Time-Series Forecasting Models for 5G Mobile Networks: A Comparative Study in a Cloud Implementation

Service requirements and the increased complexity of Fifth Generation (5G) applications and use cases, along with the transition towards virtualization and cloudification, are generating a strong interest in network traffic analysis. Network management and orchestration can be used to proactively tackle complex and data-driven environments to improve the performance across the entire network and meet the stringent Quality of Service (QoS) requirements for the different use cases. Time-series forecasting is one of the most important proactive approaches in communication systems. Thus, in this work, the performance of different prediction models - statistical models, Machine Learning (ML) and Deep Learning (DL) models - is evaluated using a virtualized 5G mobile network implemented in a cloud environment. The data used to train and validate the models is gathered while an operational User Equipment (UE) is connected to the core network and performing various activities. Performance evaluation of the implemented models is conducted using mathematical metrics and a graphical comparison. The results indicate that DL models outperform the statistical and ML models.

I. Introduction

Softwareization and cloudification give the Fifth Generation (5G) mobile network, among others, flexibility, scalability, and reliability in the overall network performance [1]. Utilizing concepts such as Network Function Virtualization (NFV) and Software-Defined Networks (SDNs), is essential to implement fundamental technologies for both current and next generation mobile networks, such as network slicing [2], [3]. Together with cloudification, they form a solid foundation for Next Generation Mobile Networks (NGMNs), giving networks the ability to differentiate the level of service for each use cases. The particular strength lies in the ability to adapt the system differently depending on the application, e.g. in terms of data rate, latency or security constraints [4].