Federated Learning Cellular Traffic Prediction Based on Multi-time Scale Information

This study focuses on the field of cellular traffic prediction and proposes a federated learning method for cellular traffic prediction based on multi-time scale information combined with network pruning. It utilizes periodic time factors and the application of the Informer model within a federated learning framework. This paper integrates temporal features such as weeks, months, and holidays to optimize the model’s ability to model time series. To address the non-independent and identically distributed problem in federated learning, a data augmentation based method is introduced to enhance the model’s adaptability to heterogeneous data. Additionally, it optimizes the model structure through gradient priority iterative pruning techniques, applying lightweight processing to the Informer model to reduce computational complexity. The experimental results verify the good performance of the proposed method in cellular traffic prediction tasks, enhancing prediction accuracy and significantly reducing the model’s computational resource requirements. This research provides robust theoretical support for optimizing resource utilization efficiency and performance in wireless communication systems and offers valuable references for resource management and traffic scheduling in practical applications.