Use of Machine Learning for energy efficiency in present and future mobile networks David Sesto-Castilla * , Eduard Garcia-Villegas * , George Lyberopoulos + and Eleni Theodoropoulou + * Universitat Politècnica de Catalunya (UPC), Barcelona, Spain + COSMOTE Mobile Telecommunications S.A, Maroussi, Greece Abstract— Given the current evolution trends in mobile cellular networks, which is approaching us towards the future 5G paradigm, novel techniques for network management are in the agenda. Machine Learning techniques are useful for extracting knowledge out of raw data; knowledge that can be applied to improving the experience in the operation of such systems. This paper proposes the use of Machine Learning applied to energy efficiency, which is set to be one major challenge in future network deployments. By studying the cell- level traces collected in a real network, we can study traffic patterns and derive predictive models for different cell load metrics with the aid of different machine learning techniques. Such models are applied into a simulation environment designed to test different algorithms which, according to cell load predictions, dynamically switch on and off base stations with the aim of providing energy savings in a mobile cellular network. Keywords—5G; machine learning; energy efficiency I. INTRODUCTION Wireless communication is, by far, the fastest growing segment of digital communications, probably due to its cost- effectiveness and its ability to provide (near) ubiquitous access to a plethora of services and applications. Mobile cellular networks constitute one of the main pillars of the aforementioned growth: nowadays, the number of mobile subscribers exceeds the world population (7,600 million at the time of writing) and different studies predict a 5% annual growth rate (CAGR), reaching 9,000 million subscribers by 2021 [1]. Moreover, not only the number of subscribers is expected to keep growing, but also the amount of data traffic generated by each connected device is expected to show a ten-fold increase in the same period due to the proliferation of bandwidth-intensive applications (e.g. HD video streaming, virtual/augmented reality, etc.). This scenario prompts the academia and the industry to work towards a new generation of wireless network, i.e. 5G. While the specification of 5G targets a time horizon beyond 2020, the requirements those new technologies are expected to meet have already been discussed and identified in different works [2], which foresee 10-100x device density, 10-100x typical user data rate and 5 times reduced latency. Those requirements are mainly tackled by new developments in radio access technologies (RAT), such as massive MIMO or mmWave. However, expectations of ~30% reduction in infrastructure energy consumption need a different approach. What is more, the aforementioned requirements should be met while keeping sustainable costs. This “implicit” requirement can be achieved through an innovative network architecture and intelligent management of the available resources. In this regard, techniques such as network function virtualization (NFV), software-defined networking (SDN) and the application of big data-driven intelligence were also identified as key enablers of the future 5G [3]. In this context, the present paper explores the idea of applying machine-learning techniques to improve management procedures of future 5G networks towards a fully self-organized network (SON). Through the analysis of huge amounts of network-generated data, the proposed intelligent system is capable of anticipating the future state of the network so that appropriate actions can be taken in a timely manner. As a proof of concept, this paper studies the application of such intelligence to produce energy savings: we propose a mechanism to switch off/on cells according to their expected traffic conditions. These kind of techniques are highly relevant to mobile operators if we consider that densification (increasing number of deployed cells is needed to increase area capacity) will carry an increase in energy consumption and the fact that 70-80% of energy consumption (and gas emissions derived thereof) of mobile operators comes from their network infrastructure [4]. The remaining of this paper is structured as follows: section II motivates the use of machine learning applied to the management of mobile networks and reviews related literature. Next, section III discloses the proposed mechanism. Then, section IV describes the evaluation environment and V discusses the results. Finally, conclusions are provided in VI. II. MACHINE LEARNING APPLIED TO MOBILE NETWORKS A high level view of current mobile networks is given in Fig. 1, which shows three distinctive components: i) the user equipment (UE) is the device connecting to the access network and providing subscribers with access to their services; ii) radio access network (RAN) provides a wireless link to UEs through one or multiple cells served by multi- This work has received funding from the European Union under grant agreement 762057 (H2020 5G-PICTURE) and by the ERDF and the Spanish Government through project TEC2016-79988-P. Fig. 1: generic architecture of a Mobile network