect early service when arriving close to the scheduled departure time. The initial results of the models further support these observations. Operational reliability indicators were significant predictors of users’ perceptions. For example, larger delays, higher proportions of late trips, and greater schedule variability significantly increased the likelihood that riders reported a bus as later than expected. In contrast, stronger on-time performance and lower deviation magnitudes significantly increased the proportion of reports indicating that service was more or less on time. Models predicting perceptions of early arrivals demonstrated weak explanatory power, consistent with the weak correlation results. Although some route-level and trip-level performance indicators contributed to explanatory power in certain models, stop–route level reliability measures remained the most influential predictors, reinforcing the premise that transit reliability is experienced at a precise, detailed point of service. Future work will involve applying and testing more advanced models, such as multilevel mixed-effects models at the route or neighborhood level, random coefficient models, and conducting sensitivity analyses to quantify differences among reliability measures in predicting user perceptions. These efforts will help transit agencies identify which metrics most effectively capture rider experience and guide improvements in reliability monitoring and service planning. In summary, the initial findings indicate that rider perceptions align closely with observed reliability conditions and that app-based surveys effectively capture key elements of operational performance. This research will underscore the value of integrating user-generated reliability perceptions with detailed operational metrics at the stop–route level. By aligning analytical units with how riders actually experience transit service, the study will provide a stronger foundation for identifying reliability issues, prioritizing operational improvements, and enhancing rider-centered performance monitoring frameworks. Making rider perceptions visible and actionable offers transit agencies a powerful tool for improving service quality, increasing passenger satisfaction, and strengthening public confidence in transit systems.

ABSTRACT. Public transit agencies increasingly rely on automated data sources to understand operational performance and improve service delivery. While the General Transit Feed Specification (GTFS) and GTFS-Realtime formats provide unprecedented access to scheduled and real time information, significant challenges remain in transforming raw feeds into reliable indicators that support planning, performance monitoring, and operations control. This study develops and applies a reproducible methodology for analyzing bus service reliability using both static GTFS and GTFS-Realtime data, demonstrating how automated transit data can be used to evaluate schedule adherence, headway regularity, and spatio temporal operating patterns. The proposed framework is applied to the D8 route operated by the Washington Metropolitan Area Transit Authority (WMATA) in Washington, DC. The D8 route is a high demand urban bus service connecting the Washington Hospital Center and Union Station, traversing dense activity centers and complex traffic environments. These characteristics make it an ideal test case for evaluating how automated data can reveal reliability issues that are not apparent through scheduled data alone. Using fifteen consecutive days of real time data (June 1–15, 2024), the study integrates approximately 28,000 static stop time records and more than 170,000 real time vehicle location updates. A preprocessing workflow was implemented to prepare the raw datasets for analysis. A custom service_date construction was applied to assign trips that begin after midnight or span multiple calendar days. In addition, scheduled direction identifiers and shape geometries from static GTFS were merged with real time feeds to produce a unified dataset containing both spatial and operational attributes. Filtering procedures, such as removing trips with incomplete stop sequences or unrealistic movement patterns, were applied to ensure that interpolated trajectories represent actual service conditions. These technical refinements were necessary to remove error-prone trips and to prevent the propagation of anomalies during subsequent interpolation and headway calculations. A spatio temporal interpolation algorithm was applied to estimate actual stop level arrival times from continuous waypoint streams. For each trip, waypoints were projected onto the corresponding route shape to compute distances along the path, enabling linear interpolation between nearest trajectory points. This produced approximately 50,000 interpolated arrival times across more than 1,200 valid trips, fully disaggregated by direction, stop, day type, and time period. By anchoring all calculations to the scheduled route geometry, the method generates consistent and interpretable arrival time estimates that can be compared to schedule expectations. Service reliability was evaluated using two primary metrics: on time performance (OTP) and headway regularity. OTP was measured as the difference between scheduled and interpolated arrival times, with arrivals within ±5 minutes classified as on time. Headways were computed as time differences between consecutive trips at each stop, allowing assessment of variability, bunching behavior, and extreme service gaps. These indicators collectively capture both schedule based and frequency based reliability dimensions. Results of OTP analysis show notable and recurring discrepancies between scheduled and actual service. Across all observations (n = 50,625), the mean deviation between scheduled and interpolated arrival times was 2.44 minutes late, with a standard deviation of 5.49 minutes. Directional differences were prominent: southbound trips exhibited a mean lateness of 3.09 minutes and greater variability, compared to 1.85 minutes for northbound trips. Stop level analysis indicated that on time performance declined toward the terminal stops in both directions, suggesting accumulated delays along the route. Day of week comparisons showed that Saturday service was consistently the least reliable, while weekday northbound trips were the most stable. Headway analysis revealed irregularity. Across 49,425 observed headways, the mean interval was 25.49 minutes, exceeding the scheduled 20 minute target, with a standard deviation of 15.28 minutes. Approximately 24 percent of headways exceeded 30 minutes, while nearly 6 percent fell below 10 minutes, indicating occurrences of bus bunching. Maximum headways exceeded 150 minutes across all major stops, and midday hours exhibited high variability. Trajectory plots using time-space graphs illustrated bunching patterns where slower trips were overtaken by following vehicles, as well as divergence patterns where fast trips created expanding service gaps. This study demonstrates multiple operational insights: reliability varies substantially by direction, location along the route, and time of day; late night and weekend service are particularly unstable; and systemic headway volatility suggests that schedule based dispatching may be insufficient for maintaining consistent service. These findings underscore the importance of integrating automated data into continuous monitoring tools rather than relying solely on scheduled information. The research contributes a practical, open methodology for transforming GTFS data into actionable performance indicators. The workflow is fully scalable, transparent, and suitable for adoption by transit agencies, researchers, or third party analytics platforms. It directly supports applications in strategic planning, service evaluation, operations control, and equity analysis. As agencies continue to expand their use of real time sensors, the techniques introduced here can help bridge the gap between growing data availability and operational decision making. Future extensions could incorporate crowding or passenger load data, apply machine learning methods for anomaly detection, or link external data sources such as roadway congestion or weather information to identify causal factors behind unreliability.

ABSTRACT. Bus travel times are highly dependent on traffic conditions, which can vary daily and create significant stress for drivers and frustration for passengers. Implementing an Automatic Vehicle Location (AVL) system gives transit agencies the ability to better monitor and manage travel times. Collecting billions of data points is valuable—but only if we can transform them into actionable insights. In this presentation, we will showcase a series of Power BI visualizations designed for a funnel-style analysis of schedule adherence: • Step 1: Identify the main reliability gaps across the network. • Step 2: Pinpoint recurring causes behind problematic trips (e.g., insufficient time allocation, delays caused by previous trips). • Step 3: Drill down into specific details such as weekday patterns, variation by driver, or dates. With this visualization tool, we are now equipped to perform a comprehensive diagnosis and take immediate, informed action!

(Do not hesitate to contact us if you want more details or visuals).

ABSTRACT. This paper presents a stop-level simulation framework for measuring transit line reliability using AVL and APC data. The framework quantifies reliability from the passenger perspective by estimating expected waiting times under observed headway irregularity and crowding conditions. For each stop, AVL data provide vehicle arrival times and APC data supply boarding and alighting counts, enabling a microsimulation of passenger arrivals, queue formation, and capacity-constrained boarding. The resulting waiting times yield a stop-level reliability index that reflects both headway variability and the effective loss of capacity due to crowding.

The framework is demonstrated using a TTC bus stop on Line 29, with simulations performed for multiple time periods and benchmarked against fully regular, Erlang-2, and exponential headway distributions. Results show that the observed headway pattern exhibits moderately irregular behavior, lying between the Erlang-2 and exponential benchmarks. The study also describes how the estimated reliability parameters can be incorporated into transit assignment models through a crowding-dependent effective headway formulation. The full paper will extend the analysis to the entire TTC network and evaluate the impacts of reliability-adjusted headways on regional transit assignment outcomes.

ABSTRACT. Bus services in Indian cities operate under persistent mixed-traffic conditions, where schedule-based performance measures provide limited insight into passenger-experienced reliability. This study proposes a GPS/Automatic Vehicle Location (AVL)–based analytical framework to quantify headway reliability and bus bunching on high-frequency corridors in an Indian megacity. The analysis is designed around commonly observed operational conditions, focusing on corridors with scheduled headways of 3–10 minutes, where reliability failures are most consequential for passengers.

The framework reconstructs vehicle trajectories from raw GPS feeds to compute headway-based reliability indicators, including the coefficient of variation of headways, the share of headways exceeding ±50% of scheduled values, and the spatial extent of bunching episodes, defined as headways falling below 50% of the scheduled headway for sustained segments. Space–time analysis is used to track the formation and persistence of bunches, while terminal recovery effectiveness is evaluated by comparing inbound and outbound headway dispersion. These metrics enable systematic assessment of congestion effects, terminal constraints, and limited headway control under mixed-traffic operations. By explicitly addressing data noise, missing pings, and irregular dispatching, the study offers a transferable, agency-ready approach for diagnosing reliability failures using automated data, with direct relevance for operations control and performance monitoring in large, mixed-traffic transit systems.

ABSTRACT. The electrification of public transit systems has emerged as a pivotal strategy for promoting sustainable urban mobility, mitigating greenhouse gas emissions, and reducing the environmental impacts of conventional diesel-powered bus fleets. Transitioning to electric buses (Ebuses) represents a fundamental shift in urban transit operations, offering the dual benefits of lower operational emissions and quieter service. However, this transition also introduces a set of unique operational challenges that transit agencies must address to ensure efficient and reliable service. Key among these challenges are the inherent limitations of Ebus battery capacity, variability in energy consumption due to route characteristics and operational conditions, and the need to optimize fleet utilization to avoid unnecessary capital expenditures and service disruptions. Effective management of these factors is essential for minimizing both operational costs and energy waste (unused) while maintaining the service quality expected by urban passengers. One promising operational strategy that has the potential to address these challenges is bus interlining. Interlining refers to the practice whereby a single bus sequentially serves multiple routes during a single operational block, rather than being dedicated to a single route for the duration of its service period. This operational approach can enhance energy efficiency by reducing residual (i.e., unused) battery energy. By allowing buses to complete cycles on different bus routes before returning to a depot or charging station, interlining leverages residual battery capacity that would otherwise remain unused, thereby increasing the overall utilization of each bus and reducing the total fleet size required to maintain service. Despite the recognized theoretical benefits of interlining, the literature lacks a systematic and quantitative evaluation of its impact on Ebus fleet sizing, particularly when considering the complex interactions between route characteristics, energy consumption, and vehicle-specific parameters. This study addresses this gap by developing a data-driven framework to evaluate the effects of interlining on Ebus fleet requirements. Central to this framework is the formulation of an Integer Linear Programming (ILP) model designed to minimize the total fleet size necessary to meet daily service demands while maximizing the effective use of each bus’s battery energy. The model explicitly accounts for operational constraints, route-specific characteristics, and vehicle attributes, providing a realistic representation of the operational environment faced by transit agencies during the electrification process. Two operational scenarios are analyzed: Dedicated Operations (No Interlining) and Interlined Operations (Optimized Interlining). In the dedicated operations scenario, each bus is exclusively assigned to a single route and continues service on that route until the battery is depleted and the bus energy is not sufficient to complete a cycle on that route, after which the bus must be withdrawn from service for charging. Under this scenario, fleet size requirements are determined by summing the minimum number of buses needed to satisfy the service demand for each route independently, without considering opportunities for energy sharing or operational overlap. This conventional approach provides a baseline against which the benefits of interlining can be evaluated. In contrast, the interlined operations scenario enables buses to serve multiple routes sequentially, thereby optimizing residual battery usage and reducing the need for additional vehicles. The ILP model incorporates key parameters such as route length, energy consumption rate, headway, battery capacity, and usable capacity factors, as well as external considerations such as seasonality and the operational characteristics of individual buses. The optimization objective is to minimize the fleet size required to complete all scheduled service cycles while adhering to battery limitations. Importantly, the model also accounts for the strategic placement of operational hubs, such as subway stations or major transfer terminals, which facilitate interlining by minimizing deadheading distances and avoiding any unnecessary energy consumption due to deadheading. To rigorously examine the interactions between operational and vehicle-level factors, a full factorial analysis is conducted, assessing the combined effects of route length, energy consumption rate, battery capacity, and service headway on fleet size requirements under interlined operations. The factorial design enables a comprehensive evaluation of both main effects and interaction effects, offering insights into how different combinations of factors influence fleet efficiency using real-world data for Ebuses operations in the City of Toronto. The analysis incorporates operational data for three Ebus types (BYD, Proterra, and New Flyer), capturing their distinct energy consumption profiles. The study focuses specifically on interlining opportunities for routes that have a subway station as a terminal station, evaluating all terminals along Lines 1, 2, and 4 within the city. The findings of the study reveal that interlining can yield substantial reductions in fleet size (10-15% compared to dedicated operations) while maintaining service levels. Notably, pairing long routes with short routes produces the greatest efficiency gains, as buses are able to utilize remaining battery capacity from the longer route when transitioning to a shorter route. Similarly, interlining routes with differing energy consumption profiles (high energy consumption with low energy consumption) also reduces fleet requirements, highlighting the importance of considering complementary operational characteristics when designing interlining schemes. Additionally, the analysis demonstrates that interlining routes with low service frequency and routes with high service frequency further enhances fleet efficiency, providing opportunities for transit agencies to optimize resource allocation across diverse service demands. These results highlight the potential for strategic interlining to reduce both capital investment and operational costs, while simultaneously improving the energy efficiency of Ebus operations. This study provides a structured methodology that integrates interlining into Ebus fleet planning. By combining ILP optimization with empirical analysis of route-specific energy consumption, the framework offers an approach for assessing fleet size requirements under complex operational scenarios. The research highlights the critical role of operational hubs in facilitating interlining. Moreover, the analysis provides insights for transit planners, demonstrating that optimal interlining strategies depend on careful consideration of route length, energy consumption rates, and service frequency, and that suboptimal route pairings can significantly limit potential efficiency gains. Overall, the findings demonstrate that interlining represents a viable and effective approach for enhancing the operational efficiency of Ebus networks. By reducing fleet size requirements, minimizing energy waste, and optimizing battery utilization, interlining supports cost-effective transit electrification and advances the sustainability of urban mobility systems.

ABSTRACT. Battery-electric buses (BEBs) are becoming a core strategy for transit agencies aiming to reduce carbon emissions while promoting mobility. However, natural hazards such as coastal flooding and power outages can undermine these benefits by disrupting transit bus operations. This study proposes a Garage-Level Resilience Index based on the cost of system impact due to disruption and the total recovery effort associated with mitigation measures. This index quantifies and compares how individual bus garages responsible for BEB charging would withstand and recover from the impacts of natural hazards. A case study is conducted for the impact of Hurricane Sandy, assuming a full fleet conversion to BEBs. Historical data of GTFS (General Transit Feed Specification) from Sandy period are used to estimate how flooding affects bus garages and their associated routes in service disruption, canceled trips, and affected riders. On the other hand, the post-flooding power outage poses more severe constraints for BEB fleets. Within this context, the benefits of mitigation options, including diesel generators and solar panels with battery storage are evaluated using the Garage-Level Resilience Index. The results can be used for prioritizing resource allocation and recovery effort across garages for hazard-aware planning of BEB transitions.

ABSTRACT. This study develops a comprehensive four-stage optimization framework to address the joint siting and operational management of opportunity charging stations shared by electric buses and private electric vehicles (EVs) within an urban environment. The proposed model captures the hierarchical interactions between strategic infrastructure planning, operational pricing, priority-based bus charging scheduling, and responsive private EV user behavior.

