PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on 5G Radio Access Network Energy Efficiency: Massive MIMO, Lean Carrier Design, Sleep Modes, and Machine Learning

David Lopez-Perez, Antonio De Domenico, Nicola Piovesan, Harvey Bao, Geng Xinli, Song Qitao, Merouane Debbah

arXiv:2101.11246v2cs.NI

TL;DR

The paper addresses the energy-efficiency challenge in 5G networks. It surveys power-consumption models, energy-efficiency metrics, and four 3GPP NR enabling technologies, supporting progress toward sustainable communication systems.

  • Problem

    Understanding and practically addressing the energy-efficiency challenge in 5G networks remains an important research focus.

  • Method

    The paper surveys base-station power-consumption models and metrics, then reviews mMIMO, lean carrier design, ASMs, and machine learning as 3GPP NR enabling technologies.

  • Results

    The survey provides an overview of energy-efficiency research and reviews the impact of four 3GPP NR enabling technologies.

  • Takeaways & Limitations

    The synthesis supports continued progress toward a sustainable communication system.

Abstract

from arXiv · show

Cellular networks have changed the world we are living in, and the fifth generation (5G) of radio technology is expected to further revolutionise our everyday lives, by enabling a high degree of automation, 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 provides tools to significantly decrease the energy consumption and the green house emissions of next generations networks, thus contributing towards information and communication technology (ICT) sustainability targets. In this survey paper, we thoroughly review the state-of-the-art on current energy efficiency research. We first categorise and carefully analyse the different power consumption models and energy efficiency metrics, which have helped to make progress on the understanding of green networks. Then, as a main contribution, we survey in detail -- from a theoretical and a practical viewpoint -- the main energy efficiency enabling technologies that 3GPP NR provides, together with their main benefits and challenges. Special attention is paid to four key enabling technologies, i.e., massive multiple-input multiple-output (MIMO), lean carrier design, and advanced idle modes, together with the role of artificial intelligence capabilities. We dive into their implementation and operational details, and thoroughly discuss their optimal operation points and theoretical-trade-offs from an energy consumption perspective. This will help the reader to grasp the fundamentals of -- and the status on -- green networking. Finally, the areas of research where more effort is needed to make future networks greener are also discussed.

I. INTRODUCTION

5G is presented as both an energy-efficiency opportunity and a sustainability challenge: it can enable more efficient resource use, but greater traffic, network density, and demanding services increase energy pressures. The survey reviews the technologies, models, metrics, and operating trade-offs relevant to greener 5G RANs.

  • Survey Scope: The survey focuses on power models, energy-efficiency metrics, and four 3GPP NR technologies: massive MIMO, lean carrier design, advanced sleep modes, and machine learning.It examines their theoretical and practical operation, benefits, challenges, and energy-efficiency trade-offs.
  • The Enabling Role of 5G: 5G communication capabilities support flexible and efficient resource use across government, industry, and business applications.The paper links this enabling effect to smart energy management, reduced office space and travel, and more efficient supply chains.
  • The Enabling Role of 5G: 2,135 million tones of CO2e was the estimated enabling effect of mobile communications in 2018, while the ICT-wide effect was predicted to reach 15 % of global emissions by the end of 2020.The cited estimates are used to illustrate the scale of mobile communications’ indirect sustainability contribution.
  • The 5G Energy Efficiency Challenge: The RAN accounted for about 57 % of total network energy consumption and was projected to remain the largest consumer at 50.6 % in 2025.Within a base station, transceivers and cables typically used about 65 % of energy, followed by cooling, signal processing, and AC-DC conversion.

C. Comparison to previous 5G Energy Efficiency Surveys

Earlier surveys covered energy efficiency through generic, LTE-focused, or specialized perspectives, but did not address 3GPP NR in detail. This survey provides a detailed, up-to-date 5G NR-focused synthesis of enabling technologies, models, trade-offs, and research directions.

  • Limitations of prior work: Some earlier treatments were qualitative and lacked detail, often using power models that accounted only for transmit power.Such models neglected equipment-hardware consumption, limiting their treatment of practical energy efficiency.
  • Previous survey coverage: Earlier surveys examined energy efficiency through trade-offs, deployment strategies, resource management, and green-network technologies.Their scopes included cellular and Wi-Fi algorithms, energy harvesting, RAN resource allocation, and several efficiency metrics.
  • Gap: Several prior works were generic or focused on 3GPP LTE rather than the assets of 3GPP NR.Consequently, they did not survey the energy-saving opportunities introduced by the new technology generation.
  • This survey: This survey reviews mMIMO, lean carrier design, 5G sleep modes, and machine learning for sub-6 GHz 5G RAN energy efficiency.It addresses these technologies from both theoretical and practical perspectives.
  • This survey: The survey presents power-consumption models, energy-efficiency metrics, theoretical bounds, trade-offs, and optimal operation points.It also distinguishes practical concepts inherited from earlier generations from energy-efficiency enhancements specified by 3GPP NR.
  • Open research directions: It highlights spatio-temporal traffic prediction and machine-learning approaches as open research directions for maximizing energy efficiency.The survey also provides more detailed descriptions than earlier surveys, including relevant base-station power formulations.

