Multi-scale temporal feature-enhanced federated learning framework for network traffic prediction

Abstract

With the rapid development of mobile internet, network traffic has shown exponential growth. Accurate traffic prediction has become a key technology for ensuring stable network performance and optimized resource allocation. However, existing methods fail to fully integrate periodic features and neglect the persistent impact of non-periodic temporal features such as holidays on network traffic. This oversight makes it challenging for models to effectively capture both periodic patterns and sudden fluctuations in traffic. To address this issue, this paper introduces a Multi-Scale Temporal Feature Enhanced Federated Learning Framework for Network Traffic Prediction (MTFE-FL). The framework proposes a Holiday Impact Factor to comprehensively measure the persistent impact of holiday characteristics on network traffic data. High-quality predictive models are trained collaboratively in multiple edge clients, each using an iTransformer model to process time series data. By encapsulating the entire time series into variable tokens, the iTransformer provides a global perspective, enabling the effective identification of complex patterns and dependencies evolving over time. In addition, multivariate attention mechanisms are utilized to deep explore the relationships between network traffic data and temporal information. To further enhance the generalization ability of the global model and mitigate the “client drift” caused by client heterogeneity, Stochastic Controlled Averaging is introduced to correct the gradients of the local models at each edge client. The aggregated corrected models then generate the global model. Experimental results demonstrate that the proposed framework achieves superior performance on two real-world network traffic datasets, significantly improving the accuracy of network traffic predictions.