<p>Traffic flow prediction is a fundamental aspect of intelligent traffic scheduling. Existing methods often struggle to effectively capture deep spatiotemporal dependencies across time scales, particularly in long-term predictions during dynamic periods such as morning and evening peak hours, limiting their practical application. This paper proposes a multi-period spatiotemporal fusion model (MPST-Mamba) based on the Mamba architecture, aimed at improving the fusion of spatiotemporal features for long-term traffic flow prediction, especially for enhancing accuracy during complex peak periods. First, a parallel encoding structure for multi-periods (week, day, hour) is constructed to strengthen the modeling of cross-period traffic patterns. Second, a graph fusion selective state-space module (Graph-SSSM) is designed based on the selective state-space model, dynamically integrating graph structure information to enhance the model’s ability to capture the spatiotemporal dynamics of traffic flow. Finally, a decoding mechanism that considers node associations is developed to improve the non-linear mapping capability in long-term prediction. Experimental results on multiple public datasets, including PeMS03, PeMS04, PeMS08, and METR-LA, demonstrate that the proposed model significantly outperforms existing baseline methods in long-term prediction tasks, especially during high-dynamic traffic periods like morning and evening peaks, providing reliable technical support for urban dynamic traffic management and scheduling optimization.</p>