Traffic flow forecasting method based on multiscale wavelet decomposition and enhanced time-frequency modeling

With the advancement of intelligent transportation systems, traffic flow prediction faces growing challenges in modeling spatio-temporal dependencies and multi-scale dynamics. Existing graph neural network models mostly adopt single-scale spatio-temporal structures, limiting their ability to capture multi-level traffic patterns. Moreover, traditional time–frequency analysis suffers from trade-offs between frequency resolution and temporal sensitivity, making it difficult to account for both local fluctuations and global trends. To address these issues, this paper proposes STFD-WaveNet, a traffic flow prediction model integrating multi-scale wavelet decomposition with an enhanced time–frequency modeling module. The model first applies multi-scale wavelet decomposition to separate traffic sequences into low-, mid-, and high-frequency bands, capturing dynamic variations across temporal scales. Then, an enhanced time–frequency module strengthens trend learning in low-frequency channels and fluctuation modeling in high-frequency channels, while a cross-scale attention mechanism adaptively fuses multi-frequency information. Finally, a serial spatio-temporal structure ensures consistency between spatial dependencies and temporal evolution. Experiments on multiple real-world datasets show that the proposed method outperforms existing models, achieving higher accuracy and stability, especially under complex traffic conditions and in long-term forecasting.