II. 3GPP NR ENERGY EFFICIENCY RELATED ENABLING TECHNOLOGIES

3GPP NR introduces a beam-centric design, ultra-lean signalling, sleep modes, and analytics capabilities that support more energy-efficient RAN operation. The survey examines their mechanisms, benefits, trade-offs, and optimization challenges, especially for mMIMO, carrier design, and dynamic component shutdown.

  • 3GPP NR beam-centric design: 3GPP NR’s beam-centric design uses many antenna elements for beamforming and larger spatial multiplexing.These gains can reduce transmit power for a target distance or end-user QoS demand.
  • Machine learning: 3GPP NR’s distributed architecture and centralized data gathering support machine-learning policies for network-wide energy-efficient operation.NWDAF and MDAF can analyze collected data or enhance network functions with statistics and prediction capabilities.
  • Massive MIMO: mMIMO can increase network capacity with the number of spatial streams multiplexed, but requires careful optimization.Its performance is affected by CSI acquisition constraints, finite orthogonal pilots, pilot reuse, and pilot contamination.
  • Massive MIMO: 3GPP NR supports mMIMO through beamformed control and synchronization, new CSI reports and measurements, channel reciprocity, and analog or digital beamforming.The survey analyses its energy-efficiency benefits and trade-offs in single- and multi-cell settings.
  • Lean carrier design: Ultra-lean carrier design transmits control signals more sparsely and on demand according to traffic requirements.This reduces overhead and interference while enabling longer and deeper sleep periods, including sleep periods up to 160 ms.
  • Sleep modes: 5G sleep modes dynamically deactivate base stations, carriers, channels, or antennas so underutilized resources can be shut down.More components shut down for longer periods can save more energy, while lean signalling facilitates micro-sleeps up to 160 ms.

D. Machine Learning

Machine learning is surveyed as a way to model complex, dynamic 5G behavior and optimize energy efficiency, while requiring sufficient data and incurring exploration and computing costs.

  • Machine-learning models can replace rule-based heuristics with learned parameters for configuring 5G energy-efficiency mechanisms.The surveyed mechanisms include massive MIMO, lean carriers, and 5G sleep modes.
  • Supervised and unsupervised learning use network measurements to model 5G behavior and make predictions or decisions in complex scenarios.
  • 5G energy minimization is a large-scale problem shaped by base-station and user distributions, traffic demands, wireless channels, and hidden trade-offs.
  • Reinforcement learning lets agents interact with the network environment to learn resource-shutdown policies that minimize total energy consumption.Shutdown decisions form a combinatorial problem with many variables.
  • Adding training data generally improves machine-learning performance, particularly for recent models.Figure 5 presents the relation between available training data and model performance.
  • Machine-learning approaches may require long exploration phases, substantial computing and storage, and adaptation to changing network settings.Policies derived for limited system configurations may not transfer directly as network settings vary.

1) Carrier Aggregation Power Consumption Model:

The section surveys power-consumption models for carrier aggregation, massive MIMO, distributed RAN, and centralized RAN architectures. It emphasizes that functional splits and transport-network power materially affect centralized-RAN energy characterization.

  • 1) Carrier Aggregation Power Consumption Model:: Carrier aggregation power models account for the number and bandwidth of active component carriers and shared hardware consumption.Contiguous aggregation may share one FFT and RF transceiver, whereas non-contiguous aggregation usually requires multiple units.
  • 1) Carrier Aggregation Power Consumption Model:: In the worst case, carrier-aggregation load-independent circuit power scales linearly with the number of active component carriers.
  • TRX depends on the type of beamforming architecture: Massive-MIMO models extend linear power models to represent many antenna elements, RF transceivers, digital processing, and load-dependent circuit power.RF transceiver modules need not equal the number of antenna elements and may be fewer.
  • TRX depends on the type of beamforming architecture: Massive-MIMO power consumption includes analog-front-end, digital-baseband, control and backhaul, and power-system contributions, alongside coding, channel-estimation, and beamforming terms.
  • Centralization offers stacking, pooling, and cooling gains, but lower functional splits provide larger centralization gains while increasing power consumption.
  • Transport contributes around 2% of network power for split option 6, around 30% for option 7, and around 60% for option 8.

IV. ENERGY EFFICIENCY METRICS

The survey reviews standardized and academic energy-efficiency metrics that combine network energy use with coverage, capacity, delay, traffic load, deployment scenarios, and economic cost. It highlights weighting choices and metric limitations for comparing 5G networks.

