A Survey on 5G Energy Efficiency: Massive MIMO, Lean Carrier Design, Sleep Modes, and Machine Learning
David Lopez-Perez, Antonio De Domenico, Nicola Piovesan, Harvey Baohongqiang, Geng Xinli, Song Qitao, Merouane Debbah
aa r X i v : . [ c s . N I] J a n A Survey on 5G Energy Efficiency: MassiveMIMO, Lean Carrier Design, Sleep Modes,and Machine Learning
David L ´opez-P´erez, Antonio De Domenico, Nicola Piovesan, HarveyBaohongqiang, Geng Xinli, Song Qitao and M´erouane Debbah
Abstract
Cellular networks have changed the world we are living in, and the fifth generation (5G) of radiotechnology is expected to further revolutionise our everyday lives, by enabling a high degree of automa-tion, through its larger capacity, massive connectivity, and ultra-reliable low latency communications.In addition, the third generation partnership project (3GPP) new radio (NR) specification also providestools to significantly decrease the energy consumption and the green house emissions of next generationsnetworks, thus contributing towards information and communication technology (ICT) sustainabilitytargets. In this survey paper, we thoroughly review the state-of-the-art on current energy efficiencyresearch. We first categorise and carefully analyse the different power consumption models and energyefficiency metrics, which have helped to make progress on the understanding of green networks. Then, asa main contribution, we survey in detail —from a theoretical and a practical viewpoint— the main energyefficiency enabling features that 3GPP NR provides, together with their main benefits and challenges.Special attention is paid to four key technology features, i.e., massive multiple-input multiple-output(MIMO), lean carrier design, and advanced idle modes, together with the role of artificial intelligencecapabilities. We dive into their implementation and operational details, and thoroughly discuss theiroptimal operation points and theoretical-trade-offs from an energy consumption perspective. This willhelp the reader to grasp the fundamentals of —and the status on— green networking. Finally, the areasof research where more effort is needed to make future networks greener are also discussed.
I. I
NTRODUCTION
The living world is a unique and spectacular wonder. For 10,000 years, the average temperatureof our planet has not increased or decreased by more than one degree Celsius [1]. The biodiversity of the Holocene, together with this stationary climate, brought stability, settling the living worldinto a gentle, reliable rhythm —the seasons.We invented farming, and with it, learnt how to master the seasons to produce large amounts offood crops. The history of all human civilization has followed, with each generation progressingwith respect to its preceding one, mostly because the living world could be relied upon to deliverus the essentials we needed.The industrial revolution and the automation of labour changed this natural rhythm of evolu-tion, and since then, we have significantly transformed our living world [2]. After the SecondWorld War, technology has developed faster than ever before. Automation is making our liveseasier, and until very recently, it felt like nothing would limit our progress. However, we arestarting to see the consequences of unsustainable progress now [3].From 1937 to 2019, the world population has grown from 2.3 to 7.7 billions [4], and ourmodern transport, agricultural, manufacturing and life styles have sharply increased energyconsumption. As a consequence, the amount of carbon in the atmosphere raised from 280 to 409parts per millions in such time period [5], and the levels of greenhouse gas (GHG) emissionsreached the historical record of 37.5 gigatones of carbon dioxide equivalent (CO e ) in 2018, —a1.5 % increase with respect to 2008 [6]. If our societies do not significantly change the mannerin which we consume energy, this level may reach 47.5 gigatones of CO e by 2030, and theoutcomes could be catastrophic [7].Indeed, this increase of GHG emissions, combined with current trends on deforestation, arecontributing to global warming —the rise of the average Earth surface temperature [8]. Thistemperature rose 0.6 to 0.9 degrees Celsius between 1906 and 2005, and the rate of temperatureincrease has nearly doubled in the last 50 years [9] [10]. Alarmingly, an increase in the Earth’stemperature of 1.5 to 2.0 degrees Celsius, above pre-industrial temperatures, has been estimatedto be a threat to most natural ecosystems on Earth today, with the resulting dramatic effects onour planet, and in turn, on our everyday lives [11]–[14].To address this challenge, international policymakers are targeting a dramatic increase inenergy efficiency, and a sharp shift from fossil fuels to renewable sources of energy, such assolar, wind, and water. This will entail a completely new approach to the generation and use ofenergy, which must be adopted by every government, industry, business, and individual.In the following, we review the important role that the fifth generation (5G) of mobiletechnology will play to revert the unsustainable energy consumption trend, and survey, in the rest Figure 1. Wireless industry client action goals [15]. of the paper, the technical innovations that the third generation partnership project (3GPP) newradio (NR) specification brings in terms of energy efficiency —from a theoretical and practicalviewpoint— with respect to previous ones to support such change.
A. The Enabling Role of 5G
Governments and industries have set —or are setting— ambitious targets to reduce their GHGemissions and help dealing with global warming. To date, 77 major economies have already es-tablish a net-zero GHG emission target by 2050, and their industries are accordingly (re)definingtheir energy efficiency and consumption road maps. In this regard, the telecommunication sectorhas taken the lead —and an exemplary role—, setting stringent requirements for both the energyefficiency and consumption of their networks, together with a clear road map to achieve them(see Fig. 1) [15].Enhancing the energy efficiency and reducing the energy consumption of 5G networks willhelp reducing GHG emissions. Their enabling effect, however, will be —without doubt— themost important contribution of the mobile industry to address the current climate change [16].At a macro-scale level, the new 5G technology enables a new type of networking capabilityable to connect for the first time both everyone and everything together, including machines,objects, and devices, thanks to its higher capacity, lower latency, improved reliability, and largernumber of supported connections Please refer to Fig. 2 for a comparison between 5G andfourth generation (4G) specification capabilities. These new 5G communication capabilities are
Figure 2. Comparison of key capabilities of IMT-Advanced (4G) with IMT-2020 (5G) according to [18]. already helping governments, current industries, and new forms of businesses to implement novelprocesses with improved effectiveness, by supporting a more flexible, tailored, and efficient useof resources.Importantly, 5G has already become an integral part of governmental and industrial energyefficiency and consumption programs, as it is envisioned that an intelligent exploitation ofresources will enable a significant decrease of GHG emissions through different avenues, forexample i) an improved support for smart city and building energy management, ii) reducedrequirements for office space and business travel, and iii) efficient just-in-time supply chainsenabled by predictive analytics, to cite a few [17].To give some idea of the magnitude and importance of this enabling effect, it should be notedthat, according to the International Telecommunication Union (ITU) SMART 2020 report [19],the enabling effect of mobile communications alone was estimated to be around 2,135 milliontones of CO e in 2018, and that according to [20], the scale of it will increase in the 5G era,where the enabling effect across all the information and communication technology (ICT) sectorwas predicted to be equivalent to 15 % of all global emissions by the end of 2020.Recent reports indicate that industries, such as transportation, health care, and manufacturing,are already significantly benefiting from such 5G enabling effect. For example, through smart cityprograms and 5G-related innovations, London, Berlin, and Madrid have already reduced GHG emissions of motor vehicles by 30 % each from their peak rates, and Copenhagen by 61 % [21].The ultra-reliable low-latency communication (URLLC) capabilities of 5G can also lead toautonomous driving, optimizing the route as well as the fuel and break control. which will furtherreduce the fuel consumption of vehicles [22]. In China, 5G-enabled remote computed tomographyhas eliminated the need for the road and/or air travel associated with expert consultation sessions,thus reducing by 99 % the GHG emissions of this health care sector [23]. Some smartphonemanufactures, to give another example, are also using 5G-connected artificial intelligence (AI)cameras to perform quality inspections before smartphones are packaged, which has reduced by6 % the energy consumption of the per-unit smartphone production [17].To further assess the of breath and depth of the enabling effect of 5G networks, interestedreaders are refereed to [16], and references there in. B. The 5G Energy Efficiency Challenge
Unfortunately, the 5G enabling effect is no free lunch. In fact, it comes at the expense of atremendous challenge for the telecommunication sector in terms of both carried data and energyconsumption.Allowing governments, industries, businesses, and individuals in general to increase theirenergy efficiencies and reduce their energy consumption through more flexible, tailored, andefficient operations via a telecommunication network entails • a dramatic growth of data usage in some scenarios, and • the need for more sophisticated networking to meet the required low-latency, high-reliability,and/or large volume of data connections in some others.Confirming such challenge, recent studies already indicate that, by 2030, the number of connecteddevices is expected to grow to 100 billion [24], and that 5G networks may be supporting up to1,000 times more data than 4G ones did in 2018 [17].Importantly, • to reach their established energy efficiency and consumption targets, and • reduce their energy bills, up to 40 % by 2030 [17], to make their businesses profitable,mobile network operators (MNOs) will need to meet the aforementioned more challenging trafficdemands and requirements with significantly reduced GHG emissions with respect to those oftoday’s 4G networks. Considering these two aspects, the 3GPP NR stakeholders have already called for a 90 % reduction in energy consumption compared to 4G long term evolution (LTE) [25]. However,whether these gains can be realised or not in practical networks will not only depend on whatthe new specification can do and/or the energy performance of a single site, but also on howthe actual network is deployed and operated as a whole.In fact, to support the growing use of 5G connectivity and its more stringent requirements,while reducing energy consumption on a per-bit basis through an intelligent use of the network,changes are needed at all levels of it to achieve the maximum holistic effect. MNOs mustthus embrace new approaches to network planning, deployment, management, and optimisationthat have energy efficiency at heart, and are implemented end-to-end. Without energy efficiencydriving future deployments, the study in [26] indicates that a 5G network, despite of its enhancedenergy efficiency in bits per Joule due to its larger bandwidth and better spatial multiplexingcapabilities, could typically consume over 140 % more energy than a 4G one, with similarcoverage area. This unwanted energy consumption arises from 5G’s greater density of basestations (BSs), antennas, cloud infrastructure, and user equipment (UE), among others.To address this challenge, and better understand where the energy consumption could bemeaningfully reduced in a 5G network, thus helping MNOs to take educated decisions, theauthors of [17] put together an interesting analysis, reporting that the most consuming elementsof the network in terms of energy are the BSs. They currently account for about 57 % of thetotal network energy consumption [27]. By 2025, that figure should be lower, as 5G becomesmore prevalent, but the radio access network (RAN) will still be the biggest consumer of energyin the network, a 50.6 % according to [28] (see Fig. 3).Within a BS itself, the radio frequency (RF) equipment i.e., the power amplifier plus thetransceivers and cables, have been identified as the largest energy consumer, typically using about65 % of the total BS energy. The cooling system, the digital signal and base band processingas well as the alternating current (AC)-direct current (DC) converters follow with an energyconsumption of around 17.5 % , 10 % , and 7.5 % , respectively.In this line, 5G can be —and has been— improved for a better energy efficiency. Moreefficient power amplifiers have been developed, renewable energy sources for powering on-gridand off-grid sites, including solar power, are starting to be widely adopted. Moreover, smartlithium batteries are becoming an integral part of any 5G site to enhance energy management,and liquid cooling is being implemented to reduce the need for air conditioning [17]. Figure 3. Energy consumption breakdown by network element in 2025 [28].
Importantly, since the BSs —and their RF equipment— consume most of the energy at a5G network, we must advocate for a judicious networking [29]. BSs must thus only be active—and consume energy— when handling actual data. Plainly speaking, the Joules consumedshould ’follow’ the bits transmitted. As a result, and in contrast to 4G, the amount of always-onsignalling in next generation networks must be greatly minimised at all cost. It is also equallynecessary that data —and its related signalling— are transmitted to/received from the intendedUEs while consuming the minimum possible energy to meet their required quality of experience(QoE) [30]. Avoiding the resource waste occasioned by the over-provisioning of QoE is essentialto significantly reduce the GHG emissions of 5G networks.To realise such network-wide energy efficient operation, the 3GPP NR specification [31] i) has been redesigned and developed using a new and more flexible user-centric principle [32],and ii) provides support for a number of new technology features that help reducing energyconsumption at a network level. The unique 3GPP NR power saving technologies are: • massive multiple-input multiple-output (mMIMO), • lean carrier design, • advanced sleep modes (ASMs), and • artificial intelligence.The good understanding of these features, how they enable energy efficiency and their optimal operation points in such terms are the main focus of this survey. C. Comparison to previous 5G Energy Efficiency Surveys
Several surveys have been already published on energy efficiency and related aspects duringthe last 10 years. In the following, we describe the most relevant ones.In [33], the authors have surveyed the goals set by the United Nations (UN) for sustainabledevelopment, where among others, urgent actions to combat climate change are called upon.Importantly, the wide variety of opportunities in the ICT sector for enabling energy efficiencyhave been highlighted, but although of relevance to understand the current scene, this survey didnot touch on the particularities of wireless networks. The survey in [34], instead, has focused ongreen wireless networks, and has explained at a high level —and an introductory form— howdifferent metrics such as energy efficiency, spectral efficiency, throughput, and delay relate toeach other. The discussion has been organised around the open systems interconnection (OSI)layers, and has presented an exhaustive summary of the different frameworks and techniquesthat can be used to achieve optimal networking trade-offs in practical networks. Descriptions,however, are only qualitative, and lack of detail. For example, the paper has considered a simplepower consumption model in most explanations, which only accounts for the transmit power,and has neglected the power consumption of equipment hardware. With a similar scope, theauthors in [35] have surveyed the literature around the understanding of fundamental energyperformance trade-offs, i.e., energy and deployment efficiency, energy and spectral efficiency,bandwidth and power as well as delay and power, this time, however, from a more focused cellularperspective, covering both 3GPP universal mobile telecommunication system (UMTS) and LTEaspects. Importantly, this paper has brought the attention to the importance of accurate BS powerconsumption models. Based on this work, the authors in [36] have further surveyed techniques tooptimise the above mentioned four trade-offs, while putting the spotlight on the potential benefitsof two features, i.e., multiple-input multiple-output (MIMO) and relay. With a more 3GPP LTEstandard focus, the survey in [37] has provided an in-depth review on green aspects, discussingadvancements in power amplifier technology, 3GPP LTE protocols for carrier shutdown and microsleeps, as well as the potential benefits of small cell networks. On a forward looking note, theauthors have also put forward cognitive radio and cooperative relaying as energy saving enablingtechnologies, and share the latest developments in these areas. Complementarily, the authorsin [38] have surveyed the efforts of different academic and industry projects on energy efficiency, emphasizing how to make use of daily network traffic variations and various quality of service(QoS) requirements to save energy at the network level. Several deployment strategies, such ascell-size, heterogeneous networks, cooperative communications and network coding, have beenalso discussed, together with their merits and challenges. On a similar note, in [39], a completesurvey on 3GPP metrics and power consumption models for energy efficiency analysis can befound, with a detailed formulation of the most relevant energy efficiency trade-offs. An exhaustive—and critical discussion— on standardised features for energy-aware management in 3GPP LTEcellular networks from a broadband networking point of view has been also provided. In [40],the authors have turned the focus on the survey of efficient resource management schemes,which are capable of controlling how much of the network infrastructure is actually needed ina given space-time and which parts can be temporarily powered off to save energy. The surveyincludes both cellular- and Wi-Fi-based algorithms. Focusing on more modern applications,the work in [29] has overviewed the challenges brought by the big data era, describing issuesand solutions around energy efficient data acquisition, communication, storage, computation,and analytics. Discussion on the necessity of avoiding radio resource waste to reduce energyconsumption in wireless networks has been at the core of the survey, and four types of schemeshave been surveyed in this line, i.e., power control, time-domain scheduling, spatial resourceallocation, and spectrum sharing.Unfortunately, it should be highlighted that any of the above mentioned surveys has gone intothe specifics of 3GPP NR. They are either generic or 3GPP LTE focus, and as a result, they didnot survey the assets of this new technology generation to harvest energy savings.Giving a more related 5G perspective, the authors in [41] have provided guidelines for thedevelopment of energy efficient related features in 3GPP NR, and surveyed the literature aroundthe user-centric concept of no more cells . This networking paradigm would enable energy effi-ciency through a flexible network comprised of heterogeneous cells, decoupled signal- and data-planes as well as downlink and uplink. A dynamic cloud radio access network (C-RAN)-basedconfiguration has also been proposed to handle spatial and temporal mobile traffic variationswithout energy over-provisioning. In this paper, the potential of both mMIMO and full duplexfeatures in 5G for power savings has been also surveyed, while considering hardware issues.In the same line, the work in [42] has provided an overview of the latest research on green5G techniques. The authors have explored ultra-dense sub-6GHz and millimetre wave networks,unlicensed spectrum as well as device to device (D2D) and mMIMO communications, while analysing their potential energy efficiency improvements, as well as circuit power consumptionissues. As a main contribution, energy harvesting has been presented as fundamental to meet 5Ggreen requirements, and the different lines of work in this have been discussed (e.g., renewable,RF energy harvesting). Taking a more theoretical and systematic approach, the authors in [43]have extended their work in [35], surveying energy efficient solutions for 5G networks using theaforementioned fundamental green trade-offs as a driver. This overview has been around threemain pillars, i.e., non-orthogonal access, mMIMO, and heterogeneous networks. Importantly, thepaper concludes that mMIMO is the most effective approach to enable high energy efficiencies,provided that issues around channel state information (CSI) acquisition, transceiver hardwareimpairments, and power inefficient components are addressed. A large number of referenceshas been also provided around channel and carrier shutdown techniques for dense small cellnetworks, and the implications of centralised versus distributed network architectures have beendiscussed. With a more practical —but still vanilla— 5G perspective, the authors in [44] providea comprehensive survey on how machine learning (ML) can be used to address the energyefficiency challenges encountered in generic 5G networks. Finally, the recent survey in [45]has provided the most up-to-date overview on power saving techniques supported by the 3GPPNR standard, covering developments in Release 15 and 16, and the potential upcoming onesin Release 17. Such overview, however, has mainly focused on —and evaluated— UE powersaving mechanisms, such as bandwidth parts, radio resource control (RRC) inactive state, dis-continuous reception (DRX) mechanism, wake up signaling, cross-slot scheduling, and MIMOlayer adaptation. At the network side, the lean carrier and discontinuous transmission (DTX)concepts together with that of dormant cells, have been only briefly touched upon.As it can be derived from the previous summary, most of the existing energy efficiency surveyswere written in a pre-5G era before the 3GPP NR existed/matured, or have a strong focus onthe UE —and not on the network— side. D. Objective and Structure of this Paper
In this survey paper, contrary to the previous ones, which speculatively discussed what 5Gcould be in terms of energy efficiency, we provide for the first time a detailed, up-to-date overviewof the main practical 3GPP NR features that are currently used —or can be further leveraged—to increase the energy efficiency of practical 3GPP NR networks. The survey covers the roleof mMIMO, the lean carrier design, ASMs, and ML in providing energy savings in macro Table IS
UMMARY AND COMPARISON OF THE MOST RELEVANT ENERGY EFFICIENCY SURVEYS . Ref. Year Tech. era Contribution/techniques Gaps [33] 2018 ICT Surveys UN frameworks for sustainable development No cellular or ML re-lated.[34] 2016 Generalwireless Surveys green metrics and performance trade-offs. No 5G or ML related.[35] 2011 3G/4G Surveys green metrics and performance trade-off. No 5G or ML related.[36] 2011 4G Surveys green performance trade-off optimization. ExploresMIMO and relay features. No 5G or ML related.[37] 2011 4G Surveys 3GPP LTE green protocols. Focuses on carrier shutdown,and micro-sleep features. Explores cognitive radio, and coopera-tive relaying. No 5G or ML related.[38] 2013 4G Surveys green academic and industry projects, and energy mini-mization under UE QoS constraints. Explores heterogeneous net-works, cooperative communications, and network coding features. No 5G or ML related.[39] 2014 4G Surveys 3GPP green metrics, power consumption models, and3GPP LTE protocols. Detailed formulation of energy efficiencytrade-offs. No 5G or ML related.[40] 2014 4G andWi-Fi Surveys energy efficient resource management schemes (withfocus on Network and MAC layers). No 5G or ML related.[29] 2018 Big dataera and4G Surveys the challenges brought by the big data era. With re-spect to cellular, surveys MAC layer power control, time-domainscheduling, spatial resource allocation, and spectrum sharing greenprotocols. No 5G or ML related.[41] 2014 Pre-5G Surveys potential 5G energy efficiency features. Explores the user-centric concept, downlink and uplink split C-RAN, mMIMO, andfull duplex features. Guesses what 5Gcould be. No 3GPPNR or ML related.[42] 2017 Pre-5G Surveys research on green 5G techniques. Explores ultra-densesub-6GHz and millimetre wave networks, unlicensed spectrum aswell as D2D, mMIMO, and energy harvesting features. Guesses what 5Gcould be. No 3GPPNR or ML related.[43] 2017 Pre-5G Surveys green metrics and performance trade-offs. Explores non-orthogonal access, mMIMO, and heterogeneous networks. Guesses what 5Gcould be. No 3GPPNR or ML related.[44] 2020 5Ggeneric Surveys how ML can be used to address 5G energy efficiencychallenges. No 3GPP NR focus.Lack of detail.[45] 2020 5G Surveys power saving techniques supported by the 3GPP NRstandard (Rel.15/16) with focus on UE Does not cover net-work aspects or MLtechniques.
