Resumen de Energy sustainability of next generation cellular networks through learning techniques
The trend for the next generation of cellular network, the Fifth Generation (5G), predicts a 1000x increase in the capacity demand with respect to 4G, which leads to new infrastructure deployments. To this respect, it is estimated that the energy consumption of ICT might reach the 51% of global electricity production by 2030, mainly due to mobile networks and services. Consequently, the cost of energy may also become predominant in the operative expenses of a Mobile Network Operator (MNO). Therefore, an efficient control of the energy consumption in 5G networks is not only desirable but essential. In fact, the energy sustainability is one of the pillars in the design of the next generation cellular networks.
In the last decade, the research community has been paying close attention to the Energy Efficiency (EE) of the radio communication networks, with particular care on the dynamic switch ON/OFF of the Base Stations (BSs). Besides, 5G architectures will introduce the Heterogeneous Network (HetNet) paradigm, where Small BSs (SBSs) are deployed to assist the standard macro BS for satisfying the high traffic demand and reducing the impact on the energy consumption. However, only with the introduction of Energy Harvesting (EH) capabilities the networks might reach the needed energy savings for mitigating both the high costs and the environmental impact. In the case of HetNets with EH capabilities, the erratic and intermittent nature of renewable energy sources has to be considered, which entails some additional complexity. Solar energy has been chosen as reference EH source due to its widespread adoption and its high efficiency in terms of energy produced compared to its costs. To this end, in the first part of the thesis, a harvested solar energy model has been presented based on accurate stochastic Markov processes for the description of the energy scavenged by outdoor solar sources.
The typical HetNet scenario involves dense deployments with a high level of flexibility, which suggests the usage of distributed control systems rather than centralized, where the scalability can become rapidly a bottleneck. For this reason, in the second part of the thesis, we propose to model the SBS tier as a Multi-agent Reinforcement Learning (MRL) system, where each SBS is an intelligent and autonomous agent, which learns by directly interacting with the environment and by properly utilizing the past experience. The agents implemented in each SBS independently learn a proper switch ON/OFF control policy, so as to jointly maximize the system performance in terms of throughput, drop rate and energy consumption, while adapting to the dynamic conditions of the environment, in terms of energy inflow and traffic demand.
However, MRL might suffer the problem of coordination when finding simultaneously a solution among all the agents that is good for the whole system. In consequence, the Layered Learning paradigm has been adopted to simplify the problem by decomposing it in subtasks. In particular, the global solution is obtained in a hierarchical fashion: the learning process of a subtask is aimed at facilitating the learning of the next higher subtask layer. The first layer implements an MRL approach and it is in charge of the local online optimization at SBS level as function of the traffic demand and the energy incomes. The second layer is in charge of the network-wide optimization and it is based on Artificial Neural Networks aimed at estimating the model of the overall network.
