Next-Generation Travel Demand Modelling

Traditional travel demand models heavily rely on travel surveys, which are costly, laborious, and suffer from small-sample and infrequent issues. My Ph.D. dissertation focuses on building a next-generation travel demand model which is fully big-data-driven. This includes 1) a normative pipeline for parsing multi-source travel data to derive trip rosters and multi-modal OD matrices, 2) a set of spatiotemporal neural networks for forecasting future travel demand, and 3) a set of traffic simulation tools for simulating traffic flow at both individual and aggregated level.

Big-data-driven Travel Demand Estimation 📑📑📑🚩🚩🚩

Fully relying on mobile device location data and using travel survey as validation, we build a pipeline to extract individual trip rosters. The process includes home&work identification, trip identification, mode imputation, population weighting, and result validation. These trip rosters are then aggregated at different spatiotemporal units to construct multi-modal OD matrices. The whole pipeline is employed on AWS EMR, a cloud-computing server, to timely quantify large-scale human travel patterns.

Using nationwide census block group-level trip flow derived from mobile devices as the proxy of travel demand, We examine its relations with socioeconomics, demographics, and land use. Over 6*8 machine learning models and interpretation techniques are compared to determine the best model and justify interpretation robustness. Pronounced nonlinearities, threshold effects, and interaction effects are observed.

Deep-learning-based Travel Demand Forecasting 📈📈📈📉📉📉

OD flow forecasting is challenging due to complex spatiotemporal dependencies and heterogeneous external effects. We propose a Multi-graph Multi-head Adaptive Temporal Graph Convolutional Network (Multi-ATGCN), a general deep learning framework for citywide multi-step OD flow forecasting. Experiments on two real-world tasks demonstrate its steady performance improvement over state-of-the-art baselines.

Individual mobility forecasting is more challenging because of high randomness of individual travels, multi-structure forecasting tasks, and imbalanced distributions of places and activities. We propose a hierarchical activity-based framework for simultaneously predicting the activity, time, and location of the next trip for each device. Meanwhile, loss functions in the semantic segmentation domain are introduced to address the imbalanced classification issue. The whole framework is applied on a county-level dataset covering over 18,000 residents and shows acceptable prediction accuracy.

Multi-dimensional Traffic Simulation 🚦🚦🚦🚗🚗🚗

Integrating forecasted travel demand with traffic simulators completes the last puzzle of a travel demand model. At an aggregated level, we feed the predicted OD matrices into a dynamic traffic assignment tool, DTALite, to generate link traffic speed and volume. Simulated results are compared with field observations collected from roadside sensors to validate the whole data-driven travel demand model.

Mobile device location data provide detailed individual travel information, which is more compatible with behavior-oriented agent-based simulators. My ongoing research focuses on meshing the forecasted trip itinerary with micro (Vissim, SUMO) or agent-based simulators (MATSim) with high computational efficiency to achieve fine-grained citywide simulation.