Application of machine learning for energy efficiency in mobile networks

Abstract

Future generation networks (5G) will bring a new paradigm to network management, as the networks themselves will suffer evident changes that will imply new requirements in upper layers. The 5G-XHaul project, framed under the Horizon 2020 European research and innovation programme is focused on providing dynamically reconfigurable optical-wireless backhaul and fronthaul architectures with a cognitive control plane for small cells and cloud-RANs. One of the objectives contained under that premise consists in the design of new network management strategies for mobile networks, subject to which this thesis contributes. Making use of new technologies and techniques, we can deploy a multi-tier network with a lower layer of small cell deployments that are managed through a dynamic system that can automatically perform certain operations over that network. Machine Learning is an increasing trend in this field, can help with the process by making use of the data collected from the network, obtain useful knowledge, and create predictive models that can tell us the state of the network in the near future. For the development of this project, we have collaborated with COSMOTE, one of the main telecommunications companies in Greece, who have provided us with several data sets of a real network deployment in the centre of Athens. With these data, several predictive models have been created to predict the state of the network during certain time intervals and act in consequence. Many different applications can be found for those algorithms, although one of those that is a hot topic nowadays is energy efficiency. To work on that field, the prediction models where used to create a dynamic system that turns cells on and off dynamically, depending on the expected traffic, in order to achieve notable energy savings. Finally, a simulation environment was developed, based on the real traces from the COSMOTE network, in order to test the proposed network management techniques in a large number of different scenarios. This simulator generates realistic random scenarios from which several statistics can be extracted, with the aim of measuring the performance of the algorithms developed during the earlier stages of the project. Working with different tools and environments, this project studies the best data analysis and Machine Learning techniques regarding network usage data. From that data, prediction models are created, which can be used for many different and interesting applications. The one chosen for this thesis is the design of an energy efficient management system for dense small cell deployments. Finally, results are collected, and the validity of the proposed strategies is proved.