In the first stage, a mixed-integer linear program determines optimal charging station locations, balancing investment costs, operating profits, and passenger waiting penalties while ensuring complete spatial coverage of bus routes. The second stage introduces a Stackelberg game structure, in which the charging operator (leader) sets dynamic electricity prices to maximize revenue, anticipating the price-elastic charging response of private EV users (followers) under capacity and regulatory constraints. The third stage models individual EV user charging as convex utility maximization problems, reflecting heterogeneous behavioral and technical characteristics. A coordinated, iterative solution strategy integrates dual decomposition and equilibrium modeling to ensure consistency across strategic, tactical, and user-level decisions. The fourth stage optimizes bus opportunity charging schedules through a linear programming model that minimizes additional passenger waiting time and guarantees energy sufficiency for each route.

A real world case study based on a representative urban bus network demonstrates the framework’s effectiveness in coordinating public and private charging activities. Two electric bus routes and ten random private cars are analyzed, incorporating empirical data on passenger occupancy, time-dependent energy consumption, and station accessibility. Simulation results show that the integrated optimization maintains service reliability under varying demand conditions. The results confirm the model’s capability to guide equitable and efficient deployment of shared charging infrastructure, supporting both operational and sustainability objectives. This research provides actionable insights for urban planners, transit operators, and policymakers seeking data-driven strategies for electric mobility planning and operation, bridging theoretical modeling and practical implementation in sustainable transport systems.

ABSTRACT. Context & Motivation

Fleet electrification introduces new operational uncertainties for transit agencies. Range anxiety, for example, was never a major concern with conventional or hybrid buses, whose large fuel reserves provided more than sufficient autonomy. In contrast, electric buses (EBs) require substantial upfront investment, offer a shorter and more sensitive driving range, and are supported by a market whose operational tools are still maturing. As a result, a fundamental gap is introduced which requires new decision-support systems to facilitate the deployment and increase the reliability of EBs within rapidly changing daily conditions.

From both our operational experience and recent literature, a wide range of factors influence EB energy consumption and therefore bus range. Weather conditions (such as temperatures and snowfall) (Liu et al., 2018), passenger load (Vepsäläinen et al., 2019), driver behavior (Blagojević et al., 2020), traffic congestion, and route characteristics (Li et al., 2021) can all positively or negatively affect the range of an electric bus. Thus, there is a need for real-time state of charge (SoC) supervision when EBs are in operation to limit service interruption and minimize impacts on client satisfaction.

At the Société de Transport de Laval (STL), a mid-sized transit operator in the province of Québec, Canada, we have been operating 10 New Flyer Xcelsior 1st generation electric buses since 2021, with plans to acquire only electric buses going forward, starting with 26 Nova LFSe EBs entering service in 2026 for a total of 70 EBs at the end of 2027. As the proportion of EBs in our mixed fleet grows, the need for optimized operational planning becomes increasingly critical. Currently, daily operations struggle to fully use electric buses in operation due to limited visibility of their expected energy requirements in real-world conditions.

To address this gap, STL initiated a project to develop a predictive, real-time decision-support tool capable of estimating the remaining autonomy of each EB in operation as in Dong et al., (2025). By combining the vehicle’s SoC with expected operating conditions and route characteristics, we developed a machine learning model that predicts the energy required to complete the remaining portion of the assigned run. Ultimately, this tool aims to support more confidently and proactively EBs deployment while preserving service reliability and eliminating the risk of buses depleting their battery during service.

Methodology

The primary dataset consists of detailed telemetry collected from STL’s electric buses beginning in January 2024, totaling 1,381,537,041 data points and covering approximately 12 600 completed trips over a 20-month period. These data streams are ingested daily, cleaned, and engineered to create an initial training dataset at the trip level. An XGBoost model is trained to estimate the energy required to complete each trip based on weather conditions, trip and block characteristics, time of day, and commercial speed. Weather emerges as the dominant driver of consumption variability, as temperature directly influences auxiliary loads such as heating and cooling (HVAC). In addition, snowfall and slippery road conditions trigger the ABS system, which temporarily disables regenerative braking, thereby increasing overall energy consumption. These processes are illustrated in the offline pipeline in Figure 1.

After the initial training phase, the model is deployed in production using real-time data from multiple sources. Environmental information is obtained through a live weather API that provides current and forecasted temperatures. Vehicle SoC is retrieved through a real-time MQTT stream connected to the onboard telematics system. Trip and block metadata are sourced from STL’s CAD/AVL system to contextualize each prediction. These processes are illustrated in the online pipeline in Figure 1. Energy consumption is first predicted at a fine temporal resolution (trip level with model A in Figure 1). These trip-level predictions are then aggregated at the block level to estimate the total energy needed to complete the remaining portion of the block. A second model subsequently integrates bus-specific information (namely the vehicle ID, current SoC, and predicted energy requirement) to compute the expected SoC at the end of the block (model B in Figure 1). This two-stage approach allows the system to estimate the energy needed independently of battery capacity or state of health (SoH), while the second model adjusts the final SoC prediction using intrinsic vehicle characteristics.

A dashboard displaying the predicted end-of-run SoC is provided to operations staff, along with block-level information to support timely decision-making, including the reassignment of buses when necessary.

Figure 1: High-level pipelines (offline and online) used to generate predictions and feed the users’ dashboard

Results

The current operational method, based on a static historical consumption with conservative security margins, yields errors ranging from the 30–50%. A simple baseline using average consumption per period (season) reduces error to 15–20%. However, this remains insufficient when compared to operational planning targets. Our model achieves on average sub-9% error at the block level, representing a significant improvement over both existing and baseline approaches.

The model is currently being integrated into the production ecosystem, and full operational metrics will be collected and presented at the conference. Ongoing work focuses on enhancing prediction accuracy by incorporating additional features, including driver-behavior embeddings and routes structural representations, which are expected to further improve generalization across vehicles, routes, and conditions.

Contributions

Our work makes four main contributions: 1) the development of a real-time prediction pipeline integrating telemetric, operational, and environmental data; 2) the introduction of a two-stage trip-level modeling approach suited for transit operations improving generalization across EBs; 3) a comparative evaluation demonstrating substantial gains over baseline methods (from current operational practice of 30–50% to under 9%); and 4) a real-time prototype dashboard that operationalizes these predictions for planners and dispatchers enabling proactive decision making and reduce operational risk related to EB range.

We aim to expand this work by improving feature engineering and by applying the model-generated data to other use cases, particularly at dispatch, to proactively assess a bus’s ability to complete its run with the current state of charge.

ABSTRACT. The transition to battery electric buses (BEBs) is reshaping public transit planning by introducing new operational challenges related to vehicle range limitations, charging coordination, and fleet scheduling. This study presents a data-driven framework for optimizing electric bus allocation and charging schedules, integrating detailed service data with operational and energy considerations. A mixed-integer optimization model is developed to jointly determine the assignment of BEBs to daily operating blocks and the corresponding charging strategies that ensure continuous, reliable service. Using General Transit Feed Specification (GTFS) data, the framework captures realistic operating conditions, including trip durations, interlining structures, and block configurations, while incorporating parameters such as battery capacity, charger power, and required charging times. The Toronto Transit Commission (TTC) network serves as the empirical case study, focusing on the ten routes with the highest service intensity and number of operating blocks. Results show that most blocks can be electrified using moderate-capacity batteries when opportunity charging is provided at key terminals or layover locations. However, longer or high-frequency blocks may require mid-day charging, larger battery capacities, or schedule adjustments to maintain feasibility. Sensitivity analyses demonstrate that strategic optimization of charger placement, power levels, and fleet composition can substantially reduce operational costs and minimize service disruptions.

ABSTRACT. Willingness to pay (WTP) for public transit is a major contributor to its usage as it reflects individuals’ valuation of the transit service being offered. Understanding WTP for public transit can provide policymakers with strategies to face financial difficulties experienced by public-transit agencies. Still, to our knowledge, no previous studies has employed WTP within a public-transit market segmentation approach to understand its distribution within the population. This study contribute in filling this gap by conducting a combined factor-cluster analysis and thematic analysis using WTP data from the 2023 Montréal Mobility Survey (MMS).

The MMS is a detailed survey collecting data on travel behaviour, travel attitudes, social perceptions and detailed sociodemographic characteristics in the Montréal region, Canada. It also incorporates weights based on gender, age, household income and travel behaviour to ensure representativity. In 2023, the MMS included a question asking respondents to state the maximum amount they would be willing to pay for a single transit trip. A complementary open-ended question on perceptions of transit funding and pricing and was also collected. From the complete clean MMS 2023 sample (n = 5,284) we applied various additional data cleaning steps including removing extreme WTP values (defined as above $8 based on sample distribution) as well as removing older adults (65+ years old) as they benefitted from a variety of local free fare initiatives which would have likely impacted their WTP. Our final sample of analysis was of 4,024 adults aged 18 to 64 years old. To account for actual fare paid in our analysis, we calculated individualized fare cost per trip for public-transit for each respondent based on their reported fare type used, transit usage (for monthly passes) and zone of residence. We then calculated the difference between the maximum WTP and the average fare paid per trip for each user.

To analyze the data, we first conducted a factor analysis to generate two factors: spatial access (combining distance to the nearest metro station, distance to downtown and accessibility by public transit to jobs within 60 minutes) and transit perceptions (combining overall satisfaction with public transit and willingness to recommend transit to friends and family). Both factors were then employed in a weighted k-means clustering process alongside the difference between maximum WTP and fare paid as well as public-transit mode share for the last week. The weighted clustering process revealed four unique groups: frequent transit users (25%), active travellers (34%), transit critics (21%), and suburban car drivers (20%). In terms of spatial distribution, suburban drivers were mostly distributed on the periphery of the region while the three other clusters were heavily concentrated on the Island of Montreal.

Frequent transit users have the highest average transit modal share (60.1%) and the highest WTP surplus ($1.64) driven both by having the highest maximum WTP ($3.71) and the lowest average cost of transit per trip ($2.07). This cluster was also the second highest in terms of public perceptions and had the third highest level of spatial access. In terms of sociodemographic characteristics, frequent transit users have the most student (34.3%), the most recent immigrant (11.6%), the most women (57.4%) and are lower income than the overall sample.

Active travellers have the best perceptions of transit and the highest level of spatial access with a small WTP surplus ($0.23). While they have below average transit modal share (11.6%), 97.4% of active travellers had taken transit in the last year and 95.2% of those were planning to keep using it over the next year. Active travellers had the highest active travel (e.g., walking and cycling) modal share (60.8%). Compared to the entire sample, active travellers have the lowest share of women (51.7%), they are slightly older (44.3 years old), they have less students (7.4%) and are of higher household income status than the average.

Transit critics have strongly negative perceptions of public transit, despite having the second highest levels of spatial accessibility. Transit critics have the lowest maximum WTP for public transit ($2.98) contributing to a WTP deficit of - $0.54. They have below average public-transit modal share (9.4%) and above average car modal share (65.4%). Compared to the full sample, transit critics have less women (51.8%), and students (11.1%) and are lower income than the overall sample.

Suburban drivers are primarily characterized by having the lowest level of accessibility by public transit, the lowest transit modal share (8.0%) as well as the highest average fare per trip ($4.88) driving resulting in a -$1.33 WTP deficit. This cluster is travelling predominantly by car (77.7% modal share) which they perceived the most as being essential to their daily travel (91.2%). Suburban drivers are the oldest (46.2 years), they have the least students (6.6%) and the least recent immigrants (4.6%) and they have the highest household income of all clusters.

To complement the weighted clustering, applied thematic analysis was conducted at the cluster level on open-ended question responses provided by a subset of participants (n = 702). Two primary themes were highlighted, being perceptions of transit and affordability. In terms of public transit perceptions, criticisms of service quality and transit funding management were common for suburban drivers and transit critics. Conversely, perceptions of transit being an essential, socially beneficial service were predominantly shared by frequent transit users and active travellers. These combined findings highlight improved service quality and communication as key leverages to promote higher WTP for transit critics and suburban drivers. Nonetheless, having affordable transit was the most prevalent theme across all clusters, highlighting the limits of WTP-centric approaches under increasing fare costs. Affordability is particularly important considering that the cluster with the highest WTP surplus (frequent transit users) is of lower income, meaning that increases in fare costs are likely to be vertically inequitable. This study contributes to the literature through the integration of WTP in a market segmentation approach enabling a more nuanced understanding of the drivers of WTP. Findings from this study can be of value to policymakers aiming to adapt transit pricing policies and to advocate for funding reforms.

ABSTRACT. Mobile ticketing platforms are becoming an increasingly important data source, especially crucial for smaller agencies without reliable Automated Fare Collection (AFC) or Automated Passenger Counter (APC) systems. This study demonstrates how mobile fare activations can be transformed into meaningful indicators of rider behavior through a three-fold clustering framework that operates at spatial, temporal, and stop levels. Using over 700,000 time-stamped activations from Unitrans in Davis, California, we show how mobile ticketing data, despite not capturing actual boardings, can yield robust spatial and temporal observations for a university-oriented transit system.

The first component applies DBSCAN to each rider’s activation coordinates to identify spatial activity centers such as home locations, campus destinations, and secondary nodes. We found that although some riders exhibit two dominant spatial clusters (commuting OD pairs), others show up to six or seven clusters indicative of more complex mobility patterns. Aggregating across riders highlights consistent patterns: off-campus residential zones dominate morning clusters, while campus and adjacent mixed-use areas become the main activity zones by midday.

The second component focuses on riders’ daily temporal rhythms through Dynamic Time Warping (DTW) combined with Partitioning Around Medoids (PAM). For each rider, an 18-hour activation profile (6:00–23:00) is constructed and averaged across all travel days. DTW allows profiles with peaks occurring at slightly different times to be aligned, revealing the underlying shape of temporal activity. Across solutions with two to four clusters, distinct patterns emerge for each fare type. Single-ride adult users tend to split into those with concentrated morning peaks and those with diffuse all-day activity. Paid pass holders show the greatest heterogeneity, with identifiable morning, midday, afternoon, and evening travel patterns. Undergraduate pass users display highly consistent temporal structures dominated by the academic schedule, including strong morning peaks and sustained midday activation. The third component uses k-means clustering to examine variation in stop-level usage across the transit system. We summarized each stop using intuitive indicators: the number of unique riders activating near the stop and the average frequency of activation per rider. These measures capture differences between stops that serve many riders infrequently versus stops that support repeated use by a smaller population. Across K = 2–5, several stable stop typologies emerge. High-density-residential-area stops tend to show high repeat usage driven by students residing off campus. Central campus stops exhibiting high rider turnover consistent with dense academic activity. Peripheral stops show lower overall activity, as they reflect a greater distance from key campus and residential nodes.

In conclusion, this three-fold clustering framework - the integration of spatial DBSCAN, temporal DTW–PAM, and stop-level k-means - illustrates how mobile ticketing data can produce insights even in the absence of AFC hardware or full APC coverage. Home–campus commuting patterns, class-driven temporal peaks, and land-use-dependent stop activity become clearly observable. For small agencies like Unitrans, we believe this framework provides a scalable analytical toolkit that can be used to support schedule design, service planning, and stop rationalization. More broadly, the findings show that mobile ticketing data can yield consistent and fine-grained behavioral insights that enhance, rather than replace, traditional survey and count-based approaches used in transit planning.

