TL;DR
This paper introduces a novel Bayesian model to estimate time-varying origin-destination (OD) matrices for transit services using only boarding and alighting counts. This approach offers a more accurate and detailed understanding of passenger demand at the individual bus journey level, crucial for optimizing transit operations and planning.
What Problem Does This Paper Address?
Transit agencies rely on Origin-Destination (OD) matrices to understand passenger flow, which is vital for designing routes, scheduling services, and allocating resources. Traditional methods for collecting OD data, like onboard surveys, are time-consuming and can be biased. While automated fare collection (AFC) systems offer more data, many only record boarding (tap-in) and not alighting (tap-out), making direct OD inference impossible. Existing methods that infer alighting stops from tap-in data are often inaccurate, especially for infrequent users. Automated Passenger Counting (APC) systems provide boarding and alighting counts, but these are aggregate numbers and don't directly reveal the OD matrix. Furthermore, transit demand is dynamic, varying throughout the day and even between consecutive bus journeys due to factors like bus bunching. Existing dynamic OD estimation methods are often designed for road traffic and don't account for the discrete, over-dispersed, and zero-inflated nature of transit passenger counts. This paper addresses the critical need for a method that can accurately estimate time-varying OD matrices at the bus-journey level using readily available APC data, while accounting for the unique characteristics of transit demand.
Methods
The authors propose a temporal Bayesian model that treats each row of the OD matrix (passengers boarding at a specific stop and alighting at subsequent stops) as a realization from a multinomial distribution. The key innovation lies in how they model the time-varying alighting probabilities. They introduce a latent variable matrix and factorize it into a "mapping factor matrix" and a "temporal factor matrix." This matrix factorization significantly reduces the number of parameters. To ensure temporal smoothness, they impose Gaussian process priors on the columns of the temporal factor matrix.
For model inference, they employ a two-stage Markov Chain Monte Carlo (MCMC) algorithm:
- Conditional Sampling of OD Vectors: They use a Metropolis-Hastings algorithm to sample latent OD vectors, leveraging a proposal distribution that efficiently bypasses the need to enumerate all possible OD matrices consistent with observed counts.
- Conditional Sampling of Model Parameters: They sample the model parameters (temperature, mapping factor matrix, and temporal factor matrix) using Gibbs sampling and Elliptical Slice Sampling (ESS) for the Gaussian processes. ESS is particularly useful for efficiently sampling latent Gaussian processes in a non-conjugate setting.
Datasets
The authors evaluated their model using high-quality Automated Passenger Counting (APC) data from three distinct bus routes in an anonymous city:
- A short route with 22 stops.
- A medium route with 40 stops.
- A long route with 72 stops.
This data allowed them to reconstruct "true" OD matrices for each bus journey, serving as ground truth for evaluation.
Key Insights and Ideas
- Time-varying demand is crucial: The model's performance significantly improves when temporal dynamics are considered, highlighting that transit demand is not static.
- Alighting probabilities are smoother than demand: While OD demand fluctuates significantly even between consecutive bus journeys, the underlying alighting probabilities at stops change more smoothly over time. This observation justifies modeling alighting probabilities as time-varying and forms a core assumption of the model.
- Matrix factorization for efficiency: Decomposing the latent variable matrix into mapping and temporal factors drastically reduces the parameter space, making the model more tractable and efficient, especially for routes with many stops.
- Bayesian inference for uncertainty: The Bayesian framework naturally provides posterior distributions for OD matrices, offering not just point estimates but also crucial uncertainty quantification. This is invaluable for decision-making under uncertainty.
- Gaussian Processes for temporal structure: Using Gaussian processes for the temporal factor matrix effectively encodes smooth temporal variations and allows for correlations between temporal patterns, which is more realistic than assuming independence.
Key Results
- The proposed temporal Bayesian model consistently outperformed the widely used Iterative Proportional Fitting (IPF) method and a static Poisson-Gamma model in terms of Root Mean Squared Error (RMSE) and Continuous Rank Probability Score (CRPS) across all three bus routes.
- For the short route, the proposed model achieved an RMSE of 0.45 (compared to 0.62 for IPF and 0.47 for Poisson-Gamma) and a CRPS of 0.119 (compared to 0.136 for Poisson-Gamma).
- The model demonstrated robust estimation capabilities, with estimated OD flows closely aligning with true OD flows, and provided good uncertainty quantification, with true values falling within the estimated 95% credible intervals.
- The computational feasibility was demonstrated, with run times for 10,000 iterations being manageable even for longer routes.
Limitations and Future Work
- Data quality: The model assumes that the APC data is error-free. Inconsistencies or errors in boarding/alighting counts can lead to inaccurate estimations. The authors suggest pre-processing APC data to correct errors.
- Irregular events: The model is designed for typical operating conditions and does not explicitly account for irregular events like special occasions, seasonal changes, or major traffic accidents, which can significantly impact passenger behavior. Future work could explore integrating such events.
- Scalability for networks: Applying the model to an entire bus network with numerous routes could pose significant computational challenges. The authors suggest parallel computing as a potential solution.
- Sampling efficiency: While the Metropolis-Hastings proposal for OD vectors is effective, more advanced global sampling techniques, such as dynamic Markov basis sampling, could potentially improve efficiency and reduce convergence time.
- Kernel function choice: The current implementation uses a simple squared-exponential kernel for the Gaussian process. Exploring other kernel functions could further enhance the model.
Summary updated