Dual-heterogeneity temporal-spatial network for traffic flow prediction

Abstract

Urban traffic flow is a typical non-stationary system with pronounced dual-heterogeneity. Here, dual-heterogeneity refers to both temporal and spatial heterogeneity within traffic systems. Specifically, temporal heterogeneity denotes the coexistence of stable periodic patterns and abrupt non-stationary dynamics, while spatial heterogeneity denotes the coexistence of topological reachability and functional semantic similarity. Existing spatiotemporal prediction models typically rely on a single parameter space to represent diverse patterns. This strategy limits their ability to represent dual-heterogeneity in complex traffic scenarios. To address this limitation, a Dual-Heterogeneity Temporal-Spatial Network (DHTS-Net) is proposed to explicitly decouple and dynamically model heterogeneous patterns across temporal and spatial dimensions. First, a Unified Spatiotemporal Embedding aligns multi-source observations within a shared representation space. In the temporal dimension, Temporal Dual-Path Attention and a Mixture of Experts module model structural temporal dependencies and non-stationary dynamics separately. These components are integrated through Cross-Temporal Interaction to enable complementary fusion. In the spatial dimension, Spatial Dual-Path Diffusion-Attention captures topological diffusion and semantic similarity in parallel. Cross-Spatial Interaction is then applied to form a unified spatial representation. Finally, a Multi-Step Prediction Head supports stable long-horizon prediction. Experimental results demonstrate that DHTS-Net achieves competitive prediction performance across multiple real-world traffic flow datasets. The model shows strong dynamic adaptability and robust prediction capability.