Network traffic prediction is essential for intelligent resource management in modern transportation systems, but existing methods struggle to simultaneously capture multi-scale temporal patterns, long-range dependencies, and periodic behaviors while maintaining computational efficiency. This paper presents WMF-Traffic (Wavelet-Mamba-Fourier Traffic prediction framework), a novel traffic forecasting approach that synergistically integrates wavelet decomposition, selective state space modeling, and frequency domain processing. The framework introduces four key components: Multi-scale Wavelet Decomposition for hierarchical temporal pattern extraction, Wavelet Traffic Convolution with scale-adaptive mechanisms, Traffic-aware Mamba for efficient long-range dependency modeling, and Fourier Pattern Adjustment for periodic pattern enhancement. WMF-Traffic employs a comprehensive training objective that balances reconstruction accuracy, temporal consistency, and spectral coherence. Extensive experiments on four real-world traffic datasets (METR-LA, PEMS-BAY, PEMS04, PEMS08) demonstrate consistent improvements over state-of-the-art methods, achieving 1.0-1.3% gains in MAE, 0.6-1.1% in RMSE, and 0.2-1.0% in MAPE across different prediction horizons. Ablation studies reveal that Traffic-aware Mamba provides the largest individual contribution (10.2% MAE reduction), while the complete framework achieves up to 27.1% improvement over baseline approaches. The proposed uncertainty-based fusion mechanism further enhances robustness with 3.2-4.1% additional improvements.
Jiale Song (Fri,) studied this question.