Planning-oriented travel demand forecasting for evolving transportation systems using deep neural networks

Abstract

Forecasting travel demand for an evolving transportation system is crucial for urban and transport planners to evaluate and refine the network's design, thus improving urban mobility. Traditional prediction methods face precision challenges due to oversimplified assumptions, while recent deep learning approaches, focusing on short-term predictions from historical data, assume static system structures and are unsuitable for new infrastructure design. To address these limitations, this thesis focuses on enhancing travel demand forecasting for transportation system planning through the use of deep learning techniques.

The thesis comprises four main studies on travel demand forecasting: 1) Trip generation, predicting demand at new locations; 2) Temporal distribution, assessing demand variations over time; 3) Spatial distribution, mapping trip flows between origins and destinations; and 4) Route choice modeling, analyzing route preferences. These components are interrelated, and their aggregate findings contribute to a comprehensive understanding of both spatial and temporal mobility patterns within a proposed transportation network.

The first study introduces a Spatially Multi-Graph Attention Network (Spatial-MGAT) to solve the trip generation problem. Spatial-MGAT constructs localized graphs around each target site and applies attention mechanisms to learn site correlations. Using multi-year bike sharing data from New York City, Spatial-MGAT demonstrates superior performance and interpretability, particularly in understanding the effects of the built environment on demand and spatial interactions between sites.

The second study presents a multi-task memory-augmented graph neural network (MMGT) for the temporal trip distribution problem. MMGT utilizes urban contexts and historical demand features of nearby existing sites to uncover temporal demand patterns. Its innovative design includes a memory network for feature extraction and a multi-task demand predictor, distinguishing the impacts of urban density and functionality on temporal demand. Validation with New York City bike sharing data shows its superior performance and interpretability.

In the third study, a zero-inflated negative binomial graph neural network (ZINB-GNN) addresses the spatial trip distribution challenge. This model constructs localized graphs for origin-destination (O-D) pairs and designs an O-D-based graph learning method for spatial feature encoding. A ZINB prediction layer is proposed to manage the sparse and uncertain nature of O-D matrices. The ZINB-GNN's performance and interpretability are confirmed using the same New York City dataset.

The fourth study introduces a deep inverse reinforcement learning framework for route choice modeling (RCM-AIRL). Compared to traditional discrete choice models, the RCM-AIRL effectively incorporates diverse features and uses an adversarial IRL model for efficient estimation. The model's superiority is validated with Shanghai taxi GPS data, demonstrating better prediction performance for unseen destinations with reasonable interpretability.

Collectively, this thesis enhances travel demand forecasting in transportation system planning by introducing deep neural network models that adeptly navigate four interrelated challenges. These models not only excel in predicting outcomes for new transportation locations but also offer deep insights into how urban characteristics influence mobility patterns. This facilitates the evaluation and improvement of network designs to meet diverse demand scenarios. While initially applied to bike sharing and taxi trip data, the methodologies developed are general and can be easily adapted across different transport modes and networks.