A BIG-DATA-DRIVEN FRAMEWORK FOR SPATIOTEMPORAL TRAVEL DEMAND ESTIMATION AND PREDICTION

Abstract

Traditional travel demand models heavily rely on travel surveys, simplify future demand forecasting, and show low sensitivity in response to spatiotemporal dynamics. This study proposes a deep-learning-driven framework based on mobile device location data (MDLD) for estimating and predicting large-scale travel demand at both individual and aggregated levels. This study first introduces how raw MDLD should be parsed to distill trip rosters and estimate population flow. Based on derived information, this study reexamines relations between average population flow and its determinants such as built environment, socioeconomics, and demographics, via a set of explainable machine learning (EML) models. Different interpretation approaches are employed and compared to understand nonlinear and interactive relations learned by EML models. Next, this study proposes a Multi-graph Multi-head Adaptive Temporal Graph Convolutional Network (Multi-ATGCN), a general deep learning framework that fuses multi-view spatial structures, multi-head temporal patterns, and various external effects, for multi-step citywide population flow forecasting. Multi-ATGCN is designed to comprehensively address challenges such as complex spatial dependency, diverse temporal patterns, and heterogeneous external effects in spatiotemporal population flow forecasting. Last, at an individual level, this study proposes a Hierarchical Activity-based Framework (HABF) for simultaneously predicting the activity, departure time, and location of the origin and destination of the next trip, incorporating both internal (individual characteristics) and external (calendar, point-of-interests (POIs)) information. For each individual, HABF first predicts activities via an Interpretable Hierarchical Transformer (IHTF). IHTF can efficiently handle big data benefiting from its transformer-based design to avoid recursion. Meanwhile, loss functions used in semantic segmentation are introduced into IHTF to address imbalanced distributions of activity types. Then, a local plus global probabilistic generator is designed to generate locations based on predicted activities and historical places, allowing individuals to visit new or historically-sparse places. Analyses are performed on several real-world datasets to demonstrate the model's capability in forecasting large-scale high-resolution human mobility in a timely and credible manner. Altogether, this study provides sound evidence, practically and theoretically, of the feasibility and reliability of realizing data-driven travel demand estimation and prediction at different spatiotemporal resolutions and scales.