ABSTRACT. For operations and planning purposes, transit operators and authorities need to have a detailed understanding of users’ travel patterns in space and in time. However, aggregate flows provided by passenger counts mask the distinctive travel patterns of various user segments. While data from origin-destination surveys may contain additional travel details such as the fare product purchased, they often lack precision and accuracy. To address this weakness, this paper proposes to integrate the concept of “fare cohorts” into transit network analysis. Fare cohorts are closely tied to social and user segments. They can handily be defined by using or combining fare characteristics, such as statutory discount, loyalty discount and fare medium. This allows transit professionals to better identify and model the target segments. For example, it is possible to reveal spatial and temporal network usage by frequent and occasional transit users; by the senior, adult and student populations; or by users who make onboard cash or card payments.

Montreal is home to a large multi-modal transit network and has a diverse user base. Its smart-card fare and ticketing system generates detailed anonymized fare validation data. With a high resolution transit network model and data processing techniques that have been maturing over the past decade, combined with contextual information, this paper aims to demonstrate the methodology and the value of characterizing travel patterns in a transit network by fare cohorts.

Data and methodology The paper features an empirical study: we are interested in measuring and understanding how various cohorts behave spatially and temporally in the multi-modal transit network. We propose the following datasets and methodology:

Datasets and processing Here are the dataset and data processing requirement for the study: Contextual data containing: Detailed fare information on statutory discount, loyalty discount, fare medium and validity and eligibility requirements of each fare in order to define fare cohorts. Fare validation records from the OPUS smart card automated fare collection system: this dataset remains the most detailed source of information on transit network usage and is central to this study. The system is entry-only, meaning that processing is required to derive full trip information (origin-destination) from individual smart card geolocated (with equipment location or GTFS data) tap-in records. Many papers (including Trépanier et al., 2007; Chu & Chapleau, 2008, etc.) have already addressed and applied the concept. This paper will mostly focus on the applications with the enriched dataset.

Figure 1 Fare validations obtained from entry-only smart card AFC system

Figure 2 Trips derived by combining trip sections

Figure 3 Deriving origin-destination information by incorporating GTFS data

Detailed georeferenced transit network which was built and constantly maintained by the Polytechnique Montreal research group. Trips containing fare cohorts information are then loaded into the transit network. The geospatial nature of the network allows us to perform visualization, spatio-temporal selection and analysis on full trips derived in the previous step.

Figure 4 Load profile of a bus route from enriched AFC data

Key indicators to represent travel behaviour and fare use Here are the proposed indicators that would be able to highlight the travel patterns of each fare cohort: Number of transit modes used (heavy modes: commuter rail, light rail, metro; bus, etc.) Number of trip sections (first boarding and subsequent transferts) Trip by location (origin and destination sectors) Trip by time of the day and day of the month Trip symmetry The study measures and compares the patterns of various cohorts using descriptive statistics and spatial representations by GIS.

Expected results We expect significant differences in travel patterns for various fare cohorts: The proportion of occasional users would be higher in heavy modes and during off-peak periods The mean distance traveled by frequent users and commuters (work and study) would be longer than that by occasional users Occasional users, often associated with discretionary activities, would be concentrated in centre rather than in the peripheral of the network

Anticipated conclusions and next steps The proposed methodology and expected results tie together two important aspects of transit service realities: network use and fare use. It can be accomplished by enriching existing datasets and using a detailed transit model. There are numerous applications, from adjustments in service planning, refining fare level, proposing more equitable fare structure and funding mechanism to initiatives that aim to improve customer experience of targeted segments. For example, for locations and time with a strong presence of occasional users, one can prioritize the implementation of direct contactless payment, improve signage and increase staff availability. The next steps would consist of fine-tuning the methodology and materializing the proposed applications.

References Trépanier M., Tranchant N., and Chapleau R. Individual Trip Destination Estimation in a Transit Smart Card Automated Fare Collection System. Journal of Intelligent Transportation Systems: Technology, Planning and Operations, Vol. 11, No. 1, 2007, pp. 1–14.

Chu, K. K. A., and Chapleau, R. Enriching archived smart card transaction data for transit demand modeling. Transportation Research Record 2063, 2008, pp. 63–72.

ABSTRACT. Author’s preference for presentation: In-person

Introduction Worldwide, urban environments are experiencing rapid growth and transformation, driven by rising populations and increasing traffic volumes. As car-oriented cities transition toward safer, greener, and more livable streets, planners must continuously balance competing demands for limited roadway space. These transformations – particularly those involving roadway and curbside management – face both design constraints and operational capacity limits. While improving urban quality of life is a central objective, dynamic and often conflicting mobility demands must be accommodated efficiently to ensure system performance and network stability. Delivering balanced and effective urban mobility, therefore, requires management approaches consistent with the principles of Complete Streets. Within this context, the central role of public transport (PT), including tramway and bus services, remains fundamental as the backbone of urban mobility. However, PT is frequently routed in mixed traffic operations with private vehicles and cycling to clear space for alternative uses, potentially compromising service quality and reliability. In practice, planners must balance the operational vulnerability of mixed traffic operations with the higher spatial requirements of separated infrastructure. This trade-off requires empirically robust evidence and decision support tools that help practitioners identify the most suitable configuration for their specific planning context. Based on a research project examining tramway operations in mixed-traffic environments, substantial deficiencies in data provision were identified. These findings underscore the importance of improving the foundations of transit data with respect to data formats, access, robustness, and transferability.

Methodology In this study, we evaluated the performance of ten tramway routes across Germany. We conducted an extensive analysis of tramway performances under mixed-traffic conditions. In total, more than 22,000 tramway trips were researched. As the primary data source, the Intermodal Transport Control System (ITCS) data, which includes vehicle speeds, dwell times, planned and actual arrival/departure times, and interactions with traffic signals, was harnessed. Based on the collected data, suitable key performance indicators – such as travel speed indices, speed profiles, dwell times, and disruption/incident metrics – were applied to benchmark the different corridors. The main objective was to assess the extent to which tramway performances can be maintained under mixed-traffic operating conditions. Besides, the ITCS-datasets and additional sources (e.g., traffic volume data derived from camera-based field surveys or accident data) were used to support microsimulation studies aimed at identifying operational boundary conditions and critical thresholds that arise in mixed-traffic operations.

Results In the study, tramway operations on street-embedded tracks in mixed traffic were evaluated using ITCS data from multiple transit agencies, and German manual-based assessment procedures were utilised. Passenger transport speed indices, dwell times, and operating speeds were calculated for two operational variants (operational variant 1: tramways in mixed traffic with cars; operational variant 2: tramways in mixed traffic with cars and bicycles) and compared over several weeks in summer and winter. Across all routes, the analyses revealed substantial differences between corridors and operational variants. Although the routes in variant 1 showed higher speed indices, their actual operating speeds were often comparable to – or even lower than – those in variant 2. Seasonal differences in speed indices were negligible, indicating that higher bicycle volumes in summer do not necessarily impair tramway performance. Median operating speeds above 20 km/h as a threshold were not achieved across the full length of the variant 1 routes, despite spatial separation between tramways and bicycles. In contrast, sections of variant 2 often exceeded 20 km/h. One route even achieved a level of service of A in the passenger transport speed index, indicating that high-quality tramway operations can be maintained under mixed traffic conditions with cars and bicycles. Microscopic traffic simulations further showed that optimised signal control with dynamic road-space allocation can improve tramway performance under varying demand conditions for variant 1 routes. The results are summarised in a decision-support flowchart that structures the sequence of contextual and operational criteria. This stepwise decision process enables planners to evaluate local conditions and identify the appropriate form of street-embedded track deployment, with dynamic road-space allocation governed by signal control.

Discussion/Conclusion The resulting findings provide robust performance indicators from a traffic-planning and engineering perspective, illustrating how tramway operations work under mixed-traffic conditions. Moreover, the insights are, to a certain extent, transferable from tramway operations to bus operations. The core analyses, conducted through large-scale, script-based processing of complex datasets, highlight a key methodological challenge in relying on automated collected data. Worldwide, transit agencies increasingly collect operational data automatically, yet in the case of tramways, the substantial data generated daily remains only partially utilised for practice, research, or public benefit. In Germany, one reason is that many transit agencies lack suitable software tools to fully exploit these data volumes; therefore, practical, user-oriented tools are urgently needed. Furthermore, standardised data formats are essential for enabling cross-city and cross-operator analyses. Emerging approaches, such as the Transit Operational Data Standard (TODS), need further development and should be embedded within conceptual frameworks, such as the Transit ITS Data Exchange Specification (TIDES), to ensure consistency, interoperability, and scalability.

Acknowledgements The authors thank the Research Training Group (GRK 1931) of the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation - 227198829), which partly funded and enabled this research. Additional acknowledgement is given to the anonymised transit operator for supporting the provision of ITCS data, and the authors highly appreciate this willingness to share. Large language models were used to improve language and readability.

ABSTRACT. Attendance preference: undecided

Public transport is crucial to the development of sustainable traveling and commuting. Countries and cities all around the world are aiming at developing their public transit offer and making it as efficient as possible. The functionality of bus terminals is central to this efficiency: a large proportion of all journeys in a bus network pass through a terminal, which often provides the link between regional and local public transport.

This large volume of traffic makes it easy for terminals to experience capacity issues. Yet, as they are often located near the city centre, expanding terminals is not always possible nor cost-efficient [1]. A solution is then to optimize their operations, which can start with identifying the origins of congestion. To do so, this study aims at utilizing open data, notably public transport data stored in the General Transit Feed Specification (GTFS) format.

The two GTFS components are used in this project. The "Schedule” component consists of information about the bus routes and timetables, while the "Realtime” component consists of historical Automatic Vehicle Location (AVL) data. Both types of data come from KoDa, a database containing historical data from, at the time of writing, 11 Swedish transit operators (approximately 50 when considering only static data). The AVL data consists of the position of each vehicle for each second, also including its bearing and speed.

Working with such a dataset implies several challenges. One of them is that Global Navigation Satellite Systems (GNSS) data is known for being imprecise [2]; this can potentially have a negative effect on the accuracy of the results. Elements that influence this precision include weather- and atmosphere-related factors, nearby tall buildings, and the type of positioning equipment used by the vehicles [3]. Thus, one of the first aspects to consider is how reliable (or unreliable) the data is to work with.

The research question addressed in this study is the following: - How can historical AVL data be used for performance evaluation of bus terminals? This research question is divided into several sub-questions: - How can the imprecision of positioning data be quantified for buses driving through a terminal? - How does the imprecision in the utilized data vary within one terminal, and from one terminal to another? - What are the determining factors inducing this imprecision? - How does this imprecision affect the evaluation of the performance of the bus terminal?

A method is developed to quantify the imprecision of the data given a bus terminal and a time frame. This is partly done by first clustering the vehicle trajectories so that each vehicle in a cluster goes on the same path and stops at the same berth, then assessing the trajectory noise in the cluster using metrics such as the Round Mean Square Error (RMSE) for the measurement of positioning precision relatively to real-life driving paths, and Dynamic Time Warping (DTW) for the measurement of trajectory similarity in a cluster. Additionally, different types of positioning errors are distinguished, and the data is further explored by analysing the noise for several bus driving situations: stopped at a berth, driving on a path in the terminal, or driving on a road outside of the terminal. The data is also compared from one operator to another, from one day to another, and from one hour to another.

A few terminals in Sweden are used as case studies. The selection of these terminals has been made by considering several factors: the size of the terminal must be large enough to be interesting to study, all the buses going through the terminal should be represented in the available data, there should not be any car traffic, the terminal should not be underground, and ideally there is a known presence of congestion in the terminal. Drone videos have been filmed at those locations to validate the algorithms and their results. This is done by comparing the actual vehicle positions to the positions in the data, which allows assessing how accurate our evaluation of imprecision is for specific times.

The results include the quantified uncertainty in the raw data for the studied terminals and periods of times, as well as a statistical analysis of the factors influencing this unreliability. Lastly, we identify performance metrics for the traffic efficiency at the station, such as travel delay distributions, and we present an analysis of how the estimated uncertainties in the data affect the accuracy of these metrics.

These first methods and results are part of a larger project. On the longer-term, future research will involve numerical analysis of the performance of the terminal, congestion detection, as well as the construction of a cost-efficient framework for microsimulation of the bus terminals, all while using open data.

REFERENCES

[1] Lindberg, T., 2019. Discrete Event Simulation of Bus Terminals, Linköping Studies in Science and Technology. Licentiate Thesis. Linköping University Electronic Press, Linköping.

[2] Dovis, F., Muhammad, B., Cianca, E., Ali, K., 2015. A Run-Time Method Based on Observable Data for the Quality Assessment of GNSS Positioning Solutions. IEEE J. Select. Areas Commun. 33, 2357–2365. https://doi.org/10.1109/JSAC.2015.2430513

[3] Shah Sadman, A.A.M., Hossam-E-Haider, M., 2021. Study of GNSS Parameters and Environmental Factors over Bangladesh Intended for Selecting Ideal Ground Station Location for SBAS, in: 2021 2nd Global Conference for Advancement in Technology (GCAT), IEEE, Bangalore, India, pp. 1–6. https://doi.org/10.1109/GCAT52182.2021.9587629

ABSTRACT. Preferred Presentation Format: In-person

Transit service reliability remains a critical challenge for high-frequency bus and streetcar networks, particularly in large metropolitan systems where irregular vehicle spacing leads to bus bunching and gaps. These phenomena result in overcrowding, increased passenger waiting times, and operational inefficiencies. While transit agencies have traditionally relied on static schedules and manual control methods, there is a growing demand for data-driven, real-time strategies that can proactively detect and mitigate service instability.

This research develops an integrated framework for real-time detection, prediction, and mitigation of bunching and gapping using General Transit Feed Specification (GTFS) and GTFS-Realtime data. Our system monitors vehicle locations, scheduled and actual headways, and live vehicle occupancy to detect emerging disruptions and quantify their severity across the network. Building on this detection layer, an optimization-based holding model recommends which vehicle should hold, at which stop, and for how long to restore even spacing while minimizing passenger delay. Equity considerations are incorporated by prioritizing holding decisions for lower-occupancy vehicles to avoid imposing additional delay on already crowded buses.

In parallel, a short-term prediction model forecasts the number and location of bunching and gapping events in advance. The predictive engine continuously updates its estimates based on recent incident patterns and system behavior, enabling proactive rather than reactive operational control.

By processing live GTFS feeds alongside historical performance data, this system provides a scalable and practical pathway toward real-time headway management, early detection of service disruptions, and AI-assisted transit control. The framework demonstrates how open data, predictive analytics, and intelligent holding strategies can work together to support more reliable, efficient, and equitable transit service delivery.

ABSTRACT. Urban bus systems routinely suffer from bunching, crowding, and excessive idling, all of which erode reliability and passenger experience. Predicting bunching before it fully develops would give controllers a critical window to intervene proactively rather than reacting once service has already deteriorated. However, the mechanisms that drive bunching are often distributed across disparate data sources: increasing in demand captured by automatic passenger counting (APC) systems, prolonged idling at intersections, and slow-moving traffic observed in automatic vehicle location (AVL) streams, among others. Building on our previously proposed Integrated Road Operation and Anomalies Monitor (iROAM), which fuses multi-source real-time data (AVL, APC, and inferred anomalies) to flag operational irregularities, this paper introduces iROAM+, a short-term bus bunching prediction framework that exploits the full richness of this integrated data. iROAM+ employs advanced spatio–temporal deep learning architectures to learn how demand patterns, dwell-time anomalies, and upstream traffic conditions jointly shape near-future headway dynamics. We evaluate iROAM+ on real-world multi-source data collected from high-frequency routes integrated by iROAM, training models to forecast stop-level headways and identify emerging bunching episodes over short horizons (e.g., 5–20 minutes ahead). We examine both overall predictive accuracy and the marginal contribution of different feature groups, including anomaly indicators and contextual variables such as time of day, day of week, and location. Experiments show that iROAM+ improves short-term headway prediction and early detection of bunching events compared with strong statistical and deep learning baselines built on single-source feeds. By coupling anomaly-aware data integration with state-of-the-art deep learning, iROAM+ moves from simply detecting problematic conditions to forecasting them, providing transit controllers with earlier, more informative signals to support pre-emptive control actions.

