MSESTA: Multi-scale evolving spatial-temporal attention for long-term traffic prediction

Accurate long-term traffic prediction is crucial for the planning and scheduling of Intelligent Transportation Systems. The previous work has sufficiently investigated short-term traffic prediction within a one-hour horizon, achieving success in capturing the intrinsic spatial topology and temporal patterns of traffic data. However, when extending to longer horizons, both computational efficiency and the capacity to model long-term spatial temporal dynamics become critical challenges. In this paper, we propose a Multi-Scale Evolving Spatial Temporal Attention (MSESTA) network, which enhances the model's capacity to learn long-term spatial temporal dependencies while improving efficiency through a cross-scale evolution mechanism and explicit modeling of global spatial temporal patterns. Specifically, we first partition traffic data into multi-scale sequences of varying patch sizes to uncover their latent multi-scale representations. Then, at each scale, we design an evolving spatial temporal attention mechanism with linear complexity, consisting of three stages: summarization, aggregation, and broadcasting, where "summarization" bridges adjacent scales and the latter two stages perform global spatial temporal information aggregation and broadcasting. Moreover, we develop a global spatial information aggregation strategy to effectively integrate prior spatial knowledge. Extensive long-term prediction experiments on four real-world traffic datasets demonstrate that MSESTA achieves significant improvements in both prediction accuracy and computational efficiency. Our code will be released soon.