Towards integrated and fine-grained traffic forecasting: A Spatio-Temporal Heterogeneous Graph Transformer approach

Abstract

Fine-grained traffic forecasting is crucial for the management of urban transportation systems. Road segments and intersection turns, as vital elements of road networks, exhibit heterogeneous spatial structures, yet their traffic states are interconnected due to spatial proximity. The heterogeneity and interrelationships arising from different road network elements pose major challenges to accurate traffic forecasting. However, existing forecasting studies focus solely on bidirectional road segments, disregarding the relationships between roads and turns. To achieve integrated traffic forecasting that considers both road segments and intersection turns, we propose a novel Spatio-Temporal Heterogeneous Graph Transformer (STHGFormer). For road network representation, we innovatively define a Heterogeneous Road network Graph (HRG), which provides a comprehensive depiction of the complete traffic network and emphasizes its inherent heterogeneity. Then, we propose a Heterogeneous Spatial Embedding (HSE) module to encode road network information, including heterogeneous attributes and interactions in the HRG. Based on the spatial information encoded by HSE, a unified SpaFormer, serving as the spatial module of STHGFormer, captures the interdependencies between roads and turns across the entire traffic network. To mitigate the impact of high temporal fluctuation, we embed the Adaptive Soft Threshold (AST) module into TempFormer, which dynamically adjusts the threshold to enhance the analysis capability of complex temporal correlations. Experiments conducted on a real-world dataset from Wuhan, China, demonstrate that STHGFormer outperforms state-of-the-art methods, achieving a 6.1 % improvement in road forecasting and an 8.5 % improvement in turn forecasting.