ABSTRACT. Accurately predicting how many crew members will be absent in the coming days is crucial for crew planners and dispatchers to align available capacity with operational needs. Reliable forecasts enable better decision-making, such as responding to requests for extra days off. Poor capacity management can lead to overtime costs or even train cancellations due to staff shortages.

We propose a hybrid CNN–LSTM model to predict the number of employees absent on a given date due to sickness or other uncertain reasons. The model was trained and tested using data from a Northern European railway operator. Results indicate that the predictions are sufficiently accurate to support capacity management and that our deep learning approach outperforms alternative models. While hybrid CNN–LSTM architectures have been applied in other domains, to the best of our knowledge, this is their first application to crew absence prediction.

The dataset used in our study includes historical employee activity, employee characteristics (role, age, seniority, and home depot), and national holiday information (from 2013 to 2022).

Historical activity covers planned and realized activities of train drivers and guards from 2013 to 2022. It is provided as a set of records, where each record specifies, for a specific employee and date, what the employee was supposed to do on that date (take a day off or perform a duty with a given schedule) and what he really did (whether he followed the plan or was absent, and the reason for the absence).

All data were anonymized by replacing real names and identifiers with synthetic counterparts.

The preprocessing phase encompassed the removal of outliers, imputation of missing values from historical records, and the application of min–max normalization to numerical variables.

Outliers were filtered at employee level based on insufficient total or average working days.

Missing duty durations were imputed using employee-specific mean values to preserve individual-level integrity.

Categorical variables were one-hot encoded (with one dummy category dropped to avoid multicollinearity), and numerical features were normalized to min-max scaling fitted on training data.

Employee sequences were obtained as the result of grouping the records by employee and sorting them in chronological order.

Chosen features characterize records, in the sense that, for each record (corresponding to a specific employee and date), they characterize the employee, the date and the work assigned to the employee on that date, which will hereafter be referred to as the target date. They include: 3 features categorizing temporally the assigned work, 1 absence indicator (used only for training the model), 4 vacation/day-off indicators for adjacent days, 2 holiday indicators for the target and previous day, 4 employee attributes (age, seniority, role, home depot); 2 calendar features (day of week, month).

After cleaning, the training set contained 3,467,971 records and the test set 412,773, covering 1,517 drivers and 1,649 guards.

In terms of architecture, based on the ideas of M. Salvado (https://run.unl.pt/handle/10362/163573), we deploy a CNN–LSTM hybrid model where stacked 1-D convolutional layers extract spatial and short-term temporal patterns, and two LSTM layers capture longer-term dependencies, producing absence probabilities per day and employee.

Model tuning is applied to this dynamic CNN–LSTM model, adjusting hyperparameters such as the number and width of convolutional filters, kernel sizes, LSTM units, dropout rates, and the choice of optimizer. The hyperparameter optimization yields the best-performing configuration based on the validation loss on held-out data.

For the training we considered data from 2013 to 2021. We grouped the corresponding employee sequences into batches of 256 records. With a total of six batches, we randomly selected five (approx. 85%) for fitting the network, while the remaining batch was reserved for validation.

For the purpose of testing, data concerning the year 2022 is kept completely separate until the final evaluation.

For the evaluation we used the model to produce forecasts for each day of 2022. The forecast for each day is computed as the sum of the absence probabilities obtained by the model for each employee on that day, more precisely the value of the sum rounded to the nearest integer. In order to assess the predictive quality of the model we use the MAPE metric, obtained by computing the absolute percentage error of the forecast for each day and then by calculating the average over 2022.

With the model trained only with train driver (guard) data, and the testing done with train driver (guard) planned activity that includes the last 200 records of each employee from 2013 to 2021 and the 2022 records until the forecasted days, we obtained a MAPE of 10.39% (8.46%).

With the model trained both with train driver and guard data, and the testing done with train driver (guard) planned activity that includes all records from 2013 to the forecasted days, we obtained a MAPE of 11.72% (10.50%). A possible explanation to the fact that these results are slightly worse than the former is that guards and drivers follow different patterns in terms of absence occurrence.

With the model trained only with train driver data, and the testing done with planned activity of train drivers from the largest home depot (from 2013 to the forecasted days), we obtained a MAPE of 16.51%. This result outperforms what was produced by other models with the same data (see MSc thesis by C. Afonso, https://run.unl.pt/handle/10362/163574), namely 19.96% with the DHR regression model, 20.70% with the Profet regression model, and 21.09% with the NeuralProphet regression model.

These results confirm our initial claim that the model proposed can be used in practice to help planners and dispatchers improve capacity management.

ABSTRACT. Service adjustments are a routine part of transit operations, yet systematic evaluation of their impacts remains limited. This study demonstrates a comprehensive pipeline that agencies can adopt to assess schedule changes using existing operational data sources. Using a recent schedule revision implemented by the Chicago Transit Authority (CTA) as a case study, we illustrate how this framework can be applied in practice. Passenger-side evaluation considers ridership response, headway variability, and the resulting excess waiting time and crowding. On the operator side, we quantify actual runtime, recovery, and their variability to examine whether schedules provide realistic running and layover conditions. These dimensions are inherently coupled: improving reliability often requires additional recovery time, but excessive padding can reduce productivity or encourage slow driving. To account for resource efficiency, total bus hours are included as a cost metric. Finally, to isolate causal effects from broader network fluctuations, we employ a Bayesian Structural Time Series–based synthetic control method that constructs counterfactual routes unaffected by the intervention, ensuring that observed differences can be attributed to the schedule change itself.

On passenger experience, crowding improved slightly, while headway variability remained close to pre-intervention levels and excess waiting time was essentially unchanged, with up to 5 minutes at peak period. Interpreting the observed headway variability in time-equivalent terms, a peak scheduled headway of about 8 minutes corresponds to a headway near 12 minutes with normal variability; riders therefore face average waits comparable to a reliable service operating less frequently. Despite limited improvement in passenger-experienced reliability, ridership along the Route22 corridor increased notably following the schedule revision. The largest gains occurred in the inbound peak directions, particularly on the northern segment exclusive to Route22 (+20%) and on the shared downtown segment with Route36 (+30% during rush periods). These patterns likely reflect broader post-pandemic ridership recovery and renewed activity along the Clark Street corridor rather than direct service-quality improvements.

Operationally, recovery became more realistic, the variance decreased, and running time before the control point remained essentially unchanged, while full-cycle running time increased modestly by 4–8 minutes. This pattern follows the design: elevating the scheduled runtime percentile and concentrating slack after the control point preserved speeds upstream while adding buffer downstream, which absorbed stochastic delay and reduced infeasibilities. However, drivers responded heterogeneously at timepoints, often extending dwell when early, so a portion of the variability was shifted into dwell rather than converted into tighter headways. Moreover, because headway variability at the terminal remained high and origin dispatch control was unchanged, the added recovery did not fully translate into more on-time departures. Late starts persisted, and early-cycle dispersion propagated along the route.

Resource use increased as expected. Operator pay-hours rose by about 10%. Taken together, the intervention met the objective of improving feasibility and driver experience with more realistic recovery and buffered delay propagation, while passenger outcomes moved only at the margin: crowding eased slightly and peak ridership rose, but headway variability and excess waiting time remained largely unchanged.

The results imply a pragmatic lesson for agencies considering similar changes. Runtime percentiles and recovery allocation are effective when they are targeted to periods and segments where delay accumulates; uniform padding risks inflating cycle time without compressing headways. To translate added recovery into smaller passenger waits, schedule adjustments should be coupled with terminal departure discipline and light-touch headway control, and timepoint design should be revisited where holding generates excessive dwell without downstream regularity benefits.

Beyond this corridor, the approach constitutes a general pipeline for service change evaluation. The Chicago Transit Authority serves as an example: integrate AVL, APC, and scheduling data; construct route–period outcomes; form a donor pool of unaffected services; estimate counterfactuals with Bayesian synthetic control; and report stratified effects by time of day, direction, and segment, alongside diagnostics. Embedding this pipeline enables agencies to turn pilots and routine schedule adjustments into cumulative, evidence-based decisions, and to calibrate where percentile targets, recovery, timepoint policy, and control strategies should be refined to deliver measurable passenger benefits per unit of resource.

ABSTRACT. In-person preference

1. Introduction Public transport (PT) systems continue to face significant uncertainty in the aftermath of Covid-19. Although patronage has partially recovered, demand in many cases has stabilised below pre-pandemic levels, resulting in reduced fare revenues and heightened financial pressure on operators and local transport authorities (Buehler et al., 2025). At the same time, structural changes in post-pandemic travel behaviour and patterns further deepen the uncertainty about whether and when PT demand will recover to pre-Covid levels (Hensher et al., 2024). The adaptability of bus networks can help cope with drops in PT ridership. Compared to fixed urban rail systems, bus services offer greater operational flexibility and shorter lead times for modifying service levels. Thus, bus frequency adjustments represent one of the most immediate tactical strategies available to local transport authorities seeking to realign supply with reduced demand and budget constraints. However, deciding about where and how to reduce (or increase) bus frequencies is not trivial, as this may have substantial impacts on passenger travel times, accessibility/equity, crowding, operator costs, and the overall ridership.

The frequency-setting problem has been examined extensively in the literature, with studies proposing optimisation models and heuristic approaches, often incorporating elements such as network design or vehicle capacity constraints (Caetano et al., 2025). However, to the best of our knowledge, research assessing the marginal impacts of frequency adjustments across PT networks remains scarce. To address this gap and support evidence-based decision-making, we leverage the potential of automated data to develop a modular line-prioritisation framework to iteratively evaluate the impact of bus frequency adjustments on PT system welfare and identify line interdependencies. The overarching research question is: “What is the impact of line frequency adjustments on PT system welfare and residual capacity in the network?” The proposed method is applied to a case study of the Stockholm Region’s PT network.

2. Methodology To address the research question, we propose a modular, iterative framework composed of two main components. The first module evaluates the impact of line bus frequency adjustments on residual capacity, consumer surplus and operational costs. Consequently, the second module uses these impact indicators to iteratively refine the number of frequencies per bus line, aiming to derive a line-prioritisation framework for capacity reduction.

Data and Network Construction The analysis leverages automated data to reconstruct passenger travel patterns, estimating OD flows, and analysing how service changes affect generalised costs and demand. Passenger journeys are derived from AVL, GTFS and APC data using an existing inference framework to infer tap-out locations (Kholodov et al., 2021). The destinations of journeys for which a tap-out location could not be inferred are proportionally reallocated based on the observed destinations’ distribution within the origin area. The resulting inferred journeys are used to construct an OD matrix under baseline frequency conditions. GTFS data is used to build a simplified hybrid PT graph for route-choice modelling and passenger assignment. In this network representation, each node corresponds either to an OD stop, used as an origin or destination, or to a stop-line pairing, representing a specific line at a stop. Edges can represent (1) in-vehicle links connecting adjacent stops on the same line, (2) transfer links connecting two lines serving the same stop, or (3) unidirectional waiting links connecting OD stops with all the respective stop-line nodes. Scheduled in-vehicle times are used as weights for in-vehicle edges, frequency-dependent waiting times for waiting edges, whereas transfer edges are weighted by a transfer time component derived from the frequencies of the two connecting lines, which is added to a transfer penalty term. This representation allows frequency adjustments to propagate efficiently into generalised travel costs per route, supporting the iterative assessment of different scenarios.

Bus Frequency Adjustments Impact Assessment For each OD pair, a small but behaviourally meaningful route choice set is constructed by running a K-shortest paths search on the hybrid graph. To avoid generating unrealistic or highly overlapping alternatives, we apply simple overlap-based filtering to retain a limited set of distinct alternatives and validate the route choice sets based on the observed route choice set. Utilities for each alternative route are computed, and the probabilities are derived from a path-size logit route choice model to account for the overlap of alternative routes. OD flows are probabilistically assigned to the network based on the route choice probabilities per OD. OD flows are then updated by applying elasticity terms in relation to the consumer surplus based on the logsum of the route choice set per OD pair. The model explicitly avoids imposing full equilibrium and capacity constraints, and using crowding as a route choice determinant, as the objective of the assignment is to evaluate marginal flow shifts resulting from frequency adjustments and identify capacity allocation needs. Thus, residual capacity is only incorporated ex post using the Volume Over Capacity (VOC) ratio per line segment (and line aggregates) to identify interdependencies and complementarities between lines.

Iterative Line Prioritisation Heuristic Outputs from the impact assessment module, including consumer surplus, VOC per line-segment, and frequency-based operational costs, are fed into a heuristic module. This component iteratively adjusts the frequencies of each line, updates OD matrices based on exogenous demand elasticities, recomputes route choice utilities, assigns the updated OD flows, and evaluates the resulting impacts on the network. Through the iterative exploration of line frequency adjustments, the module constructs line-level indicators of their impact on the PT system, providing a structured framework for line prioritisation.

3. Expected Results Frequency-setting in practice remains largely incremental, relying on expert judgement and organic adjustments. Building on pre-processed smart card data, this study provides a framework to support evidence-based bus frequency adjustments. The iterative line-prioritisation framework is anticipated to enable the identification of lines where frequency adjustments have a disproportionate (or very limited) impact on public transport system welfare, measured as consumer surplus and operational costs, and line interdependencies, highlighting lines with strong (or weak) demand propagation effects. The line-level indicators generated from single-line frequency adjustments can be used to guide the selection of plausible multi-line frequency adjustment scenarios, which can then be evaluated using the impact assessment module.

