TLAST: A Time-Lag Aware Spatial-Temporal Transformer for Traffic Flow Forecasting

Traffic flow forecasting is a strongly supportive component of intelligent transportation services. While in light of the expanding road networks or city grids, there is a critical concern to enhance both the accuracy and efficiency of prediction models. Despite the remarkable improvements in prediction accuracy, existing research continues to face three limitations in practical engineering scenarios. Firstly, current research often overlooks the time delay characteristics when capturing spatial relationships between global nodes. Secondly, most approaches have a quadratic computational complexity with respect to the number of nodes, resulting in significant training overhead and poor scalability. Furthermore, studies that do consider dynamic spatial relationships typically require complex model structures, resulting in higher computational costs. To address these issues, we propose a Time-Lag Aware Spatial-temporal Transformer (TLAST), a lightweight yet effective traffic flow forecasting model. TLAST introduces a cross-time strategy into the embedding stage and the attention extraction to capture the time-lag aware spatial-temporal features. Furthermore, we propose a Spatial Proxy Attention (SPA) module. It utilizes proxy representations to efficiently capture time-varying spatial dependencies with linear complexity, significantly reducing computational overhead. Extensive experiments on seven real-world traffic datasets demonstrate that TLAST consistently outperforms state-of-the-art baselines, achieving up to 7.84% improvement in prediction accuracy (MAE) while reducing memory usage and time cost by 85.21% and 75.14%, respectively. Results from the empirical analysis not only demonstrate the model’s efficiency and scalability but also highlight its practical usability in real-world traffic forecasting scenarios.