5G deployments, from both a theoretical and practical perspective, and highlights their still tobe addressed challenges. Importantly, this survey provides detailed descriptions, including theformulation of BS power consumption models and energy efficiency metrics currently used forenergy efficiency optimisation, as well as the bounds, the trade-offs, and the optimal operationpoints derived in the literature, making this a self-contained paper. The important role to beplayed by ML and data-driven in terms of energy efficiency is also surveyed for the first time.Areas of research which require further efforts are also discussed.The rest of this survey paper is organised as follows (see Fig. 4): • In Section II, we introduce mMIMO, the lean carrier design, ASMs and ML as energyefficiency enabling features in 3GPP NR; • In Section III and Section IV, we present and discuss in detail existing BS power consump-tion models and metrics used for energy efficiency optimisation, respectively; • In Section V, we overview the current theoretical understanding of mMIMO in termsof energy efficiency from a single and a multi-cell perspective and from a uplink anddownlink viewpoint. Current bounds and trade-offs with other key performance indicatorsare formulated and explained, and the most relevant optimisation frameworks to enhanceenergy efficiency via mMIMO are surveyed; • In Section VI, we dive into the details of the lean carrier design, and highlight the importanceof ASMs and their optimisation at different levels (i.e., micro sleeps, carrier, and channel(antenna) shutdown); • In Section VII, we highlight the potential of spatio-temporal traffic predictions and MLapproaches to maximize energy efficiency, and overview the research in this area; • Finally, in Section VIII and IX, we discuss future research directions and draw the conclu-sions, respectively.II. 3GPP NR E
NERGY E FFICIENCY R ELATED F EATURES
To reach the ambitious targets of 5G networks in terms of capacity, latency, reliability, numberof supported connections, and energy efficiency, the 3GPP NR specification presents a paradigmshift with respect to any preceding cellular technology [32].In comparison with 3GPP LTE and with regard to energy efficiency, 3GPP NR introduces anew beam-centric —or mMIMO-centric— design, which enables both I. IntroductionII. 3GPP NR energy efficiencyrelated featuresIII. Power consumption m odels
IV. Energy efficiency m etrics V. mMIMO energy efficiency theoretical understandingVI. Network adaptation to QoS requirementsVII. ML and data-driven energy efficiency optimizationVIII. Open research directionsIX. Conclusions The energy saving enabling role of 5GThe 5G energy efficiency technical challengeComparison to previous energy efficiency surveysObjective and structure of this paperMassive MIMOThe lean carrier designAdvanced sleep modesArtificial intelligenceDistributed RAN-based modelsCentralized RAN-based models Statistical methodsML-based methodsInterference-limited network-based metricsNoise-Limited network-based metricsNetwork delay-aware metricsUE link performance-aware metricsThe global energy problemSingle cell scenarioMulti cell scenario BoundsTrade-offs Available mathematical toolsUplink mMIMO network deployment perspectivesDownlink mMIMO network deployment perspectives Time - domain- based energy saving solutionsCarrier-domain-based energy saving solutionsAntenna-domain-based energy saving solutionsML for traffic predictionML for 5G energy efficiency optimizationEnergy efficiency-driven network planning toolsMulti-carrier and heterogeneous network analysisData-driven optimizationGreen AIMulti cell energy efficiency theoretical modellingRenewable energy sources Recurrent neuronal networksConvolutional neuronal networksGraph neuronal networksExternal inputs: Point of InterestsModel re-usability Figure 4. Outline and structure of this survey. • an extensive use of beamforming through a massive number of antenna elements, not onlyfor data transmissions, but also for control-plane procedures, such as the initial access [46],and • a larger spatial multiplexing of information on a given time-frequency resource.Such beamforming gains allow for a reduced transmit power to reach a given targeted distanceand/or meet a given QoE at the UE, with the potential consequent benefits in terms of interferencemitigation and energy savings [47].Moreover, 3GPP NR follows a new ultra-lean design principle, in which control signals are notconsistently transmitted in every radio frame, but on demand, based on traffic requirements [48].This ultra-lean design allows for a more efficient operation of the mMIMO-centric design, andfacilitates ASM support, i.e., BS wake-up/sleep, including symbol, channel, or carrier shutdown,which can significantly reduce energy consumption [49].3GPP NR enhancements also allow for a more distributed architecture, which facilities the usage of AI to assist network optimisation through a centralised data gathering and processing—the mentioned required intelligence [50]. In this line, the current 3GPP NR architecture alreadyintroduces new functions in the core and the management domains, i.e., the network data analyticsfunction (NWDAF) and the management data analytics function (MDAF), which can either runanalytics on collected data or enhance the already supported network functions with statisticscollection and prediction capabilities [51] [52]. Given the increased complexity of 5G networks,further leveraging 5G network data and ML will be necessary to derive optimum network-wideenergy efficient operation policies [53].In the following, we introduce the main concept behind these 3GPP NR features, whose properoptimisation will be key to minimise network energy consumption. We will further elaborate ontheir details and potential energy saving abilities in the rest of this paper. A. Massive MIMO
Full dimension MIMO —generally refereed to as mMIMO— is probably one of the mostimportant developments in 3GPP NR [46]. Plainly speaking, mMIMO refers to a technologyfeature where BSs are equipped with antenna arrays comprised of a large number of antennaelements [54]. At higher frequency bands, due to their more challenging propagation conditions,the large number of antenna elements are primarily used for beamforming to extend coverage.At lower frequency bands, the focus of this survey, in addition to leverage beamforming gains,the large number of antenna elements is readily used to enable extensive spatial multiplexingand interference mitigation by spatial separation [55] [56]. In more detail, the large mMIMOantenna array can excite a plurality of channel sub-spaces to support multiple simultaneoustransmissions to —or receptions from— several UEs. In this way, the network capacity canpotentially linearly grow with the number of spatial streams multiplexed. Due to the channelhardening effect, mMIMO is also more robust against wireless channel fluctuations, whichsimplifies its operation [57].3GPP NR provides extensive support for mMIMO operation. Channels and signals, speciallythose used for control and synchronization, have been re-designed with respect to 3GPP LTEto natively use beamforming. The acquisition of CSI for the large number of antenna elementsin a mMIMO BS is now supported through i) new UE CSI reports estimated over channelstate information-reference signals (CSI-RS) in the downlink, or ii) channel measurements over sounding reference signalss (SRSs), exploiting channel reciprocity, in the uplink. Among others,3GPP NR is also providing new functionalities to support analog beam-forming as well as digitalprecoding [58].Although mMIMO was primarily designed with the objective of maximising achievable rates,studies have shown that mMIMO is also suited for maximising energy efficiency and minimisingenergy consumption. For example, the large number of antenna elements at mMIMO BSscan be used to increase the transmission rates, while keeping constant the transmit powerat the mMIMO BS. This results in shorter active transmission periods, and is possible sincethe large number of antenna elements facilitates achieving unprecedented spatial resolutions,simultaneously diminishing harmful communication effects, such as the channel noise or theinter-user interference [47] [59]. Thanks to its beamforming gain, mMIMO is also able tominimise the transmit power required to achieve a given UE QoE, e.g., average UE throughput,at a given distance, which also reduces the BS energy consumption.In Section V, we provide a detailed survey and analysis of the energy efficiency benefits andtrade-offs of this important 5G feature —mMIMO— in both single- and multi-cell setups. B. The Lean Carrier Design
In previous generations, signals for BS detection, broadcast of system information, and channelestimation were always-active and transmitted over the air, regardless of whether the BS wasserving UEs or not [60]. It is important to note that, while these always-active signals facilitateUE operations —as UEs always have signals to rely on—, they also • result in a large overhead in dense deployments, • introduce inter-cell interference to other cells, thus reducing the achievable throughput, • reduce the battery lifetime of the UE, and • worsen the energy efficiency,thus, becoming a burden to efficient network operation.In 3GPP NR, many of these procedures have been revisited, following a new lean carrierdesign [48]. In general, following the DTX concept [61], when the traffic load of a cell —orgroup of them— is low, or the mobility conditions of the UE allows, larger signaling cyclescan be selected. This makes the carrier signalling transmission more sparse —leaner—, thusreducing overhead, mitigating interference, and saving energy. To accommodate for such longer signalling cycles, cell search and association as well as CSI procedures at both the BS and UEside have been accordingly redesigned in the new specification.In Section VI, together with the use of ASMs, we survey and discuss how the lean carrierbenefits energy efficiency. C. Advanced Sleep Modes
Cellular networks are usually planned and deployed to meet certain requirements, where copingwith the traffic needs at the peak hours usually leads to an over-dimensioning of the network forthe less challenging traffic loads [62]. As the traffic pattern fluctuates over both time and space,underutilized BS resources can be dynamically turned off to save energy. The more networkcomponents that are shutdown and the longer the time that they are shutdown, the more energycan be saved [63].Importantly, the less required always-active signalling of a lean carrier allows for longer micro-sleeps in the presence of bursty traffic. This micro-sleeps can last for hundreds of microsecondsor event seconds, and when coupled with traffic shaping and efficient hardware at the BS ableto power up and down in fractions of a millisecond, they can be leveraged to allow deeper sleepmodes, further saving energy [64].Most of the improvements that a network can achieve by appropriate management, however,do not lay on the micro-sleep space, as these sleep periods are only opportunistic and generallyshort. To enable longer BS resource deactivation times, and achieve larger energy consump-tion reductions, more advanced mechanisms are required, able to leverage knowledge on UEdistributions and their requirements in terms of QoE.Such tailored resource management according to the UE demands at a macro-time scale willavoid the resource waste emanating from the over-dimensioning of the network to meet peakhour needs. Provided with such UE-related information, not a single BS, but the entire networkcan be configured —and reconfigured— to operate with the minimum set of resources necessaryto satisfy the active UEs’ QoE on a semi-dynamic basis [65] [66].In Section VI, together with the lean carrier design, we also survey and discuss how theadvanced network functionalities can operate in time, carrier, and antenna domains to enhanceenergy efficiency. D. Artificial Intelligence
The heterogeneous and stringent service requirements of 3GPP NR networks, together withtheir increasing complexity —a pinch of which has been depicted in previous sections— aremaking traditional approaches to network operation and optimization no longer adequate. Suchmethods use a significant level of expert knowledge and theoretical assumptions to characterizereal environments. Thus, they do not scale well, and cannot handle the complexity of realscenarios with their many parameters and imperfections as well as stochastic and non-linearprocesses. To bridge this gap, and provide 3GPP NR networks with the intelligence requiredto strike optimum operation points, equipment vendors and MNOs have started to equip theirproducts with ML-based functionalities [29] [50].Fed by network measurements, supervised and unsupervised learning tools [67], two differentbranches of ML, are being extensively used nowadays to model 5G network behaviour first,and subsequently, take educated decisions and/or make predictions on complex scenarios [68].This is particularly relevant to energy efficiency. As one can infer from the previous discussions,minimising 5G energy consumption is a large-scale network problem, which highly dependson complex BS and UE distributions, varying traffic demands and wireless channels as well ashidden network trade-offs. Thus, understanding and predicting UE behaviours and requirements,as well as their evolution in time and space, is critical to tailor the 5G network configuration—mMIMO, lean carrier, and advanced sleep modes—, and address UE specific communicationsneeds with the minimum possible energy consumption.Additionally, due to the dynamic nature of wireless networks, and the lack of network mea-surements data for all network procedures and on all the possible configurations they canadopt, reinforcement learning (RL) [69] is also being widely explored to optimise 5G networkperformance in general, and energy efficiency in particular. For example, shutting down networkelements is a combinatorial problem with a large number of variables. RL agents may be used tolet the network interact with the environment, and learn optimum resource (de)activation policiesto minimise the total network energy consumption. However, such learning may come at theexpense of both i) an undesirably long exploration phase, where 5G network performance maybe highly suboptimal, and ii) large computing powers and storage capabilities [70]. Moreover,continuously adapting an optimal policy, derived from and for a limited set of specific systemconfigurations, to variations of network settings is a challenge. (a) (b)Figure 5. (a) NG-RAN overall architecture [72]; (b) 3GPP options for the function split between gNB-DUs and gNB-CU [71]. In Section VII, we will review how ML is being used to tackle the energy efficiency problemin 3GPP NR networks, and discuss both its benefits and main problems to tackle.III. P
OWER C ONSUMPTION M ODELS
To assess the impact of the different technology features presented earlier on the energyefficiency of a 5G network, it is necessary to define models that provide a good estimation of theirenergy consumption. Importantly, such energy consumption models need to offer the right balancebetween accuracy and tractability, while embracing different network and BS architectures, toempower 5G system performance characterisation and optimisation.Fig. 5a shows the 3GPP NR RAN logical architecture. This architecture, denoted as nextgeneration radio access network (NG-RAN) architecture, consists of a set of next generationNodeBs (gNBs) —the 3GPP NR BSs in the 3GPP terminology— connected i) amongst themthrough the Xn interface and ii) to the 5G Core Network (5GC) through the next generation(NG) interface. To take advantage of virtualization technologies and provide more implementationflexibility, a gNB may also consist of a central unit (CU) and multiple distributed units (DUs),connected to each other through the F1 interface. The 3GPP has studied eight functional splitoptions between CU and DU (see Fig. 5b), and current RAN implementations are focusing onoption 2 [71].From an implementation perspective, each functional split corresponds to a distinct deploy-ment option, which needs an appropriate power consumption characterisation. Specifically, it is necessary to take into consideration both the power consumption of the sites where the DUs andCUs are deployed, as well as the transport network, also referred to as fronthaul, that connects theDUs and the CU. In addition, with the advent of Network Functions Virtualization (NFV), DUsand CUs functions can be implemented either through standard dedicated hardware or virtualnetwork functions (VNFs) in a network cloud. Therefore, the overall RAN power consumptionmodel may need to include the contribution of the cloud server, where the RAN VNFs aredeployed. Accordingly, we can generally express the aggregated RAN power consumption of a5G network as follows: P RAN = X i P BS i + X j P FH j + X k P VBBU k , (1)where P BS i , P FH j , and P VBBU k are the power consumption of the i -th gNB, the power consumptionof the j -th fronthaul, and the power consumption of the k -th virtualized baseband unit (BBU),respectively. Depending on the specific RAN architecture, distributed or centralised, some ofthese components may not be considered.In the following, we survey the most relevant power consumption models for both distributedand centralised RAN, while considering their most relevant characteristics. A. Power Consumption Model for Distributed RAN
In case of a fully distributed radio access network (DRAN), the network power consumptioncan be modelled by taking into account only the BS contribution.A widely used BS power consumption is that defined in [73], where the power consumption ofa non-mMIMO BS is computed as a function of the power consumption of all its active antennas in all its sectors, each one including a power amplifier (PA), an RF transceiver module, a BBUwith a transmitter and a receiver section, a DC-DC power supply, an active cooling system andan AC-DC unit for connection to the electrical power grid. Such model is formulated as follows:P BS = N TRX P out η PA (1 − σ feed ) + P RF + P BB (1 − σ DC ) (1 − σ MS ) (1 − σ cool ) , (2)where N TRX is the overall number of RF transceiver modules in a BS, P out is the transmit power,P RF and P BB are the RF transceiver module and the BBU power consumption, η PA is the PA Examples of active antennas are an antenna element, a dipole, or a macrocell column in a radome operated by a BBU. power efficiency, and σ DC , σ MS and σ cool are the power losses in the DC-DC power supply, mainssupply, and active cooling, respectively.It should be noted that the model in eq. (2) is widely represented in the literature by a simplifiedversion of it, which explicitly shows the linear relation between the BS power consumption, P BS ,and the transmit power, P out , as follows:P BS = N TRX · ( P + δ p P out ) , (3)where P and δ p are cell-type dependent parameters, which indicate the power consumption atthe minimum non-zero output power and the slope of the load-dependent power consumption,respectively. Note that this model is general, and accommodates to macro, micro and small cells.For example, the parameters of eq. (3) for different types of small cells are provided in [73].Importantly, in the last decade, in addition to appearance of the aforementioned smallercells [74], two other main solutions have emerged as key enablers to boost the mobile net-work capacity, i.e., carrier aggregation (CA) (see Section VI-C) and mMIMO (see Section V).Accordingly, a number of works have evolved the previous presented model to capture the impactof these technology features on the BS power consumption.
1) Carrier Aggregation Power Consumption Model:
CA is a 3GPP flagship feature primarilyintroduced to increase the cell throughput in 3GPP long term evolution advanced (LTE-A). Thefirst version of CA allowed to aggregate up to 5 component carriers (CCs) of up to 20 MHz,and currently, in 3GPP NR, it has been extended to support up to 16 CCs and 1 GHz ofbandwidth. The CA framework is flexible by design. It enables to use continuous or discontinuousintra-band CCs as well discontinuous inter-band CCs, which can be characterized by differentbandwidth or coverage. It is also a key technology to enable licensed Assisted Access (LAA)[75], heterogeneous network (HetNet) deployments [76] and dual connectivity [77]. For eachUE, a CC is defined as its primary cell (PCell) [78], which acts as the anchor CC, and is thusused for basic functionalities, including mobility support and radio link failure (RLF) monitoring.Additionally configured CCs are denoted as secondary cells (SCells), and they can be added,changed, or removed to optimise the BS performance.When modelling the power consumption of a system using CA, it is necessary to take intoaccount how the power consumption scales with the number of active CCs, Ncc. In this line, the work in [79] has presented the following power consumption model for CA:P BS = Ncc X j =1 (cid:16) P TX j + B j P CACP j (cid:17) + P CAiCP , where P TX j = P out j η PA , B j , and P CACP j are the effective transmit power used by CC j , the bandwidth ofCC j , and the variable circuit power consumption, which scales linearly with both the number ofactive CCs, Ncc, and their bandwidth, B j , respectively, while as in the general model, P CAiCP is theload independent circuit power consumption of the CA system, i.e., of the hardware componentsshared by the distinct CCs.Embracing the complexity of CA, the variables, P
CACP and P
CAiCP , may take different values,depending on the specific CA implementation. For instance, contiguous CA can be realized witha single fast Fourier transform (FFT) and a single RF transceiver module for all CCs, whilenon-contiguous CA usually requires multiple of them [78]. In the worst case, each CC wouldrequire a fully dedicated hardware, and thus the load independent circuit power consumption ofthe CA system, P
CAiCP , would scale linearly with the number of active CCs, Ncc [45].
2) mMIMO Power Consumption Model:
With regard to mMIMO, and similarly as for theCA case, the linear power consumption model in eq. (3) has also been extended to take intoconsideration the large number of antenna elements and the new architecture of a mMIMO BS.In this line, the work in [80] proposed the following power consumption model for mMIMO:P BS = P TX + N ATRX P ACP + P liCP , (4)where N ATRX is the number of RF mMIMO transceiver modules, which do not need to benecessarily equal and may be smaller than the number of antenna elements, P
ACP is the powerconsumption of the RF and digital processing needed to support each RF mMIMO transceivermodule, and P liCP is the load-independent circuit power consumption.Importantly, it should be noted that N
ATRX depends on the type of beamforming architectureimplemented at the BS [81]. Digital beamforming provides high flexibility, but requires a largenumber of RF mMIMO transceiver modules. In contrast, analog beamforming decreases thepower consumption by significantly reducing the number of RF mMIMO transceiver modulesat the cost of a lower spatial resolution capability of the beamforming and a larger latency toselect the proper beam weights/configuration. Hybrid beamforming combines the advantages ofthe two architectures. The linear mMIMO power consumption model in eq. (4) provides a simple description ofthe relation between the number of RF mMIMO transceivers and the power consumption ina mMIMO system. However, more advanced works have highlighted that it is of paramountimportance to also take into account the impact of multi-user scheduling. Specifically, the researchin [82] has described the steps to derive a more complete model for mMIMO BSs, whichaccounts for both downlink and uplink communications, under the assumption of zero forcing(ZF) processing . In more detail, the authors describe the mMIMO BS power consumption asthe sum of the effective transmit power, P TX = P out η PA , and the circuit power, P CP , as follows:P BS = P TX + P CP , (5)and then decomposed the latter term as:P CP = P FIX + P TRX + P CE + P C/D + P BH + P SP = P liCP + P ldCP , (6)where P FIX is the fixed power required for site-cooling, control signalling, and load-independentbackhaul and signal processors, P
TRX , P CE , P C/D , P BH and P SP are the power consumption ofthe transceiver modules, the channel estimation process, the channel coding and decoding units,the load-dependent backhaul, and the beamforming processing, respectively. For clarity, let usdenote by P liCP the sum of both the variable, P FIX , and the load-independent part of the variable,P
TRX , and by P ldCP all the rest, i.e. the load-dependent part of the circuit power consumption, P CP .Accordingly, the mMIMO BS power consumption model in eq. (6) can be expressed asfollows [83] [84]:P BS = K · P UE η PA + P liCP + C K + D M + D M · K + D M · K + A K · R UE , (7)where K is the number of simultaneously multiplexed UEs at the BS, P UE is the downlinkoutput power per UE (i.e., P out = K · P UE ), C is the part of the beamforming processing, P SP ,which scales linearly with K , D is the power consumed by the transceiver module attachedto each antenna, D is the part of the beamforming processing, P SP , which scales linearly withM · K, D is the sum of the contributions of the channel estimation process, P CE , and thebeamforming processing, P SP , which scale linearly with M · K , R UE is the UE throughput, and A is the aggregated power consumption per bit of information required by the coding/decodingoperations and by the load-dependent part of the backhaul. Table III describes typical values ofthe parameters in eq. (7). Models for other specific precoding and combining schemes are discussed in [56]. Table IIT
YPICAL VALUES FOR M
MIMO
POWER CONSUMPTION MODEL PARAMETERS [82] [83] [84].
Parameter Value Parameter Value η PA liCP
20 [W] C − [W] D D · − [W] D · − [W] A B. Power Consumption Model for Centralised RAN
Considering a more sophisticated RAN architecture, the work in [85] has studied the powerconsumption modelling of a centralized radio access network (CRAN) considering differentfunctional splits between the gNB-DUs and the gNB-CU. In terms of energy efficiency, central-isation enables three main benefits with respect to a decentralised architecture: stacking gain,pooling gain, and cooling gain [85]. The stacking gain refers to the capability of deploying lessprocessing units (in the central node) to serve the same amount of cells in a given area, whilethe pooling gain refers to the capability of using a limited amount of centralised resources tooperate a large amount of cells, by exploiting the load variations in the network. The coolinggain appears due to the reduced amount of energy required to cool the cell site and the moreadvanced cooling solutions that can be implemented at the central node [86].In this centralized architecture, the aggregated power consumption at the RAN can be com-puted as [85]: X i P BS i = X i P DU i (cid:18) − P co i + P NF i (cid:19) + P BBU , (8)where P DU i is the power consumption of the site where the i -th DU is deployed, which can becomputed as, e.g., eq. (7), P BBU is the power consumption of the BBU host where the CU islocated , P NF i is the fraction of power consumption that corresponds to the Network Functions(NFs) moved to the CU, and P co i is the fraction of power consumption that corresponds to thecooling related to the i -th DU. Importantly, note that the higher the number of NFs moved to the This model can be easily generalised to the case where a larger network with multiple CUs is considered. CU, the higher the values of both parameters, P co i and P NF i . Moreover, the power consumptionof the BBU host, P BBU , can be computed as:P
BBU = X i P DU i (cid:18) / G co
100 + P NF i / N BS (cid:24) N BS G st G po (cid:25) P add (cid:19) , (9)where G st , G po , G co , and P add are the stacking gain, the pooling gain, the cooling gain, and theadditional power consumption in the BBU needed to enable resource pooling, respectively.It is important to note that eq. (9) assumes that the CU is deployed on a dedicated hardwareplatform, which has a similar architecture to the one used for a single DU but with largercomputational capacity. In the case of a virtualized radio access network (VRAN), the goal isto move the processing to general purpose processors (GPPs), which are capable to providereal-time processing to maintain the timing in the RAN protocols, while equipped with efficientand elastic resources —CPU, memory, and networking— to perform intensive digital processing.This approach promises further improved energy savings, which depend on the specifically usedarchitecture [87]. One option is to use dedicated hardware (e.g., system on chip) for managingthe layer 1 functions on dedicated hardware, while higher-layer functions are implemented ina software-based architecture. This solution may enable, however, limited additional energysavings with respect to CRAN. To reduce the power consumption and increase flexibility, analternative option is to deploy only the most computationally-intensive functions on dedicatedhardware, such as turbo decoding and encryption/decryption. As an extreme option, in the fullGPP architecture, all the RAN functions are implemented in a virtual environment, which mayenable large power savings at the expense, however, of performance if the GPPs and infrastructurearound cannot cope with the workload.Considering the GPP architecture, the work in [88] proposes a power consumption model forvirtualized BBUs in a cloud node, which takes into account the impact of the cooling system,the workload dispatcher switch, and the GPPs as follows:P VBBU = P Vco + P Dis + X i P GPP i , where P Vco , P
Dis and P
GPP i are the power consumption contributions in the virtualized BBU dueto the cooling system, the switch, and the i -th GPP, respectively.To characterize the GPP power consumption, a linear power consumption model can be used[72] [89], i.e., P GPP = P GPP + ∆ GPP P GPPm ρ GPP , where P GPP and P GPPm are the power consumption of the GPP when it is in sleep mode and atmaximum load, respectively, and ∆ GPP and ρ GPP are the slope of the power consumption model,which is related to the specific GPP architecture, and the load of the GPP, respectively, wherethe latter depends on the GPP computational capacity and the computational resources requiredby the hosted VNFs.To characterise of the GPP computational load, the work in [89] has used an experimentalplatform to infer the relationship between the downlink throughput and the percentage of centralprocessing unit (CPU) usage at the BBU. Instead, the authors in [90] propose an analyticalcharacterisation of the GPP computational load, which jointly depends on the functional split,the number of used transceiver modules/antennas, the bandwidth, the rate and the number ofspatial MIMO-layers used at the virtualized BS. Importantly, the authors in [91] have furtheranalytically investigated the computational complexity of lower layer functions in 3GPP NR,and this research has shown that, today, due to the form factor and power consumption of GPP,dedicated hardware is the only feasible option for deploying full 3GPP NR capabilities.Complementing the above, the authors in [92] have also recently developed an empirical modelto describe the computational requirements of the RAN NFs, focusing on the physical (PHY)layer, which includes the most computationally expensive functionalities. Their results highlightthat NFs can be classified according to their complexity into three classes: • PHY NFs, whose computational complexity only depends on the system configuration, anddoes not change with time (e.g., FFT), • PHY NFs, whose computational complexity depends on both throughput requirements andchannel quality (e.g., encoding/decoding), and • higher layer NFs, whose computational complexity only depends on throughput require-ments (e.g., packet scheduling).Finally, to connect each DU to the CU, a fronthaul link, whose capacity fits the functional splitthroughput and latency requirements is needed. Therefore, the total power consumption of thefronthaul is a function of the transport network technology, its topology, the capacity requiredby each BS, and the number of BSs deployed in the RAN. For instance, when the fronthaulis based on optical dense wavelength division multiplexing with a ring architecture and opticalswitching, the power consumption of the network connecting the i -th DUs to the CU can be modelled as [85]: P FH i = (cid:24) C FS i R tr (cid:25) (cid:18) · P tr + P W N W (cid:19) , where C FS i is the transport capacity required by the functional split at the i -th gNB, P W isthe power consumption of a port used to interconnect access and metro rings in the transportnetwork, N W is the number of wavelengths per fiber, and R tr and P tr are the rate and the associatepower consumption of the transport network nodes, respectively.To provide a general view, and highlight the impact of the transport network on the overallsystem power consumption, we show in Fig. 6a the network power consumption with respect tothe BS deployment density for a classic DRAN and different CRAN architectural options, usingthe parameters indicated in Table III. Moreover, Fig. 6b shows the impact of each contributor,i.e., DUs, fronthaul nodes, and BBU, with respect to the aggregated CRAN power consumption.Fig. 6a highlights that only a CRAN with functional split option 6 (see Fig. 5b) has similar powerconsumption to the classic DRAN. CRAN architectures with lower functional splits provide largercentralisation gains, but lead to higher power consumptions. In addition, Fig. 6b shows that thecontribution from the transport network cannot be neglected in the overall power consumptioncharacterisation, particularly when some architectures are adopted. Specifically, while for splitoption 6, it only amounts for around 2 % of the network power consumption, in split option 7and 8, it contributes for around 30 % and 60 % of the network power consumption, respectively.To conclude this section, let us also highlight that the research and industry communitieshave spent notable efforts to characterize the RAN power consumption —whether distributed orcentralised— while considering the most relevant architectures. However, this is still an openresearch field. In particular, we argue that due to the large ecosystem of equipment vendors andtheir different implementation of solutions, there is a need for further data and experimentallydriven research to validate the proposed models and find their appropriate parameters on real3GPP NR equipment. In addition, further research is also needed to characterise CA and mMIMOBS power consumption in the presence of multiple configurations and more complex features,such as coordinated multi-point (CoMP), and the characterisation of the power consumption ofsites where multiple radio access technologys (RATs) co-exist. (a) (b)Figure 6. Network power consumption (a) and its breakdown (b) as a function of the BS deployment density for classic DRANand different CRAN architectural options. Table IIICRAN POWER CONSUMPTION MODEL PARAMETERS [85] [93].
