Bandwidth Prediction in 5G Mobile Networks Using Informer

Abstract

Fifth-generation (5G) mobile networks aspire to deliver exceptionally high data rates with ultra-reliable and low-latency connectivity. With the growing popularity of mobile Internet and the increased bandwidth requirements of mobile applications, user Quality of Experience (QoE) is becoming increasingly critical. 5G networks demand predicting the real-time bandwidth of a channel to satisfy the QoE for bandwidth-savvy applications such as video streaming/conferencing, vir-tual/augmented/mixed reality, and autonomous driving. If future bandwidth can be forecast in advance, the bandwidthhungry applications may utilize the estimates to adapt their data transmission rates and dramatically enhance user QoE. By analyzing a publicly available 5G dataset comprised of the channel, context, and cell-related metrics with throughput information, existing work has used Long Short Term Memory (LSTM) based mechanisms to predict future bandwidth. We applied the Transformer-based model, namely ‘Informer,’ to the 5G dataset and found significant improvement of about 95% error decrease for bandwidth prediction. In addition, we combined some new feature analysis approaches (LASSO and Random Forest with new hyper-parameters) in addition to the the existing Random Forest with Informer to find out the most accurate prediction approach.