关键词: traffic flow forecasting, frequency-domain graph learning, Mamba, graph convolutional network(GCN)

Abstract: With the rapid advancement of intelligent transportation systems (ITS), urban traffic prediction faces significant challenges in effectively modeling complex spatio-temporal dynamics while maintaining computational efficiency. Existing approaches are often limited by static graph structures and the high computational cost of self-attention mechanisms when processing long sequences. To overcome these limitations, this paper proposes a novel framework, termed spatio-temporal frequency-domain mamba network (STFD-MambaNet). Specifically, the framework integrates a frequency-domain dynamic graph learner to capture evolving traffic topologies and employs the Mamba structured state space model to efficiently extract long-range temporal dependencies with linear complexity. Furthermore, a hierarchical spatial modeling module is developed to characterize multi-scale spatial correlations. Experiments conducted on four real-world datasets demonstrate that STFD-MambaNet consistently outperforms state-of-the-art methods in both accuracy and efficiency. The results further demonstrate the effective complementarity between frequency-domain dynamic graph learning and structured state space modeling, providing a robust solution for spatio-temporal traffic forecasting.

Key words: traffic flow forecasting, frequency-domain graph learning, Mamba, graph convolutional network(GCN)