CRAN split 8 Value CRAN split 7 Value CRAN split 6 ValueP co
10 P co
10 P co NF
15 P NF NF po po po st st st add add add co co co FS FS
882 Gbps C FS
10 GbpsR tr
100 Gbps R tr
100 Gbps R tr
10 GbpsN W
40 N W
40 N W W W W tr tr tr IV. E
NERGY E FFICIENCY M ETRICS
To evaluate the impact of energy efficiency mechanisms on energy savings, energy efficiencymetrics are as important as the used power consumption models. These metrics must be com- prehensive, reliable and widely accepted to allow comparisons. In addition, they have to captureboth the energy consumed by the system under investigation as well as the performance measuredat network level (such as coverage, capacity, and delay).To achieve these goals, the European Telecommunications Standards Institute (ETSI) Envi-ronmental Engineering technical committee, the ITU-T Study Group 5 and the 3GPP TechnicalSpecification Group RAN have specified metrics to assess mobile network energy efficiencyunder different operating conditions. The contribution of these standard development organi-zations (SDOs) has mainly focused on global system performance, while considering differentdemographic areas, load scenarios, and radio access technologies. In contrast, the academicresearch community has contributed to this effort by proposing energy efficiency link-specificmetrics, which enable a more tailored energy efficiency network optimisation [94].In the following, we will first overview the contributions from different SDOs, consideringboth interference- and noise-limited scenarios, and then touch on energy-delay related metrics.Subsequently, we also describe alternative metrics proposed by the academic research communityto drive energy efficiency optimisation problems, indicating the merits and the drawbacks of eachone of them. A. Energy Efficiency Metrics for Interference-Limited Networks
As one of the main targets of 5G networks to enhance energy efficiency is to adapt the systemcapacity —and the associated power consumption— to the network load, load-aware metrics arekey for the next generation of green communication networks. In this context, ETSI has definedthe mobile network data energy efficiency metric, EE DV [ bit/J ] [95], which is the ratio between thedata volume, DV, delivered in the network and the network energy consumption, EC, observedduring the time period required to deliver such data, i.e.,EE DV = DVEC , (10)where EC should be computed by integrating eq. (1) over an observation period that includesdistinct load levels . Note that the data volume, DV, includes both downlink and uplink traffic forboth circuit switched services and packet switched services, and that this metric can be used tocharacterise a single or multiple BSs, operating in urban and/or dense-urban areas, in which the the 3GPP recommends that the performance should be evaluated considering at least 3 load levels [96]. network experiences highly variable loads, i.e. interference-limited scenarios. However, it is notappropriate for scenarios where the traffic load is low. Specifically, since the energy efficiencymetric, EE DV , is not weighted according to the global reference, a small energy saving in thelow energy consumption region in a low load scenario may comparatively lead to apparentlylarge energy efficiency gains. while a large energy saving in the high energy consumption regionmay comparatively result into limited energy efficiency gains.To characterise the energy efficiency, while considering distinct deployment scenarios, e.g.,dense-urban, urban, sub-urban, rural, or deep rural areas, ETSI has extended the previous metricto the total energy efficiency metric, EE Total [ bit/J ] [95], which can be defined as the weightedsum of the energy efficiency in each deployment scenario, i.e.,EE Total = P m PoPP m · EE DV m P m PoPP m , where m is the index of the m -th scenario, EE DV m is the energy efficiency of the m -th scenario,and PoPP m is the weight (percentage) representing the typicality of the m -th scenario in thenetwork under test.To jointly consider deployment scenarios and traffic loads, the 3GPP has further complementedthe work from ETSI, introducing the network energy efficiency metric, EE global [ bit/J ] [96], whichcan be defined as the sum of the energy efficiencies in multiple deployment scenarios and underdifferent traffic loads, i.e.,EE global = X m b m · EE DS m = X m X l b m · a l · EE DV m,l , where EE DV l is the energy efficiency of the network for an observed deployment scenario, m , andtraffic load level, l , while b m and a l are the weights for the corresponding deployment scenario, m , and traffic load level, l , respectively. The 3GPP recommends to compute b m considering theproportion amongst the different deployment scenarios in terms of i) power consumption, ii) traffic load, or iii) connection density [97]. With respect to a l , and giving an example, if 10 % ,30 % , and 50 % traffic load levels are investigated, the corresponding weights based on a dailytraffic model could be 6/24, 10/24 and 8/24, respectively [97].Although the above presented metrics provide a useful indication on how the energy con-sumption scales with the increase of the data rate requirements, it does not give any informationabout actual economic costs. To address this challenge, the work in [98] has introduced the economical energy efficiency metric, E [ bit/J ] , which is defined as the ratio of effective systemthroughput to the associated energy consumption, weighted by a cost coefficient, i.e.,E = P k ∈K α k R k P n ∈N P T n + P n C n , where K and N are the sets of UEs and serving BSs, respectively, R k is the effective throughputperceived by the k -th UE, α k is the priority weight related to the k -th UE, P n and P T n are the static and the load-dependent power consumption, respectively, and C n is the costcoefficient of the n -th BS. Note that the cost coefficient, C n , is calculated as the ratio of thecorresponding device cost to a predefined benchmark cost, such that the metric, E , has thesame unit as the energy efficiency. In scenarios where multiple access, fronthaul, computing, andother mechanisms coexist to provide efficient services to the end-users, this economical energyefficiency metric, E , is able to discriminate amongst solutions, not only in terms of providednetwork throughput and energy consumption, but also in terms of additional (deployment andoperational) costs. This allows the analysis of advanced network deployment and resourcemanagement schemes. B. Energy Efficiency Metrics for Noise-Limited Networks
To complement the energy efficiency metrics presented in the previous section and deal withscenarios with sustained low data volumes, in particular in rural or in deep rural areas, ETSI alsointroduced the mobile network coverage energy efficiency metric, EE
MN,CoA [ m /J ] [95], whichis the ratio between the area covered by the network, CoA, and the network energy consumptionobserved during one year, EC, i.e., EE CoA = CoAEC . (11)In addition to the above, the 3GPP has extended the coverage energy efficiency metric ineq. (11) to consider a global metric, the mobile network data energy efficiency metric, EE DV [72],which, similarly as presented earlier, spans over distinct deployment scenarios, i.e.,EE global, CoA = X i C i · EE CoA,i , where C i is the weight for each of the corresponding deployment scenarios, which is computedby taking into account the area covered per deployment scenario and the relevance of the scenarioitself to the total power consumption, e.g., the percentage of power consumption in dense urban areas to that in rural areas. The energy computation in the actual area under coverage —and howit is affected by a network feature—, however, can be complex to estimate, and it may requirethe collection of a large number of UE measurement reports. In this context, ETSI has defined acoverage quality factor, CoA Q [%] [95], to estimate the quality of the coverage and measure theamount of connection failures due to coverage issues, load congestion or significant interferenceeffects, i.e., CoA Q = (1 − FR RRC )(1 − FR RABS )(1 − FR RABR ) , where FR RRC , FR
RABS , and FR
RABR are the radio resource control setup failure ratio, the radioaccess bearer setup failure ratio, and radio access bearer release failure ratio, respectively.
C. Delay-aware Energy Efficiency Metrics
With respect to other important performance metrics, such as latency, which is also a keyrequirement of 5G systems due to the emergence of URLLC, ETSI has introduced the latencymetric, EE L [ s − /J ] [95], to characterize 5G energy performance when these use cases arepredominant. This metric is the inverse of the product between the end-to-end user-plane latency,T e2e , and the energy consumption, EC, in the observation time period, i.e.,EE L = 1 T e2e · EC , where T e2e is the overall end-to-end latency between the transmitter side and the receiver side,including the delays due to, e.g. packet queuing, scheduling and the hybrid automatic repeatrequest (HARQ) process. D. Link-aware Energy Efficiency Metrics
As mentioned earlier, to aid a finer grain performance optimization, the research communityhas extended the network energy efficiency metric, EE DV , presented in eq. (10) by taking intoconsideration the different requirements of its distinct links. In fact, since the energy efficiencymetric, EE DV , can be seen as the aggregated sum of the energy efficiencies of each networklink, its optimisation through radio resource management may tend to favour the links that canprovide the largest throughput, which may limit the QoS of e.g., cell edge UEs, and thus thenetwork fairness. To address this issue, and denoting by L the set of links in the network, the weighted sum ofthe energy efficiencies (WSEE) metric, WSEE [ bit/J ] [94], is defined as the weighted mean ofthe different energy efficiencies measured at each link i ∈ L , i.e.,WSEE = X i ∈L w i · EE DV,i , (12)where EE DV,i is the link energy efficiency, defined as in eq. (10), and w i is the weight of i -thlink. Importantly, note that the use of different weights for different links enables assigning themdifferent priorities during the resource allocation to increase the system fairness.To enable an even more fair resource allocation, researchers have also proposed the weightedproduct of the energy efficiencies (WPEE) metric, WPEE [ bit/J ] [94], which is defined as theexponentially weighted product of the different energy efficiencies measured at each link i ∈ L ,i.e., WPEE = Y i ∈L ( EE DV,i ) w i . In particular, the WPEE metric maximisation ensures that no link experiences a zero throughput,and has been shown to converge to the Nash Bargaining solution [99]. Following this approach,however, it is not possible to improve the energy efficiency of the i -th link, EE DV,i , withoutdecreasing the energy efficiency of other link, e.g. the j -th link, EE DV,j .To achieve a better trade-off among overall system performance and fairness, a max-min fairresource allocation policy can be adopted. An energy efficiency resource allocation scheme thatis max-min fair can be designed by maximising the weighted minimum of the energy efficiencies(WMEE) metric, WMEE [ bit/J ] , i.e.,WMEE = min { i ∈L} ( w i · EE DV,i ) . Note that when the WMEE metric is optimised, the resource allocation achieves the same product, w i · EE DV,i , for all links i ∈ L . Thus, if the weights are set equal, this max-min fair resourceallocation will provide the same energy efficiency for each link.To facilitate the reader’s understanding, Fig. 7 provides a qualitative comparison of theperformance achieved by a network with two links when optimising its energy efficiency usingthe WMEE, the WSEE and the WPEE metrics with respect to the Pareto boundaries. Each of thethree approaches leads to a solution that belong to the energy efficient Pareto region. However,the WSEE and the WPEE metrics allocate more resources to the link 1, which is characterizedby a better energy efficiency in order to get closer to the global optimum. In contrast, the WMEE Figure 7. Operating regions of the energy efficient optimisation based on the NEE, the WMEE, the WPEE, and the WSEEmetrics [94], where NEE = P i ∈L EE DV,i . metric shares the resources between the two links such that they are characterized by the sameenergy efficiency.To conclude this section, Table IV provides a summary of the most relevant energy efficiencymetrics for 5G system optimisation, where it is important highlighting that these metrics havebeen mainly designed for assessing 4G networks, focusing on enhanced mobile broadband(eMBB) services, and considering only data rate, latency, and coverage requirement. In contrast,with respect to other 5G use cases, e.g. massive machine type communication (mMTC) andURLLC, there is a lack of specific and well understood metrics, and further research is needed inthis area. mMTC applications call for energy efficiency metrics that takes into account the numberof connection handled by the network in combination with the area covered by the network,e.g., an extension of the mobile network coverage energy efficiency metric, EE MN,CoA . Similarly,URLLC applications require a metric able to capture the system reliability in combination withthe end-to-end delay, e.g., an extension of the latency metric, EE L .V. T HEORETICAL U NDERSTANDING OF E NERGY E FFICIENCY : M
ASSIVE
MIMO3GPP NR networks are targeted at a 100 times higher energy efficiency with respect to previousgenerations (see Fig. 2), and mMIMO has been identified as a key technology to reach suchtarget. As discussed in previous sections, by leveraging its extensive beamforming and spatial Table IVS
UMMARY OF EE METRICS . Metric Unit Calculation KPI Pros ConsEE DV [95] [ bit/J ] DVEC Load-aware It is a simple metric Nothing to reportEE
Total [95] [ bit/J ] P m PoPP m · EE DV,m P m PoPP m Load-aware Appropriate for overall It does not allow to tunenetwork assessment the EE of each linkEE global [96] [ bit/J ] P m P l b m · a l · EE DV l Load-aware It captures multiple load scenarios It does not allow to tunethe EE of each linkEE
CoA [95] [ m /J ] CoAEC Coverage-aware It can compare performance The effective coveragein cells with different size is not well-definedEE global, CoA [72] [ m /J ] P i C i · EE CoA,i
Coverage-aware It can capture multiple The effective coveragedeployments is not well-definedPowerSubscriber [100] [W/UE] P k ≤ NUE P k NUE QoS-aware It enable evolutionary comparison Need to model systemsamongst systems with a large UE numberEE L [95] [ s − /J ] e2e EC Latency-aware It is a simple metric Nothing to reportE [98] [ bit/J ] P k ∈K α k R k P n ∈N P Tn + P n C n Load & cost-aware It can capture the cost The cost isof different solutions not well-definedWSEE [94] [ bit/J ] P i ∈L w i · EE DV,i
Link & Load-aware It captures link priorities It cannot prevent linkswith zero throughputWPEE [94] [ bit/J ] Q i ∈L ( EE DV,i ) w i Link & Load-aware It prevent links It cannot achievewith zero throughput max-min fairnessWMEE [94] [ bit/J ] min { i ∈L} (cid:16) w i · EEDV,i (cid:17)
Load-aware It leads to max-min fairness Nothing to report multiplexing capabilities, mMIMO can significantly reduce the transmit power required at the BSto achieve a targeted capacity, given a frequency band of operation and coverage area. However,running such larger number of antennas at the BS, together with the more signal processingrequired to handle the larger capacity in a mMIMO cell, also increases the energy consumptionof the BS, and in turn, that of the network.Multiple studies have set out to fundamentally understand the challenges of mMIMO networksin general, and the above presented energy efficiency trade-off in particular. Mainly concentratingon providing insights for network design, a large body of research has focused on derivingtheoretical bounds on the capacity and the energy consumption of mMIMO systems, as wellas the interplay between different network parameters. More practical research, on the otherhand, has tackled mMIMO-based network design, BS deployment, as well as radio resourcemanagement and optimisation problems, while taking into account more realistic QoS constraints. One of the main challenges faced in these fundamental studies is around the accuracy andthe tractability of the models used to characterise performance at the network-level. mMIMOcapacity and rates, for instance, significantly degrade in the presence of channel correlation and/orpilot contamination, which is a function, among others, of the BS and the UE distributions, thescenario topology and the wireless channel, as well as independent BS scheduling decisions andthe resulting interference. However, this is hard to capture in tractable models. Same issues alsorevolve around the accuracy and complexity of the power consumption models, as discussed inSection III.In light of these tractability issues, most theoretical findings on energy efficiency relatedto mMIMO networks are limited to single-cell scenarios, where model simplifications can beeasily justified. In contrast, the estimation of energy efficiency metrics and trade-offs in multi-cellscenarios tends to be addressed through computer aided numerical evaluations.In the following, we provide an overview of fundamental studies addressing the problem ofenergy efficiency in the context of mMIMO systems, in both single- and multi-cell scenarios.We focus on the relevant findings with regard to the energy efficiency bounds and its trade-off with respect to the spectral efficiency, and discuss the assumptions taken to extract suchknowledge. We also summarize useful insights for an energy efficient mMIMO network designand deployment.
A. Single-cell scenario
In this section, we focus —and survey— energy efficiency bounds and trade-offs whenconsidering a single-cell mMIMO scenario, of which there is a good understating in the literature.Explicit closed-form expressions are available, and the impact of different mMIMO parametersinto energy efficiency and power consumption has been carefully analyzed.
1) Bounds:
To understand the energy efficiency scaling law with respect to the systembandwidth, the work in [101] considered a single-cell MIMO case —not a mMIMO one—with a single UE, where both the transmitter and the receiver were equipped with M antennas.The channels were assumed to be deterministic, and as consequence, perfect CSI was availableat both the transmitter and the receiver. The adopted BS power consumption model included • the BS transmit power, P TX , • a term which corresponds to the power consumption of processing M parallel signals atthe transmitter and the receiver, and • a term which relates to the power consumption of encoding and decoding the correspondinginformation.The results of such analysis showed that the energy efficiency of this single-cell MIMO ismaximized for a non-zero finite ratio of the BS transmit power, P TX , to the bandwidth, B , whichtypically corresponds to a low signal to noise ratio (SNR). Importantly, the optimal mentionedratio depends on the propagation characteristics and the transceiver hardware, but not on thepower consumption of encoding and decoding information. The latter only affects the quantitativemaximum value of the energy efficiency, but not the optimal operating point in terms of the BStransmit power, P TX , and the bandwidth, B . Given such optimal ratio, any bandwidth, B , canthus be selected to achieve any data rate. In other words, there is no trade-off between theenergy efficiency and the data rate when the BS transmit power, P TX , and the bandwidth, B ,are not constrained by external factors. This is an important insight, which may also applyto the mMIMO case. However, it is important to note that both the BS transmit power, P TX ,and the bandwidth, B , are scarce resources, which cannot be infinitely abused, and usuallymore complex hardware is generally needed to handle more bandwidth. Turning the attentionto mMIMO systems, where data rates can be scaled up by the usage of more antennas, thepioneering work in [102]–[104] provided a first analysis of the energy efficiency in a single-cell scenario, mostly based on the assumption of perfect CSI being available at the BS. Theresearch in [102] showed that the performance of a mMIMO system with M antennas at the BSand a BS transmit power, P TX /M , is equal to the performance of a single input single output(SISO) system with a BS transmit power, P TX , without any intra-cell interference. This resultindicated that, by using a large number, M , of BS antennas the BS transmit power, P TX , can beproportionally scaled down by a factor, /M . This work also suggested that the spectral efficiencyin a mMIMO system can be increased by a factor, K , when serving K UEs in the same time-frequency resource. The findings in [103], [104] also resonated with these conclusions, reportingthat a power reduction proportional to /M can be achieved in time division duplexing (TDD)systems , while maintaining non-zero rates, as the number, M , of antennas grows to infinity. Inthis way, the energy efficiency of the system can be monotonically increased with the number, M , of antennas, without any trade-off. This was a promising result, but it is important to mentionthat it was obtained with significant assumptions in the BS power consumption model. The power reduction is proportional to / √ M in the case of imperfect CSI. Soon after, further developments in this area of research showed that the energy efficiencybounds in mMIMO systems are hidden behind more complex BS power consumption models,and that inaccuracies and/or oversimplifications, not reflecting the essence of mMIMO hardwareimplementations, can lead to misleading practical insights. For example, when not consideringthe circuit power consumption related to running a larger number of antennas, as it was the casein [102]–[104], one can be let to believe that an unbounded energy efficiency can be achieved byadding more and more antennas. Deploying more and more antennas, however, requires additionalcircuitry, incurring a larger power consumption, which needs to be accurately captured by theanalysis to draw the appropriate conclusions. In this line, the pioneering work in [82] considereda more sophisticated —and realistic— mMIMO BS power consumption model, and in turn, foundsignificantly different conclusions to the previously stated ones. Since this paper has become alandmark on the subject, let us explore it further in the following, and analyse the theoreticalbounds derived.Using the more detailed BS power consumption model developed by the authors in [82], andalready presented in eq. (6), the total energy efficiency of the uplink (UL) and downlink (DL)in a single-cell mMIMO network was analyzed in such work with respect to • the total transmit power, P totTX , accounting for the downlink and the uplink transmit powers, • the number, K , of simultaneously multiplexed UEs, and • the number, M , of antennas, where M ≥ K + 1 .Remarkably, closed-form expressions for the energy efficiency optimal operating point withrespect to each of these parameters, when the other are fixed, were provided for the casewhere both ZF processing and perfect CSI knowledge are considered. In particular, the authorsconsidered a TDD system, in which the pilot signaling occupies τ (ul) K and τ (dl) K symbolsin the uplink and downlink, respectively, where the inequality, τ (dl) , τ (ul) ≥ , must be true toenable orthogonal pilot sequences among UEs, and with this model, they provided a formulationof the gross downlink rate, expressed in bit per second, as follows: ¯ R = B log(1 + ρ ( M − K )) , (13)where ρ is a design parameter proportional to the received SNR. Following such formulation,the total transmit power, P totTX , required to serve each UE with a gross rate, ¯ R , was shown to be: P totTX = Bσ ρ S x η K, (14) where σ is the channel noise power, S x is a term that accounts for the UE distribution andpropagation environment, and η is a term that accounts for the efficiency of the PAs.From the above formulation, it is important to restate that the value of the design parameter, ρ , is proportional to the UE SNR, which in turn, is directly proportional to the total transmitpower, P totTX , when considering ZF processing. Thus, finding the optimal total transmit power, P tot ∗ TX , which maximizes the joint UL and DL energy efficiency involves deriving the optimaldesign parameter, ρ ∗ . After some manipulations, in [82], the authors showed that the optimaldesign parameter, ρ ∗ , is lower bounded by: ρ ∗ ≥ η ( C ′ + M D ′ ) Bσ S x − ln (cid:16) η ( M − K )( C′ + M D′ ) Bσ S x − (cid:17) M − K ln (cid:16) η ( M − K )( C ′ + M D ′ ) Bσ S x − (cid:17) − , (15)with C ′ = P i =1 C i K i + P liCP K and D ′ = P i =0 D i K i K , (16)where the terms C , P liCP , and D i,i ∈{ , , } are BS power consumption model coefficients, definedas in Section III-A, whereas C and C are UE power consumption model coefficients. Morespecifically, C is the power required by all the circuit components of each single-antenna UE,and C is the term that accounts for the linear processing at the UE. The bound in eq. (15) showsthat the optimal total transmit power, P tot ∗ TX , increases with the load-independent circuit powerconsumption, P liCP , the power required by all the circuit components of each single-antenna UE, C , and the power consumed by the transceiver module attached to each antenna, D . In fact,if the circuit power consumption, P CP , is large, then higher total transmit powers, P totTX , —andin turn, higher UE rates— can be afforded in the system, since the total transmit power, P totTX ,has a small impact on the overall power consumption. This indicates that the optimal strategy toimprove the energy efficiency is to increase the total transmit power, P totTX , with the number, M ,of antennas, not in an arbitrary manner, but while considering the circuit power consumption, P CP .These conclusions are in stark contrast with those in [102]–[104], which as depicted earlier,concluded instead that the BS transmit power, P TX , —the downlink part of the total transmitpower, P totTX — can be decreased with the increase of the number, M , of antennas, while remainingenergy efficient, thus implying a linear increase of the energy efficiency with the number, M , ofantennas. This optimistic finding only applies, however, in the idealistic case where the circuitpower consumption, P CP , plays no —or a negligible role— in the overall power consumption. Importantly, the analysis in [82] showed that, for a moderate and/or large number, M , ofantennas, the optimal value of the parameter, ρ ∗ , can be approximated by: ρ ∗ ≈ η D ′ Bσ S x M ln( M ) , (17)which shows a quasi-linear relation between the number, M , of antennas and the total transmitpower, P totTX . This more clearly indicates that, since the circuit power consumption, P CP , growswith the number, M , of antennas, the total transmit power, P totTX , can be increased to improvethe UE rates up to a given extend, before it becomes the limiting factor of the energy efficiency.Moreover, the study on the optimal number, K ∗ , of multiplexed UEs, provided in [82],highlights that, if the power consumption for the beamforming processing, P SP , and channelestimation, P CE , are negligible, then the optimal number, K ∗ , of multiplexed UEs can beapproximated as follows: K ∗ ≈ $ µ s U ( τ (ul) + τ (dl) ) µ − !' , (18)where U is the channel coherence time (in symbols), and µ = P liCP + Bσ S x η ρK C + ¯ β D , (19)where ¯ β is the ratio of the number, M , of antennas to the number, K , of multiplexed UEs, (i.e., ¯ β = M/K ),From eq. (18), we see that the optimal number, K ∗ , of multiplexed UEs decreases with thepower required by all the circuit components of each single-antenna UE, C , and the powerconsumed by the transceiver module attached to each antenna, D . On the contrary, the optimalnumber, K ∗ , of multiplexed UEs increases with the load-independent circuit power consumption, P liCP , the total transmit power P totTX (proportional to ρ ), the noise power, σ , and the parameter, S x . Note that this last term, S x , increases proportionally with the coverage radius of the cell,meaning that a larger number, K , of multiplexed UEs must be served when the coverage areaincreases in order to maximize the energy efficiency.In addition, the study on the optimal number, M ∗ , of antennas, provided in [82], also indicatesthat such number, M ∗ , is lower bounded by: M ∗ ≥ Bσ S x η D ′ ρ + C ′ D ′ + K − ρ ln( ρ ) + ln (cid:16) Bσ S x η D ′ ρ + C ′ D ′ + K − ρ (cid:17) − − ρ . (20) This bound indicates that the optimal number, M ∗ , of antennas increases with the load-independent circuit power consumption, P liCP , and the power required by all the circuit com-ponents of each single-antenna UE, C , whereas it decreases with the power consumed by thetransceiver module attached to each antenna, D .Importantly, when the design parameter, ρ , becomes large, the lower bound given in eq. (20)can be approximated as: M ∗ ≈ Bσ S x η D ′ ρ ln( ρ ) , (21)which shows that there is an almost linear scaling law of the optimal number, M ∗ , of antennaswith respect to the design parameter, ρ , in its high value regime, and thus with respect to thetotal transmit power, P totTX .It should also be noted the linear dependence of the optimal number, M ∗ , of antennas with theUE distribution and propagation environment, captured by the variable, S x , which implies thata larger optimal number, M ∗ , of antennas is needed, as the size of the coverage area increases,since the variable, S x , increases with the cell radius.For the sake of clarity, Fig. 8 shows the achievable energy efficiency for four differentvalues of the number, M , of antennas, when varying the number, K , of multiplexed UEs. Notethat the energy efficiency curves are characterized by a concave shape in all the consideredantenna configurations, and that the maximum achievable values are highlighted in red. Thisplot highlights that the optimal number, K ∗ , of multiplexed UEs increases sub-linearly with thenumber, M , of antennas.In a similar way, Fig. 9 shows the achievable energy efficiency for different values of thenumber, K , of multiplexed UEs, when varying the number, M , of antennas. The energy efficiencycurves have, in this case, a quasi-concave shape, and the maximum achievable values are alsohighlighted in red. These results confirm that, for a given value of the number, K , of multiplexedUEs, augmenting the number, M , of antennas increases the energy efficiency up to a maximumenergy efficiency value, where the UE rate gain due to the further increasing the number, M , ofantennas is not sufficient anymore to counterbalance the cost incurred by their associated powerpower consumption.Finally, it should be noted from Fig. 8 and Fig. 9 that deploying hundreds of antennas to servea large number of UEs is the optimal solution from an energy efficiency perspective, confirmingthe energy efficiency-enabler role of mMIMO. Figure 8. Energy efficiency with respect to the number, K , of multiplexed UEs for a given number, M , of antennas. Themaximum energy efficiency values are highlighted in red.
