MACHINE LEARNING ASSISTED MULTI-CELL BASE STATION SLEEP MODE MANAGEMENT

Abstract: Fifth generation (5G) communication technology provides higher data rates, better connectivity, and many new facilities. However, there are some challenges that need to be addressed. Denser 5G base station deployment, more power-hungry communication equipment, and an exponentially growing number of mobile users are expected to result in high energy consumption, which is a critical issue for the future in terms of sustainability and potential threats to the environment. These issues raise the importance of energy-efficient network design in an aim to increase energy saving. Energy saving could be potentially achieved on the most energy-consuming component of the mobile networks: base stations. By switching off a number of components while not serving any users, base stations could switch to a sleeping mode thus, save energy. This thesis focuses on maximizing the sleeping opportunity of a base station while it is not serving any users by exploiting novelties brought by 5G networks namely 5G NR numerology and designing a sleep management algorithm for base stations to reduce redundant energy consumption by using Advanced Sleep Modes. To maximize the duration of sleep, 5G numerology with finer time granularity was used and an increase in the sleeping opportunity was observed compared to the baseline numerology in 4G systems. Based on that, a Q-learning algorithm has been designed in a reference base station in a multi-cell environment in order to manage how long and how deep to sleep by using Advanced Sleep Modes. The results have shown that compared to a non-intelligent sleep mode management scheme, which takes advantage of only the shallowest sleep mode, up to 80% of the potentially wasted energy when the BS is inactive could be saved. With a balanced trade-off policy between power consumption and additional latency induced to the users, 51% of the potentially wasted energy is saved while the users experience an additional latency of 2.1 ms on average. The algorithm adapts to the changing conditions in the network by making decisions according to the varying sleeping opportunities.