ABSTRACT. Presentation Preference: Undecided

Introduction Public transit agencies are continuously facing the challenge of balancing operational efficiency and fulfilling passenger satisfaction. Within this context, time points play a crucial role in maintaining bus service reliability and performance consistency. Time points are specific and defined stops, where the arrival and departure times of buses are rigidly monitored and controlled. However, the criteria used to select these stops vary across agencies and are often based on experience and day-to-day operations rather than on empirical analysis. This study aims to address this gap by using General Transit Feed Specification (GTFS) datasets from four Canadian cities: Vancouver, Calgary, Edmonton, and Ottawa. The results help to identify the operational and contextual factors influencing time points placement and can help transit agencies use these criteria, considering their operations. Data and Methodology For each of the four Canadian cities, GTFS static datasets were processed to get information about routes, trips, and stop sequences for each bus network. The datasets included scheduled stops, travel times, and travel distances. The validation process of the time points involved cross-checking the GTFS-labeled time points with the official transit agency’s documentation, achieving a match accuracy of 95.3%. After the validation, key variables were defined to consider route conditions or characteristics, such as: total number of stops, average route length, average trip duration, average headway, and single and multi-transfer stops (within a 300m radius), and stops with key locations close to schools, hospitals, and major activity hubs that vary between cities. After cleaning and preparing the dataset for each city, correlation matrices were made for each city to explore the relation between the defined key variables and the number of designated time points. The variables that showed high correlation were evaluated for multicollinearity using Variance Inflation Factors (VIF) to ensure a robust model. Then, simple and multivariate regression models were developed to quantify the influence of the variables on an individual and a combined level regarding the time points allocation. Model performance was measured using the coefficient of determination (R^2) and the root mean square error (RMSE). Results From all the resulting models, several common patterns were identified: In Calgary, trip duration, number of stops, and presence of key locations showed strong positive correlations with the number of time points. This indicated that longer and denser routes seem to require more time points to maintain schedule reliability. In Vancouver, transfer stops and headways were the most influential variables. The negative correlation found between headways and the number of time points suggests that high-frequency routes tend to require fewer control points, due to the high frequency of service. In Ottawa, single transfer stops and stops with key locations were the significant variables. These results reflect a transit system organized around main transfer hubs, since many routes pass through these key locations, influencing where time points are located. In Edmonton, the route length has the strongest relationship with the number of time points, which is consistent with its large urban area and the longer average trip distances. Across all cities, transfer-related variables emerged as the most consistent predictors of the number of time points. This reaffirms that network connectivity plays a decisive role in determining where time monitoring is most critical. The analysis also revealed the presence of multicollinearity between route length and trip durations and between single and multi-transfer stops. As such, the final multivariable model used only one variable from each correlated pair. The refined model, combining transfer stop, trip duration, and route length, achieved higher explanatory power while also maintaining interpretability across cities. Discussion These findings emphasize the importance of integrating automated transit data into the service planning and time points designs. Transfer stops represent both operational bottlenecks and opportunities for performance improvement. Routes with longer trip durations tend to accumulate more travel time variability due to the longer exposure to traffic, passenger transfers, and operational delays; so they require more strategic placing of time points to maintain adherence to the schedule. And, high-frequency services with short headways can operate effectively with fewer time points because frequent service stabilizes regularity. This emphasizes the need for flexible and data-driven adjustments rather than static scheduling, like the transit agency in Toronto does. Moreover, characteristics of cities are also important to consider. Calgary and Ottawa’s higher sensitivity to trip duration indicated that routes with different headways depend more on time point control to maintain reliability, while Edmonton’s emphasis on route length suggests spatial coverage challenges. Vancouver’s network benefits directly from short headways, demonstrating how operational design can offset the need for extensive time point monitoring. These differences demonstrate the potential value of developing algorithms that can adapt to context for time point selection rather than applying uniform rules. Implications and Future Work The research provides actionable recommendations for transit agencies that are constantly seeking the improvement of reliability and responsiveness of bus operations, such as: Prioritizing transfer hubs for time points monitoring and performance tracking. Optimizing headways to reduce passenger waiting times and minimize schedule deviation. Using GTFS data to conduct real-time and periodic assessments of time points adaptation. Refining validation procedures to ensure continuous improvement in time-point selection through automated data quality control. Future extensions of this work can be done using real-time GTFS data (GTFS-RT) and applying methods to predict the optimal time points selection under varying demand and congestion conditions. Expanding the analysis to include smaller or bigger cities could provide further insights into the scalability of the framework. Including data sources such as Automated Vehicle Location (AVL) data and Automated Passenger Counting (APC) systems can also enable deeper analysis of time points and passenger impacts. Conclusion Through the four Canadian cities, we identified that transfer stops, trip duration, and route lengths are the most influential variables for time point selection. The results highlight the potential for developing standardized but also flexible data-driven frameworks that can adapt to the context of every city. In conclusion, by using data analytics, transit agencies can have more efficient, reliable, and user-oriented scheduling practices.

ABSTRACT. Unpredictability is inherent in transit systems due to varying passenger demand and traffic congestion. However, providing a reliable service is essential to meet passenger needs and is recognized as a key indicator of transit service quality. Effective bus scheduling and operational planning are critical for reducing passenger costs and improving service reliability. A well-designed system increases passenger confidence in transit reliability, improves their experience, and can save transit agencies considerable resources. Transit agencies often struggle to balance trip speed with on-time performance, and frequently rely on general rules of thumb when allocating slack time to schedules. Striking an appropriate balance is crucial in maintaining a reliable and efficient system, as early departures or insufficient slack time can increase passenger waiting costs, while excessive slack time can increase riding and delay costs. Additionally, quantifying the benefits of adherence to holding policies is challenging for many transit agencies, and compliance rates vary in practice. However, passenger demand data and vehicle-tracking technologies at the microscopic level now enable detailed evaluation of these trade-offs and more effective scheduling and planning improvements. Prior research on optimal slack time and stop placement has primarily focused on single routes. There is a need for comparison between routes with significantly different characteristics, exploration of temporal variation of timepoint locations based on demand, and insights into the effects of passenger arrival distributions.

This study uses bus routes in Calgary, Canada and Kyoto, Japan to compare transit systems with significantly different operational characteristics. Kyoto has much higher passenger demand, rarely skips stops, and treats every stop as a timepoint, whereas Calgary has approximately 10% of stops designated as timepoints. These differences allow for evaluation of how schedule design performs under distinct operational strategies. The study aims to determine the most suitable schedule structure and operational plan that minimizes total passenger costs. Two timepoint strategies are compared: one where all stops are timepoints and holding controls prohibit early departures, and a second where only selected stops are timepoints, allowing buses to depart early from non-timepoint stops. Both are evaluated assuming perfect policy adherence and compared to actual trip observations. The effects of slack time and schedule tightness on performance are explored by varying the original schedule. Different passenger arrival distributions are also modeled to assess the influence of passenger demand on user cost. Finally, the study investigates whether timepoint locations should vary spatially or temporally based on demand.

The effectiveness of different bus schedule designs and holding strategies is evaluated by simulating alternative operational plans and schedule flexibilities and quantifying their impacts on total passenger cost. The aforementioned two timepoint strategies are simulated using AVL and APC data from observed trips: the first, S_all, treats all stops as timepoints, while the second, S_sel, uses only designated stops as timepoints. Both simulations use observed link travel times and passenger demand, assume buses never depart early from timepoints (100% holding policy compliance), and estimate dwell time as the greater of observed dwell time or minimum service time. Arrival, departure, and dwell times for each stop are generated iteratively along the route. A simple regression model estimates the minimum service time based on the recorded dwell times associated with critical passenger demand.

To assess schedule flexibility, scheduled travel time between stops is used as a baseline and is multiplied by a factor, k, between 0.5 and 2.0. This simulates tighter and more relaxed schedules which can be evaluated to determine how changes in slack time impact trip costs. A schedule multiplied by k = 0.5 would mean that the scheduled run time is cut in half, while k = 2.0 means the simulated scheduled run time would be twice as long as the actual schedule. The study formulates a user cost function comprised of four components: extra waiting cost for delayed arrivals, riding cost for additional on-board time, early departure cost for passengers who miss the bus, and delay penalties for late arrivals at alighting stops. Passenger arrival rates at stops prior to boarding, combined with the schedule deviation, determine the expected number of passengers as well as those who miss the bus. The passenger arrival rate is modelled using two distributions representing extremes: the beta distribution, which assumes that passengers know the schedule and plan their arrival to minimize waiting time, and the uniform distribution, which assumes passengers arrive randomly without considering the schedule.

Preliminary results show that S_all and S_sel both accumulate similar total dwell time by the end of a trip but distribute dwell time differently across the route. Actual trips sometimes exhibit much lower dwell time overall, highlighting potential schedule inefficiencies and the need for schedule flexibility adjustments. User cost analysis reveals that S_all typically produces higher user costs than S_sel, while both simulations generally outperform actual observations for trips with high user cost. Schedule flexibility tests indicate that a schedule scaled by approximately k = 0.9 best fits actual conditions, though the ideal k-value varies by route segment, suggesting schedule adjustments should be segment-specific rather than uniform across the entire route. S_all is best suited for tighter schedules, while S_sel is better for slower trips where a longer schedule with more slack is appropriate. These observations are pretty consistent throughout the day, though time-dependent timepoints may better reflect demand variations.

Stop-level cost breakdowns indicate that early departure cost is the primary contributor to user cost, especially at stops with high boarding demand. S_all eliminates early departure cost, with most cost stemming from on-board time, while S_sel can incur extremely high costs when excessive slack time causes buses to arrive far too early.

Overall, the study demonstrates the effectiveness of different bus schedule designs and holding strategies by simulating alternative operational plans and schedule flexibilities, and by quantifying their impacts on total passenger cost. Enabled by high-quality transit data, results highlight how schedule design and timepoint strategies can be adjusted to improve passenger experience and enhance the system performance.

ABSTRACT. This presentation shows how modified General Transit Feed Specification (GTFS) schedules and open-source accessibility tools can be used to evaluate the equity impacts of transforming Boston’s commuter rail into an all-day regional rail network. The analysis is powered by the TransitCenter Equity Scenario Comparison Application (TESCA), a lightweight scenario engine built by the author to compare access-to-opportunity outcomes across different transit futures. A brief overview of the TESCA tool will be included in the presentation.

Using the MBTA’s November 2024 GTFS as a baseline, we created a modified set of GTFS timetables representing a possible regional rail future: 15-minute all-day service on core corridors and 30-minute headways on branch segments, with simplified stopping patterns and a variant including modest speed improvements consistent with electrification. These GTFS feeds were processed through TESCA, which standardizes the steps needed for equitable scenario analysis: validating GTFS, fetching recent Census demographics and OpenStreetMap data, computing travel times with the R5 routing engine, and generating cumulative and travel-time accessibility measures. The results show that even without major capital projects, shifting commuter rail to a frequent, all-day pattern dramatically expands access to jobs, healthcare, and education for many parts of the region. Gateway Cities and Environmental Justice communities—including Worcester, Lynn, Brockton, and portions of the North Shore—see some of the largest gains, often moving from just outside to well inside a 45–60 minute job-access envelope. When minor speed improvements are added, the magnitude of these gains increases further. Importantly, the analysis also captures the share of residents who cannot reach key destinations at all within a specified travel time, revealing additional equity impacts not visible in regional averages. In parallel, we applied the same GTFS processing pipeline to map “dependable bus” networks—routes running at least every 30 minutes, seven days a week—across Boston and three peer cities (Calgary, Toronto, and Houston). This comparison highlights the importance of modest frequency on the overall perception of transit access and reach in a regional context.

Taken together, the findings demonstrate the value of pairing modified GTFS feeds with transparent, reproducible accessibility calculations. TESCA’s role in this project is not as a standalone product but as the underlying workflow that makes the analysis possible: it ensures consistency across scenarios, connects the data sources needed for equity analysis, and produces clear, interpretable outputs suitable for planners and advocates alike. The combination of scenario-based GTFS modeling and equity-focused accessibility metrics provides a practical and replicable approach for evaluating major service changes in other metropolitan regions considering regional rail transformation.

ABSTRACT. (in-person presentation is preferred) Public transit disruptions reveal how strongly daily life depends on it, leaving some users with limited or inadequate alternatives. Major disruptions in public transit services can be caused by unforeseen events (such as system failure or user incidents) or planned activities (such as maintenance work or labour strikes). When public transport is unavailable, travelers can face long delays, driving them to change modes, departure times, and in some cases, cancel or reschedule planned activities. In June 2025, Société de transport de Montréal (STM) maintenance workers announced a nine-day strike, severely reducing or suspending regular public transit service in Montréal, Canada. In the first three days, the strike caused the most disruption, with bus and metro service operating only during peak hours in the early morning and late afternoon, and late at night. On the subsequent days of the strike, transit service either returned to normal or operated at reduced levels (50%) outside of peak hours. This paper uses data from an online bilingual survey (N = 860) conducted in June 2025 during the strike to examine how residents managed their work and non-work trips. The survey gathered data on residents’ travel strategies, travel experiences, attitudes, and sociodemographic characteristics. To better ensure population representativeness, we determined respondent weights from the income, gender, and age profiles of their residential census tract (CT). Using the collected data, this study aims to answer the following research questions (1) who was able and unable to maintain access to work and non-work activities during the strike, and (2) what strategies did travelers use to adapt to the reduced service, including mode shifts, rescheduling, and trip cancellation. Using the weighted data, we conducted two k-means cluster analyses using the kcca function from the flexclust package in R: one grouping workers (N = 644) based on their stated adaptive behaviours for commute trips, and another grouping all respondents (N = 860) based on their adaptive behaviours for non-work trips. To identify the optimal number of clusters for each analysis, we use silhouette analysis in combination with an assessment of cluster characteristics, alignment with prior literature, and interpretability. For the work-trip clustering, the variables included whether respondents adjusted their departure or return times, teleworked instead of commuting, or used alternative modes of transport, as well as their car ownership and income level. For the non-work-trip clustering, the variables included whether respondents cancelled or rescheduled activities, used alternative modes, car ownership, income level, and employment status. Five clusters were identified for both the commuting and non-work-trip analyses. The commuting clusters are Mode-constrained Transit Reliants (39%), Multimodalists (16%), Active Travel Adopters (9%), Teleworkers (9%), and Auto-commute Switchers (24%). The Mode-constrained Transit Reliant cluster had the lowest rate of car ownership and income level and primarily adjusted the time they left for work, demonstrating their strong reliance on public transit for their commute. They found transport alternatives to be unsuitable but maintained a more positive general view of public transit during the strike compared to the other four groups. Multimodalists, comprised primarily of women, students, and lower-income individuals, relied on various transport alternatives, including carpooling, carsharing, ride-hailing (e.g., Uber), walking, and bikeshare services (i.e., Bixi) to get to work. They were the group most likely to report dissatisfaction with their travel-related stress. Active-travel Adopters used bike sharing services, walked, and in great majority, cycled to reach work during the strike. In fact, almost all respondents in this group own a bicycle, and stated higher satisfaction with the transport alternatives available to them during the strike compared to the other clusters. Teleworkers largely opted to work from home instead of commuting during the strike and were generally higher income. Finally, Auto-commute Switchers, with the highest levels of income and car ownership, stated primarily shifting to their personal car for their commute to adapt their commute during the strike. The non-work-trip clusters include Constrained Non-workers (22%), Multimodalists (17%), Shared-mode Adopters (12%), Unimpacted Higher-income Workers (34%), and Circumstantial Drivers (13%). Constrained Non-workers, who are generally women, older, people with disabilities, and lower-income respondents, primarily canceled or rescheduled their activities due to the transit strike. Multimodalists, similarly to the synonymous group from the work-trip clustering, carpooled, used ride hailing services, walked, and cycled to get to their non-work-related activities during the strike. Despite using multiple alternatives, they were highly dissatisfied with the options available to them. Shared-mode Adopters primarily used car sharing and bike sharing services to get to their non-work activities. They were the most satisfied with their transport alternatives compared to other groups and did not believe that public transport was a reliable way to get around during the strike. Unimpacted higher-income Workers did not report needing to use adaptive strategies to get to their non-work activities. They found transit to be more reliable than other groups and were generally satisfied with their punctuality in getting to their non-work activities. Finally, Circumstantial Drivers, with the highest level of car ownership and income of the five groups, primarily used their private vehicle to reach their non-work destinations. They still reported higher levels of travel-related stress, and many mentioned that the vehicular congestion worsened during the strike. This research highlights the inequitable impacts of Montréal’s June 2025 public transit strike. Underserved groups, including lower-income individuals, women, and people with disabilities, faced more severe disruptions and were more likely to walk for excessive distances, use ride hailing, or cancel trips, suggesting personal and financial burdens, and unmet needs. While the right to strike is a fundamental labour right and an important bargaining tool, strikes in essential services such as public transit should not disproportionately harm society’s most vulnerable groups. Findings from this research can be of value for public transit professionals and union members as it highlights the impact of a labour strike on the served population.