20 40 60 80 100 120 140 160 180 200 22005101520253035
Figure 9. Energy efficiency with respect to the number, M , of antennas for a given number, K of multiplexed UEs. Themaximum energy efficiency values are highlighted in red.
2) Trade-offs:
The spectral efficiency, defined as the system throughput per unit of band-width, has been historically adopted as the key optimisation metric to optimise mobile networkdeployments. As energy efficiency has become a new important key performance indicator (KPI)to the operation of 5G networks, the relation between these two metrics is thus of fundamentalimportance. How much spectral efficiency is trade to realize some energy efficiency gain?Early works in the literature, not accounting for the circuit power consumption, P CP , in the BS power consumption model, concluded that a 10x spectral efficiency improvement could beachieved for a given BS transmit power, P TX , when adopting hundreds of antennas at the BS [54].In this idealized scenario, since the energy efficiency grows proportionally to the number, M , ofantennas, the corresponding energy-spectral efficiency curve is pushed outwards, meaning thatboth metrics can be simultaneously maximized. However, this insight is not practical, since thecircuit power consumption, P CP , plays a major role today, as already discussed.Indeed, as shown earlier, the circuit power consumption, P CP , linearly grows with the number, M , of antennas, and when this number is large, it dominates the BS overall power consumption,implying that the energy efficiency cannot be enhanced, unless the efficiency of the BS hardwareis improved accordingly [82]. In this way, the circuit power consumption, P CP , breaks themonotonic relation between energy efficiency and spectral efficiency, and makes these twometrics not consistent and conflicting with each other. For this reason, their trade-off must becarefully analyzed, and proper network optimisation techniques need to be developed to strikethe right balance between these two metrics.Studies on such energy and spectral trade-off are often carried out based on optimisationproblems, aiming at maximizing the energy efficiency given a spectral efficiency requirement.However, in some more sophisticated approaches, the balance between the energy and the spectralefficiency is achieved by maximizing the resource efficiency (RE) metric [105], where this metricis defined as the weighted sum of the energy efficiency and spectral efficiency, and the weightsassigned to the two terms allows to give more or less importance to any of them. It is importantto highlight, however, that all such frameworks do not generally provide explicit equations forthe energy-spectral efficiency trade-off. Instead, they mostly built on the top of results obtainedfrom tailored optimisation algorithms aimed at solving such problem, which usually turns outto be intricate [106]–[109].In this line, the fundamental trade-off between the energy and the spectral efficiency hasbeen carefully studied in [109], when considering a single-cell mMIMO system with linearprecoding, i.e. ZF and Maximum Ratio Transmission (MRT), and transmit antenna selection. Inparticular, the BS transmit power, P TX , and the number, M a < M , of active antennas are jointlyconsidered as a resource to balance the energy and the spectral efficiency, and the adopted BSpower consumption model is equivalent to the one introduced in eq. (6). Note, however, thatthis study has not taken into account the impact of the number, K , of multiplexed UEs, whichmay have a significant influence to the trade-off. Importantly, this study has shown that different number, M a , of active antennas leads to different energy efficiency-spectral efficiency curves. Inmore details, for a given number, M a , of active antennas, the energy efficiency is a quasi-concavefunction with respect to the spectral efficiency, which confirms that the existence a clear trade-offbetween these two metrics. Their numerical results have also shown that in the low SNR region,which corresponds to a low spectral efficiency, MRT achieves a higher energy efficiency thanZF due to its lower complexity. In the high spectral efficiency regime, ZF outperforms MRTin terms of spectral efficiency, for a given energy efficiency, owing to its ability of cancelingintra-cell interference.As briefly mentioned when discussing the single-cell mMIMO energy efficiency bounds inthe previous section, most of the works in literature assume spatially uncorrelated channels andperfect knowledge of the CSI due to tractability reasons, as in the aforementioned study. However,in practical scenarios, spatial channel correlation is likely to appear, favouring more some spatialdirections than others due to the propagation environment. In this case, the CSI knowledge atthe transmitter becomes fundamental. However, its acquisition is challenging, especially whendealing with the large antenna arrays in the mMIMO case. In TDD systems, the acquisitionof downlink CSI can be performed via uplink training by taking advantage of the channelreciprocity. However, the CSI may still be inaccurate due to practical hardware limitations, suchas calibration errors in the transceivers [110], and in high mobility scenarios, it can quicklybecome outdated.Tackling this challenge, the authors in [107], [108] have explored this problem, and providedanalyses of the energy and the spectral efficiency trade-off, considering the knowledge of sta-tistical CSI, instead of an instantaneous one. This type of feedback has the advantage of beingstable during longer time periods and to be more easily obtainable by the BS, through long-termfeedback or covariance extrapolation at the expense of a lower network capacity.Particularly, the work in [108] has analysed the single-cell mMIMO downlink case, where oneBS with M antennas simultaneously transmits signals to K UEs, the channel spatial correlationis modelled using a jointly correlated Rayleigh fading model [111], statistical CSI is available atthe transmitter, and the considered BS power consumption model accounts for the BS transmitpower, P TX , and the static circuit power consumption, P CP . The energy-spectral efficiency trade-off has been investigated by maximizing the RE metric [105], which strikes for a energy-spectralefficiency balance.Fig. 10 shows the derived trade-off when considering different values of the BS transmit
45 50 55 60 65 70 753.23.43.63.844.24.44.64.8 10 Figure 10. Trade-off between energy efficiency and spectral efficiency for different values of the BS transmit power, P TX power, P TX . Note that the energy-spectral efficiency curve is characterized by a concave shape.In particular, the energy and the spectral efficiency can be jointly augmented by increasing theBS transmit power, P TX , until reaching, in this case, the optimal energy-spectral efficiency pointwith a BS transmit power, P TX = 35 dBm. After reaching such optimal point, which correspondsto the maximum achievable values of both energy and spectral efficiency, the BS transmit power, P TX , becomes the dominating consuming factor. Thus, increasing the BS transmit power, P TX ,allows to increase the spectral efficiency only at the expense of a reduced energy efficiency. Table VS
UMMARY OF SINGLE - CELL ENERGY EFFICIENCY BOUND AND TRADE - OFFS LITERATURE
Paper Type KPI Parameters[101] bound EE bandwidth, BS transmit power[82] bound EE transmitting antennas, multiplexed UEs, BS transmit power[109] trade-off EE-SE transmitting antennas, BS transmit power[108] trade-off RE BS transmit power[107] trade-off RE BS transmit power
Finally, before concluding this section, it should be noted that large research efforts have alsobeing spent on the understanding of both practical precoding and UE scheduling techniques, whileconsidering single-cell scenarios, to enhance the energy efficiency of mMIMO systems. The opti-mal precoding to achieve optimal energy efficiency under ideal and known channel conditions has been derived in [112]. Optimal energy efficient precoding schemes, while considering imperfectCSI at the transmitter, have also been studied in [113]. UE scheduling algorithms that takechannel orthogonality into account, avoiding to schedule two nearby UEs, have been proposedin [114]–[117]. Other more experimental approaches considered to increase energy efficiencyhave been modulation diversity [118], cognitive radios [119], and spatial modulation [120].
B. Multi-cell scenario
In this section, we focus —and survey— the latest most representative developments on theunderstating of the energy efficiency in large-scale multi-cell mMIMO networks, introducingthe most relevant tools used to carry the analyses out and differentiating among the uplink anddownlink case. In contrast to the single-cell mMIMO case, it should be noted that the multi-cell mMIMO one is not so well understood as of today, due to its much higher complexity.Large-scale networks are in general hard to model (see Fig. 11), due to their • complex topology, • evolving UE distributions, • dynamic traffic demands, • fluctuating wireless channels, • sophisticated protocols and algorithms, and • large number of parameters to tune.This poses a real challenge on their theoretical comprehension. As a result, there are no availableexplicit —and general— closed-form expressions that describe the energy efficiency bounds andtrade-offs of a multi-cell mMIMO network in a holistic manner. Instead, the research is scatteredand focused on particular aspects of the system, which together with some of the assumptionstaken, have helped to increase the tractability of the problem, and derive some initial insightson the energy efficiency problem. Further research on the topic is needed.
1) Available Mathematical Tools:
Given that the comprehension of the power consumptionof a mMIMO BS has reached some degree of maturity (see Sections III and V-A), one of themain difficulties to unlock a fundamental understanding of the energy efficiency in multi-cellmMIMO networks is the derivation of its capacity. Stochastic geometry [121], the state-of-the-arttool for the theoretical analysis of multi-cell networks, particularly small cell ones [122]–[125],has been recently started to be used to address this issue. Figure 11. Example of a large-scale muti-cell mMIMO network.
Embracing the randomness of today’s deployments, the work in [126] has analysed the uplinkmMIMO performance, considering a large-scale multi-cell network and the pilot contaminationproblem. This paper has presented a mMIMO network capacity scaling law as a function ofthe number of antennas and multiplexed UEs per BS, and investigated the performance gainsattainable through a practical fractional uplink pilot reuse, used to mitigate pilot contamination. Itis important to note, however, that significant assumptions were taken for tractability reasons. Thestudy assumed that there was a large number of UEs in each BS at all times, and thus that all BSsin the network were active and that all uplink pilots available were in use in all BSs. This makesit difficult to study off-peak hours were energy savings are more likely to be harvested. Moreover,a pure non-line-of-sight (NLoS) path loss model with Rayleigh multi-path was considered,which cannot capture the important effect of channel directionality on mMIMO performance, asdescribed in the previous section. Unfortunately, even though these important assumptions andsimplifications, the resulting expressions are still not tractable, requiring few folds of integrals,making difficult to infer relationships among system parameters. This is particularly true whencapacity expressions need to be coupled with complex BS power consumption models. as thosepresented in Section III. Using similar stochastic geometry tools, the authors in [127] studied the downlink mMIMOperformance, while considering a heterogeneous network comprised of mMIMO macrocells andsingle-antenna small cells, together with the effect of line-of-sight (LoS) and NLoS transmissions,pilot contamination, and cross-tier interference. However, similar as in the previous case, a fullyloaded network is considered with all uplink pilots at use, which significantly affects the pilotcontamination, and does not allow to analyse low and medium traffic load scenarios. In addition,important assumptions also revolve around the use of Rayleigh fading for the LoS transmissions.Some of these issues have been recently addressed by the research in [128], where pilot allocationschemes are considered with the objective of reducing the pilot contamination. However, theexpressions are still too complex to infer parameter relationships without a numerical evaluation.
2) Uplink mMIMO network deployment perspectives:
Aware of such complexities, the authorsin [129] proposed a tractable stochastic geometry-based analysis of the energy efficiency ofa TDD multi-cell mMIMO network subject to QoS constraints, while using a simplified butyet complete system model. In more detail, the authors targeted at maximizing the uplinkenergy efficiency of such multi-cell mMIMO network, while ensuring a minimum uplink spectralefficiency to the average UEs, and used this formulation to obtain some practical insights intothe problem. The adopted homogeneous Poison point process (HPPP)-based system model accounted for • the BS density, λ , • the number, M , of antennas per BS, • the number, K , of multiplexed UEs per transmission time interval (TTI) at each BS, • an idealised uplink fractional power control, • imperfect channel estimation through pilot contamination , • hardware impairments, modelled as a reduction of the signal power by a factor, − ǫ , • maximal ratio combining (MRC) received filters at the BS, • single-antenna UEs, and • an uplink pilot reuse scheme. It should be also noted that this general system model allows to compare mMIMO setups, with few BSs and many antennasper BS, to small cell ones, with many BSs and few antennas per BSs, thus providing guidance on the design of future greenwireless networks, from the uplink perspective. Rayleigh fading was still assumed, and thus the effect of spatial correlation was ignored in this analysis. Importantly, the complexity of calculating the average spectral efficiency of the network as thesum of the spectral efficiencies of all UEs through this framework was acknowledged, indicatingthe need for heavy numerical evaluations of integrals. To address this issue, and obtain explicitexpressions, the authors focused on the performance of the average UE instead, and derived thefollowing tractable —but yet tight— lower bound for its uplink signal to interference plus noiseratio (SINR):
SINR UL = M (1 − ǫ ) (cid:0) K + σ ζ (cid:1)(cid:0) β ( α − + σ ζ (cid:1) + Kα − (cid:0) σ ζ (cid:1) + Kβ (cid:0) α − + α − (cid:1) + M (1 − ǫ ) (cid:0) β ( α − + ǫ (cid:1) , (22) where α is the path loss exponent, β is the pilot reuse factor, ζ is the path loss compensationpower control coefficient, and σ is the noise power.Leveraging this expression, the authors formulated the uplink average UE spectral efficiencyand the resulting uplink area spectral efficiency (ASE), and with that, they derived the uplinkarea power consumption (APC) of the multi-cell mMIMO network, using a linear version of theBS power consumption model presented in (7), i.e., APC UL = λ (cid:18) − βK − S (cid:19) ζ ωη Γ( α + 1)( πλ ) ( α ) K + P liCP + C K + D M + D M K ! + A ·
ASE , (23)where S is coherent block length, which is related to the channel coherence time, U , in eq. (18), η is the PA power efficiency, Γ() is the Gamma function, and
ASE is the area spectral efficiency.With these formulations, the authors defined an optimisation problem to find the most uplinkenergy efficient network deployment, while provisioning the average UE with a minimum SINR.Through some mathematical manipulations detailed in [129], the authors derived • the uplink spectral efficiency feasibility, • the optimal uplink pilot reuse factor, β ∗ , • the optimal BS density, λ ∗ , • the optimal number, K ∗ , of multiplexed UEs per TTI at each BS, subject to a given ratio, MK , of the number, M , of antennas to the number, K , of multiplexed UEs, and • the optimal number, M ∗ , of antennas per BS, subject to a given number, K , of multiplexedUEs,and demonstrated that reducing the cell size is beneficial for energy efficiency, but that suchpositive effect —increasing the BS density— saturates when the circuit power consumptiondominates over the transmission power. Their results also showed that adding antennas to the BS to bring it to the mMIMO regime also enhances the energy efficiency . In more detail,their numerical examples on a typical scenario resulted in the maximum energy efficiency whenhaving 91 antennas and 10 UEs multiplexed per BS, which resembles a mMIMO —and not asmall cell— setup. It should be noted that the energy efficiency gains, in this case, mostly camefrom the intra-cell interference suppression provided by mMIMO, and by sharing the circuitpower costs among the multiple multiplexed UEs. Moreover, the analysis showed that a largepilot reuse factor can be used to protect the network against inter-cell interference, and that itcan be tailored to guarantee a certain average UE spectral efficiency.Using a different modelling tool than stochastic geometry, based on numerical evaluations,but also focusing on the uplink energy efficiency, the authors in [130] provided an analysis ofthe area energy efficiency (AEE) and the ASE trade-off for different system parameters, such asthe pilot reuse ration and number of antennas and multiplexed UEs per BS. Importantly, evenif this work does not focus on minimising power consumption, but maximising the AEE, itreaches similar general conclusions than those of [129]. The multi-cell mMIMO network alwaysperforms better in terms of ASE with an increasing number of antennas. However, adding moreantennas or multiplexed UEs might not achieve the optimal AEE. The reason is that the powerconsumed by the transceiver module attached to each antenna and the related signal detectionand processing have a non-negligible effect on the total BS power consumption. In more detail,their results have shown that, when considering smaller ASE targets, a lower number of antennas(see Fig. 12a) and multiplexed UEs (see Fig. 12b) at the BS suffices to achieve the optimal AEE.However, more antennas and multiplexed UEs are required to satisfy higher ASE requirements,which makes the network less energy efficient, resulting in a smaller AEE, as soon as thetransmit power dominates the circuit and processing power consumption. Finally, in this work,the authors have derived the pilot-to-data power ratio that maximizes the AEE, and studied howsuch parameters affects the AEE and the ASE trade-off (see Fig. 12c). The study has also shownthat in relevant scenarios, the optimal number of multiplexed UEs for maximizing the AEE ismuch smaller than half of the coherent clock length, as it is for maximizing the ASE [54]. These conclusion are inline with that presented in the analysis of the single-cell mMIMO case in the previous sections,indicating that the optimal strategy to improve the energy efficiency is to increase the total transmit power, P totTX , with thenumber, M , of antennas, not in an arbitrary manner, but while considering the circuit power consumption, P CP . (a) AEE and ASE trade-off w.r.t. thenumber of antennas, M . (b) AEE and ASE trade-off w.r.t. thenumber of multiplexed UEs, K . (c) AEE and ASE trade-off w.r.t.thepilot-to-data power ratio, β .Figure 12. Multi-cell AEE and ASE trade-offs
3) Downlink mMIMO network deployment perspectives:
Using a similar stochastic geometryapproach and BS power consumption model as in [129], the authors in [131] extended theprevious work to the downlink case, aiming at optimising the downlink energy efficiency of aTDD multi-cell mMIMO network, while providing a minimum spectral efficiency to the averageUE. It is important to note that perfect CSI was assumed in this case due to the complexityof modelling in the same framework both the channel estimation phase in the uplink and thedata transmission phase in the downlink . As a result, the effect of pilot contamination wasneglected, decreasing the accuracy of the model. ZF precoders at the BS and single-antennaUEs were considered. Importantly, it should be noted that embracing the same methodology asin [129], the authors also focused their analysis on the performance of the average UE, andderived the following tractable lower bound for its downlink SINR: SINR DL = (1 − ǫ )( M − K ) K ( α − + ǫ ( M − K ) + Γ( α/ πλ ) ωσ ρ , (24)where ρ now represents the downlink transmit power allocated by the BS to the average UE.With such expression, and using the corresponding downlink APC model, shown in thefollowing for the sake of clarity: APC DL = λ (cid:18) Kρη + P liCP + C K + D M + D M K (cid:19) + A ·
ASE , (25) Such downlink dependency on the uplink is the main reason why the theoretical downlink energy efficiency of mMIMOnetworks has been less rigorous studied. the authors derived the following optimal operation points by following similar steps as in theuplink counterpart work, i.e., • the optimal downlink transmit power per UE, ρ ∗• the optimal BS density, λ ∗ , • the optimal number, K ∗ , of multiplexed UEs per TTI at each BS, subject to a given ratio, MK , of the number, M , of antennas to the number, K , of multiplexed UEs, and • the optimal number, M ∗ , of antennas per BS, subject to a given number, K , of multiplexedUEs.The analysis of the obtained closed-form expressions showed that the same conclusions obtainedfrom the uplink analysis apply to the downlink one. The optimal energy efficiency is achieved bya mMIMO-like deployment, where in their particular example, the optimum number of antennasand multiplexed UEs per BS are 193 and 21, respectively. Note that the results for the downlinkcase in the section and those for the uplink one in the previous section differ due to the differentnetwork requirements and power consumption values.These uplink and downlink results advocate for mMIMO deployments and their dimensioningoptimisation, according to end-users’ QoS, as an important tool to increase energy efficiency in5G networks. Using an inadequate number of BS or antennas per BS to meet a given end-users’QoS can result in highlight suboptimal energy efficiency performances. In the following section,we provide more details on how antenna selection and advanced sleep modes can be also befurther leveraged to increase energy efficiency in a large-scale network.Finally, before concluding this section, it is important to highlight that, from an optimisationperspective, a large body of research exists on the development of power control algorithms forenergy efficiency maximization in multi-cell mMIMO networks. In [94], systematic approachesto solve energy efficiency maximization problems are extensively discussed. In this regard, theframework presented in [132] has provided network- and UE-centric downlink power controlalgorithms, where minimum rate constraints are imposed and the SINR takes a general form,able to deal with complex mMIMO systems/configurations. Centralised algorithms are alsodeveloped, which are guaranteed to converge, with affordable computational complexity, to aKarush–Kuhn–Tucker point of the considered non-convex optimisation problem. Building onsuch framework, the work in [133] has proposed a framework to compute suboptimal powercontrol strategies with even more affordable complexity. This is achieved by jointly using fractional programming and sequential optimisation. Numerical evidence has shown that suchsequential fractional programming framework achieves global optimality in several practicalcommunication scenarios.VI. N ETWORK A DAPTATION TO Q O S R
EQUIREMENTS
As motivated previously, future wireless communication systems require efficient hardware andmechanisms that enable the adaption of the network functionalities and parameters to the loadvariations in an on-line manner in order to avoid excessive energy consumption, while ensuringend-users’ QoS. These enabling technologies can be classified, for example, by observing thetime-scale and the domain in which they operate, e.g., time, frequency, or spatial (antenna)domains. Time-domain mechanisms refer to those solutions that operate in the time scale ofthe 3GPP NR frame, and take advantage of its lean carrier design, to rapidly (de)activate thehardware components of the BS according to the absence or not of traffic. Frequency domainmechanisms include solutions that adjust the status of the cell —or the number of CCs— tomatch the cell capacity with the slow variations of the traffic across the day. Finally, consideringthe abilities of the new energy efficiency enabler in 3GPP NR, i.e., mMIMO, spatial domainsolutions can be used to adapt the number of active antenna elements and channels as well asthe associated hardware to the cell load.
A. Time-domain energy saving solutions
In this section, we focus on the energy savings that can be harvested by considering theshort-term traffic load variations in a cell, which depend on several factors, such as the numberof active UEs, the traffic types, and the interference. To realise such savings, fast adaptationmechanisms, which operate in the 3GPP NR frame time scale, are thus required to dynamicallymeet the momentary rate requirement with the minimum energy consumption.In 3GPP LTE, cell DTX can be used for deactivating BS hardware components, such as thePA, when transmissions are absent in a given frame [61]. The benefits of cell DTX are, however,limited by the control signalling required by 3GPP LTE to drive UE cell camping procedures,even in the absence of traffic. In more detail, the 3GPP LTE frame lasts 10 ms, and it is composedby 10 sub-frames, each one including 14 orthogonal frequency division multiplexing (OFDM)symbols with a duration of 71.4 µ s. In the unicast mode, cell-specific reference symbols (CRSs)are transmitted in every sub-frame, primary synchronisation channels (PSSs) and secondary Figure 13. LTE frame for (left) unicast cell; (middle) Rel. 14 MBSFN-dedicated cell; (right) Rel. 14 FeMBMS/unicast-mixedcell. synchronisation channels (SSSs) every 5 ms, and broadcast channels (BCHs) are repeated inevery first sub-frame of a frame (see the left side of Fig. 13). Therefore, the BS is only ableto sleep for a few of the OFDM symbols, and needs to wake-up one OFDM symbol before thetransmission of any control signal. In this case, cell DTX can achieve at most a 33 % sleep ratioat zero load [64].To allow longer sleep periods in 3GPP LTE, the use of the MBSFN frame was proposed,which is a feature introduced to enable mobile television broadcasting, characterized by theneed of less frequent signalling. To standardise such feature, the 3GPP evaluated in [134] boththe sleep ratio (at symbol level and sub-frame level) and the sleep duration, when using • Rel. 14 FeMBMS/unicast-mixed cells, in which two out of ten sub-frames, i.e. sub-frames0 and 5, contain unicast control signaling (see the central plot in Fig. 13), and • LTE Rel. 14 MBSFN-dedicated cells, where only one non-MBSFN sub-frame with standardunicast signaling is transmitted every 40 ms (see the right side of Fig. 13).The results from the 3GPP studies showed that, in FeMBMS/unicast-mixed mode, a BS canstay in energy saving mode for up to 4 ms, which leads to an 80 % sleep ratio. Importantly, inMBSFN-dedicated mode, a BS can sleep even further up to 39 ms, which results into a 93.75 % sleep ratio.As discussed in Section II-B, in contrast to 3GPP LTE, 3GPP NR is characterized by auser/data-specific signalling instead of a cell-specific one —the lean carrier (see Fig. 14). Specif-ically, • The CRS is not used anymore in 3GPP NR, and a synchronization signal (SS) burst set, Figure 14. NR frame structure example [49]. including one or multiple SS block(s) —each one of them in turn comprised of PSS, SSS,and BCH— is transmitted to support UE cell (re)selection and handover procedures with alarger periodicity, i.e., 5, 10, 20, 40, 80, and 160 ms [32]. • Since the CRS is not used, CSI acquisition procedures has also been redesigned, and ondemand CSI-RSs are reused —and further extended— to provide support for beam andmobility management as a complement to the SS-block. • The minimum required system information (SI) broadcast in 3GPP NR has also beenreduced with respect to 3GPP LTE, and the part not strictly necessary for network entry istransmitted on-demand now [135]. • Moreover, in 3GPP NR, a single-antenna port can be used to transmit the mandatory controlsignals, while, in LTE, all the antenna ports are used to transmit such mandatory controlsignals [64].This lean carrier design enables both longer sleeping periods and larger sleep ratios, whoseparticular values depend on the specific system numerology, i.e., subcarrier spacing, the numberof SS blocks per SS burst set, and the SS burst set periodicity. In more detail, when consideringa subcarrier spacing of 15 KHz and two SS blocks per SS burst set, a 3GPP NR cell can stay in sleep mode up to 19 ms, which leads to a sleeping ratio of 95% for a SS burst set periodicityof 20 ms. Importantly, when the SS burst set periodicity is maximised to 160 ms, a 3GPP NRcell can stay in sleep mode up to 159 ms, which leads to a sleeping ratio of 99.38% [134].It is important to note that when a BS enters in long sleeping periods, larger energy savings canbe achieved, also because more hardware components can be switched off. According to [136],multiple BS sleep states can be defined in this line, where each sleep mode is associated to agiven sleeping period duration. As a rule of thumb, all the hardware elements that can enter andexit the sleep mode fast enough with respect to the sleep mode duration can be easily deactivatedin such period, while the components having a longer latency need to remain active. In moredetail, and to give an example, • in sleep mode 1 (i.e., cell DTX [61]), characterized by a duration (deactivation plus reac-tivation time) of 71 µ s, the PA and some components of the digital BBU and the analogfront-end (FE) are deactivated. • In sleep mode 2, which has a minimum duration of 1 ms, additional components of theanalog FE are switched off. • In sleep mode 3, which has a minimum duration of 10 ms, the BS additionally deactivatesthe PA, all the digital BBU processing, and almost all the analog FE (except the clockgenerator). • Finally, in sleep mode 4, which has a minimum duration of 1 s, only the wake-up function-alities are maintained [137].Interested readers should note that the complete list of BS elements that can be switched off foreach sleep mode can be found in [138].To assess the positive impact of such different sleep modes on energy efficiency, a new BSpower consumption model was proposed in [139], i.e.,P BS = P TX + P CP , if P TX > ,δ s P liCP , if P TX = 0 and sleep mode is active , (26)where P TX , P CP , and P liCP are the transmit power, the circuit power, and the load independent partof the circuit power consumption, respectively, and δ s is the fraction of the load-independentcircuit power consumption, P liCP , required by the cell in sleep mode. In this case, the authorsassume that the fraction, δ s , is equal to 0.84, 0.69, and 0.29 for sleep mode 1, sleep mode 2,and sleep mode 3, respectively. For sleep mode 4, the fraction, δ s , can be lower than 0.1 [136]. Figure 15. Power consumption trend as the BS enters in successive sleep modes [136]. Activation time is assumed to be equalto the half of the minimum sleep duration [137].
