Decoupled spatiotemporal graph convolution with probabilistic sparse self-attention for traffic flow forecasting

Accurate traffic flow forecasting is critical for developing intelligent transportation systems. While Graph Neural Networks have achieved notable success in short-term forecasting, their performance significantly degrades over long-term horizons due to the complex entanglement of heterogeneous temporal patterns and the dynamic, non-local nature of spatial dependencies in road networks. To tackle these challenges, this study proposes a novel framework, termed Decoupled Spatiotemporal Graph Convolution with Probabilistic Sparse Self-Attention (DSGC-PSSA). In the temporal domain, a decoupling module is utilized to disentangle and highlight heterogeneous traffic flow patterns through frequency-domain analysis. Subsequently, a time-frequency fusion mechanism extracts complementary information from the decomposed features, thereby enhancing the model’s capability to capture complex temporal dynamics, particularly in long-term forecasting. To model the inherently dynamic nature of road networks, the spatial module of DSGC-PSSA adopts a hybrid graph learning strategy that integrates adaptive graph convolution and dynamic graph attention mechanisms. This design enables effective modelling of intricate spatial dependencies and supports flexible representations across both global and local scales. Additionally, the spatiotemporal convolution layer captures complex temporal dependencies through multi-scale modelling. In parallel, the probabilistic sparse self-attention mechanism facilitates dynamic integration of spatiotemporal features, thereby enabling more efficient long-term traffic flow forecasting. Extensive experiments conducted on five real-world traffic datasets demonstrate that DSGC-PSSA significantly outperforms existing state-of-the-art models, particularly in long-term forecasting.