Spatiotemporal memory decoupled transformer for traffic flow forecasting

Traffic flow forecasting is crucial for effective urban planning and the development of intelligent transportation systems. However, accurately capturing the complex spatiotemporal dependencies in traffic flow data remains a significant challenge. Current approaches typically face two major limitations: (i) insufficient modeling of long-term historical traffic flow patterns, which hinders the capture of global trends and periodic behaviors; and (ii) inadequate consideration of spatial structures and temporal response heterogeneity, which limits the model’s representational capacity. To tackle these challenges, this paper proposes a novel forecasting framework named Spatiotemporal Memory Decoupled Transformer (STMDFormer). First, a spatiotemporal embedding module is designed to fuse topological node structures with multi-scale periodic information. Then, a spatial memory attention mechanism is introduced, incorporating a learnable historical memory bank and a biased adjacency matrix for dynamic modeling of long-term traffic flow patterns. Furthermore, a decoupled learning module is designed to separate spatial and temporal features through normalization techniques, while employing multi-head attention to better capture spatiotemporal heterogeneity. Extensive experiments on four real-world traffic datasets demonstrate that STMDFormer consistently outperforms various state-of-the-art baselines in forecasting accuracy, validating its effectiveness in dynamic spatiotemporal modeling.