However, current 3GPP NR implementation cannot realise this mode of operation, as it requiresone second of continuous sleeping period.In periods with absence of traffic, a BS can go through the presented sleep modes subsequentlyto reduce its energy consumption. This process is often referred as ASM (see Fig. 15). Whenthe cell load rises and UE traffic appears, such UE traffic is buffered, and the BS has toimmediately switch on its functionalities, and serve the required data to satisfy the end-userQoS requirements. However, it should be noted that, since hardware activation and deactivationtimes are not negligible, and their lengths increase with the number of involved hardwarecomponents [137], ASM may increase the UE perceived latency, and this can accordingly affectsthe UE throughput [140] [141]. Therefore, there is the need for optimising the path betweendifferent sleep modes, and proactively activate the BS components in order to limit performancelosses.The work in [141] has proposed the use of dedicated timers to control when to deactivatecomponents and go into deeper sleep modes. The authors have highlighted that these timersshould depend on the type of traffic carried out by each single BS, and to make a more flexibleusage of the ASM, they have designed a RL model, based on Q-Learning, to optimise the duration of each sleep state. Energy savings and experienced delay are balanced using thistechnique, using as enabler the average packet inter-arrival time. Importantly, their results haveshown that it is possible to implement ASM and achieve significant energy savings, even withstringent delay constraints, for medium and low load scenarios. In more detail, up to 80 % energysavings can be obtained when replacing 3GPP LTE with 3GPP NR technology and using theproposed ASM.Nevertheless, it should be noted that this type of scheme relies on a continuous exchange ofnetwork signalling, and may impact, e.g., the performance of cell (re)selection processes [142].More work is thus needed in this direction to understand of how ASM impacts the network loaddistribution, the resulting inter-cell interference, the related radio resource management (RRM)procedures, and finally the end-users’ QoS. B. Carrier-domain energy saving solutions
In this section, in contrast to the previous one, we focus on the energy savings that can beharvested by considering the long-term traffic load variations in the network. For example, inmost of the cells, the traffic daily profile shows a regular trend, with low load periods early inthe morning, medium loads during the work-hours and high data rate in the late evening (seeFig. 16). Weekends may be characterized by lower traffic demands with respect to workdays.As a result, since mobile networks are sized to satisfy peak time traffic, energy may be wastedduring low and medium load periods, and thus energy saving mechanisms able to adapt, thenetwork configuration to this long-term traffic load variations are necessary.Overall, to realise these longer-term energy saving mechanisms, different approaches may beconsidered, i.e. intra-site, Inter-site, and Inter-RAT energy saving me [143], which are furtherdiscussed in the following.
1) Intra-site energy saving mechanisms:
In the first case, intra-site energy saving, a BS mayactivate energy saving mechanisms to locally adapt its capacity to traffic requirements. At thislevel, one possibility is to control the number of active cells/sectors [144], the number of CCsactive in each sector, if CA is implemented [79], and/or the transmission bandwidth [145].The work in [144] has investigated the impact of cell-level and sector-level switch off patternson the network performance, considering hexagonal tri-sectorial cell layouts and UEs uniformlydistributed in the network, according to a Binomial point process. Specifically, the authors havecompared the performance of different switch off patterns in terms of number of served UEs, Figure 16. Normalized weekly load for a typical cell in a dense urban scenario. spectral efficiency, and energy efficiency. Results indicate that the energy savings of sector-level patterns importantly depend on the hardware shared among the sectors of each site, suchas cooling and BBU processing equipment. The authors’ analysis has shown that, out of thetwenty-six investigated patterns, those that activate only one out of three sectors are particularlybeneficial.To improve the system energy efficiency, besides switching off the overall cell —or one ormore sectors—, it is also important to dynamically control the number of active CCs, if CA isavailable. As discussed in Section III-A1, CA is a 3GPP flagship feature introduced in 3GPPLTE-A, which allows a BS to simultaneously operate on different bands. In 3GPP NR Rel. 15,the dormant state was introduced in CA, such that (de)activation delay for SCells could bereduced, and the set of active CCs could be rapidly adapted to match UEs requirements [45].In addition, the concept of Bandwidth Part (BWP) was also introduced in Rel. 15 [146]. Withthis mechanism, when the cell load is reduced, the BS can configure only a part of a givenCC for actual transmission/reception, which is refereed to as a BWP . Importantly, controland data signalling only occur within this part of the spectrum, enabling thus a reduced powerconsumption at both the BS and UE sides, as they need to handle/monitor smaller bandwidths.BWP can be (de)activated by a timer, downlink control information (DCI), and/or RRC signalling,which can enable faster bandwidth adaptation with respect to the CA framework. In this context, For each CC, at most four BWPs can be defined for the downlink and the uplink communications. the work in [79] investigated the optimal transmit power and CA configuration to optimise theWSEE metric (see eq. (12) in Section III), while satisfying the downlink and the uplink UEsrequirements.It should be noted, however, that these intra-site energy saving mechanisms —whether cell,sector, or CC (de)activation— may have unwanted network side effects. This is because, whilereducing energy consumption, they may significantly impact the network layout (i.e., coverage)as well as the load and interference distributions across the network [147]. Turning off a BS,for example, may leave some UE without coverage, and increase the traffic loads of multipleneighbouring BSs. If these BSs operate on the same spectrum band, BS (de)activation will alsoimpact the inter-cell interference pattern, which will also have important consequences on therank and the modulation and coding scheme (MCS) selection, and thus on packet success rate.To provide a proper network-wide optimisation, and ensure the satisfaction of the end-users’QoS, it is necessary to take these effects into account when adjusting a BS configuration.
2) Inter-site energy saving mechanisms:
Inter-site mechanisms serve this purpose, operatingover multiple neighbouring BSs, in a centralised or a distributed manner, to jointly optimise theuse of radio resources and provide energy savings without affecting end-users’ QoS. This resourcemanagement problem is typically challenging, as it involves multiple network functionalitiesand complex mathematical formulations. For instance, the authors of [148] investigated howto optimise the network energy efficiency by jointly managing BS activity, UE association andpower control in a HetNet with mMIMO capabilities. They modelled this problem using mixed-integer programming, and to address complexity, proposed a sub-optimal centralised schemewhere the integer variables (i.e., the cell (de)activation and the UE association variables) arerelaxed, and ii) the BS activity, UE association, and power control problems are iterativelysolved. Since their centralised solution was still characterized by a large complexity, they alsoproposed a distributed solution based on game theory, which provided lower performance, butit is proven to converge to the Nash equilibrium.In addition to affect the network performance, energy saving mechanisms can have a detri-mental effect on the BS hardware life-time. Specifically, deep and frequent transients betweendifferent status lead to large temperature gradients in the involved hardware components, whichincrease their failure rates, thus augmenting the maintenance costs to fix or replace the BS. Theimpact due to these long-term energy saving mechanisms can be measured as the accelerationfactor, which is defined as the ratio between the failure rate observed by applying such mecha- nisms over time and the one experience when keeping the BS always active. Recently, the workin [149] has modelled hardware failure rate due to cell switch on/off, and proposed a heuristicto control the statuses of BSs, which minimises the acceleration factor growth over time, whilesatisfying the end-users’ QoS requirements. In contrast to baseline solutions, which maximizethe power saving at the cost of increasing the BS failure rate over time, the proposed approachachieves around a 30 % of power savings in a 3GPP LTE scenario, while keeping the accelerationfactor close to one.To further avoid excessively frequent status changes of the BSs and associated networkperformance losses, the BS control policy in charge of (de)activation decisions could consider theload distribution and the manner in which it varies in time and space. To this end, load can eitherbe characterized statistically or using data-driven approaches. For instance, the authors of [150]have considered a dense HetNet, where small cells and UEs are randomly deployed, followinga HPPP distribution, and packets arrive to the transmission buffers according to an exponentialdistribution. Using this model, they have characterized the probability density function of the cellload using a gamma distribution, and used this information to elaborate multiple (de)activationstrategies, comparing them in terms of complexity, blocking rate probability, throughput, andenergy efficiency.In the same line, the authors of [151] have proposed a stochastic game, where distinct BSinstances in a CRAN platform take advantage of spatial correlation, and jointly estimate thenetwork traffic, exploiting past observations. The CRAN then uses these estimates to decide thestatus of the remote radio heads (RRHs) in the network, distribute the UEs among the cells,control the RRHs transmit power, and setup cooperative transmission schemes to guarantee thatcoverage requirements are satisfied. The authors also demonstrated, through a CRAN experimen-tal platform, that the proposed solution using traffic estimations leads to large energy savingswith respect to a dynamic BS switching solution [66], which is not aware of the traffic evolution.Cell zooming has also become a popular inter-cell energy saving mechanism, which consistsin reducing the cell coverage of lightly loaded cells, while simultaneously increasing the areacovered by neighbouring compensating ones [152] (see Fig. 17). When using this mechanism,topology changes should be smoothly implemented to limit service outage [153]. To face thischallenge, the work in [154] has proposed a data-driven approach to optimise the cell zoomingmechanism. Specifically, this framework has represented the network through a graph, and useda BS connectivity metric related to user-level data to construct the graph adjacency matrix. Figure 17. Cell zooming procedure [154].
The authors have run a Markov process on the graph to identify network communities on whichimplementing the cell zooming. This process is realized through two steps. The first step consistson using a polynomial model to predict the expected community traffic load in the next hour, andthe second step operates the cell zooming to identify the BSs to deactivate and how to distributethe load among the active BSs. With respect to baseline solutions based on the knowledge ofinstantaneous traffic [66], [152], the authors have highlighted that the proposed solution leads to20 % energy saving gains at the cost of increasing the blocking rate only by 0.1 % . They have alsoshown that the achieved blocking rate is greatly affected by the large prediction noise, whichresults in inaccurate forecasts.To make the prediction more robust, researchers have also designed more complex frameworksthan the previous one based on artificial neural networks (ANNs). The work in [155] hasinvestigated multiple prediction models and evaluated how their errors affect the (de)activationcontrol. In particular, they have considered ANN and long short term memory cell (LSTM)together with other simpler models. Their results have highlighted that the considered modelsachieve energy savings close to the ideal strategy, which is aware of future traffic trends, leadingto a blocking rate lower than 5%. The authors have also noted that prediction errors affect thenetwork QoS only at the time of the day when the (de)activation control is activated. Accordingly,average errors may lead to misleading conclusions at this important time periods, indicatingthat prediction frameworks should be designed considering the specifically implemented energy saving algorithm.
3) Inter-RAT energy saving mechanisms:
It should also be noted that, in the early stageof 5G deployments, 3GPP NR BSs are not uniformly distributed across a city area, and thusthere is a need for a tight inter-working between the 3GPP NR network and the underlying3GPP LTE one. As previously mentioned, 3GPP NR BSs are characterized by a larger powerconsumption with respect to 3GPP LTE ones, as they integrate more complex hardware tooperate on a wider bandwidth and use a larger number of antennas and transceiver modules (dueto mMIMO). Therefore, the wireless community is also currently developing inter-RAT energysaving solutions to switch off capacity booster cells i.e., 3GPP NR cells, when the traffic demandis low [156]. In its simpler form, the 3GPP NR cell can autonomously decide to switch off basedon its own load. However, as highlighted earlier, this process would allow better coordinationacross 3GPP NR and LTE BSs to perform mobility management and reactivate capacity boostercells on a need basis [157]. More research is needed in this area.
C. Antenna-domain energy saving solutions
While the mMIMO frameworks presented Section V have provided general important insightson the deployment of energy-efficient mMIMO networks in both the uplink and the downlink, itshould be noted that, once the network is deployed, different approaches can be used to minimiseenergy consumption. When the traffic load is low, the energy consumption of a mMIMO systemmay be reduced by using only a subset of the available BS antennas and/or transceiver modules,according to traffic requirements and avoiding resource waste. This type of approaches arereferred to as antenna selection or channel shutdown [158], [159].A number of frameworks have investigated channel shutdown subject to QoE constraints on thebasis of multi-path fast fading variations, i.e., activating those BS antennas with favorable channelconditions at each —or a small number of— TTIs [160]–[165]. However, in wideband systemssuch as 5G with many subcarriers per carrier, it is unlikely that a BS antenna is simultaneouslynot selected on all such subcarriers. Moreover, antenna selection based on multi-path fast fadingalso requires all antennas to be activated at least for channel estimation, thus limiting theirsleeping time.Taking a more practical approach, the authors in [166] investigated the antenna selectionproblem in the downlink based on large-scale fading instead, targeting at finding both the optimalnumber of BS antennas and their transmit powers to minimise the downlink power consumption of a mMIMO network. QoS constraints were also considered in the form of a minimum SINRper UE. Both the single-cell and multi-cell scenarios were analysed, where the system modelin the latter accounted for multiple BSs, a number of antennas and simultaneously multiplexedUE per BS, and imperfect channel estimation through pilot contamination. It should be noted,however, that the authors adopted a basic BS power consumption model, which depends onthe PA efficiency, and is only linear with the number of BS antennas. Signal processing powerconsumption, for example, as a function of the number of simultaneously multiplexed UEs ineach TTI is not considered. For the single-cell case, the work derived the optimal number ofBS antennas and their transmit powers in closed-form. Importantly, these expressions prove that,only when the circuit power consumption per BS antenna is small, the minimum BS powerconsumption can be attained by activating all the BS antennas. Otherwise, the BS can saveenergy by deactivating some of them. For the multi-cell case, and contrary to the single-cellone, since pilot contamination was considered, a coherent interference term appeared in theSINR formulation, which scales with the number of BS antennas in the pilot-sharing cells, thuslimiting the achievable SINRs of UEs. As a consequence, the results indicated that increasing thenumber of BS antennas still leads to lower transmit powers, but this is not necessarily the optimalto minimise the BS power consumption, as there is a cost associated with using such BSantennas.Unfortunately, the authors concluded that it is hard to obtain closed-form expressions for theoptimal number of BS antennas and their transmit powers for this more complex multi-cellcase. Instead, they showed that the joint optimisation problem can be relaxed as a geometricprogramming problem that can be solved efficiently, and suggested that their algorithm can beused to optimally (de)activate antennas depending on the traffic load variations.The work in [83] has extended these studies investigating how to solve the antenna selectionproblem, while considering daily load variations in multi-cell mMIMO networks. The authorshave computed the distribution of the active UEs in a cell for different network loads usingqueuing theory. They have then modelled the distributed daily energy efficiency maximizationproblem, using the active number of BS antennas in each cell as a variable, and solved itthrough a game theory framework. As a result, a best response algorithm has been proposed,where each BS iteratively selects the strategy that produces its most favorable outcome givenother BS strategies. This selfish approach has not achieved the global optimum, but lead to aNash equilibrium without the need for coordination across cells. It should be noted that theauthors have not considered specific UE rate requirements. Therefore, their analysis have shown that, at low loads, their adaptive antenna selection scheme can achieve around 250 % of energysavings at the expense, however, of around 50% reduction of the average UE data rate.Considering both rate and latency requirements, the work in [167] has also recently investigatedthe power consumption minimisation problem in multi-cell mMIMO networks, by implementingcell DTX (see Section VI-A) in conjunction with precoding and antenna selection. The authorshave considered a multiple frame optimisation window, and have proposed a strategy to selectthe proper precoding technique for each transmission frame such that the total transmission timeand latency are minimised. Moreover, they authors have proposed a technique to trade the UElatency for additional energy savings, by reducing the number of active BS antennas used ineach frame. Numerical results have highlighted that the proposed adaptive system provides largeenergy efficiency gains in lightly loaded scenarios without impacting the end-users’ QoS.VII. M ACHINE L EARNING AND D ATA - DRIVEN E NERGY E FFICIENCY O PTIMIZATION
As previously discussed in Section II-D, energy efficiency optimisation highly depends on theaccuracy of the embraced models, and unfortunately, many of the current models are rigid, mostlythe theoretical ones, unable to adapt to specific channel characteristics, technology features, orenvironment changes. This may yield a considerable theory to practice gap. Instead, data-drivenoptimisation may be able to close this gap, learning the practical state of the network andinferring optimum network operation policies by means of AI to enhance the energy efficiencyof mobile networks [29]. In this line, recently, several works in the literature are aiming atimproving energy efficiency by exploiting state-of-the-art ML algorithms.In Fig. 18, we illustrate the relationships between the main concepts developed in the frame-work of AI and ML [168]. Specifically, today, within the AI world, ML comprises the family ofalgorithms which uses data to develop intelligent systems. We can identify three main models:supervised learning, unsupervised learning, and RL.Supervised and unsupervised learning models have been investigated to characterise andforecast network traffic and predict 5G network behaviour, leveraging rich data-sets. Thesemodels are, for example, becoming increasingly popular to define new solutions for PHY layerfunctions [169]. In contrast to supervised and unsupervised learning, the essence of RL concernslearning to make online decisions through interactions with the environment to control networkoperations. Therefore, ML models are key to enable an intelligent mobile network able to Figure 18. Relationships among deep reinforcement learning, deep learning, reinforcement learning, supervised learning,unsupervised learning, machine learning, and AI. characterize its environment, predict system changes in time and space, and react accordinglyin a real time manner.In the following, we review the literature related to the use of ML techniques for trafficprediction and network optimization in green 5G networks.
A. ML for Traffic Prediction
One of the fundamental challenges along the path to enable full network adaptation to end-users’ QoS requirements is the accurate forecasting of the network traffic. Such data forecastingcan help driving energy efficient network decisions, e.g., carrier shutdown and others, as it willbe shown in the next section.The forecasting of the network traffic presents, however, important challenges: • End-users have different QoS requirements at different moments of the day and in differentplaces. Therefore, traffic demands change in time and space, making the prediction taskdifficult. • The mobility of UEs introduces spatial dependencies among neighboring cells. Moreover,spatial dependencies can occur between distant cell towers, as efficient urban transportationsystems easily enable UEs to travel across cities within half an hour. • The spatial distribution of UEs at the urban scale is further influenced by many factors,including land use, population, holidays, and various social activities. These further com-plicate the spatio-temporal dependencies among traffic in distinct cell towers. • The prediction time scale should match the decision periodicity of the energy savingmechanism. For instance, when adjusting every hour the number of active carriers, theforecasting model should provide predictions of the cell load every hour. In contrast, if themechanism works on a daily basis, a longer prediction, i.e., 24 hours, is required, makingthe task more challenging.The studies in the literature aiming at predicting the network traffic can be differentiated intotwo groups, according to the adopted methods, i.e., statistical-based and ML-based approaches.
1) Statistical-based methods:
Statistical-based methods rely on capturing the statistics of thenetwork traffic. One of the most popular statistical approaches when predicting network traffic isautoregressive integrated moving average (ARIMA) [170], which originates from three models:the auto-regressive model, the moving average model, and their combination (ARMA). Thepredictions performed by this model are based on considering the lagged values of a given time-series, while accommodating for non-stationarity. The main limitation of ARIMA is its inabilityto capture the seasonality —a time series with a repeating cycle— of network traffic. To overcomesuch limitation, an extension of this algorithm, named seasonal autoregressive integrated movingaverage (SARIMA), has been proposed [171]. SARIMA adds three new hyper-parameters tospecify the auto-regression, the moving average, and the differencing for the seasonal componentof the series, as well as an additional parameter for the period of the seasonality.Statistical methods like this, however, are not able to capture rapid traffic variations, sincethey rely on the mean value of the historical data. Moreover, they are mainly linear, and it hasbecome clear that they cannot provide high accuracy when predicting network traffic, especiallywhen considering complex network traffic behaviors observed in real scenarios [172].
2) ML-based methods:
In contrast to statistical methods, data-driven approaches based onML have been recently investigated as a solution for network traffic prediction, as they allowto model non-linearities, while taking advantage of the big amounts of data currently beingcollected by the BSs.Traditional ML algorithms such as k -nearest neighbours (KNN) [173] and support vectorregression (SVR) [174] are able to model non-linear relationships. However, they require well-tuned parameters to achieve accurate prediction results. Moreover, these methods are known Figure 19. Example of RNN composed of three stacked LSTM units. The terms x t and h t are respectively the input and theoutput at time time t . Moreover, the sigmoid and hyperbolic tangent activation functions are represented by the σ and tanh symbols, respectively. to have short memory due to their limited parameter set and inefficient computing, which isdetrimental for improving the prediction accuracy. a) Recurrent Neuronal Networks: Research has then moved into recurrent neural networks(RNNs) to model more complicated nonlinear sequence patterns, which has provided promisingresults in many fields, such as speech recognition, image caption, and natural language process-ing. In particular, LSTM has been proposed as a solution to the problem of vanishing gradient intraditional RNNs [175]. This neural network architecture allows to learn long-term dependenciesfrom the time series provided in the input. A LSTM unit is characterised by three gates, i.e.,input gate, forget gate, and output gate. These gates control the unit operations by consideringthree inputs, i.e., the input vector, the memory of the previous time-step, and the output ofthe previous time-step. The non-linearity is modeled through a sigmoid unit and a hyperbolictangent unit, which implement the respective functions. As an example, Figure 19 shows a neuralnetwork architecture composed of three stacked LSTM units.Based on such initial definition, the authors in [176] improved the LSTM state-of-the-art withan LSTM architecture using an encoder-decoder model based on gated dilated causal convolution.In the encoder, the long-range memory capacity is enhanced by gated dilated convolutions withoutincreasing the number of model parameters in order to learn a vector representation of the inputsequence. Subsequently, different temporal-independent and temporal-depended features, such asthe daytime, holidays, weather, are fused with the representation vector. This allows to provideto the model additional relevant information with respect to the network traffic time series.In the decoder, the model applies a RNN with multiple LSTM units to map the fused vector representation back to the variable-length target sequence.Attention mechanisms are also often used when adopting a LSTM architectures to weight theimportance of previous observations [177]. However, experiments have highlighted that simplyusing lagged inputs (i.e., data points from one year ago, half a year ago, and a quarter before)allows reaching better prediction accuracy than using complex attention mechanisms due to thestrong periodicity characterizing the network traffic time series.It should be noted, however, that the aforementioned LSTM methods do not take into con-sideration the spatial dependencies between the traffic experienced by different BSs, althoughit has became apparent that capturing this information may be fundamental to provide accurateforecasting of network traffic [178]. Different extensions of the previous presented approacheshave been proposed in this direction.The authors in [179] have, for example, proposed a novel prediction model to forecast trafficcongestion, such that the uplink to downlink resource ratio can be adjusted to improve networkingefficiency. The proposed model is composed of a tree-based deep model, followed by an LSTM.This tree-based model uses convolutional layers, which are useful to capture spatial information.Moreover, using a deep model allows to reduce the computational cost, because the convolutionoperations are performed in parallel in a tree-like structure.In [180], a hybrid deep learning model for spatio-temporal prediction is proposed insteadto incorporate spatial correlation, in which the temporal dependence is captured by a LSTM,whereas the spatial dependence is encapsulated by auto-encoders. Specifically, an auto-encoderis a neural network architecture used in unsupervised learning to represent a set of data, whilereducing its dimensionality [181]. The auto-encoder learns to compress data from the input layerinto a short code, i.e., the embedding, and then decompresses that code into a data structurethat closely matches the original data i.e., the output layer. With this architecture, the auto-encoders are used to model historical information from the neighboring BSs, and capture spatialdependencies.Even though the aforementioned attempts to capture spatial information, LSTM is not fun-damentally adequate for it. In particular, the gates that characterize this model are usually fullyconnected, and as a consequence, the number of parameters is large, requiring high memoryand computation time for training the model. This model is thus highly complex and frequentlyturns overfitted. Figure 20. Example of a GNN architecture. The activation function unit is indicated by the f a symbol. b) Convolutional Neuronal Networks: An evolution of LSTM, named Convolutional LSTM(ConvLSTM) has been proposed to solve this problem, by replacing the inner dense connectionswith convolution operations [182]. This architecture significantly reduces the number of param-eters, and enhances the ability of capturing spatio-temporal information. Indeed, convolutionalneural networks (CNNs) are widely adopted now to deal with image classification problems,and capture spatial information. Similarly, when considering the network traffic prediction case,network traffic data is treated as images, where the geographical space is modeled by a matrix,and the traffic distribution in different areas of the city is described by the elements of suchmatrix.As a good example, the authors in [183] provide a traffic prediction architecture, in whichspatio-temporal dependencies are captured by utilizing densely connected CNNs. In a similarway, the authors in [184] have proposed a prediction algorithm, which can model both tem-poral and long-distance spatial dependencies. The proposed model follows an encoder-decoderparadigm, where a stack of ConvLSTM and CNN elements are combined.A main limitation of this type of approaches, however, is that they only work with regular grid-based region partitions, which are not practical for cellular networks, and limits the predictionperformance. c) Graph Neuronal Networks:
To overcome this limitation, an architecture based on graphneural network (GNN) has recently been proposed to model the network traffic spatio-temporaldependencies using a graph representation. In particular, given a direct graph, each BS is modeledas a vertex, and each edge defines the spatial relations between adjacent BSs. The authors in [178] have adopted such GNN-based architecture, and decompose the totaldata traffic volume into in-tower traffic and inter-tower traffic, which corresponds, respectively,to the traffic serviced to the UEs residing within the coverage of a BS and the traffic serviced tothe UEs moving among areas covered by different BSs. In the proposed architecture, each edgehas a weight that depends on the total data traffic moving from the corresponding BSs vertexes.Importantly, it should be noted that complete directed graphs can contain a huge amount of edges,which hinder the efficient learning of model parameters. Therefore, low weights are treated asnoise, and the corresponding edges are pruned by defining a threshold that allows to balanceprediction accuracy and computing efficiency. The presented numerical results have highlightedthat the traffic mobility induced by the roaming of humans plays a large role in the predictionaccuracy, and showed that a combination of in-tower and inter-tower traffic patterns can beapplied for network or social event forecasting. d) External inputs: Point of Interests:
While the aforementioned research has mainly fo-cused on the network traffic itself, it is also well understood that external factors, such as thepoint of interest (POI) distribution, may influence the demand of network traffic. In particular,the analysis provided in [185] reveals that the dynamic urban network traffic usage exhibits fivebasic time domain patterns, which are correlated to the city functional zones, e.g., residential,transport, office.In this line, the work in [186] has targeted network traffic prediction by exploiting cross-domaindata. Specifically, three types of data sets are considered, i.e., BSs location, POIs distribution,and social activity level. In particular, the latter contains information generated by the end-userswhen using social networks, such as location and keywords, which may allow to better captureparticular social events, such as concerts and football matches. The correlation between thesedata sets and the network traffic is analyzed and used to improve the prediction accuracy. A noveldeep learning based traffic prediction architecture is then proposed. This architecture can fusethe cross-domain data sets into a unified representation. Spatial, temporal, and external factorsare then captured and processed by ConvLSTMs. In order to consider the pattern diversity andsimilarity of the network traffic of different city functional zones, the authors have also proposedan algorithm for grouping city areas into different clusters. Then, inter-cluster transfer learningis proposed to capture regional similarities and differences. The achieved results have shown thatcross-domain data sets have high correlation with the network traffic, and thus the introductionof the aforementioned data sets benefits the prediction accuracy. e) Model re-usability: Another important issue related to traffic prediction is that of re-usability. Prediction algorithms generally lack of re-usability, which require them to be re-trainedto learn a new representation of the spatio-temporal information, when adopted in a new —ordramatically changing— scenario. The generalization problem of prediction algorithms has beenrecently discussed in [187], where the authors have proposed a model based on auto-encoders,which learns the embedding of BSs based on raw data. In this framework, the embeddings arevectors, which contain spatio-temporal information of the BSs, and their size is much smaller thanthe raw data, which allows to improve the generalisation capabilities of the prediction algorithm,while also reducing its computational cost. The architecture is composed of three main modules:an encoder, a spatial adder and a decoder. In more detail, the encoder is designed to extractinformation from the BS, and infer its embedding. The spatial adder is in charge of building therelation among different BSs, whereas the decoder restores original data from the embeddings.In this way, the training phase makes the encoder learn how to generate an embedding thatconforms to the spatial relation with neighboring BSs. After training the model, the encoder isable to use the raw data from the BS itself to infer its embedding, which contains informationabout how this BS influence other BSs. Numerical results have shown that this approach helpstemporal models to achieve similar performance as spatio-temporal ones, at the cost of a smallincrease in the training time.