ABSTRACT. Presentation preference: in person (dependent on travel funding availability) Social equity is a key theme in transportation. TransLink – Metro Vancouver’s transportation authority – adopted a long-range transportation strategy – known as Transport 2050 – that includes a commitment to integrating an equity lens in transportation planning, policy, and decision-making. To support this effort, TransLink’s Research and Analytics team created a Transportation Equity Data Analytics Program to develop equity data analytics tools that quantitatively integrate social equity into transportation planning and policy projects such as transit service planning, multimodal transportation planning, infrastructure evaluations, and broader transportation regulation and policy work. This presentation will provide an overview and applications of the current transportation equity data analytics tools at TransLink, focusing on the Equity Data Analytics Tool (EDAT) and two scenario evaluation tools. EDAT is a self-serve dashboard enabling planners and policy makers to look at the current state of transportation equity in the Metro Vancouver region by visualizing socio-demographic indicators, equity composite indices, and access to opportunities metrics. The tool also combines demographic data from the census and present-day accessibility as estimated by TransLink’s Regional Transportation Model (a four-step trip generation model) to create indices of transportation poverty – areas with high populations of equity deserving groups and low access to opportunities by transit. The scenario evaluation tools are designed to visualize and compare transit accessibility to opportunities under different transportation planning scenarios, including changes for different equity-deserving groups compared to the general population, offering valuable insights for decision-making in transportation planning. Their difference lies in the data sources used to calculate travel times; the Equity and Access Scenario Evaluator (EASE) uses GTFS data and the R5 routing engine to calculate travel times between traffic analysis zones, while the Scenario Tool uses TransLink’s Regional Transportation Model to calculate travel times between the same zones, which also accounts for external factors like congestion. This difference also has implications for project applications: EASE allows for more iterations of different transit routing and scheduling scenarios and is used for near-term projects that assume travel behaviour will largely remain the same, while the Scenario Tool is used for major projects that will impact future travel behaviour and land development. The presentation will showcase these tools, highlight their functionalities and project applications, discuss challenges and limitations of each, and provide an outlook on future enhancements and updates.

ABSTRACT. *In-person presentation*

Equitable public transit is essential for enabling access to opportunity, reducing social exclusion, and supporting sustainable urban mobility. Yet these benefits are not consistently distributed, and vulnerable populations may reside in areas with limited service, unreliable operations, or reduced connectivity. Accessibility remains a central concept in transportation equity research. First conceptualized as the “potential of opportunities for interaction” (Hansen, 1959), accessibility measures an individual’s ability to reach desired destinations such as jobs, schools, healthcare, retail, and other essential amenities. There are several platforms that visualize transit accessibility and equity in North America, including Mobilizing Justice and TransitCenter’s Transit Equity Dashboard. While these tools provide detailed multidimensional insights, they often require users to interpret multiple overlapping indicators.

This presentation introduces the Transit Equity Score (TES), a single and interpretable 0–100 index that integrates social priority, transit accessibility, and service reliability to evaluate transit equity across geographic areas. Inspired by the success of American Forests’ Tree Equity Score, the TES provides an actionable metric designed to highlight neighbourhoods where targeted investment can most effectively improve transportation equity. The methodology is demonstrated using data collected for six major Canadian transit operators: TransLink, Edmonton Transit Service, Calgary Transit, OC Transpo, Société de transport de Montréal (STM), and Toronto Transit Commission (TTC). TES incorporates three components at the Statistics Canada Dissemination Area (DA) level: priority, accessibility, and reliability.

The Priority Index captures socioeconomic vulnerability using four variables drawn from the 2021 Census: median household income, unemployment rate, recent immigration (within five years), and housing cost burden (share of renter households spending over 30% of income on rent). Variables were normalized and weighted to produce a 0–1 priority value for each DA.

The Accessibility Index draws from Statistics Canada’s 2023 Spatial Access Measures dataset, which provides gravity-based accessibility indices for multiple amenity types at the Dissemination Block level. Accessibility is modeled using an exponential decay function to account for travel impedance, with closer opportunities weighted more heavily. Scores were normalized, aggregated across amenity categories, and spatially joined to DA boundaries.

The final component is reliability. A key contribution of this work is the creation of a cloud-based, multi-city data pipeline for collecting and processing real-time vehicle position data, enabling a standardized evaluation of transit reliability across all six transit operators. Using GTFS-realtime vehicle positions, the system continuously collected location data every 30 seconds. A geospatial matching algorithm identified observed arrival events within 15 metres of scheduled stop locations, which were then used to determine schedule deviation for each stop and trip. On-time performance was defined as arrivals between 1 minute early and 5 minutes late. This dataset provides one of the most comprehensive multi-city reliability baselines available for Canadian transit systems and fills a major gap in equity-focused evaluations, where reliability is often omitted due to lack of standardized data.

To compute the final Transit Equity Score, accessibility and reliability values were compared against their respective median values within each city. These “gap” values quantify where service conditions fall below a reasonable baseline. The TES multiplies the Priority Index by the average of the accessibility and reliability gaps to generate a 0–100 score, where higher values indicate neighbourhoods in which socially vulnerable populations face disproportionately limited access and unreliable transit service. The presentation will highlight how equity patterns differ across the six participating transit operators. This includes examining whether certain cities exhibit consistently higher reliability gaps, whether accessibility challenges cluster differently depending on network design, and whether socioeconomic vulnerability aligns differently with transit performance across regions. The presentation will also explore how using a unified methodology allows transit operators of varying size and geography to be evaluated side-by-side, offering insights into national-scale disparities in transit equity.

The TES provides a novel, data-rich way to assess transit equity that moves beyond traditional accessibility-only approaches. By combining national accessibility datasets, real-time transit feeds, and socioeconomic indicators, the score produces a simplified, communicable tool that supports research, advocacy, and capital planning. The presentation will demonstrate both the methodological framework and the feasibility of integrating diverse spatial and real-time data sources into a comprehensive equity measure.

ABSTRACT. Passenger satisfaction in public transport is strongly influenced by the quality of disruption information, yet current messages in Region Stockholm remain generic and insufficient. Personalized communication, particularly proactive disruption notifications targeted to passengers who regularly use affected lines or stations, offers a promising improvement. Identifying these passengers can be done either through self-selection or automated prediction based on historical travel patterns, with automated systems enabling more efficient and timely information delivery.

However, predictive models inevitably generate false positives and false negatives, creating potential costs due to irrelevant notifications or missed beneficiaries. No prior study has quantified these costs and benefits. This research addresses that gap by evaluating the system-level costs and benefits of proactive notifications using a regularity-based travel probability model derived from smartcard data of Region Stockholm.

ABSTRACT. A broad consensus has emerged that agencies should share service information relevant to customer trip-planning in standardized GTFS and GTFS-RT formats so that third-party platforms — Transit app, Google Maps, Apple Maps or Citymapper — are able to ingest and display this data to local riders.

The reality of universal open data has created a situation of fierce competition for any agency’s “official” platform: dominant platforms provided by the world’s largest corporations — Google and Apple — who express little interest in direct collaboration with transit agencies, in addition to dedicated consumer-grade transit apps that seek to convert large local audiences into direct commercial partnerships with agencies (e.g. Transit app, Citymapper).

This fierce competition raises the stakes for the quality of customer-facing apps that are built in-house or made-to-order: increasingly, in-house customer apps must be sufficiently well-funded and offer a constantly-evolving feature set to attract and retain rider loyalty. In this context, an “official app” partnership with a high-adoption existing platform can appeal to agencies: a customer trip-planning app is as valuable as the number of riders who choose to use it. Partnership thereby avoids the risk of heavy upfront investment in an in-house solution that does not compete successfully with existing platforms, leading to low adoption.

However, concerns remain about any strategic approach to improving the quality of customer information via an official third-party platform: commercial and brand risk; dependence on a third-party private corporation to serve the needs of local riders, and questions about the value of paying for an “official app” when so many zero-cost options are available to customers.

Amidst the overhaul of its Wayfinding and Digital Customer Experience strategies, TTC undertook a 1.5-year paid pilot (Sept 2024-April 2026) with Transit app to help improve customer information by using Transit’s AI-driven approach to detecting and displaying unplanned detours, while also improving the quality of TTC’s open data.

This joint session from Transit app and TTC staff will briefly present the results of the Pilot before overviewing the potential benefits and challenges of partnering with a third-party platform from the local perspective of Canada’s largest transit system.

ABSTRACT. To be considered as an attractive travel alternative, public transit must not only provide efficient and reliable service under regular operations but also demonstrate resilience (i.e., the ability to absorb and quickly recover from unplanned disturbances). Yet, these disruptions regularly challenge system resilience, forcing passengers to seek immediate substitutes such as agency-provided shuttles, ride-sourcing, or increasingly, shared micromobility options like bikeshare. Although large-scale subway delays are common, the causal role of bikeshare—as a flexible, first/last-mile resilience buffer around disrupted stations—remains insufficiently quantified, particularly through rigorous quasi-experimental methods.

In recent years, bikeshare has emerged as a promising alternative for travelers affected by transit service disruptions. For example, 37% of surveyed riders in Lyft’s U.S. operated markets report using shared micromobility when public transit is unavailable—with even higher shares in major metropolitan areas such as the Bay Area (42%), Chicago (45%), Portland (48%), Boston (50%), Washington, DC (48%), and New York (56%) (Lyft, 2025). Prior research using system-level bikeshare data has also documented statistically significant increases in bike-share ridership during planned, longer-term transit disruptions such as maintenance works (Cheng et al., 2022; Kaviti et al., 2020; Younes et al., 2019) and labor strikes (Fuller et al., 2012; Saberi et al., 2018; Zaltz Austwick et al., 2013). However, these studies largely focus on disruptions that are scheduled, prolonged, and communicated to riders in advance. Far less is known about whether bikeshare effectively supports resilience during unplanned, short-duration disruptions—typically triggered by unexpected incidents such as signal failures or track issues that cause sudden service reductions and unpredictable wait times.

This study addresses this gap by examining the extent to which bikeshare supports public transit resilience during unplanned subway disruptions. We apply a staggered and repeated difference-in-differences (DiD) design to estimate the causal impact of unplanned transit delays on bikeshare demand in Toronto, Canada during 2023. A key objective is to determine whether riders use bikeshare as a first/last-mile connector or as a full bypass of the disrupted subway segment.

Two datasets obtained from the City of Toronto’s Open Data Portal are used for the analysis. The first is the subway disruption dataset provided by the Toronto Transit Commission (TTC), which reports the location, cause, and duration of each unplanned delay. Delay is measured as the excess headway between consecutive departing trains relative to the scheduled headway (typically 2.5–5 minutes). We restrict the sample to significant disruptions, defined as delays exceeding 15 minutes. The second dataset consists of historical bike-share trip records from Bike Share Toronto, including trip start and end times, duration, origin and destination stations, bike ID, and user type (member vs. casual).

We define a 500-meter Station Resilience Zone around each affected subway station and operationalize treatment at the bikeshare-station level. The Treatment Group includes trips originating from bikeshare stations within this buffer during an active disruption. The Control Group comprises all trips originating outside the buffer, including stations unaffected by any delay at that moment. For each disruption, the Treatment Period (Post) spans the full delay duration plus a 60-minute decay window to capture lingering crowding and continued diversion behavior, while the Baseline Period (Pre) is the immediately preceding window of normal operations.

Given the count nature and over-dispersion of bikeshare trip data, the primary estimation employs a Negative Binomial regression. To ensure unbiased causal inference in a staggered, repeated-event DiD setting, we employ a modern, decomposition-based estimator (Callaway & Sant’Anna, 2021). We further assess the plausibility of the parallel trends assumption through a dynamic event-study specification.

The results of this study will offer insights into the role that bikeshare plays in promoting transit system resilience during sudden service disruptions. These insights can be used to inform coordinated strategies between transit agencies and bikeshare operators—such as targeted, real-time bike rebalancing around affected stations, integrated service alerts, and more efficient shuttle deployment. Strengthening these multimodal connections can help alleviate shuttle crowding, reduce reliance on costly ride-sourcing options, minimize curbside congestion due to ride-sourcing passenger pickup that further slows shuttle operations, and ensure riders have reliable alternatives during system shocks. Ultimately, this work contributes to building a more adaptive, resilient, and passenger-centered transit ecosystem.

ABSTRACT. Metro service disruptions alter urban mobility patterns, yet comprehensive understanding of passenger behavioral responses across disruption phases remains limited. This research develops an analytical framework to examine passenger behavior, specifically mode choice and path choice decisions, before, during, and after metro service disruptions. By analyzing both planned Revenue Service Adjustments (RSAs) used for maintenance activities and unplanned events, the framework provides a complete picture of how passengers adapt their travel patterns in response to different types of service interruptions. The research aims at providing an understanding of passenger responses across the full disruption timeline. Rather than examining disruptions as isolated events and capture snapshots of behavior during disruptions, we trace the crucial transitions: how normal travel patterns evolve when disruptions begin, adapt during service interruptions, and either recover or stabilize into new patterns after restoration. By examining mode choice (whether passengers stay with metro, switch to car, bus, or other modes) and path choice (how passengers adjust routes within the transit network) as parallel decision processes, we capture the full spectrum of behavioral adaptation. Our framework compares passenger behavior across three distinct phases. The "before" phase establishes baseline travel patterns under normal service conditions, capturing regular mode choices, route selections, and travel frequencies. The "during" phase tracks how these patterns change when disruptions occur, identifying immediate responses versus evolved adaptations over multi-day events. The "after" phase examines recovery patterns, revealing whether and how quickly behaviors revert to baseline, partially return, or adopt new patterns. This temporal analysis applies to both planned disruptions, where advance notice allows passenger preparation, and unplanned incidents requiring immediate response. We utilize Automatic Vehicle Location (AVL), Automated Fare Collection (AFC), and comprehensive infrastructure data from the Washington Metropolitan Area Transit Authority (WMATA). This core dataset is supplemented with additional sources to capture behaviors beyond the metro system: Transportation Network Company (TNC) data provides information on ride-hailing, bikeshare records show potential changes near affected stations, and Metrobus data captures bus substitution patterns. By combining these sources, we can track not only passengers who remain within the transit system but also those who temporarily or permanently shift to other modes. Our analysis examines behavioral patterns across different passenger segments and disruption contexts. By using panels of users with distinct patterns (e.g. frequent and infrequent users), we investigate how travel behaviors evolve from stable baselines through various adaptation strategies. This will also help identify factors that influence whether passengers maintain transit use, temporarily adapt, or permanently shift to other modes, with particular attention to variation across passenger characteristics and trip purposes. The differences between planned and unplanned disruptions are also investigated. For planned RSAs, passenger behavior can be tracked from the announcement period through service changes to eventual recovery. For unplanned incidents, the focus is on capturing immediate responses and subsequent adaptation patterns. The comparison also helps determine how information delivery can impact passenger experience, providing practical insights for communication and mitigation strategies. Understanding the whole behavioral picture across disruptions offers practical value for transit management. These before-during-after behavioral trajectories enable phase-appropriate interventions: preparation and information before planned disruptions, real-time alternative routing during disruptions, and recovery incentives after service restoration. Recognizing that mode and path choices interact suggests coordinated strategies addressing both decisions simultaneously. Identifying passenger groups that are most likely to abandon transit versus those who temporarily adapt, can inform targeted retention efforts. As transit systems face increasing maintenance needs and service challenges, comprehensive understanding of passenger behavioral adaptation provides useful insights into strategies to retain ridership through inevitable service interruptions.

