Modelling and minimization of energy use in 5G networks

与4G长期演进（LTE）相比，5G网络的设计目标是每单位面积提供100倍以上的数据，尽管以比特/焦耳为单位的能效总体提高，但5G基站（BS）的能耗仍是4G基站的两倍多。考虑到这一点，网络能量效率正在成为通信网络中的主要品质因数。该研究项目是与英国电信公司合作进行的，旨在利用模型和数据驱动的方法优化通信网络的能源效率。例如，机器学习（ML）方法已被用于预测网络流量和/或优化BS参数，以最大限度地提高能源效率，同时保持服务质量要求。然而，传统的BS通常不会收集这些方法在实践中工作所需的训练或输入数据。一个很有前途的方法，已被建议用于处理这个问题是使用分析模型结合机器学习。这种方法可能会结合模拟和迁移学习的元素，以减少对大量训练数据的需求。目前，ML方法通常不比局部优化技术更好，特别是考虑到它们苛刻的计算要求。为了使机器学习方法更加可行，本研究可以探索机器学习方法的各个方面，例如选择合适的机器学习架构，以及确定算法应该以集中式还是分布式方式部署。另一个研究方向可以是调查BS参数如何影响相邻网络组件。BS参数最经常被优化，目的是改善小区内操作。由一个BS选择的策略（休眠模式、发射功率等）可以是还例如通过移动业务或干扰电平来影响相邻节点。人们对这些关系知之甚少，因为很难对这些依赖关系进行建模。随机几何是分析大型无线网络的最先进的数学工具。这种网络最常见的模型是所谓的泊松点过程（PPP）模型，该模型将网络组件视为在空间中均匀分布的随机点集。该模型已被证明产生合理的各种网络变量（中断概率，干扰等）易于处理的表达式。然而，为了理解各种网络参数之间的关系，我们需要引入一些空间依赖性，但这往往会违反PPP中存在的同质性，导致数学上的易处理性损失。考虑到这一点，我们可以致力于开发一些方法，以一种足够简单的方式来建模可单独配置的网络组件，以提供一些设计见解。

Compared to 4G long-term evolution (LTE), 5G networks are designed to provide 100 times more data per unit area and despite an overall increase in energy efficiency in bits/joule, 5G base stations (BS) still consume more than twice as much energy as their 4G counterparts. With this in mind, network energy efficiency is becoming a major figure of merit in communication networks. This research project, which is conducted in collaboration with BT, will be aimed at using models and data-driven methods to optimize energy efficiency in communication networks.Machine learning (ML) methods, for example, have been used in order to forecast network traffic and/or optimize BS parameters to maximize energy efficiency while maintaining quality of service requirements. Traditional BSs, however, don't usually collect the training or input data required for these methods to work in practice. A promising approach which has been suggested for dealing with this problem is the use of analytic models in conjunction with machine learning. This approach would possibly incorporate simulations and elements of transfer learning to reduce the demand for extensive training data.Currently, ML methods are usually no better than local optimization techniques, especially given their demanding computational requirements. In hopes of making ML methods more viable, this research could explore various aspects of the ML approach, such as selecting the appropriate ML architecture and determining whether the algorithms should be deployed in a centralized or distributed manner.Another research direction could be to investigate how BS parameters affect adjacent network components. BS parameters are most often optimized with the intent of improving intra-cell operation. The policies chosen by one BS (sleep modes, transmit power, etc.) also affect adjacent nodes for example by shifting traffic or interference levels. These relationships are poorly understood as it is quite difficult to model these dependencies. Stochastic geometry is the state-of-the-art mathematical tool for analyzing large wireless networks. The most common model for such networks is the so-called Poisson point process (PPP) model which considers network components as a random set of points uniformly distributed in space. This model has been shown to yield reasonably tractable expressions for various network variables (outage probability, interference, etc.). However, to understand the relationships between various network parameters we would need to introduce some spatial dependence, but this tends to violate the homogeneity present in PPPs resulting in the loss of mathematical tractability. Considering this, we could aim to develop some method of modeling individually configurable network components in a way which is simple enough to provide some design insight.