B. ML for 5G Energy Efficiency Optimisation
The wireless network environment is complex and stochastic by nature as already discussedearlier, and in more detail, traffic requirements, user mobility, interference and channel variationsin time and space make system-wide optimisation a hard problem. In the past, most of solutionsproposed in the literature to configure mobile network parameters have not considered thedynamic nature of wireless networks. More specifically, state-of-the-art algorithms, as manyof the once already surveyed, are typically based on perfect —or partial— knowledge of theinstantaneous system conditions, which requires to re-compute the solution of a problem when-ever a notable change has occurred in the environment. With the complexity carried by suchapproaches, they may lead to significant computation and signaling overhead. Thus, there is anurgent need for more light-weight, flexible and adaptive solutions with respect to environmentdynamics to minimise the energy consumption of practical networks. In the last decade, RL, and more recently deep reinforcement learning (DRL), have emergedas potential tools to pave the way for artificial intelligence driven optimisation in 5G systems andbeyond. For more details on the motivation, refer to [188] and [67] and the references therein.Given their importance, in this section, we provide an overview of RL, highlighting its benefitsand drawbacks, and emphasizing how the combination of deep neural network (DNN) and RL,i.e., DRL, is leading to continuous breakthroughs in multiple research domains, with distinctchallenges. We also present several class of algorithms, and show relevant applications in theenergy efficiency optimisation domain.
1) RL Framework Overview:
In RL, the environment is often modelled as a markov decisionprocess (MDP), which consists of [189]: • a set of states, S , • a set of actions, A , • transition probabilities; T ( s t + | s t , a t ) , which map a state-action pair at time t to a distri-bution of states at time t + 1 ; • a reward function, R ( s t , a t ) , and • a discount factor, γ ∈ [0; 1) .The sequence of received rewards leads to the definition of the cumulative discounted reward,i.e., R t = ∞ X k =0 r t + k γ k . where r t + k is the expected reward at time t + k .In an MDP, a decision-maker, i.e., the agent, attempts to find the optimal policy, π ∗ , whichmaps the optimal action to each state, such that the cumulative discounted reward, R t , ismaximized. Specifically, in literature, two functions are defined to characterise the expectedvalue of a policy, and find the strategy that optimises the system behaviour: • the state-value function, V π ( s ) = E [ R t | s t = s ] , and • the state-action value, Q π ( s, a ) = E [ R t | s t = s , a t = a ] ,where V π ( s ) = max a Q π ( s, a ) [69].Fig. 21 describes the agent interactions and learning process in a MDP. This problem is charac-terised by multiple challenges. For instance, the environment is typically partially observable, i.e.,an agent does not have full knowledge of the system state, but only a partial observation. In thecontext of wireless network optimisation, this is very likely, and it can represent the case where Figure 21. The perception-action-learning loop. a BS is not aware of the load of other BSs. Moreover, the perception of the reward relatedto a state-action pair is often delayed, which makes hard to evaluate the effect of an action,and thus to learn the optimal policy. This effect is known as the temporal credit assignmentproblem [69]. Similarly, in a multi-agent system, a perceived reward depends on the actionsof multiple agents behaving independently from each other, i.e., a cell throughput depends onhow each neighbouring BS schedules its own resources. In addition, the environment transitionprobabilities are likely to be unknown, as the system is too complex to be modelled.When the environment can be fully modelled, dynamic programming can be used to solve thelearning problem through algorithms whose complexity is polynomial in the size of the set ofstates, S [69]. In RL, an agent tries to achieve a full characterisation of the transition probabilities, T , and the reward function, R , through continuous interactions with the environment. Throughthis learning by interaction loop, a exploration-exploitation trade-off arises, i.e., exploiting theinformation collected so far to benefit the locally optimal decision, or exploring for achieving abetter characterisation of the environment and achieve higher long-term gains.
2) RL and Energy Efficiency:
Recently, in the context of energy efficiency, new RL schemeshave been proposed to manage ASMs in 5G BSs. In [137], the authors have mapped the time-domain ASM control —see Section VI-A—- as a decision making problem in which a BSsequentially sets the sleeping level length. Specifically, when the cell becomes idle, this approachfirst puts the BS in the deepest level of sleep, and then gradually switches it on. At each stage,the BS decides the number of slots during which the current sleep mode status will be kept. Ifa UE request arrives during a sleep period, the associated data is saved in a buffer until the BSwakes up. Accordingly, the state set includes the possible states in which a BS can operate, i.e.,idle, active, or one of the sleep modes enabling energy saving. The action set includes the valuesof the possible number of time slots that can be associated to a given sleep mode. The reward is defined as the weighted sum of the energy saving gain due to the ASM and the additionallatency experienced at the UEs due to the buffering of their traffic. In this way, different ASMpolicies can be defined according to whether the operator wants to trade end-users’ QoS forenergy savings. The authors have used a popular RL scheme, named as Q-Learning, to findthe optimal policy. This scheme is a model-free algorithm, which uses transition experiences toiteratively construct an estimate, ˆ Q ( s, a ) , of the optimal state–action value function, Q ∗ ( s, a ) ,also referred to as Q-function, as follows: ˆ Q ( s t , a t ) = ˆ Q ( s t , a t ) + α (cid:18) r t + γ max a ′ ∈A ˆ Q ( s t +1 , a ′ ) − ˆ Q ( s t , a t ) (cid:19) , (27)where (cid:16) r t + γ max a ′ ∈A ˆ Q ( s t +1 , a ′ ) − ˆ Q ( s t , a t ) (cid:17) is the temporal difference (TD) error betweenthe predicted Q-value, r t + γ max a ′ ∈A ˆ Q ( s t +1 , a ′ ) , and its current value, ˆ Q ( s t , a t ) , and α is alearning-rate parameter, which controls how new estimates are iteratively blended together overtime. If each state-action pair is visited infinitely often, and the learning rate is decreased overtime, the estimate, ˆ Q ( s, a ) , converges to the optimal value, Q ∗ ( s, a ) [190]. Note that the optimalstate-action pairs are stored into a look-up-table. The results in [137] have shown that if delayis critical, ASM should not be activated for a cell load larger than 30 % . In contrast, for verylow loads, up to 55 % of energy savings can be achieved, even when prioritizing the end-users’QoS.A well known problem of RL algorithms is the so-called curse of dimensionality, meaningthat their computational requirements grow exponentially with the size of the state and actionspaces [69]. To deal with this challenge, function approximation can be used to approximatethe state–action value function when the state and/or action spaces are large or continuous, i.e., ˆ Q π ( s, a, w ) ≈ Q π ( s , a ) , where w is a set of parameters defined by the function approximators.Multiple function approximation methods have been investigated in the literature for RL, e.g.,linear functions or ANNs.In [191], the authors proposed a fuzzy Q-learning model to deal with complexity, wherea network controller jointly optimises the DTX of the underlying cells and backhaul nodesto minimise the energy consumption and satisfy the end-users’ QoS. Specifically, to reducecomplexity, the controller maintains a distinct model for each cell. The state space of eachcell characterises its buffer state in terms of rate and latency requirements, the BS capacity,and the estimated spectral efficiency loss due to the interference of the nearby BSs, which areexpected to be active. Then, in each time slot, the controller observes individually each state space, and decides in a distributed manner which cells (and associated backhaul nodes) to keepin energy saving mode or activate. The authors have associated to each state-action pair a costfunction, which models the weighted sum of the BS power consumption and the packet lost,either due to latency constraints or interference. As the state space is composed by continuousvariables, this would prevent a classic RL algorithm to converge to an optimal policy in afinite time. Accordingly, the authors have integrated a fuzzy inference system (FIS) to theirframework, and reduced the state space by mapping the state representation in fuzzy sets [192].The authors have shown that the proposed framework is able to coordinate the activation anddeactivation of neighbouring BSs, thus limiting the inter-cell interference. Moreover, this schemetakes advantage of the latency-energy trade-off, and achieves up to 38 % of energy savings withrespect to a baseline DTX, which does not exploit data buffering.To manage curse of dimensionality, DNNs are currently widely used as a powerful globalfunction approximator, where a neural network is used to compress the Q-table [193]. However,the combination of DNN and RL, i.e., DRL can lead to instability, and even divergence duringthe training process [189]. To address these issues, recently, the authors of [193] have proposeda deep Q-learning (DQL) framework that leverages two main ideas, i.e., the usage of experiencereplay, and the introduction of the target network. Experience replay consists in the usage ofa buffer, where tuples of experiences (i.e., interactions with the environment) are saved andcontinuously replayed to break the correlation across subsequent observations during training.Moreover, during training, two distinct deep networks are used. One that is continuously updated,and another one, updated less frequently, which represents the target network on which the TDerror is computed (see eq. (27)). These modifications make the algorithm training more stable.These enhancements have led to continuous research innovations and DRL architectures thatcan successfully deal with problems that were previously considered intractable. For instance,the authors of [194] have proposed a DQL model to dynamically (de)activate BSs based ontraffic requests. This framework has introduced few enhancements with respect to the baselineDQL in [193]. First, they have observed that non-stationary traffic leads to oscillation betweenwaking- and sleeping-dominating regimes. To break this correlated sequence of actions, theauthors propose an action-wise experience replay, where experiences related to different actionsare saved into distinct buffers, which are uniformly sampled during the training process.In the literature, other mechanisms have been proposed to improve the effectiveness of theexperience replay process, e.g., the well know prioritized experience replay [195]. Moreover, and although the reward is clipped to [ −
1; 1] in classic DQL, to capture the strong variationscharacterising the wireless environment, this work has also proposed an adapting reward re-scaling scheme, which consists into dividing the instantaneous reward by a positive adaptingscaling factor, and summing a saturation penalty to the DQL loss function, i.e., the square ofthe TD error in (27). In addition, the authors of this work have also used an interrupted Poissonprocess to model the traffic requests, and generate additional pseudo experiences, which, usingthe DynaQ framework [69], are periodically stored into the replay memory along with realexperiences, and used indiscriminately for training. Empowered by these innovations, their DQLalgorithm attempts to learn the optimal policy that control the BS status, based on a reward thattakes into account the served requests, the queued or re-transmitted requests, and the failed ones.The reward considers the cost to wake up the BS and the one for changing the BS status. Theirexperiments have shown that modelling the traffic requests and generating pseudo experiencesdoes not lead to large gains. In contrast, action-wise experience replay and adaptive rewardscaling improve the stability and adaptability of the proposed framework. Overall, the proposedscheme achieves large gains with respect to a baseline Q-learning approach, in terms of energysaving and QoS.Similarly, the authors of [196] have proposed a DRL model to control the small cell (de)activationin dense HetNets. In this framework, the system state comprises the status of each small cell andthe estimate of its traffic arrival rate, while the action set includes the (de)activation actions. Thecost function provides a qualitative description of the small cell network power consumption, theadditional latency experienced due to deactivating BSs, and the switching cost due to the changeof status of the small cells (i.e., from off to on and vice-versa). This work has improved thestate-of-the-art solutions by considering an actor-critic DRL scheme. In the actor-critic algorithm,the actor selects the action given the state of the environment, and the critic estimates the valuefunction, given the state and the action. Then, it delivers a feedback to the actor (see Fig. 22).Importantly, it should be noted that this type of DRL based on actor-critic has emerged as apowerful solution to deal with continuous action spaces [197]. Conventionally, the actor providesa probability distribution of the possible actions at a given state. In [196], the action space hasthe size, N SC , where N SC is the number of small cells in the network. Therefore, the outputof the actor is defined as a single vector of continuous values. To compensate for the lack ofexploration in the actor’s side, this work proposed to add noise to the output action vector. Thenoisy vector is then converted in a hard decision (i.e., the proto-action), and then, the algorithm Figure 22. The actor-critic framework. explores the set of actions close to the proto-action, and selects the action with the minimumestimated cost. For training the proposed model, the authors have used a deep deterministicpolicy gradient framework, in which the policy and value function are both approximated byDNNs. The authors have shown for that this approach limits the cumulative network cost overtime with respect to baseline RL algorithms, achieving up to 30 % of gain with respect to aQ-learning model, and provides larger stability in case of non-stationary traffic. Moreover, theyhave indicated that the proposed action exploration method reduces the convergence time.To conclude this section, let us highlight that, although DRL has allowed great progresses inthe context of system optimization in a stochastic environment, many challenges are still opened,such as enabling distributed optimization in the context of multiple competitive or collaborativeagents, or designing fast and low complex methods to update the learned policy after notablechange in the system, which have not been observed during the training phase. More challengesfaced by RL with respect to the energy efficiency are described in the next section.VIII. O PEN R ESEARCH D IRECTIONS
In this section, we identify lines of research, which according to the authors’ understanding,still require further efforts to aid increasing the energy efficiency of 5G and future networks.
A. multi-cell Energy Efficiency Theoretical Modelling
As discussed throughout the survey, serving the end-users’ QoS requirements with the mini-mum power consumption is key to energy savings, and while the bounds and trade-offs to drive such optimisation in a single-cell case may be well understood, the fundamental understandingof energy efficiency in multi-cell networks is still limited, due to complexity issues.Large-scale multi-cell networks are intricate to model (see Section V-B), and as a result,for tractability reasons, most current theoretical understanding of energy efficiency for wide-area networks have been derived based on, for example, the performance of the average UEin uniform networks with simplistic channel, operational and BS power consumption mod-els [129] [130]. These models do not generally capture, however, relevant features, such as BS andUE distributions, NLoS and LoS transmissions, directional channels, antenna correlations. Noveltheoretical analyses embracing such complexities are thus required to characterise still unknownenergy efficiency trade-offs, which may exist and allow further technology breakthroughs in realdeployments.In this line, the work in [198]–[200] have represented a step forward, accounting for non-uniform BS and UE distributions, while being able to estimate local performance, i.e., not onlythe performance of the average UE, but also its distribution. On the same note, the work in [201]has characterised the cell load distribution for a given traffic density, and studied how this affectthe network performance. These frameworks, however, are still in their infancy, and have beenmostly applied, up to now, to the analysis of simpler single-antenna small cell networks. Furtherresearch is needed for their application to more complex networks and features, such as mMIMO.With regard to channel models, to give another example, it is widely accepted that mostmMIMO performance bounds used today work well when the useful signal coefficients behavealmost deterministically, i.e., they have a non-zero mean and a small variance. However, theresearch in [202] has recently proven that under highly directional channels, for instance, thechannel hardening effect does not so clearly appear, and that new bounds are thus required,in terms of both capacity and energy efficiency, for a more accurate performance evaluation inthese more realistic setups.It is also important to mention that there is a gap in the literature with respect to sophisticatedenergy efficiency performance analyses, via detailed numerical and/or system-level simulationtools, able to capture the complexity of large-scale multi-cell networks. Gaining understandingof the interplay among complex features such as mMIMO, CA, and coordinated transmissions,together with their power consumption, which is hard to derive through a pure theoreticalanalysis, can provide new road maps to fundamental network deployment and operation. Theintuition gained via these tools can also lead to new theoretical research avenues. B. Energy Efficiency-driven Network Planning Tools
Deploying and operating a large-scale multi-cell network is expensive, and thus requirescareful network dimensioning and planning to ensure an optimum radio resource utilisation,e.g., spectrum and bandwidth, number of BSs, their location, architecture and transmit power,number of antennas, and transceivers per BS, to cite a few [203], [204]. Importantly, networkplanning tools must ensure that the deployed system has a sufficient amount of radio resources,and can use it in an effective manner, to achieve the required level of network performance atthe appropriate cost. Unfortunately, however, such tools are mostly network capacity-driven asof today, and not yet designed to derive optimal energy efficient deployments.Once the network is planned and deployed, such implementation also imposes hard constraintson the future network performance and its energy consumption. It is thus of imperative impor-tance to equip MNOs with sophisticated network planning tools for wide-area network designwith energy efficiency at heart. For example, when should an MNO deploy less BSs and CCswith larger mMIMO arrays, or in contrast, use more BS and CCs with smaller mMIMO arrays?The applicability of optimisation algorithms in the network planning phase to find such typeof practical answers is crucial, as the answer is local, and should rely on accurate topologicaldescriptions of the deployment scenario, knowledge of current site deployments and performance,as well as UE and required traffic distributions and accurate BS power consumption models,among others [63].It is also important to highlight that optimisation performance strictly depends on a reasonabletrade-off between the channel and network functioning modelling accuracy with respect tocomplexity, and thus MNOs should chose their network models carefully on a per problembasis. This reinforces the need for flexible and efficient numerical radio propagation and system-level simulation tools and tailored optimization theories and algorithms for energy efficiencyproblems [205].More developments in this area are needed at a professional level to make sure future networksare optimally dimensioned, and energy waste is avoided.
