In the dynamic landscape of contemporary metropolitan areas, the role of transit travel is increasingly pivotal. Transportation modeling practices, crucial for understanding both traffic and transit movements, require meticulous representation of transit-related frameworks, including schedules, networks, and parameters regularly updated by transit agencies. Larger Metropolitan Planning Organization (MPO) agencies, responsible for integrating multiple transit systems into their models, face unique challenges.

To address the hurdles encountered by transportation modelers, this presentation introduces an innovative automated tool designed specifically for determining route headways and itineraries. Leveraging General Transit Feed Specification (GTFS) data, the tool significantly reduces the manual burden on modelers, bridging the gap between the urgent needs of transportation modelers and the supply of automation.

Background

Understanding transit users’ preferences for premium services such as Light Rail Transit (LRT) and Bus Rapid Transit (BRT) is crucial for improving service design, accessibility, and long-term investment strategies. However, traditional discrete choice models often struggle to fully capture the complexity of choices, especially when dealing with rich non-path attributes like demographics and trip contexts (e.g., trip purpose) available in standardized on-board surveys. These valuable data sources have historically been underutilized due to the complexity, non-linear effects, and high cardinality they introduce, making their integration into conventional models challenging.
This study applies an advanced modeling framework, Discrete Choice Model with Segmentation and Embedding via AI-Learning (DCM-SEAL), to address these limitations on investigating the user preference towards the above premium transit. The resulting two latent classes identified from DCM-SEAL and their preliminary parameter estimates provide a richer and more nuanced understanding of who is more likely to use premium transit and by how much.

Description

Using the Minnesota Metropolitan Council’s 2022 Transit On-Board Survey data from the Twin Cities Area and General Transit Feed Specification (GTFS), we recovered realistic binary choice sets for each survey respondent, specifically comparing Transitway (LRT/BRT) vs. Bus-Only paths.
The DCM-SEAL framework uniquely combines several AI/ML components to analyze user preferences on this choice set data:

1. A Feedforward Neural Network (FNN) is utilized to uncover heterogeneous user groups (latent classes) and capture complex interactions among demographic variables, which are directly fed into this component.
2. Variable embeddings are incorporated to efficiently handle and interpret high-cardinality categorical variables, particularly trip contexts. These embeddings allow for a compact, interpretable, and comparable representation across different levels, significantly decreasing utility formulation complexity.
3. A traditional linear utility and multinomial logit formula is maintained for travel path-related variables like travel time and cost. This preserves the direct interpretability of core insights, such as the value of time.

By integrating these advanced methods, we identify two distinct latent classes: one group more likely to use/prefer Transitways and another less likely. Our findings confirm distinct latent user groups, revealing that specific travel contexts and demographic profiles often characterize likely users of premium transit services (e.g., BRT/LRT). Furthermore, integrating neural networks and embedding techniques significantly enhances model capability, demonstrating that advanced AI methods can provide deeper insights into class composition and travel behavior. These insights are particularly valuable for transit agencies seeking to optimize service design, allocate resources efficiently, and develop targeted policies that enhance both equity and operational effectiveness.

Statement on Why This is Noteworthy

This study introduces a scalable and adaptable framework (DCM-SEAL) applied to a real-world survey for the transit planning, balancing predictive accuracy with interpretability. The DCM-SEAL effectively leverages complex surveyed features with minimal manual effort, delivering nuanced and rich insights into travel behavior. For example, our preliminary findings pinpoint distinct latent user groups: one segment perceives transitway use as saving 5.6 minutes of in-vehicle time and finds out-of-vehicle activities less burdensome, while another perceives less time savings. Some interpretations on the preliminary model result show that transitway-preferring riders are often younger, choice riders, or identify as female, and both groups tend to favor transitways for summer and weekend trips. These precise, data-driven insights are invaluable for transit agencies to optimize service design, allocate resources efficiently, and develop targeted policies that enhance both equity and operational effectiveness, thereby shaping the future of public transportation.

The Colorado General Assembly directed the Colorado Department of Transportation (CDOT) to study alternatives for the Mountain Passenger Rail service in a corridor stretching 200 miles west from Denver to the city of Craig.  Major attractions along the corridor include two of the largest ski resorts in the state, at Winter Park and Steamboat Springs.  While ski-related travel is concentrated in the winter, there are year-round recreational activities in the corridor.  Most of the overall demand is recreational, but there is a segment of local travel that would be served by the proposed rail service, including work commuters—many of them traveling to the resort areas.  Existing passenger service by Amtrak, including a winter seasonal service to the Winter Park resort, is limited to the eastern portion of the corridor.  Limited bus service is also available.

A ridership forecasting approach was developed for three rail service options between Denver and Craig, with differing service frequencies for various portions of the route.  A challenge was the availability of data on travel demand in the corridor.  Ridership estimates were available for the existing train and bus services, but the majority of trips in the corridor are made by auto.  Statewide travel survey sample sizes were insufficient to estimate demand.  Therefore, “big data,” specifically LOCUS location-based services data obtained by CDOT, was used.  This data not only has a large sample size at a relatively fine spatial level of detail (the zones in the Colorado Statewide Travel Model), but it also is segmented by season.

The forecasting approach estimates total travel demand—segmented by recreational and local travel, season, and day of week—by origin-destination in the corridor and applies an incremental mode choice model to estimate travel by rail, auto, and bus.  Estimates of rail ridership by route segment were produced for a winter Friday, with factors based on observed auto and rail travel, to expand to other days of the week, the winter season, and the entire year.  Travel time by mode, access time to/from stations, fares, and service frequency are considered.  To deal with uncertainty (e.g., auto travel time variations due to congestion and weather), a range of forecasts was produced for each scenario. The forecasting approach was implemented quickly to meet tight deadlines for required ridership estimates.  The forecasts are being used as planning continues for the Mountain Rail service.

For this presentation, we will show how big data was used to develop an understanding of potential station catchment areas and in turn demand along the project corridor. We will demonstrate how a baseline mode share was established using available data, present on the modeling approach, and finally show forecasts developed for three scenarios.

Exploring the benefits, data needs, and applications of regional STOPS models in transit planning.