Abstract Title

Creating an ML-based Open-source Tool for Estimating Ridership Based on Transit Network and Operation Data

Topic

Machine learning/AI; Transit; Travel demand model development; Ridership prediction

Abstract Background

Machine Learning (ML) methods have gained attention in predicting transit ridership because of their ability to uncover complex feature relationships. Yet, ML approaches come with two challenges. First, they require substantial data, making it difficult for transit agencies to develop their ML models. Second, they have poor generalization ability, as models often fail to account for variations in features across different agencies. As a consequence, treating diverse agencies as homogeneous can lead to misleading patterns and inaccurate predictions. To address these two issues, this research introduces an open-source two-step ML tool designed to enhance the predictive performance of ML models by accounting for variations in agencies’ network designs.

Description of Abstract

This study develops a new open-source tool for enhancing transit ridership predictions based on key network features (such as stop spacing, route directness, and service area) and operational data (such as vehicles operating on maximum service, speed, frequency, fare, on-time performance.). Exploiting 2022-2023 data from 64 U.S. transit agencies, we proposed a two-step ML approach that not only can lead to more accurate ridership predictions, but also can help transit agencies identify the potential network/service changes that they can implement to reach a desirable ridership level. The tool first clusters agencies using the Fuzzy c-Means algorithm based on their network. Then, it estimates ridership for each cluster based on their operational and sociodemographic variables, using three ML models: Random Forest, Gradient Boosting Regression, and Support Vector Regression. The performance of each model is evaluated against a baseline ridership model without clustering, using metrics such as Mean Squared Error, Mean Absolute Error, R-squared, and Mean Absolute Percentage Error. Our results indicate that clustering improves prediction accuracy and robustness. Also, the analysis of the important variables shows meaningful and interpretable results, with key features differing across clusters. These findings confirm that transit networks with distinct characteristics require tailored ridership models, highlighting the value of clustering in enhancing both reliability and interpretability. The developed tool (including the dataset, clustering algorithm, and prediction ML models) is hosted on GitHub and publicly available. This tool enables any U.S. bus agency not only to predict their ridership by simply inputting their network and operational data but also to explore the impacts of potential/planned network or service adjustments on their ridership.

Statement on Why Abstract is Noteworthy

This work provides an open-source tool and dataset, making ridership prediction models accessible to transit agencies of all sizes and characteristics. In particular, agencies, which often lack the resources to develop their models, can use this tool to predict their ridership and identify network configurations or operational strategies that maximize ridership. Additionally, we proposed a two-step approach that improved the accuracy and robustness of ML models.

Project is Complete

UrbanSim is family of computer systems designed for simulation of changes in urban development. Classical implementations, beginning with the Eugene-Springfield, OR Metropolitan Area (Waddell, 2002) used an agent-based microsimulation (ABM) approach. While ABMs have many strengths, they can also be overelaborate and difficult to calibrate. Partially calibrated outputs from one module can feedforward paradoxical inputs for calibration of other models, and the system can remain biased after many dozens or hundreds of human interventions by trial and error. Therefore, the creators of UrbanSim have, in recent years, recast the land use model as a differentiable program. In the differentiable version, all discrete choice model parameters can be jointly and longitudinally calibrated by machine learning and automatic differentiation. By selecting a suitable multicriteria loss function, and tracing local rates of change in the loss function with respect to the various model parameters by repeated application of the chain rule, an optimal parametrization can be found for the entire system simultaneously by gradient descent. (Waddell, 2024). This presentation will focus on early results from practical application of a cloud-based implementation of this new differentiable version of UrbanSim by and with Oregon Metro, the metropolitan planning organization for the greater Portland area. Major milestones to be completed by Fall 2025 will include draft results from the new model’s allocation of Metro’s 2024 regional population and employment forecasts to small area geographies within our planning area.

Waddell. P. (2002). “UrbanSim: Modeling Urban Development for Land Use, Transportation, and Environmental Planning”, Journal of the American Planning Association 68(3), pp. 297-314.

Waddell, P. (2024). “Using AI Algorithms to Train Simulation Models Longitudinally”, Presented to the Oregon Model Users Group, Oct. 10: https://www.youtube.com/watch?v=j1jidHVshtk (Accessed Feb. 13, 2025)

The evolution of transportation planning has long been constrained by traditional static models, lacking the adaptability to fully integrate modern AI-driven computational frameworks. In this presentation, we propose a paradigm shift by leveraging computational graph-based methods, inspired by TensorFlow and deep learning architectures, to rethink foundational transportation models such as land use modeling, activity-based models (ABM), and dynamic traffic assignment (DTA).

We demonstrate how computational graphs serve as a unifying framework, allowing not only for efficient network design and service network optimization but also for addressing transient efficiency and the complex interplay of transportation dynamics, equity, and resilience in response to disruptive events. By embedding discrete choice models into computational graphs and utilizing automatic differentiation (symbolic gradient methods), we illustrate how AI-driven scenario sensitivity analysis can enhance decision-making in dynamic travel models.

Our approach integrates dynamic programming techniques, scenario-based evaluation, and simulation-guided optimization into large-scale transportation planning. By employing forward and backward propagation, we show how these techniques can calibrate travel demand models and evaluate network interventions in an automated, data-driven manner. Through open-source tools and GPU-accelerated computing, we explore how planners can move beyond CNN-based flow predictions to develop multi-resolution, multi-modal demand-supply models.

In this talk, we will present:

• How to construct computational graphs for transportation networks, incorporating DTA, ABM, and real-time traffic state estimation.

• How AI tools, such as automatic differentiation and deep learning frameworks, can optimize scenario evaluation and policy interventions.

• The role of physical-informed and simulation-guided AI models to support real-time decision-making.

• The integration of open-source, scalable computation using TensorFlow and C++ engines to enhance efficiency.

• Applications using global and regional datasets, demonstrating large-scale AI-driven multi-resolution planning.

This workshop aims to empower transportation planners and researchers with the necessary AI knowledge to rethink transportation models beyond conventional paradigms. By integrating TensorFlow-driven computational graphs, we bridge the gap between classical transportation science and modern AI-based methodologies, unlocking new frontiers in predictive modeling, operational optimization, and policy design.

Test Cases & Real-World Applications

To illustrate the scalability and real-world applicability of this AI-driven framework, we present two major case studies:

Phoenix Metropolitan Area Planning Model – A large-scale dynamic traffic assignment model using computational graphs to integrate network equilibrium, demand elasticity, and real-time traffic state estimation. 

For individual cities within this region, such as Gilbert, we implement a multi-resolution model that leverages TensorFlow-based auto-differentiation and simulation-based optimization to enhance policy scenario evaluation, effectively capturing urban-rural interactions and transit planning challenges.