The Traffic Prediction Method Based on Wavelet Decomposition and Structural-Enhanced Graph Attention
Kang Xu, Bin Pan, Xuan Zhang, Mingxin Zhang, Jingxian Yu, Zhizhu Lu, Xiao Zeng, Qingqing Jia · IEEE Sensors Journal · 2025
With the widespread deployment of sensors in the transportation field, traffic flow data has become easily accessible, providing a foundation for high-precision traffic time series prediction. However, traffic data exhibits significant non-stationarity and dynamic spatio-temporal correlations, which pose numerous challenges for modeling. Existing methods have the following shortcomings: (1) directly modeling the entire traffic sequence, making it difficult to effectively address distribution shifts caused by non-stationarity; (2) traditional time attention mechanisms rely on pointwise matching, lacking the ability to model trends; (3) graph attention networks (GATs) only utilize topological information for spatial modeling, making it difficult to capture semantic relationships between roads. To address these issues, this paper proposes a structurally enhanced decoupled spatio-temporal modeling framework. The method first applies wavelet decomposition to decompose the original traffic sequence into approximation signals (stable trend sequence) and detail signals (sudden event sequence), achieving explicit decoupling of long-term trends and short-term disturbances. In trend sequence modeling, a trend-aware temporal attention module (TTAM) is constructed to capture long-term dynamic features; in event sequence modeling, a causal convolution module is introduced to efficiently extract short-term drastic changes. To enhance spatial modeling capability, this paper proposes a structurally enhanced graph attention network (SGAT), which combines geographic adjacency relationships and time behavior similarity calculated by dynamic time warping (DTW) to construct Dual-perspective spatio-temporal graphs, applied uniformly to spatial modeling of both types of sequences. Additionally, a dual-supervision decoder is designed to jointly optimize the trend prediction and traffic prediction losses, further improving the model’s accuracy and generalization ability. Experiments conducted on six real-world traffic datasets show that the proposed method outperforms the second-best models, reducing MAE, RMSE, and MAPE by an average of 6.50%, 3.75%, and 3.79% on three traffic flow prediction datasets; and reducing by 2.23%, 1.81%, and 1.48% on three traffic speed prediction datasets. Ablation experiments further validate the important role of the trend modeling module and the graph structure enhancement mechanism in improving prediction performance. The code will be released at https://github.com/CR818-web/TFPWD.