Network traffic prediction via four-fold seasonal-trend decoupling

Accurate network traffic prediction is crucial for cybersecurity, network management, and resource optimization. However, it faces two major challenges: the non-stationary nature of traffic data with complex patterns such as bursts and periodicity, and the intricate dynamic spatiotemporal dependencies that are difficult to capture using conventional methods. Traditional graph neural networks, while effective in modeling static spatial relationships, struggle to represent long-term spatiotemporal dynamics. To address these issues, this paper proposes a Hierarchical Network Traffic Representation method via Component Decoupling (HTSR-CD). This approach employs a moving average to decompose traffic sequences into long-term trend and seasonal components, separating stable signals from short-term fluctuations. Fourier transform is then applied to further analyze these components in the frequency domain, extracting both globally shared and locally specific features for multi-level temporal representation. Experimental results on multiple public datasets demonstrate that integrating HTSR-CD into various time series prediction models consistently improves performance, validating its generality and effectiveness in temporal modeling.