ABSTRACT. In recent years, climate change has been imposing a growing direct impact on the functioning of transport infrastructure. For example, in 2022, extreme heat resulted in speed restrictions and the closure of stations and lines on the London Underground (Transport for London, 2022). In 2024 in the region of Bavaria in Germany, extreme rainfall caused flooding which severely affected large sections of the Autobahn highway network (Deutsche Welle, 2024). As a result of the complex interdependent nature of transport networks, the extents of disruption to transport services can be difficult to predict and manage (Zhu & Levinson, 2011). This can result in unanticipated prolonged degradation of services, potentially leading to adverse outcomes for human health (e.g. due to limited access to medical facilities), substantial financial losses for transport operators, and durable degradation in perceived transport service quality (Yap & Cats, 2021). The current key gaps in the literature are as follows: (i) the vast majority of disruption quantification studies focus on operational disruptions, while the analysis of extreme weather related disruptions is relatively scarce; and (ii) due to lack of access to or availability of empirical data, the majority of studies in the field adopt theoretical simulation methods and disruption scenarios, and therefore the "real-world" impacts of climate-related disruptions are rarely quantified.

In this work, we aim to robustly quantify the impact of one of the most severe flooding events experienced in Geneva, Switzerland on public transport operations in the city using big data. In mid November 2023, the Arve River reached its highest recorded discharge flow levels exceeding 1000 cubic metres per second (Swiss Broadcasting Corporation, 2023). The flooding resulted in severe disruptions to traffic and public transport services in the city. Main tram lines were reported to be severed, preventing the crossing of passengers from one river bank to the other, and 5 out of 8 bridges crossing the river were closed (Swiss Broadcasting Corporation, 2023).

The flooding event took place on 14 November 2023. The dataset used in this analysis covers 1-30 November 2023, and consists of 190 unique line/direction combinations which comprise bus and tram services in Geneva. The data have records of vehicle demand at each stop, along with disaggregate values of the number of boards and number of alights, and the scheduled and actual arrival/departure times of each service at each stop. Of the 190 total routes, we have identified 20 routes in the city which were most likely directly affected by the flooding event and these routes are the subject of the analysis presented here.

We model the impact of the flooding event using the causal inference method of regression discontinuity design (Imbens & Lemieux (2008) and Lee & Lemieux (2010)). The aim of the method is to quantify whether a statistically significant discontinuity is observed at the point of the flooding event, which indicates evidence of a causal effect (Imbens & Lemieux (2008) and Lee & Lemieux (2010)). In order to do this, we create a binary assignment indicator in the dataset (refer to equation 1 in the attached pdf)

The average treatment effect for a sharp discontinuity at the treatment threshold is defined as given in equation 2 in the attached pdf. We model whether the flooding event causes a change in demand at a public transport stop level. The general form of the RDD to estimate the average treatment effect at a stop level is given in equation 3 in the attached pdf. All modelling has been undertaken using R statistical analysis software.

We estimated a total of 846 models, one for each of the bus stops on the potentially affected routes. Adopting a level of statistical significance of 90%, the average treatment effect is statistically significant in 405 models. We therefore see that the flooding event has a causal impact on approximately 48% of public transport stops on the 20 identified routes which were considered to be most affected. Table 1 in the attached pdf summarises the distribution of the statistically significant average treatment effects, and Figure1 in the attached pdf illustrates the geographical distribution of the statistically significant average treatment effects.

As shown in Table 1, there are reductions and increases in demand. The average reduction in demand observed is approximately 28 passengers per stop while the average increase in demand is approximately 41 passengers per stop. The stops which experience the largest reductions in demand correspond to the busiest lines in the centre of city, while the stops which experience increases in demand and minor changes are located in outer zones (refer to Figure 1 in the attached pdf). We are currently undertaking further investigations to ascertain the potential drivers of this distribution of effects, and will include further analysis in our presentation and the full version of the paper.

ABSTRACT. Public transit systems in suburban areas face persistent challenges arising from low and spatially dispersed demand, which limits the efficiency of fixed-route transit (FRT). Conventional bus networks often operate with low occupancy, high per-passenger subsidy, and limited accessibility for residents in low-density neighborhoods. These limitations motivate transit agencies to explore flexible, data-driven solutions that can adapt supply to variable demand conditions. On-demand transit (ODT), enabled by digital platforms, offers such flexibility by dynamically routing shared vehicles based on passenger requests. This research develops and applies a simulation-based, data-driven analytical framework to evaluate the operational and financial performance of ODT compared to FRT in three suburban service zones within the City of Mississauga—Meadowvale, Airport Corporate Centre, and Clarkson–Port Credit. Each zone represents a unique urban form, land-use pattern, and travel behaviors, allowing for comparative assessment of ODT’s suitability under different spatial and temporal conditions. A key feature of this study is its use of multi-source real-world data, including Automatic Passenger Counts (APC, 2024), General Transit Feed Specification (GTFS, 2024), and Transportation Tomorrow Survey (TTS 2022) to ensure realistic representation of travel demand, service supply, and operating environments. In the selection of candidate zones, APC (2024) and GTFS (2024) data were jointly used to identify underutilized bus routes based on observed ridership and load factors. GTFS schedule data were used to delineate transit deserts—areas with limited route coverage during all time periods of weekdays and weekends. Accessibility analyses conducted using GTFS travel-time matrices further quantified how travel times to key destinations such as employment centers, hospitals, schools, and grocery stores varied across zones. These analyses collectively established the empirical basis for selecting Meadowvale, Airport Corporate Centre, and Clarkson–Port Credit as representative case studies for ODT deployment. In the development of scenarios, the study analyzed origin-destination survey data known as the Transportation Tomorrow Survey (TTS 2022), to understand household, population characteristics, trip purposes, and estimate intra-zonal and First-Mile/Last-Mile (FMLM) connectivity patterns within each zone. The information guided the design of service configurations. In the modelling phase, APC data for all the routes in consideration to be partially or completely replaced with ODT were used to build empirical origin–destination (OD) matrices reflecting both zone-to-zone and zone-to-hub travel flows across time periods. The transit network layer was generated from OpenStreetMap (2024) to reflect up-to-date street geometry and routing constraints, while TTS (2022) data was used to create sub-zones and population-weighted demand centroids within each service area. These steps ensured that the simulated network structure and demand distribution mirrored real-world conditions. Three distinct mobility scenarios were evaluated for each zone: (1) baseline fixed-route transit (FRT), (2) ODT operated by the transit agency (ODT-Agency), and (3) ODT operated by a transportation network company (ODT-TNC). The ODT scenarios were modelled using PTV Visum, which enabled simulation of demand-responsive services using the modules Trip Requests Generation and Tour Planning. Vehicle capacities, detour tolerances, and service response times were parameterized based on empirical observations and planning guidelines. Multiple demand seeds were simulated to capture stochastic variability, and results were averaged across runs for statistical robustness. Performance evaluation was conducted along three main dimensions: cost efficiency, level of service, and community impact. Cost metrics included total system cost per trip and per passenger. An Equivalent Uniform Annual Cost (EUAC) analysis was applied to estimate capital costs for ODT and FRT, incorporating fleet costs and platform fee (only for ODT by the agency). The operating cost was calculated as a function of vehicle operating expenses, driver wages, and maintenance costs. Service quality was assessed using passenger journey time, wait time, and experienced detours, all extracted from detailed trip-level outputs of Visum. Community-level outcomes were measured through accessibility indicators, fleet size requirements, and total vehicle-kilometers travelled (VKT). The results show that data integration across all phases substantially enhanced the realism and policy relevance of the simulations. ODT consistently outperformed FRT in passenger experience, primarily through reductions in waiting time and improved routing efficiency. In Meadowvale, where residential demand is spatially dispersed, ODT–Agency achieved up to 25% shorter average journey times relative to FRT. In the Airport Corporate Centre, where demand is concentrated and directional, the improvement was around 10–12%. In Clarkson–Port Credit, which features higher baseline FRT efficiency and a larger geographic area, travel-time gains were modest, and ODT required longer detours. Cost outcomes were strongly zone-specific: ODT was less costly than FRT in Meadowvale, cost-comparable in the Airport Corporate Centre, and more expensive in Clarkson–Port Credit. These variations were closely linked to zone characteristics derived from APC and GTFS data, including the spatial and temporal distribution of demand. Furthermore, ODT–TNC required the largest fleet size and generated the highest VKT. Sensitivity analyses, supported by APC-based demand scaling, revealed a clear demand threshold beyond which FRT becomes more cost-effective. Below that threshold, ODT remains financially advantageous while providing higher service quality. Overall, this study demonstrates how the structured integration of APC, GTFS, TTS, and other relevant data can support evidence-based transit planning and simulation. The coordinated use of these datasets across zone selection, scenario development, and simulation stages ensured empirical grounding, enabling a realistic assessment of ODT’s operational and economic viability. The framework not only provides analytical insights for transit agencies but also establishes a replicable methodology that other municipalities can adopt using their own ridership and network datasets.

ABSTRACT. In Demand-Responsive Transit (DRT) systems, zoning is a fundamental design element that determines how the service area is partitioned into operational districts within which vehicles must serve trips. Unlike fixed-route transit, DRT requires real-time scheduling and routing, making unconstrained citywide operations difficult to scale, costly, and prone to high rejection rates. Dividing the region into DRT operation zones allows the system to limit vehicle circulation to manageable subareas, balance workload across the fleet, and ensure predictable service quality for users. In practice, zoning means assigning each Basic Spatial Unit (BSU) in the city to one of several contiguous service regions, where each zone independently handles requests that originate and end within its boundaries. The zoning optimization aims to form clusters of BSUs that minimize operational costs while satisfying the requirement that each zone remains a contiguous and connected subgraph of the network. The origin–destination (OD) demand used in this study is estimated from 2019 daily boarding counts of elderly passengers at the TAZ (Traffic Analysis Zone) level in Winnipeg, which reveal spatial and temporal travel patterns that shape how zones should be delineated. Consistent with these observed elderly mobility flows, the downtown core is treated as a shared hub accessible to all zones and representing the most common destination cluster for this population group. While downtown is universally reachable, each zone otherwise operates independently; trips that begin and end in different zones—except those destined for the downtown hub—are rejected. Consequently, the quality of the zoning structure directly affects accepted-trip ratios, waiting times, and the operational burden on the DRT fleet. This creates a well-defined but computationally overwhelming optimization problem. With 1,154 BSUs (i.e., the total number of TAZs) and a fixed number of zones (i.e., five predetermined zones), each BSU has five possible assignments. Planners typically provide an initial zoning structure based on their domain expertise and observed demand patterns, but improving upon this baseline requires searching through an immense combinatorial landscape. Traditional optimization tools, including Genetic Algorithms (GA), face steep performance barriers in this environment. They struggle to meaningfully explore high-dimensional boundary configurations, especially when the starting configuration is already expert-generated and near a local optimum. Moreover, ensuring contiguity and maintaining downtown connectivity add structural constraints that cannot be handled efficiently by random mutation or uninformed crossover. To address these challenges, this study develops and applies a Reinforcement Learning–guided Genetic Algorithm (RL-GA) framework that embeds a Double Deep Q-Network (DDQN) into the evolutionary search process. Instead of allowing the GA to blindly test boundary modifications, the DDQN learns which zoning adjustments are feasible, meaningful, and likely to yield improvement. The agent masks actions that would break contiguity, split zones into disconnected components, or leave boundaries unchanged, thereby reducing the effective action space to a manageable subset focused on productive adjustments. The RL component therefore acts as a structural guide, directing crossover and mutation toward boundary modifications that preserve connectivity while potentially reducing rejected trips and user-cost metrics. In essence, clustering BSUs into DRT operations zones becomes a guided graph-partitioning problem, where the DDQN prioritizes changes that improve operational and service performance. The developed RL-GA framework is applied to Winnipeg to as a realistic and highly relevant test environment. The combination of city’s distributed urban form, severe winter conditions, and elderly population group characteristics create strong geographic and temporal variations in demand. The baseline zoning used in this study is generated through a demand-informed clustering of elderly origin–destination flows. Improving such a high-quality baseline is considerably more challenging, emphasizing the need for an intelligent search strategy. A key contribution of this work is demonstrating that even with these constraints and the extremely large action space, the RL-GA framework can produce improvements within practical computational limits. While brute-force or naïve GA approaches would require many hours of computation without guarantee of meaningful progress, the proposed RL-guided search is able to produce improved zoning configurations within roughly two hours on a standard computer environment. This is due to (1) masking infeasible or low-value actions, (2) focusing on boundary nodes where improvements are most likely, and (3) avoiding unnecessary evaluation of zoning assignments that violate contiguity or operational constraints. These characteristics illustrate that application of RL can greatly enhance the efficiency of DRT zoning optimization even before any large-scale distributed computing resources are introduced. Overall, this research contributes a novel, AI-driven framework for zoning design in DRT systems by showcasing a real-world scenario closing the gap between theoretical AI methods and practical implementations. By translating the zoning problem into a BSU-level clustering task with connectivity constraints and using RL to intelligently navigate the action space, the proposed method provides a scalable decision-support tool for transit planners. The approach is particularly applicable for systems serving elderly riders or other population groups with specific needs, where minimizing rejected trips, waiting times, and unnecessary travel is essential. While demonstrated using Winnipeg data, the methodology is generalizable to other cities and can be adapted for micro-transit, paratransit, and mobility-on-demand systems seeking to design contiguous, equitable, inclusive, and operationally coherent service zones.

ABSTRACT. Introduction Mobility-as-a-Service (MaaS) aims to provide seamless multimodal travel by combining conventional Public Transport (PT) with flexible Transport On Demand (TOD). Despite advances in trip-planning applications, operational coordination between these services remains minimal. PT relies on structured automated datasets—such as smart-card transactions, Automatic Vehicle Location (AVL) feeds, and scheduled timetables—while TOD platforms operate on continuously updated, highly granular user data. Achieving operational MaaS integration requires some form of information exchange, but sharing detailed user trajectories or proprietary operator datasets raises significant privacy, confidentiality, and competition concerns, if different companies are involved. This creates a fundamental tension: better coordination requires more information, yet practical deployment requires strict limits on what can be shared. This work investigates whether coarse, privacy-preserving signals can support meaningful multimodal interaction without exposing sensitive data or requiring centralized control.

Problem The research question addressed in this work is whether anonymization and data-minimization mechanisms can enable operational PT--TOD coordination while satisfying the privacy expectations of users, the confidentiality constraints of operators, and the broader societal goals associated with sustainable mobility. In the proposed setting, user inputs may range from detailed information to anonymized or coarse descriptions—such as approximate pickup areas or broad time windows—rather than the detailed, continuous location traces typically required by TOD platforms. TOD processes these inputs internally using its proprietary demand estimation and routing logic, while PT continues to operate with its existing automated datasets and fixed-route structure. Because neither operator is permitted to disclose raw data or proprietary algorithms, only coarse, low-sensitivity indicators may be shared. These constraints arise from structural features of the sector, including competitive dynamics between operators, legal and regulatory limits on inter-organizational data exchange, and the coexistence of multiple, institutionally independent TOD actors with heterogeneous governance arrangements. Consequently, any coordination mechanism must function under incomplete information, asymmetric visibility of demand, and model incompatibilities, while also preserving strict privacy guarantees. The central methodological question is whether such limited and noisy signals can nevertheless support coordination schemes that yield measurable operational improvements.