C. Multi-carrier and Heterogeneous Network Analysis
In most scenarios, new 3GPP NR deployments will coexist with existing ones, e.g., 3GPPLTE. In some cases, these deployments may be orthogonal in frequency with, e.g., 3GPP NRin the 3.5 GHz band and 3GPP LTE in the 2 GHz one. In some other cases, due to the scarcity of spectrum, 3GPP NR deployments will have to take place in the same spectrum already usedby 3GPP LTE [32], [58]. In the latter case, the fundamental tool to enable such 3GPP NR/LTEspectrum coexistence is the dynamic time scheduling of both 3GPP NR and LTE, for which the3GPP NR specification provides tools [206].Given that 3GPP NR and LTE sites have very different characteristics (coverage, bandwidth,antennas, etc.), leading to distinct performance and energy consumption, it would also be de-sirable to inter-work and (de)activate these two technologies in a coordinated manner, whilesatisfying end-users’ QoS requirements with the minimum energy consumption. In some cases,3GPP LTE may operate at lower carrier frequencies than 3GPP NR, and thus be able to providea better blanket coverage at smaller energy consumption. This may be the most energy efficientat low load periods. In contrast, at medium loads, 3GPP NR may be sufficient to provide therequired capacity to the active UEs, and 3GPP LTE can be deactivated. If operated at differentfrequencies, at high loads, both technologies may be aggregated via dual connectivity [207].In scenarios with the mentioned frequency imbalance, and because the BS can avail of largertransmit power than the UE, it may also make sense to simultaneously activate both technologies,and use downlink/uplink split [208], i.e., 3GPP NR for downlink transmissions and 3GPP LTEfor the uplink [58]. However, this should only be done where and when necessary, under energyefficient conditions, avoiding potential energy waste due to having both technologies activated.The optimisation of the inter-working between 3GPP NR and 3GPP LTE is also of highimportance when 3GPP NR appears in the form of small cells [209] or millimetre wave [210]access points. These types of cells have a much smaller coverage radius, and can be (de)activatedto provide boosted capacity where and when needed. Since some implementations of this typeof cell, e.g., millimetre wave, may be power hungry, the usage of coordinated ASMs across alarge number of this type of cells is critical. This can be facilitated through the separation ofthe data plane and control plane [211], where the latter is continuously provided by underlayingmacro cells to ensure robust connectivity and mobility support, while capacity cells (i.e., smallcells) allow for enhanced capacity and high rate data transmissions locally and on demand.In general, there is a lack of studies covering the understanding and optimisation of thesetechnology inter-working practical use cases from a energy efficiency point of view. D. Data-Driven Optimisation
State-of-the-art data-driven approaches for network traffic prediction are mainly related tomeasurements biased by the observation point, i.e., the BSs, which usually does not reportthe effective traffic demand but rather the cell throughput or resource usage, which dependon the network deployment, interference, current RRM parameters, and running energy savingschemes. The use of this potentially biased measurements may be an issue when adoptingprediction algorithms as enablers for improving the mobile network energy efficiency. In fact,the implementation of any energy saving scheme will impact the load distribution across thenetwork, and make related forecasting unreliable. As a result, further research is encouragedwith regard to the estimation and prediction of unbiased metrics, whose characterization is notaffected by the algorithms implemented using these observations.Moreover, there is currently a lack of understanding on how prediction errors, e.g. trafficforecasting errors, affect the gains provided by energy saving schemes. In this line, most of thecurrent literature focuses on measuring the performance of the prediction in the metric space,(function of the difference between the predicted value and ground-truth, e.g., physical resourceblock (PRB) usage, throughput). However, it is generally not straightforward to derive howimprovements in the prediction of such metrics help to minimise the network power consumption.It is thus recommended that prediction accuracy is investigated also in terms of energy savingswhen developing data-driven optimisation schemes [155].As discussed in Section VII, there is also a lack of end-to-end ML frameworks that jointly usesupervised learning and RL to characterise and optimise the 5G system. Specifically, supervisedlearning models can provide multi-step traffic predictions, achieving a comprehensive forecast offuture status of the mobile network environment. Using this information can help RL algorithmsto converge faster to optimal operational policies, and enhance the performance of the explorationphase, e.g., optimally deciding the moment to (de)activate BS functionalities without affectingnetwork performance.Importantly, it should be stressed that recent progresses in the areas of computational process-ing and data storage, as well as the increased availability of big data, have made the use of AImore practical than ever in many challenging fields. However, the acquisition of large datasetsin wireless networks is currently challenging, and their processing energy demanding, whichlimits the opportunities for implement data-driven optimization in 5G systems. To address this issue, the use of joint data-driven and model-based approaches is being widely explored [212],[213], and is becoming the foundation of new optimization mechanisms for large-scale multi-cellnetworks. Moreover, these techniques can be used for bootstrapping ML models, thus reducingtheir need for data, computational complexity, power consumption, and latency. E. Green AI
Most of the recent breakthroughs led by novel ML solutions have been possible thanks to theever-increasing computational capacity of dedicated hardware platforms. The work in [214] hashighlighted that, in the last decade, while ML models evolved from AlexNet —an image recogni-tion DNN presented in 2012 [215]— to AlphaZero —a RL algorithm proposed in 2018 [216]—,the associated computational cost trend increased by 300,000x. In this line, the work in [217]has also analyzed the energy consumption issues arising from the need of exponentially largercomputational resources to continue marginally increasing the model accuracy, and has estimatedthat the carbon footprint of the current brute force trend is environmentally unfriendly.The training of DNNs on mobile devices in a distributed, computationally and energy efficientmanner is also an ongoing research topic, which can brake down the aforementioned complexity.Collaborative learning schemes, such as federated learning [218], should be considered to miti-gate the energy inefficiencies resulting from traditional, centralised ML approaches. Moreover,notable efforts are being made towards hardware design and software accelerators, which makepossible to also move part of the ML process to the UE itself to reduce the overall energyconsumption (see [219], [220] and references therein). In addition, tailored early stopping canalso be used to terminate the training process when a near-optimal solution has been found thusreducing the required number of iterations —and the associated energy consumption— neededto train the ML model [221].However, to pave the way for the successful integration of ML in 5G systems and beyond, itis key to consider computational and energy consumption aspects also during the design phaseof data-driven optimisation algorithms. To achieve this goal, while evaluating an ML model,in addition to accuracy and optimisation-related metrics, computational efficiency and energyconsumption must also be considered. Using these metrics, ML architectures that converge fasterand/or need to be updated less frequently can be designed, prioritising energy consumption overmodel accuracy. One possible approach towards this goal is to investigate yet undiscovered MLparadigms. For instance, as discussed in [212], [213], model-based RL solutions, which exploit a-priori expert knowledge to characterise physically/mathematically the system evolution can beintegrated with model-free RL architectures, which interact with the environment to identify theoptimal policy. Similarly, as explained in Section VII-B, RL can leverage information about thefuture status of the system obtained by supervised learning forecasting to speed up the trainingspeed and converge towards improved operating conditions. F. Renewable Energy Sources
In the last years, the use of renewable energy sources for supplying network elements hasattracted the attention of the research community and industry, due to the increased efficiency andthe decreasing costs of energy harvesters and storage devices. Gathering environmental energythrough dedicated harvesting hardware to supply 5G BSs will translate into operational expen-diture (OPEX) savings and a reduction of the environmental footprint of mobile networks [222].The adoption of renewable energy sources, however, entails a higher management complexity.In fact, environmental energy, such as solar and wind, is inherently erratic and intermittent,which may cause fluctuating energy inflow and produce service outage. A proper optimisationof how the energy is drained and balanced across network elements is thus necessary for aself-sustainable network design.Operational policies for sustainable mobile networks can be found by solving optimisationproblems, which have the objective of maximizing the end-users’ QoS, while dealing withthe constraints coming with the adoption of renewables. In particular, traditional approachesassuming infinitely backlogged data buffers and unlimited energy storage cannot be adoptedsince they would provide unrealistic results [223].Literature works focus on two classes of algorithms: offline and online. Offline algorithmsenable computing optimal operational policies by adopting tools from optimisation theory, suchas dynamic programming, when the meaningful processes (e.g., energy arrivals, CSI, data traffic)are known [224], [225]. However, using such tools in a real scenario is generally impracticabledue to the high complexity of these approaches, and the optimal solutions obtained by offlinealgorithms are generally adopted as a benchmark to measure the optimality of online algorithms.On the contrary, online algorithms are designed to derive sub-optimal policies without needingany a priori information. In particular, ML approaches have been widely developed as a tool tosolve these optimisation problems [226], [227]. In general, most of the literature relies on simplistic assumption. For instance, there is alack of accurate models for batteries, whose capacity reduces with time, affecting the overallenergy efficiency and costs [228]. Moreover, batteries cannot store all the harvested energy inparticular periods of the year. Therefore, research targeting architectures and algorithms enablingan efficient use of the exceeding harvested energy is needed.IX. C
ONCLUSIONS
In this paper, we have provided an overview of the state-of-the-art on the fundamental un-derstanding and practical considerations of the energy efficiency challenge in 5G networks.We have surveyed in detail the available BS power consumption models and metrics for theoptimization of the energy efficiency. We have also reviewed the impact on energy efficiency offour 3GPP NR key features, i.e., mMIMO, the lean carrier design, ASMs, and ML, presentingthe findings in the literature with respect to their available bounds, trade-offs and/or practicalachievable energy savings. Importantly, we have highlighted the need for adapting the networkresources to meet the end-users’ QoS demands, while minimizing network power consumption,and have surveyed the related research differentiating among different algorithm time scales andclasses (i.e. micro-sleeps, carrier and channel shutdown). We have also stressed the role thatspatio-temporal predictions and online optimisation via ML will play in the previous networkpower consumption minimization task, and discussed state-of-the-art ML related approaches insuch energy efficiency field. To conclude, we have also provided discussion around the lines ofresearch that need further work to make 5G networks greener.As a final note, given the enabling effect of that telecommunications systems and the impactthat they can have in meeting the requirements for a sustainable development, we encourage theresearch community to continue making progress toward a sustainable communication system.R
EFERENCES [1] D. Kaufman, N. McKay, C. Routson et al. , “Holocene global mean surface temperature, a multi-method reconstructionapproach,”
Sci Data , vol. 7, no. 21, Jun. 2020.[2] J. Rifkin,
The Third Industrial Revolution: How Lateral Power Is Transforming Energy, the Economy, and the World ,1st ed. St. Martin’s Press, Sep. 2011.[3] S. R. Weart,
The Discovery of Global Warming: Revised and Expanded Edition [5] NASA, “Global cimate change: Vital signs of the planet,” Available at https://climate.nasa.gov/vital-signs/carbon-dioxide/(2020/10/27).[6] United Nations Environment Programme, “UN Emissions Gap Report 2019,” Tech. Rep., Nov. 2019. [Online]. Available:https://wedocs.unep.org/bitstream/handle/20.500.11822/30797/EGR2019.pdf[7] Secretariat of the United Nations Conference on Trade and Development, “UN Trade and Development report 2019,”Tech. Rep., Aug. 2019. [Online]. Available: https://unctad.org/en/PublicationsLibrary/tdr2019 en.pdf[8] T. F. Stocker et al. J. Geophys. Res. Atmos. , vol. 124, no. 12, 2019.[11] A. Cazenave, “How fast are the ice sheets melting?”
Science , vol. 314, no. 5803, pp. 1250–1252, Nov. 2006.[12] R. Seager, M. Ting, I. Held, Y. Kushnir, J. Lu, G. Vecchi, H. Huang et al. , “Model projections of an imminent transitionto a more arid climate in Southwestern North America,”
Science , vol. 316, no. 5828, pp. 1181–1184, Nov. 2007.[13] G. B. Bonan, “Forests and climate Change: Forcings, feedbacks, and the climate benefits of forests,”
Science , vol. 320,no. 5882, pp. 1444–1449, Jun. 2008.[14] A. Clement, R. Burgman, and J. Norris, “ Observational and model evidence for positive low-level cloud feedback,”
Science
Nat Electron , vol. 3, pp. 182––184, 2020.[23] G. Kouropoulos, “A predictive model for the estimation of carbon dioxide emissions of magnetic resonance imaging unitsand computed tomography scanners,”
Journal of Urban and Environmental Engineering [26] A. Abrol and R. K. Jha, “Power Optimization in 5G Networks: A Step Towards Green Communication,” IEEE Access ,vol. 4, pp. 1355–1374, 2016.[27] M. Usama and M. Erol-Kantarci, “A Survey on Recent Trends and Open Issues in Energy Efficiency of 5G,”
Sensors(Basel) , vol. 19, no. 14, Jul. 2019.[28] J. Lorincz, A. Capone, and J. Wu, “Greener, Energy-Efficient and Sustainable Networks: State-Of-The-Art and NewTrends,”
Sensors (Basel) , vol. 19, no. 22, Nov. 2019.[29] X. Cao, L. Liu, Y. Cheng, and X. Shen, “Towards energy-efficient wireless networking in the big data era: A survey,”
IEEE Communications Surveys Tutorials , vol. 20, no. 1, pp. 303–332, 2018.[30] L. Pierucci, “The quality of experience perspective toward 5G technology,”
IEEE Wireless Comm. , vol. 22, no. 4, pp.10–16, 2015.[31] 3GPP, “3gpp specification series: 38 series.”[32] E. Dahlman, S. Parkvall, and J. Skold,
5G NR: The Next Generation Wireless Access Technology . Academic Press, Aug.2018.[33] J. Wu, S. Guo, H. Huang, W. Liu, and Y. Xiang, “Information and communications technologies for sustainabledevelopment goals: State-of-the-art, needs and perspectives,”
IEEE Communications Surveys Tutorials , vol. 20, no. 3,pp. 2389–2406, 2018.[34] R. Mahapatra, Y. Nijsure, G. Kaddoum, N. Ul Hassan, and C. Yuen, “Energy efficiency tradeoff mechanism towardswireless green communication: A survey,”
IEEE Communications Surveys Tutorials , vol. 18, no. 1, pp. 686–705, 2016.[35] Y. Chen, S. Zhang, S. Xu, and G. Y. Li, “Fundamental trade-offs on green wireless networks,”
IEEE CommunicationsMagazine , vol. 49, no. 6, pp. 30–37, 2011.[36] G. Y. Li, Z. Xu, C. Xiong, C. Yang, S. Zhang, Y. Chen, and S. Xu, “Energy-efficient wireless communications: tutorial,survey, and open issues,”
IEEE Wireless Communications , vol. 18, no. 6, pp. 28–35, 2011.[37] Z. Hasan, H. Boostanimehr, and V. K. Bhargava, “Green cellular networks: A survey, some research issues and challenges,”
IEEE Communications Surveys Tutorials , vol. 13, no. 4, pp. 524–540, 2011.[38] D. Feng, C. Jiang, G. Lim, L. J. Cimini, G. Feng, and G. Y. Li, “A survey of energy-efficient wireless communications,”
IEEE Communications Surveys and Tutorials , vol. 15, no. 1, pp. 167–178, 2013.[39] A. De Domenico, E. Calvanese Strinati, and A. Capone, “Green cellular networks: A survey, some research issues andchallenges,”
Computer Communications , vol. 37, pp. 5–24, 2014.[40] L. Budzisz, F. Ganji, G. Rizzo, M. Ajmone Marsan, M. Meo, Y. Zhang, G. Koutitas, L. Tassiulas, S. Lambert, B. Lannoo,M. Pickavet, A. Conte, I. Haratcherev, and A. Wolisz, “Dynamic resource provisioning for energy efficiency in wirelessaccess networks: A survey and an outlook,”
IEEE Communications Surveys Tutorials , vol. 16, no. 4, pp. 2259–2285,2014.[41] C.-L. I, C. Rowell, S. Han, Z. Xu, G. Li, and Z. Pan, “Toward green and soft: a 5g perspective,”
IEEE CommunicationsMagazine , vol. 52, no. 2, pp. 66–73, 2014.[42] Q. Wu, G. Y. Li, W. Chen, D. W. K. Ng, and R. Schober, “An overview of sustainable green 5g networks,”
IEEE WirelessCommunications , vol. 24, no. 4, pp. 72–80, 2017.[43] S. Zhang, Q. Wu, S. Xu, and G. Y. Li, “Fundamental green tradeoffs: Progresses, challenges, and impacts on 5g networks,”
IEEE Communications Surveys and Tutorials , vol. 19, no. 1, pp. 33–56, 2017.[44] A. Mughees, M. Tahir, M. A. Sheikh, and A. Ahad, “Towards energy efficient 5g networks using machine learning:Taxonomy, research challenges, and future research directions,”
IEEE Access , vol. 8, pp. 187 498–187 522, 2020.[45] Y. R. Li, M. Chen, J. Xu, L. Tian, and K. Huang, “Power saving techniques for 5g and beyond,”
IEEE Access , vol. 8,pp. 108 675–108 690, 2020. [46] Q. Nadeem, A. Kammoun, and M. Alouini, “Elevation Beamforming With Full Dimension MIMO Architectures in 5GSystems: A Tutorial,” IEEE Communications Surveys Tutorials , vol. 21, no. 4, pp. 3238–3273, 2019.[47] F. Rusek et al. , “Scaling up MIMO: Opportunities and challenges with very large arrays,”
IEEE Signal Process. Mag. ,vol. 30, no. 1, pp. 40–60, Jan. 2013.[48] X. Lin, J. Li, R. Baldemair, J. T. Cheng, S. Parkvall, D. C. Larsson, H. Koorapaty, M. Frenne, S. Falahati, A. Grovlen,and K. Werner, “5G New Radio: Unveiling the Essentials of the Next Generation Wireless Access Technology,”
IEEECommunications Standards Magazine , vol. 3, no. 3, pp. 30–37, 2019.[49] M. Fuentes, J. L. Carcel, C. Dietrich, L. Yu, E. Garro, V. Pauli, F. I. Lazarakis, O. Grøndalen, O. Bulakci, J. Yu, W. Mohr,and D. Gomez-Barquero, “5g new radio evaluation against imt-2020 key performance indicators,”
IEEE Access , vol. 8,pp. 110 880–110 896, 2020.[50] D. C. Nguyen et al. , “Wireless AI: Enabling an AI-Governed Data Life Cycle,”
IEEE Commun. Surveys Tutorials , 2020.[51] R. Ferr´us, O. Sallent, and J. Perez-Romero, “Data Analytics Architectural Framework for Smarter Radio ResourceManagement in 5G Radio Access Networks,”
IEEE Communications Magazine , vol. 58, no. 5, pp. 98–104, 2020.[52] E. Pateromichelakis, F. Moggio, C. Mannweiler, P. Arnold, M. Shariat, M. Einhaus, Q. Wei, O. Bulakci, and A. DeDomenico, “End-to-End Data Analytics Framework for 5G Architecture,”
IEEE Access
IEEE Transactionson Wireless Communications , vol. 9, no. 11, pp. 3590–3600, 2010.[55] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo,
Fundamentals of Massive MIMO . Cambridge University Press,2016.[56] E. Bj¨ornson, J. Hoydis, and L. Sanguinetti, “Massive mimo networks: Spectral, energy, and hardware efficiency,”
Foundations and Trends® in Signal Processing , vol. 11, no. 3-4, pp. 154–655, 2017.[57] C. Oestges and B. Clerckx,
MIMO Wireless Communications: From Real-World Propagation to Space-Time Code Design .Academic Press, 2007.[58] H. Holma, D. A. Toskala, and T. Nakamura,
5G Technology : 3GPP New Radio . John Wiley & Sons Ltd., Feb. 2020.[59] K. Zheng et al. , “Survey of large-scale MIMO systems,”
IEEE Commun. Surveys Tutorials , vol. 17, no. 3, pp. 1738–1760,Apr. 2015.[60] E. Dahlman, S. Parkvall, and J. Skold, . Academic Press, Oct. 2013.[61] P. Frenger, P. Moberg, J. Malmodin, Y. Jading, and I. Godor, “Reducing energy consumption in lte with cell dtx,” in , 2011, pp. 1–5.[62] O. Blume, D. Zeller, and U. Barth, “Approaches to energy efficient wireless access networks,” in , 2010, pp. 1–5.[63] Ajay R. Mishra,
Fundamentals of Network Planning and Optimisation 2G/3G/4G: Evolution to 5G , 2nd ed. John Wiley& Sons Ltd., Jul. 2018.[64] P. Frenger and R. Tano, “More Capacity and Less Power: How 5G NR Can Reduce Network Energy Consumption,” in , 2019, pp. 1–5.[65] Z. Niu, “TANGO: traffic-aware network planning and green operation,”
IEEE Wireless Communications , vol. 18, no. 5,pp. 25–29, 2011.[66] E. Oh, K. Son, and B. Krishnamachari, “Dynamic base station switching-on/off strategies for green cellular networks,”
IEEE Transactions on Wireless Communications , vol. 12, no. 5, pp. 2126–2136, 2013. [67] N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y. Liang, and D. I. Kim, “Applications of deep reinforcementlearning in communications and networking: A survey,” IEEE Communications Surveys Tutorials , vol. 21, no. 4, pp.3133–3174, 2019.[68] S. Hu, Y. Ouyang, Y. Yao, M. H. Fallah, and W. Lu, “A study of lte network performance based on data analytics andstatistical modeling,” in , 2014, pp. 1–6.[69] R. S. Sutton and A. G. Barto,
Reinforcement Learning: An Introduction , 2nd ed. The MIT Press, 2018. [Online].Available: http://incompleteideas.net/book/the-book-2nd.html[70] N. C. Thompson, K. Greenewald, K. Lee, and G. F. Manso, “The computational limits of deep learning,” 2020.[71] 3GPP TSG RAN, “TR 38.801, Study on self evaluation towards IMT-2020 submission,”
V16.1.0 , Sept. 2010.[72] 3GPP TSG SA, “TR 32.972, Telecommunication management; Study on system and functional aspects of energy efficiencyin 5G networks,”
V16.1.0 , Sept. 2019.[73] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M. A. Imran, D. Sabella, M. J. Gonzalez, O. Blume,and A. Fehske, “How much energy is needed to run a wireless network?”
IEEE Wireless Communications , vol. 18, no. 5,pp. 40–49, 2011.[74] D. L´opez-P´erez, M. Ding, H. Claussen, and A. H. Jafari, “Towards 1 Gbps/UE in Cellular Systems: UnderstandingUltra-Dense Small Cell Deployments,”
IEEE Communications Surveys & Tutorials , vol. 17, no. 4, pp. 2078–2101, 2015.[75] B. Chen, J. Chen, Y. Gao, and J. Zhang, “Coexistence of lte-laa and wi-fi on 5 ghz with corresponding deploymentscenarios: A survey,”
IEEE Communications Surveys Tutorials , vol. 19, no. 1, pp. 7–32, 2017.[76] Hua Wang, C. Rosa, and K. I. Pedersen, “Dedicated carrier deployment in heterogeneous networks with inter-site carrieraggregation,” in , 2013, pp. 756–760.[77] C. Rosa, K. Pedersen, H. Wang, P. Michaelsen, S. Barbera, E. Malkam¨aki, T. Henttonen, and B. S´ebire, “Dual connectivityfor lte small cell evolution: functionality and performance aspects,”
IEEE Communications Magazine , vol. 54, no. 6, pp.137–143, 2016.[78] K. I. Pedersen, F. Frederiksen, C. Rosa, H. Nguyen, L. G. U. Garcia, and Y. Wang, “Carrier aggregation for LTE-advanced:functionality and performance aspects,”
IEEE Communications Magazine , vol. 49, no. 6, pp. 89–95, 2011.[79] G. Yu, Q. Chen, R. Yin, H. Zhang, and G. Y. Li, “Joint Downlink and Uplink Resource Allocation for Energy-EfficientCarrier Aggregation,”
IEEE Transactions on Wireless Communications , vol. 14, no. 6, pp. 3207–3218, 2015.[80] S. Tombaz, P. Frenger, F. Athley, E. Semaan, C. Tidestav, and A. Furuskar, “Energy Performance of 5G-NX WirelessAccess Utilizing Massive Beamforming and an Ultra-Lean System Design,” in , 2015, pp. 1–7.[81] A. De Domenico, R. Gerzaguet, N. Cassiau, A. Clemente, R. D’Errico, C. Dehos, J. L. Gonzalez, D. Ktenas, L. Manat,V. Savin, and A. Siligaris, “Making 5g millimeter-wave communications a reality [industry perspectives],”
IEEE WirelessCommunications , vol. 24, no. 4, pp. 4–9, 2017.[82] E. Bj¨ornson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design of energy-efficient multi-user mimo systems: Ismassive mimo the answer?”
