Abstract

Predicting short-term passenger flow in urban bus networks is a crucial task for optimizing transit operations, reducing congestion, and enhancing transit commuter experience. This thesis introduces several innovative deep learning models aimed at addressing the unique challenges of bus networks, including temporal dynamics, spatial variability, and the impact of real-time traffic conditions. To model the complex relationships in transit networks, we leverage Graph Neural Networks (GNNs), which are particularly well-suited for capturing the non-Euclidean structure of bus networks.

In the first model, a Bus Network Graph Convolutional Long Short-Term Memory (BNG-ConvLSTM) neural network is developed to forecast short-term passenger flow. This model outperforms traditional deep learning models in scalability and robustness, as validated by real-world data from the Laval bus network. Extending this, we introduce the Traffic-Aware Multistep Graph Neural Network (TMS-GNN), which integrates traffic conditions and addresses the issue of exposure bias in multistep forecasting by employing Scheduled Sampling. This model significantly improves accuracy in multistep prediction and better adapts to the realities of urban traffic patterns.

To further capture the dynamic nature of public transportation, we propose Spatial-Temporal Attention Masked Graph Encoder-Decoder (STAM-GED), which integrates real-time bus schedules to model both node and edge changes in a network. This approach provides a more accurate representation of passenger flow, reflecting the real-time operational state of bus stops.

Finally, we explore transfer learning as a solution to the challenge of data scarcity, a common issue in many cities. We develop a transfer learning framework for GNNs, which uses a novel reinforcement learning optimization-based graph partitioning method to adapt models trained on data-rich networks to cities with limited data. This framework enables the transfer of knowledge across diverse urban environments, ensuring scalability and generalizability without sacrificing predictive accuracy.

Through comprehensive experiments on real-world data from multiple cities, including Ames-USA and Laval-Canada, our models demonstrate significant improvements in passenger flow prediction over existing methods. These contributions offer solutions for enhancing the reliability and efficiency of public transportation systems, paving the way for smarter, more sustainable urban mobility.