Solution Method A distributed optimization framework is proposed to coordinate PT and TOD without centralized data collection or exposure of sensitive information. The approach integrates privacy-preserving internal processing, autonomous operator optimization, and a distributed coordination architecture. On the PT side, AVL, smart-card, and schedule data are processed using standard prediction and real-time control tools, producing short-horizon arrival forecasts, headway estimates, and capacity indicators. On the TOD side, coarse user-provided signals are transformed into internal demand estimates and routing decisions. In both cases, operators expose only aggregated indicators—such as expected arrival intervals, estimated demand levels, or ranges of available capacity—ensuring that neither detailed user data nor proprietary algorithms are revealed. Operator autonomy is fully preserved, with each party running its own optimization routines while sharing only low-sensitivity summaries. Coordination between PT and TOD is achieved through a message-passing layer inspired by distributed optimization, where iterative exchanges of aggregated indicators allow partial solutions to be aligned despite asymmetric and incomplete information. The architecture distributes computation across user devices, PT back-end systems, and TOD servers, enabling privacy-by-design through decentralized processing and minimizing the need for centralized data aggregation. Probabilistic estimation methods compensate for missing or anonymized information, allowing the system to infer likely multimodal interactions and maintain robustness under uncertainty. Ultimately, the aim is to obtain a flexible and scalable framework that can be extended to other service combinations—such as carsharing, bikesharing, carpooling, and ridesharing—supporting privacy-preserving coordination across a wide range of emerging mobility services.

Experiments The framework is evaluated using a simulation environment based on MultimodalSim, a CIRRELT-developed platform adapted to represent multimodal operations in the Québec City region. Experimental scenarios vary along three dimensions: the level and type of anonymization applied to user-provided TOD inputs, the degree of information visibility available to each operator, and the structure and strength of the distributed coordination mechanism. Performance is assessed through multimodal indicators such as TOD-to-PT transfer waiting times, alignment between TOD pickups and PT arrivals, robustness of inferred multimodal demand flows, and sensitivity to different cooperation levels. Comparing scenarios quantifies the operational feasibility of privacy-preserving PT--TOD coordination without centralized data or unified optimization algorithms.

ABSTRACT. Introduction In contrast to urban municipalities, rural and low-density areas tend to have decreased transportation accessibility due to their automobile-oriented development. This has a disproportionate impact on disadvantaged groups with limited vehicle access, especially since lower ridership and limited routes often render traditional fixed-route transit economically infeasible [1]. As a result, on-demand transit (ODT) has been used as an alternative solution to introduce, supplement, or replace public transportation in these areas. Operating with no fixed route or schedule, ODT systems dynamically adjust their routes to pick-up and drop-off passengers who book their trips in advance [2]. Despite the potential to improve mobility, around half of ODT services fail due to high costs and complicated service designs [3], suggesting a need to further study how these systems operate. Understanding the travel patterns of ODT can guide future planning by informing transit agencies how, when, and where the service is being used—enabling municipalities to create more resilient and equitable transit systems. In the case of Innisfil, Canada, previous research applied a sustainability framework to assess the efficiency, environmental impact, and accessibility of the system using empirical data collected from one day of operating the service [4]. Additionally, the impact of COVID-19 on ridership was studied in comparison to the ODT system in Chicago, Illinois from 2020-2021 [5]. Another study considered the predictors for ridership growth between 2016-2020 [6]. Currently, no research has been conducted that considers the travel patterns and distributions of ODT in Innisfil.

Objectives This study aims to investigate the travel patterns, distributions, and overall system performance of ODT systems in low-density municipalities by using the town of Innisfil, Canada as a case study to inform future transit planning.

Methodology The study is conducted using trip-level empirical data and passenger satisfaction surveys from 2018-2022 provided by the town of Innisfil. The analysis will be conducted in three parts. First, the performance of the ODT system will be evaluated by calculating key performance metrics listed in the Canadian Urban Transit Association on-demand transit toolkit [7] to assess the feasibility and effectiveness of the system. Second, statistical analyses will be applied to identify the distributions of trip distance, trip length, and wait time. Lastly, this study will investigate the travel patterns of the system and use clustering models to identify categories of travel patterns. Without labelled data, it may not be possible to validate the trip purpose [8], however, the clusters may provide some insight as to what they may be. This can also be estimated using activity inference frameworks that consider the passenger’s drop-off location, time-of-day, and probabilities of visiting nearby points of interest [9].

Expected Results and Discussion Since the data was collected, on-demand transit has continued to operate as the primary transit network in Innisfil [10], indicating a certain degree of success in implementation. In turn, we expect the calculated system performance measures to indicate mostly positive performance. With respect to travel patterns, we expect that a significant portion of trips are used as first and last mile connections to nearby train stations. The distributions of trip displacement and duration may follow either gamma or lognormal distributions, as was the case in similar studies using taxi trips [11]. By evaluating the travel patterns, distributions, and system performance of on-demand transit, this research can be used to forecast how travel demand will change in the future, guide the use of existing infrastructure, and plan for future transit developments [8].

References [1] V. D. Pyrialakou, K. Gkritza, and J. D. Fricker, “Accessibility, mobility, and realized travel behavior: Assessing transport disadvantage from a policy perspective,” Journal of Transport Geography, vol. 51, pp. 252–269, Feb. 2016, issn: 0966-6923. doi: 10.1016/j.jtrangeo.2016.02.001. Accessed: Nov. 18, 2025. [2] W. Klumpenhouwer, A. Shalaby, and W. Lee, “The State of Demand Responsive Transit in Canada,” Tech. Rep., 2020. [3] G. Currie and N. Fournier, “Why most DRT/Micro-Transits fail – What the survivors tell us about progress,” Research in Transportation Economics, Thredbo 16 Conference, vol. 83,p. 100 895, Nov. 2020, issn: 0739-8859. doi: 10 . 1016 / j . retrec . 2020. 100895. Accessed: Nov. 18, 2025. [4] N. Alsaleh and B. Farooq, “Sustainability analysis framework for on-demand public transit systems,” Scientific Reports, vol. 13, no. 1, p. 13 488, Aug. 2023, issn: 2045-2322. doi: 10.1038/s41598-023-40639-y. Accessed: Nov. 18, 2025. [5] N. Alsaleh and B. Farooq, “The impact of COVID-19 pandemic on ridesourcing services differed between small towns and large cities,” PLOS ONE, vol. 17, no. 10, e0275714, Oct. 2022, issn: 1932-6203. doi: 10.1371/journal.pone.0275714. Accessed: Nov. 20, 2025. [6] A. Benaroya, M. Sweet, and R. Mitra, “On-demand ride hailing as publicly subsidized mobility: An empirical case study of Innisfil Transit,” Case Studies on Transport Policy, vol. 11, p. 100 944, Mar. 2023, issn: 2213-624X. doi: 10.1016 j.cstp.2022.100944. Accessed: Nov. 20, 2025. [7] CUTA and Metrolinx, “On-demand Transit Toolkit,” 2022. [8] N. Breyer, Methods for Travel Pattern Analysis Using Large-Scale Passive Data. May 2021, isbn: 978-91-7929-665-0. doi: 10.3384/diss.diva-175347. [9] L. Gong, X. Liu, L. Wu, and Y. Liu, “Inferring trip purposes and uncovering travel patterns from taxi trajectory data,” Cartography and Geographic Information Science, vol. 43, no. 2, pp. 103–114, Mar. 2016, issn: 1523-0406. doi: 10.1080/15230406.2015.1014424. Accessed: Nov. 18, 2025. [10] Town of Innisfil, “Innisfil Transportation Master Plan,” 2023. [11] Z. Kou and H. Cai, “Understanding bike sharing travel patterns: An analysis of trip data from eight cities,” Physica A: Statistical Mechanics and its Applications, vol. 515, pp. 785–797, Feb. 2019, issn: 0378-4371. doi: 10.1016/j.physa.2018.09.123. Accessed: Nov. 17, 2025.

ABSTRACT. Introduction: While On-Demand Transit (ODT) is increasingly considered a solution for transit deserts, a more complex dilemma arises in areas where Fixed-Route Transit (FRT) services currently exist but are underperforming due to factors such as low demand. In these scenarios, agencies face a critical "Fix or Flex" decision: determining whether to optimize the existing FRT system to improve efficiency or to partially or fully substitute it with flexible ODT services. Current practices for evaluating transit performance often rely on lists of Key Performance Indicators (KPIs) that result in absolute efficiency scores. However, absolute scores fail to reflect the true performance quality of a transit route relative to its peers. To address this, this study proposes a relative performance analysis using non-parametric approaches. Specifically, we address a major gap in the literature regarding the choice between convex (Data Envelopment Analysis - DEA) and nonconvex (Free Disposal Hull - FDH) production technologies in the context of public transit. Furthermore, this study moves beyond simple route-level analysis to conduct a granular, multi-period evaluation at the segment level, integrated with demographic analysis to provide actionable, data-driven recommendations. Methodology: This research develops a novel, integrated framework for evaluating FRT performance to support the "Fix or Flex" decision-making process. The methodology consists of several critical components: 1. Convexity Testing: A pivotal step in our framework is determining the nature of the production technology. We utilize the FEAR package to test the null hypothesis of convexity versus nonconvexity. This ensures the selection of a robust evaluation model, either DEA (convex) or FDH (nonconvex), that best aligns with the actual structure of the transit services. 2. Slack-Based Measure (SBM): Based on the convexity test results, we employ a Slack-Based Measure model. Unlike radial models, SBM identifies inefficiency sources arising from specific slack and surplus values, offering precise guidance on whether specific inputs (e.g., vehicle capacity, frequency) should be adjusted or outputs (e.g., ridership, load factor) improved. 3. Malmquist Productivity Index (MPI): To capture temporal fluctuations, we integrate the SBM model with the Malmquist Productivity Index. This multi-period analysis decomposes productivity changes into "efficiency change" (catching up to the frontier) and "technical change" (shifts in the frontier itself), highlighting whether performance issues are persistent or temporary. 4. Demographic and Coverage Analysis: The quantitative results are contextualized through a demographic analysis of the catchment areas. Variables such as student and senior population density, household income, and car ownership are examined to validate whether low-performing segments are candidates for ODT or simply require FRT redesign. Case Study and Data: The proposed framework is applied to the transit network of Mississauga, Canada (MiWay). Data was collected from January 2023 to September 2024, covering 21 operational periods. Through an interactive process with the transit agency, inputs were defined as vehicle capacity, frequency, operating kilometers, number of stops, shared stops, service duration, and planned on-time performance. Outputs included load factor, boardings per service hour, cost recovery, actual on-time performance, and user satisfaction. Results: The application of the framework yielded several key findings regarding both the methodology and the specific transit network: • Validation of Convexity: The convexity tests consistently returned high p-values across all operational periods, failing to reject the null hypothesis of convexity. This validates the use of convex models like DEA for this specific dataset, although we argue that testing remains essential as network design problems can be inherently nonconvex. • Route-level Analysis: Results highlight substantial temporal fluctuations in efficiency scores across the network. Many routes traversing the dense city center perform efficiently or near efficiently. In contrast, several peripheral zones, particularly the south, east, and northwest, contain routes with persistently low efficiency. Analysis of MPI components reveals that performance changes are driven by a combination of efficiency change and technical change, underscoring the necessity of multi-period monitoring. The integrated interpretation of efficiency scores, slacks, surpluses, and MPI allows us to classify routes into categories such as consistently efficient, acceptable but improvable, recently improving, or candidates requiring deeper investigation for ODT transition. • Segment-level Analysis: By dividing each route according to major timing points we evaluate 159 segments across all periods. Whereas 87% of routes achieve average efficiency above 0.6, only 50% of segments do. This contrast illustrates the masking effect that route-level averages can create. Segment-level results expose highly localized inefficiencies, particularly in the south and west, where many segments score below 0.5. Comparing segment-level MPI with route-level MPI shows that performance declines are often driven by a specific segment rather than an entire route. • The "Fix or Flex" Determination: Focusing on the southern area of Mississauga, the framework identified specific strategies for inefficient segments. For example: o Segments for ODT Conversion (Flex): Segments such as #4a and #4b displayed consistently low efficiency (averaging ~0.26–0.30) with high unused capacity (~72%). o Segments for Redesign (Fix): Conversely, Segments #8a and #8b showed high potential (efficiency scores >0.6). The model recommended specific "fix" strategies, such as reducing segment length by ~1.2 km, removing specific stops, and utilizing smaller vehicles to improve load factors and cost recovery. o Complex Cases: For Segment #29b, the analysis combined efficiency scores with demographic analysis. It revealed that the segment acted as an inefficient detour. The optimal solution identified was a hybrid approach: partially removing the detour segments and modifying the adjacent Route #45 to cover the gap, thereby streamlining the network without sacrificing coverage. Conclusion: This study establishes a robust, data-driven framework for transit agencies navigating the transition between fixed and flexible transit modes. By integrating non-parametric performance evaluation and demographic analysis, the framework moves beyond simple performance scoring. It provides a diagnostic tool that pinpoints the root causes of inefficiency, whether they stem from route design, operational planning, or external demographic factors, and prescribes tailored solutions. The results from the Mississauga case study demonstrate that while some FRT inefficiencies can be fixed through targeted redesign, others are structural and best addressed by replacing with ODT, ensuring resources are allocated effectively while maintaining equitable service coverage.

ABSTRACT. Transit agencies are increasingly partnering with Transportation Network Companies (TNCs) like Uber to solve the persistent first/last-mile (FLM) problem. While the potential is widely discussed, empirical evidence on how riders actually use integrated TNCs for FLM access within an operational Mobility-as-a-Service (MaaS) ecosystem remains scarce. Most existing studies rely on stated preference surveys or simulated data. This study fills this critical gap by analyzing a unique, high-resolution dataset from a public MaaS platform to provide a real-world, empirical characterization of FLM behavior. We use nine months (April-Dec 2024) of integrated telemetry data from the Vamos-EZHub MaaS platform in San Joaquin County, California. The datasets used in this study include timestamped Uber trip origins and destinations, transit ticket activations, GTFS schedules, and neighborhood characteristics. A key feature of the study context is a natural experiment: users purchasing >$5 in transit fares received a $5 Uber credit, creating a natural experiment to observe FLM behavior among active transit riders. We identify FLM-connected Uber trips using a spatiotemporal rule: an Uber ride is classified as FLM if it begins or ends within 250 meters of a transit stop and occurs within 90 minutes of a validated transit ticket activation. The analysis proceeds in two parts: 1. User-Level Patterns: We quantify the share of Uber trips that are FLM-connected versus standalone and analyze variation by time of day, day of the week, and month. This reveals whether FLM use is concentrated in specific temporal gaps (e.g., off-peak, late-night) or is spread throughout the day. 2. Stop-Level Regression Analysis: We aggregate FLM trips to each transit stop to create monthly FLM demand metrics. We then use multivariate regression models to link this demand to key predictors: (a) GTFS service characteristics (headways, span of service, number of routes), (b) built environment indicators (density, land-use mix, walkability), and (c) neighborhood socioeconomic conditions (vehicle availability, poverty rate, racial/ethnic composition). Furthermore, we analyze a linked member survey subsample to explore whether users from low-income or zero-vehicle households, or those living in transit-poor neighborhoods, are more likely to be FLM-heavy users and to rely on Uber connections at specific times (e.g., early shifts, late-night returns). Expected results will identify (1) critical temporal windows where Uber fills key transit gaps, (2) specific transit stops that function as FLM hubs, and (3) the stop and neighborhood attributes most predictive of FLM reliance. Our findings will offer transit agencies evidence-based guidance for designing targeted FLM partnerships, service adjustments, and infrastructure investments to enhance equitable, multimodal mobility.