Research and Implementation of Network Traffic Prediction Algorithm Based on Informer

Abstract

As next-generation mobile networks like 5G and 6G proliferate and internet connectivity becomes more accessible, network traffic prediction plays a crucial role in network management. Excellent traffic prediction results can enhance network management efficiency and improve network bandwidth utilization, making efficient traffic prediction essential. Traditional models primarily focus on the temporal characteristics of network traffic while neglecting its spatial characteristics, resulting in poor accuracy and stability in traffic prediction. However, the development of deep learning has provided new solutions for predicting network traffic. This paper conducts an in-depth study of the Transformer time series algorithm, identifies its shortcomings in network traffic data prediction, and introduces the ProbSparse self-attention mechanism into the model. Based on this, network traffic prediction algorithms can be implemented using Informer. This study utilizes network traffic usage data from the University of Massachusetts Amherst and its Amherst campus. Four sets of experiments were conducted under different prediction step conditions to compare the prediction results with real traffic data across four metrics: MAP, RMSE, MAPE, and R², thereby evaluating whether the Informer model can maintain prediction accuracy over extended periods.