A New Perspective on Defining Dynamic Origin-Destination Data and Predicting it Using Deep Learning Methods

Abstract

The prediction of dynamic origin-destination (OD) data is critical for facilitating real-time traffic management across traffic networks. Despite numerous efforts to integrate the temporal and spatial characteristics of OD data to capture the nonlinearity and high dynamics of traffic flow, prior studies usually rely on link-level or region-level data for model construction. The temporal relationships among origin traffic flow, destination traffic flow, and OD flow remain insufficiently understood. To address this gap, we propose a novel definition of dynamic OD data using real-world OD datasets. Our framework can incorporate different temporal distributions for each OD pair. Additionally, the framework ensures flow conservation from either the origin or the destination perspective. The performance of the proposed framework is validated through numerical studies using real-world electronic toll collection (ETC) gantry data. A multi-task long short-term memory (LSTM) model predicts OD flows, and both the predictions and the resulting destination traffic distributions are statistically indistinguishable from the observed values. Furthermore, this approach enables the prediction of arrival volumes before trip completion, offering valuable insights for real-time traffic management.