IEEE Transactions on Wireless Communications , vol. 14, no. 6, pp. 3059–3075, 2015.[83] M. M. A. Hossain, C. Cavdar, E. Bj¨ornson, and R. J¨antti, “Energy Saving Game for Massive MIMO: Coping With DailyLoad Variation,”
IEEE Transactions on Vehicular Technology , vol. 67, no. 3, pp. 2301–2313, 2018.[84] E. Bj¨ornson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Designing multi-user mimo for energy efficiency: When ismassive mimo the answer?” in , 2014, pp.242–247.[85] M. Fiorani, S. Tombaz, J. Martensson, B. Skubic, L. Wosinska, and P. Monti, “Modeling energy performance of C-RAN with optical transport in 5G network scenarios,” IEEE/OSA Journal of Optical Communications and Networking , vol. 8,no. 11, pp. B21–B34, 2016.[86] C. Rowell and Rohde & Schwarz, “Pillars of 5G: Spectral & Energy Efficiency,”
Microwave Journal , pp. 182––184,2020.[87] N. Nikaein, E. Schiller, R. Favraud, R. Knopp, I. Alyafawi, and T. Braun,
Towards a Cloud-Native Radio Access Network .Cham: Springer International Publishing, 2017, pp. 171–202.[88] T. Sigwele, A. S. Alam, P. Pillai, and Y. F. Hu, “Energy-efficient cloud radio access networks by cloud based workloadconsolidation for 5g,”
Journal of Network and Computer Applications , vol. 78, pp. 1 – 8, 2017.[89] A. Younis, T. X. Tran, and D. Pompili, “Bandwidth and Energy-Aware Resource Allocation for Cloud Radio AccessNetworks,”
IEEE Transactions on Wireless Communications , vol. 17, no. 10, pp. 6487–6500, 2018.[90] D. Sabella, A. De Domenico, E. Katranaras, M. A. Imran, M. di Girolamo, U. Salim, M. Lalam, K. Samdanis,and A. Maeder, “Energy Efficiency Benefits of RAN-as-a-Service Concept for a Cloud-Based 5G Mobile NetworkInfrastructure,”
IEEE Access , vol. 2, pp. 1586–1597, 2014.[91] J. K. Chaudhary, A. Kumar, J. Bartelt, and G. Fettweis, “C-ran employing xran functional split: Complexity analysis for5g nr remote radio unit,” in , 2019, pp. 580–585.[92] S. Khatibi, K. Shah, and M. Roshdi, “Modelling of computational resources for 5G RAN,” in
Proc. EuCNC , June 2018,pp. 1–5.[93] M. Fiorani, S. Tombaz, J. M˚artensson, B. Skubic, L. Wosinska, and P. Monti, “Energy performance of C-RAN with5G-NX radio networks and optical transport,” in
IEEE International Conference on Communications (ICC) , 2016, pp.1–6.[94] A. Zappone and E. Jorswieck, “Energy efficiency in wireless networks via fractional programming theory,”
Foundationsand Trends® in Communications and Information Theory , vol. 11, no. 3-4, pp. 185–396, 2015.[95] ETSI TC EE, “ES 203 228, Environmental Engineering (EE); Assessment of mobile network energy efficiency,”
V1.3.1 ,Oct. 2020.[96] 3GPP TSG RAN, “TR 38.913, Study on Scenarios and Requirements for Next Generation Access Technologies,”
V16.0.0 ,July 2020.[97] 3GPP TSG SA, “TR 21.866, Study on Energy Efficiency Aspects of 3GPP Standards,”
V15.0.0 , June 2017.[98] Z. Yan, M. Peng, and C. Wang, “Economical Energy Efficiency: An Advanced Performance Metric for 5G Systems,”
IEEE Wireless Communications , vol. 24, no. 1, pp. 32–37, 2017.[99] H. Peters, “Axiomatic bargaining game theory,”
Mathematical Programming and Operations Research, W. Leinfellner andG. Eberlein, Eds. Kluwer Academic Publishers , vol. 9, 1992.[100] EU FP7 EARTH, “Deliverable D2.4, Most suitable efficiency metrics andutility functions,” Jan. 2012. Accessed on 18/10/2020. [Online]. Available:https://cordis.europa.eu/docs/projects/cnect/3/247733/080/deliverables/001-EARTHWP2D24.pdf[101] E. Bj¨ornson and E. G. Larsson, “How energy-efficient can a wireless communication system become?” in . IEEE, 2018, pp. 1252–1256.[102] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral efficiency of very large multiuser mimo systems,”
IEEE Transactions on Communications , vol. 61, no. 4, pp. 1436–1449, 2013.[103] E. Bj¨ornson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive mimo systems with non-ideal hardware: Energyefficiency, estimation, and capacity limits,”
IEEE Transactions on Information Theory , vol. 60, no. 11, pp. 7112–7139,2014. [104] J. Hoydis, S. ten Brink, and M. Debbah, “Massive mimo in the ul/dl of cellular networks: How many antennas do weneed?” IEEE Journal on Selected Areas in Communications , vol. 31, no. 2, pp. 160–171, 2013.[105] J. Tang, D. K. C. So, E. Alsusa, and K. A. Hamdi, “Resource efficiency: A new paradigm on energy efficiency andspectral efficiency tradeoff,”
IEEE Transactions on Wireless Communications , vol. 13, no. 8, pp. 4656–4669, 2014.[106] Y. Hao, Q. Ni, H. Li, and S. Hou, “Energy and spectral efficiency tradeoff with user association and power coordinationin massive mimo enabled hetnets,”
IEEE Communications Letters , vol. 20, no. 10, pp. 2091–2094, 2016.[107] Y. Huang, S. He, J. Wang, and J. Zhu, “Spectral and energy efficiency tradeoff for massive mimo,”
IEEE Transactionson Vehicular Technology , vol. 67, no. 8, pp. 6991–7002, 2018.[108] L. You, J. Xiong, A. Zappone, W. Wang, and X. Gao, “Spectral efficiency and energy efficiency tradeoff in massive mimodownlink transmission with statistical csit,”
IEEE Transactions on Signal Processing , vol. 68, pp. 2645–2659, 2020.[109] Z. Liu, W. Du, and D. Sun, “Energy and spectral efficiency tradeoff for massive mimo systems with transmit antennaselection,”
IEEE Transactions on Vehicular Technology , vol. 66, no. 5, pp. 4453–4457, 2017.[110] J. Choi, D. J. Love, and P. Bidigare, “Downlink training techniques for fdd massive mimo systems: Open-loop andclosed-loop training with memory,”
IEEE Journal of Selected Topics in Signal Processing , vol. 8, no. 5, pp. 802–814,2014.[111] X. Gao, B. Jiang, X. Li, A. B. Gershman, and M. R. McKay, “Statistical eigenmode transmission over jointly correlatedmimo channels,”
IEEE Transactions on Information Theory , vol. 55, no. 8, pp. 3735–3750, 2009.[112] E. V. Belmega and S. Lasaulce, “Energy-efficient precoding for multiple-antenna terminals,”
IEEE Transactions on SignalProcessing , vol. 59, no. 1, pp. 329–340, 2011.[113] V. S. Varma, S. Lasaulce, M. Debbah, and S. E. Elayoubi, “an energy-efficient framework for the analysis of mimo slowfading channels,”
IEEE Transactions on Signal Processing .[114] L. Liu, G. Miao, and J. Zhang, “Energy-efficient scheduling for downlink multi-user mimo,” in , 2012, pp. 4394–4394.[115] J. Mao, J. Gao, Y. Liu, and G. Xie, “Simplified semi-orthogonal user selection for mu-mimo systems with zfbf,”
IEEEWireless Communications Letters , vol. 1, no. 1, pp. 42–45, 2012.[116] X. Zhang, S. Zhou, Z. Niu, and X. Lin, “An energy-efficient user scheduling scheme for multiuser mimo systems withrf chain sleeping,” in , 2013, pp. 169–174.[117] H. Li, H. Zhang, D. Li, Y. Liu, and A. Nallanathan, “Joint antenna selection and user scheduling in downlink multi-user mimo systems,” in , 2018, pp.1072–1076.[118] H. Lee and A. Paulraj, “Mimo systems based on modulation diversity,”
IEEE Transactions on Communications , vol. 58,no. 12, pp. 3405–3409, 2010.[119] A. He, S. Srikanteswara, K. K. Bae, T. R. Newman, J. H. Reed, W. H. Tranter, M. Sajadieh, and M. Verhelst, “Powerconsumption minimization for mimo systems — a cognitive radio approach,”
IEEE Journal on Selected Areas inCommunications , vol. 29, no. 2, pp. 469–479, 2011.[120] M. Di Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and L. Hanzo, “Spatial modulation for generalized mimo: Challenges,opportunities, and implementation,”
Proceedings of the IEEE , vol. 102, no. 1, pp. 56–103, 2014.[121] M. Haenggi,
Stochastic geometry for wireless networks . Cambridge University Press, 2012.[122] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular networks,”
IEEE Transactionson Communications , vol. 59, no. 11, pp. 3122–3134, Nov. 2011.[123] H. Dhillon, R. Ganti, F. Baccelli, and J. Andrews, “Modeling and analysis of K-tier downlink heterogeneous cellularnetworks,”
IEEE Journal on Selected Areas in Communications , vol. 30, no. 3, pp. 550–560, Apr. 2012. [124] T. Bai and R. W. Heath, “Coverage and Rate Analysis for Millimeter-Wave Cellular Networks,” IEEE Transactions onWireless Communications , vol. 14, no. 2, pp. 1100–1114, Feb. 2015.[125] M. Ding, P. Wang, D. L´opez-P´erez, G. Mao, and Z. Lin, “Performance impact of LoS and NLoS transmissions in densecellular networks,”
IEEE Transactions on Wireless Communications , vol. 15, no. 3, pp. 2365–2380, Mar. 2016.[126] T. Bai and R. W. Heath, “Analyzing uplink sinr and rate in massive mimo systems using stochastic geometry,”
IEEETransactions on Communications , vol. 64, no. 11, pp. 4592–4606, 2016.[127] Q. Zhang, H. H. Yang, T. Q. S. Quek, and J. Lee, “Heterogeneous cellular networks with los and nlos transmissions—therole of massive mimo and small cells,”
IEEE Transactions on Wireless Communications , vol. 16, no. 12, pp. 7996–8010,2017.[128] P. Parida and H. S. Dhillon, “Stochastic geometry-based uplink analysis of massive mimo systems with fractional pilotreuse,”
IEEE Transactions on Wireless Communications , vol. 18, no. 3, pp. 1651–1668, 2019.[129] E. Bj¨ornson, L. Sanguinetti, and M. Kountouris, “Deploying dense networks for maximal energy efficiency: Small cellsmeet massive mimo,”
IEEE Journal on Selected Areas in Communications , vol. 34, no. 4, pp. 832–847, 2016.[130] Y. Xin, D. Wang, J. Li, H. Zhu, J. Wang, and X. You, “Area Spectral Efficiency and Area Energy Efficiency of MassiveMIMO Cellular Systems,”
IEEE Transactions on Vehicular Technology , vol. 65, no. 5, pp. 3243–3254, 2016.[131] E. Bj¨ornson, L. Sanguinetti, and M. Kountouris, “Designing wireless broadband access for energy efficiency: Are smallcells the only answer?” in , 2015, pp. 136–141.[132] A. Zappone, L. Sanguinetti, G. Bacci, E. Jorswieck, and M. Debbah, “Energy-efficient power control: A look at 5gwireless technologies,”
IEEE Transactions on Signal Processing , vol. 64, no. 7, pp. 1668–1683, 2016.[133] A. Zappone, E. Bj¨ornson, L. Sanguinetti, and E. Jorswieck, “Globally optimal energy-efficient power control and receiverdesign in wireless networks,”
IEEE Transactions on Signal Processing , vol. 65, no. 11, pp. 2844–2859, 2017.[134] 3GPP TSG RAN, “TR 37.910, Study on new radio access technology: Radio access architecture and interfaces,”
V14.0.0 ,Mar. 2017.[135] W. Yang, K. Lin, and H. Wei, “5g on-demand si acquisition framework and performance evaluation,”
IEEE Access , vol. 7,pp. 163 245–163 261, 2019.[136] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof power model for cellular base stations,” in
IEEE81st Vehicular Technology Conference (VTC Spring) , 2015, pp. 1–7.[137] F. E. Salem, T. Chahed, Z. Altman, and A. Gati, “Traffic-aware Advanced Sleep Modes management in 5G networks,”in , 2019, pp. 1–6.[138] F. E. Salem, “Management of advanced sleep modes for energy-efficient 5G networks,” Theses, Institut Polytechniquede Paris, Dec. 2019. [Online]. Available: https://tel.archives-ouvertes.fr/tel-02500618[139] S. Tombaz, P. Frenger, M. Olsson, and A. Nilsson, “Energy performance of 5g-nx radio access at country level,” in ,2016, pp. 1–6.[140] F. E. Salem, A. Gati, Z. Altman, and T. Chahed, “Advanced Sleep Modes and Their Impact on Flow-Level Performanceof 5G Networks,” in , 2017, pp. 1–7.[141] R. Tano, M. Tran, and P. Frenger, “KPI Impact on 5G NR Deep Sleep State Adaption,” in
IEEE 90th Vehicular TechnologyConference (VTC2019-Fall) , 2019, pp. 1–5.[142] Orange, “RP-190326, NR enhancements for Advanced Sleep Modes,” , Mar. 2019.[143] 3GPP TSG RAN, “TR 36.927, Potential solutions for energy saving for E-UTRAN,” V16.0.0 , July 2020.[144] T. Beitelmal, S. S. Szyszkowicz, D. G. Gonz´alez, and H. Yanikomeroglu, “Sector and site switch-off regular patterns for energy saving in cellular networks,” IEEE Transactions on Wireless Communications , vol. 17, no. 5, pp. 2932–2945,2018.[145] X. Lin, D. Yu, and H. Wiemann, “A primer on bandwidth parts in 5g new radio,” arXiv preprint arXiv:2004.00761 , 2020.[146] T. Kim, Y. Kim, Q. Lin, F. Sun, J. Fu, Y. Kim, A. Papasakellariou, H. Ji, and J. Lee, “Evolution of Power SavingTechnologies for 5G New Radio,”
IEEE Access , vol. 8, pp. 198 912–198 924, 2020.[147] M. Feng, S. Mao, and T. Jiang, “Base Station ON-OFF Switching in 5G Wireless Networks: Approaches and Challenges,”
IEEE Wireless Communications , vol. 24, no. 4, pp. 46–54, 2017.[148] ——, “BOOST: Base Station on-off Switching Strategy for Green Massive MIMO HetNets,”
IEEE Transactions onWireless Communications , vol. 16, no. 11, pp. 7319–7332, 2017.[149] L. Chiaraviglio, F. Cuomo, M. Listanti, E. Manzia, and M. Santucci, “Fatigue-Aware Management of Cellular NetworksInfrastructure with Sleep Modes,”
IEEE Transactions on Mobile Computing , vol. 16, no. 11, pp. 3028–3041, 2017.[150] H. Celebi, Y. Yapici, I. Guvenc¸, and H. Schulzrinne, “Load-Based On/Off Scheduling for Energy-Efficient Delay-Tolerant5G Networks,”
IEEE Transactions on Green Communications and Networking , vol. 3, no. 4, pp. 955–970, 2019.[151] N. Saxena, A. Roy, and H. Kim, “Traffic-Aware Cloud RAN: A Key for Green 5G Networks,”
IEEE Journal on SelectedAreas in Communications , vol. 34, no. 4, pp. 1010–1021, 2016.[152] Z. Niu, Y. Wu, J. Gong, and Z. Yang, “Cell zooming for cost-efficient green cellular networks,”
IEEE CommunicationsMagazine , vol. 48, no. 11, pp. 74–79, 2010.[153] A. Conte, A. Feki, L. Chiaraviglio, D. Ciullo, M. Meo, and M. A. Marsan, “Cell wilting and blossoming for energyefficiency,”
IEEE Wireless Communications , vol. 18, no. 5, pp. 50–57, 2011.[154] H. Jiang, S. Yi, L. Wu, H. Leung, Y. Wang, X. Zhou, Y. Chen, and L. Yang, “Data-Driven Cell Zooming for Large-ScaleMobile Networks,”
IEEE Transactions on Network and Service Management , vol. 15, no. 1, pp. 156–168, 2018.[155] G. Vallero, D. Renga, M. Meo, and M. A. Marsan, “Greener RAN Operation Through Machine Learning,”
IEEETransactions on Network and Service Management , vol. 16, no. 3, pp. 896–908, 2019.[156] 3GPP TSG SA, “TS 28.310, Management and orchestration; Energy efficiency of 5G ,”
V16.2.0 , Sept. 2020.[157] 3GPP TSG RAN, “TR 37.816, Study on RAN-centric data collection and utilization for LTE and NR,”
V16.0.0 , July2019.[158] K. N. R. S. V. Prasad, E. Hossain, and V. K. Bhargava, “Energy efficiency in massive mimo-based 5g networks:Opportunities and challenges,”
IEEE Wireless Communications , vol. 24, no. 3, pp. 86–94, 2017.[159] S. Asaad, A. M. Rabiei, and R. R. M¨uller, “Massive MIMO With Antenna Selection: Fundamental Limits andApplications,”
IEEE Transactions on Wireless Communications , vol. 17, no. 12, pp. 8502–8516, 2018.[160] X. Gao, O. Edfors, J. Liu, and F. Tufvesson, “Antenna selection in measured massive MIMO channels using convexoptimization,” in
IEEE Globecom Workshops (GC Wkshps) , 2013, pp. 129–134.[161] J. Xu and L. Qiu, “Energy efficiency optimization for mimo broadcast channels,”
IEEE Transactions on WirelessCommunications , vol. 12, no. 2, pp. 690–701, 2013.[162] N. P. Le, F. Safaei, and L. C. Tran, “Antenna selection strategies for mimo-ofdm wireless systems: An energy efficiencyperspective,”
IEEE Transactions on Vehicular Technology , vol. 65, no. 4, pp. 2048–2062, 2016.[163] B. Makki, A. Ide, T. Svensson, T. Eriksson, and M. Alouini, “A genetic algorithm-based antenna selection approach forlarge-but-finite mimo networks,”
IEEE Transactions on Vehicular Technology , vol. 66, no. 7, pp. 6591–6595, 2017.[164] B.-J. Lee, S.-L. Ju, N.-I. Kim, and K.-S. Kim, “Enhanced Transmit-Antenna Selection Schemes for Multiuser MassiveMIMO Systems,”
Wireless Communications and Mobile Computing , Jun. 2017.[165] M. Arash, E. Yazdian, Mohammad, S. Fazel, G. Brante, and M. A. Imran, “Employing antenna selection to improveenergy-efficiency in massive MIMO systems,”
Wireless Communications and Mobile Computing , Aug. 2017. [166] K. Senel, E. Bj¨ornson, and E. G. Larsson, “Joint Transmit and Circuit Power Minimization in Massive MIMO WithDownlink SINR Constraints: When to Turn on Massive MIMO?” IEEE Transactions on Wireless Communications ,vol. 18, no. 3, pp. 1834–1846, 2019.[167] W. Pramudito, E. Alsusa, A. Al-Dweik, K. A. Hamdi, D. K. C. So, and T. L. Marzetta, “Load-Aware Energy EfficientAdaptive Large Scale Antenna System,”
IEEE Access , vol. 8, pp. 82 592–82 606, 2020.[168] Y. Li, “Deep Reinforcement Learning,”
CoRR , vol. abs/1810.06339, October 2018. [Online]. Available:http://arxiv.org/abs/1810.06339[169] T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”
IEEE Transactions on CognitiveCommunications and Networking , vol. 3, no. 4, pp. 563–575, 2017.[170] D. L. Jagerman, B. Melamed, and W. Willinger, “Stochastic modeling of traffic processes,”
Frontiers in queueing , pp.271–370, 1997.[171] R. J. Hyndman and G. Athanasopoulos,
Forecasting: principles and practice . OTexts, 2018.[172] A. Azari, P. Papapetrou, S. Denic, and G. Peters, “Cellular traffic prediction and classification: A comparative evaluationof lstm and arima,” in
Discovery Science , P. Kralj Novak, T. ˇSmuc, and S. Dˇzeroski, Eds. Springer InternationalPublishing, 2019, pp. 129–144.[173] F. Mart´ınez, M. P. Fr´ıas, M. D. P´erez, and A. J. Rivera, “A methodology for applying k-nearest neighbor to time seriesforecasting,”
Artificial Intelligence Review , vol. 52, no. 3, 2019.[174] C.-H. Wu, J.-M. Ho, and D.-T. Lee, “Travel-time prediction with support vector regression,”
IEEE transactions onintelligent transportation systems , vol. 5, no. 4, pp. 276–281, 2004.[175] S. Hochreiter and J. Schmidhuber, “Long short-term memory,”
Neural computation , vol. 9, no. 8, pp. 1735–1780, 1997.[176] X. Zhang and J. You, “A gated dilated causal convolution based encoder-decoder for network traffic forecasting,”
IEEEAccess , vol. 8, pp. 6087–6097, 2020.[177] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, u. Kaiser, and I. Polosukhin, “Attention is allyou need,” ser. NIPS’17. Red Hook, NY, USA: Curran Associates Inc., 2017, p. 6000–6010.[178] X. Wang, Z. Zhou, F. Xiao, K. Xing, Z. Yang, Y. Liu, and C. Peng, “Spatio-temporal analysis and prediction of cellulartraffic in metropolis,”
IEEE Transactions on Mobile Computing , vol. 18, no. 9, pp. 2190–2202, 2019.[179] M. S. Hossain and G. Muhammad, “A deep-tree-model-based radio resource distribution for 5g networks,”
IEEE WirelessCommunications , vol. 27, no. 1, pp. 62–67, 2020.[180] J. Wang, J. Tang, Z. Xu, Y. Wang, G. Xue, X. Zhang, and D. Yang, “Spatiotemporal modeling and prediction in cellularnetworks: A big data enabled deep learning approach,” in
IEEE INFOCOM 2017 - IEEE Conference on ComputerCommunications , 2017, pp. 1–9.[181] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio,
Deep learning . MIT press Cambridge, 2016, vol. 1, no. 2.[182] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-c. Woo, “Convolutional lstm network: A machine learningapproach for precipitation nowcasting,”
Advances in neural information processing systems , vol. 28, pp. 802–810, 2015.[183] C. Zhang, H. Zhang, D. Yuan, and M. Zhang, “Citywide cellular traffic prediction based on densely connectedconvolutional neural networks,”
IEEE Communications Letters , vol. 22, no. 8, pp. 1656–1659, 2018.[184] C. Zhang and P. Patras, “Long-term mobile traffic forecasting using deep spatio-temporal neural networks,” in
Proceedingsof the Eighteenth ACM International Symposium on Mobile Ad Hoc Networking and Computing . New York, NY, USA:Association for Computing Machinery, 2018, p. 231–240. [Online]. Available: https://doi.org/10.1145/3209582.3209606[185] F. Xu, Y. Li, H. Wang, P. Zhang, and D. Jin, “Understanding mobile traffic patterns of large scale cellular towers in urbanenvironment,”
IEEE/ACM Transactions on Networking , vol. 25, no. 2, pp. 1147–1161, 2017. [186] C. Zhang, H. Zhang, J. Qiao, D. Yuan, and M. Zhang, “Deep transfer learning for intelligent cellular traffic predictionbased on cross-domain big data,” IEEE Journal on Selected Areas in Communications , vol. 37, no. 6, pp. 1389–1401,2019.[187] X. Wang, T. Yang, Y. Cui, Y. Jin, and H. Wang, “Bsenet: A data-driven spatio-temporal representation learning for basestation embedding,”
IEEE Access , vol. 8, pp. 51 674–51 683, 2020.[188] P. V. Klaine, M. A. Imran, O. Onireti, and R. D. Souza, “A survey of machine learning techniques applied to self-organizingcellular networks,”
IEEE Communications Surveys Tutorials , vol. 19, no. 4, pp. 2392–2431, 2017.[189] K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, “Deep Reinforcement Learning: A Brief Survey,”
IEEE Signal Processing Magazine , vol. 34, no. 6, pp. 26–38, 2017.[190] T. Jaakkola, M. I. Jordan, and S. P. Singh, “On the convergence of stochastic iterative dynamic programming algorithms,”
Neural Computation , vol. 6, no. 6, pp. 1185–1201, 1994.[191] A. De Domenico and D. Kt´enas, “Reinforcement learning for interference-aware cell dtx in heterogeneous networks,” in , 2018, pp. 1–6.[192] L. Zadeh, “The concept of a linguistic variable and its application to approximate reasoning—i,”
Information Sciences ,vol. 8, no. 3, pp. 199 – 249, 1975.[193] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland,G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis,“Human-level control through deep reinforcement learning,”
Nature , vol. 518, no. 7540, pp. 529–533, 2015.[194] J. Liu, B. Krishnamachari, S. Zhou, and Z. Niu, “DeepNap: Data-Driven Base Station Sleeping Operations Through DeepReinforcement Learning,”
IEEE Internet of Things Journal , vol. 5, no. 6, pp. 4273–4282, 2018.[195] T. Schaul, J. Quan, I. Antonoglou, and D. Silver, “Prioritized experience replay,” in
International Conference on LearningRepresentations , Puerto Rico, 2016.[196] J. Ye and Y. J. A. Zhang, “Drag: Deep reinforcement learning based base station activation in heterogeneous networks,”
IEEE Transactions on Mobile Computing , vol. 19, no. 9, pp. 2076–2087, 2020.[197] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control withdeep reinforcement learning,” 2019.[198] M. Di Renzo, S. Wang, and X. Xi, “Inhomogeneous double thinning—modeling and analysis of cellular networks by usinginhomogeneous poisson point processes,”
IEEE Transactions on Wireless Communications , vol. 17, no. 8, pp. 5162–5182,2018.[199] Q. Cui, X. Yu, Y. Wang, and M. Haenggi, “The sir meta distribution in poisson cellular networks with base stationcooperation,”
IEEE Transactions on Communications , vol. 66, no. 3, pp. 1234–1249, 2018.[200] S. Wang and M. Di Renzo, “On the mean interference-to-signal ratio in spatially correlated cellular networks,”
IEEEWireless Communications Letters , vol. 9, no. 3, pp. 358–362, 2020.[201] G. Ghatak, A. De Domenico, and M. Coupechoux, “Coverage analysis and load balancing in hetnets with millimeterwave multi-rat small cells,”
IEEE Transactions on Wireless Communications , vol. 17, no. 5, pp. 3154–3169, 2018.[202] G. Caire, “On the ergodic rate lower bounds with applications to massive mimo,”
IEEE Transactions on WirelessCommunications , vol. 17, no. 5, pp. 3258–3268, 2018.[203] M. J. Nawrocki, M. Dohler, and A. H. Aghvami,
Understanding UMTS Radio Network Modelling, Planning and AutomatedOptimisation: Theory and Practice , 1st ed. John Wiley & Sons Ltd., Apr. 2006.[204] J. Laiho, A. Wacker, and T. Novosad,
Radio Network Planning and Optimisation for UMTS , 2nd ed. John Wiley &Sons Ltd., Feb. 2006.[205] J. T. J. Penttinen,
5G Network Planning and Optimization , 2019, pp. 255–269. [206] L. Wan, Z. Guo, and X. Chen, “Enabling efficient 5g nr and 4g lte coexistence,” IEEE Wireless Communications , vol. 26,no. 1, pp. 6–8, 2019.[207] O. N. C. Yilmaz, O. Teyeb, and A. Orsino, “Overview of lte-nr dual connectivity,”
IEEE Communications Magazine ,vol. 57, no. 6, pp. 138–144, 2019.[208] D. L´opez-P´erez, J. Ling, B. H. Kim, V. Subramanian, S. Kanugovi, and M. Ding, “Boosted wifi through lte small cells:The solution for an all-wireless enterprise,” in , 2016, pp. 1–6.[209] D. L´opez-P´erez, M. Ding, H. Claussen, and A. H. Jafari, “Towards 1 gbps/ue in cellular systems: Understanding ultra-dense small cell deployments,”
IEEE Communications Surveys Tutorials , vol. 17, no. 4, pp. 2078–2101, 2015.[210] M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial on beam management for 3gpp nr at mmwavefrequencies,”
IEEE Communications Surveys Tutorials , vol. 21, no. 1, pp. 173–196, 2019.[211] A. Mohamed, O. Onireti, M. A. Imran, A. Imran, and R. Tafazolli, “Control-data separation architecture for cellular radioaccess networks: A survey and outlook,”
IEEE Communications Surveys Tutorials , vol. 18, no. 1, pp. 446–465, 2016.[212] N. Shlezinger, J. Whang, Y. C. Eldar, and A. G. Dimakis, “Model-based deep learning,” 2020.[213] A. Zappone, M. Di Renzo, and M. Debbah, “Wireless networks design in the era of deep learning: Model-based, ai-based,or both?”
IEEE Transactions on Communications , vol. 67, no. 10, pp. 7331–7376, 2019.[214] R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green ai,”
Commun. ACM , vol. 63, no. 12, p. 54–63, Nov. 2020.[Online]. Available: https://doi.org/10.1145/3381831[215] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neuralnetworks,” in
Advances in Neural Information Processing Systems , F. Pereira, C. J. C. Burges, L. Bottou,and K. Q. Weinberger, Eds., vol. 25. Curran Associates, Inc., 2012, pp. 1097–1105. [Online]. Available:https://proceedings.neurips.cc/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf[216] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel et al. , “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”
Science , vol. 362,no. 6419, pp. 1140–1144, 2018.[217] E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in NLP,” in
Proceedingsof the 57th Annual Meeting of the Association for Computational Linguistics
Foundations and Trends® inMachine Learning , vol. 14, no. 1, pp. –, 2021. [Online]. Available: http://dx.doi.org/10.1561/2200000083[219] Y. Deng, “Deep Learning on Mobile Devices: a Review,” in
Mobile Multimedia/Image Processing, Security, andApplications , vol. 10993. International Society for Optics and Photonics, 2019, p. 109930A.[220] J. Lee, N. Chirkov, E. Ignasheva, Y. Pisarchyk, M. Shieh, F. Riccardi, R. Sarokin, A. Kulik, and M. Grundmann, “On-device Neural Net Inference with Mobile GPUs,” arXiv preprint arXiv:1907.01989 , 2019.[221] L. Prechelt,
Early Stopping — But When?
Berlin, Heidelberg: Springer Berlin Heidelberg, 2012, pp. 53–67.[222] N. Piovesan, A. F. Gambin, M. Miozzo, M. Rossi, and P. Dini, “Energy sustainable paradigms and methods for futuremobile networks: A survey,”
Computer Communications , vol. 119, pp. 101–117, 2018.[223] Q. Wu, G. Y. Li, W. Chen, D. W. K. Ng, and R. Schober, “An overview of sustainable green 5g networks,”
IEEE WirelessCommunications , vol. 24, no. 4, pp. 72–80, 2017.[224] N. Piovesan and P. Dini, “Optimal direct load control of renewable powered small cells: A shortest path approach,”
Internet Technology Letters , vol. 1, no. 1, p. e7, 2018. [225] J. Gong, J. S. Thompson, S. Zhou, and Z. Niu, “Base station sleeping and resource allocation in renewable energy poweredcellular networks,” IEEE Transactions on Communications , vol. 62, no. 11, pp. 3801–3813, 2014.[226] N. Piovesan, D. L´opez-P´erez, M. Miozzo, and P. Dini, “Joint load control and energy sharing for renewable poweredsmall base stations: a machine learning approach,”
IEEE Transactions on Green Communications and Networking , pp.1–1, 2020.[227] H. Li, H. Gao, T. Lv, and Y. Lu, “Deep q-learning based dynamic resource allocation for self-powered ultra-densenetworks,” in . IEEE, 2018, pp.1–6.[228] M. Mendil, A. De Domenico, V. Heiries, R. Caire, and N. Hadjsaid, “Battery-aware optimization of green small cells:Sizing and energy management,”