Predicting Origin–Destination Travel Flow Based on a Spatiotemporal Relational Graph Learning Approach

ABSTRACT Accurate prediction of origin–destination (OD) travel flow is essential for location‐based services such as traffic management, ride‐hailing dispatch, and emergency response. However, forecasting OD travel flow presents greater challenges than estimating inflows or outflows of individual regions, as it requires modeling complex spatiotemporal dependencies between both origin and destination regions simultaneously. Furthermore, external factors such as weather conditions and calendar dates exert heterogeneous influences on OD travel flow, yet these are rarely adequately incorporated in existing approaches. To address these issues, this paper proposes a SpatioTemporal Relational Graph Learning (STRGL) model for OD travel flow prediction. Our framework first constructs four relational graphs to model multiple relationships among travel flows, including spatial distribution, origin and destination semantics, and temporal fluctuation patterns. A spatiotemporal relational graph convolutional network is then employed to jointly model spatial and temporal dependencies across these relational graphs. Additionally, an attention‐based environmental feature‐learning module is designed to quantify the heterogeneous effects of external factors. The learned representations are finally decoded by a multilayer perceptron to predict future OD travel flows. Extensive experiments on two real‐world travel datasets show that STRGL consistently outperforms multiple state‐of‐the‐art baseline methods across most evaluation metrics. Ablation studies further validate the contribution of each component in the proposed STRGL framework.