  • Energy-efficiency metrics should be comprehensive, reliable, and accepted for comparison while capturing energy consumption and network-level performance.Relevant performance dimensions include coverage, capacity, and delay.
  • Standardization bodies including ETSI, ITU-T, and 3GPP specify mobile-network energy-efficiency metrics for different operating conditions.
  • The EEDV metric can misrepresent low-load savings because it is not weighted according to the global reference.Small savings at low consumption may appear large, while large high-consumption savings may appear limited.
  • EETotal [bit/J] weights deployment-specific energy efficiencies by the typicality of each deployment scenario.It is intended for large-scale network characterization.
  • E3 [bit/J] extends energy-efficiency evaluation by weighting effective throughput and energy consumption with deployment and operational cost.It can distinguish solutions with similar throughput and energy consumption by their additional costs.

B. Energy Efficiency Metrics for Noise-Limited Networks

The survey presents coverage-, slice-, and link-aware energy-efficiency metrics for evaluating 5G networks under differing deployment objectives and service requirements. It also compares weighted link-level objectives that balance aggregate efficiency and fairness.

  • Coverage-aware metrics: Coverage energy efficiency metrics account for covered area and network energy consumption, making them suitable for rural and low-power wide-area deployments.The mobile network coverage energy efficiency metric is measured in m^2/J over one year.
  • Coverage-aware metrics: Estimating energy efficiency over actual coverage areas can require many UE measurement reports.Coverage-area energy computation is complex because technology effects must be estimated from measurements.
  • Slice-aware metrics: 3GPP defines distinct energy-efficiency metrics for eMBB, URLLC, and mMTC slices, reflecting their different service objectives.URLLC metrics incorporate latency, while mMTC efficiency uses the number of UEs relative to slice energy consumption.
  • Link-aware metrics: Weighted-product maximization prevents zero throughput on any link and converges to the Nash Bargaining solution.By contrast, maximizing aggregate link efficiency can favor high-throughput links and limit cell-edge performance.
  • Link-aware metrics: The weighted sum energy-efficiency metric assigns link priorities through predefined weights to improve system fairness during resource allocation.Weights can be based on user-specific data plans.
  • Link-aware metrics: Max-min weighted efficiency equalizes weighted link efficiencies, and equal weights produce the same energy efficiency for every link.Figure 11 qualitatively compares WMEE, WSEE, and WPEE solutions within the energy-efficient Pareto region.

V. THEORETICAL UNDERSTANDING OF ENERGY EFFICIENCY: MASSIVE MIMO

The survey examines massive MIMO as an energy-efficiency enabler while emphasizing that antenna, user, transmit-power, and circuit-power choices create operating trade-offs. Its theoretical conclusions are strongest in simplified single-cell models.

  • System role and modelling: Massive MIMO can increase capacity and reduce required transmit power, but larger antenna arrays and signal processing increase base-station energy consumption.Energy efficiency is defined as achievable data rate divided by related power consumption.
  • Limitations: Theoretical massive-MIMO results are limited mainly by tractability, with most established findings relying on single-cell assumptions and simplified channel or power models.Channel correlation, pilot contamination, interference, and imperfect CSI are difficult to capture accurately in tractable network-level models.
  • System role and modelling: The survey focuses on energy-efficiency bounds and trade-offs for massive MIMO in both single-cell and multi-cell scenarios.It provides explicit closed-form expressions and examines how network parameters affect energy efficiency and power consumption.
  • Single-cell bounds: A 1/M transmit-power reduction can maintain non-zero rates as the antenna count M grows in TDD massive MIMO systems.Earlier idealized analyses suggested that energy efficiency could increase monotonically with M.
  • Single-cell bounds: Circuit-power-aware models yield different conclusions from idealized models because antenna and user circuit costs constrain energy-efficiency gains.The optimal multiplexed-user count decreases with per-UE and per-antenna circuit power, while coverage expansion requires more multiplexed UEs.
  • Optimal operating points: For fixed multiplexed users, increasing antennas improves energy efficiency only until rate gains no longer offset their power consumption.The resulting energy-efficiency curves are quasi-concave, with a maximum operating point.

2) Trade-offs:

The survey describes energy–spectral-efficiency trade-offs in massive MIMO and shows that circuit power can make the metrics conflict. Optimal operation depends on transmit power, active antennas, precoding, and channel knowledge.

  • Energy–spectral-efficiency trade-off: Circuit power grows with antenna count and can dominate base-station consumption, breaking the monotonic relation between energy and spectral efficiency.Hardware efficiency must improve accordingly, and the trade-off requires explicit optimization.
  • Limitations: One reviewed study omitted the number of multiplexed UEs, which may significantly influence the energy–spectral-efficiency trade-off.The survey also notes that many frameworks rely on intricate optimization algorithms rather than explicit trade-off equations.
  • Energy–spectral-efficiency trade-off: For a fixed active-antenna count, energy efficiency is quasi-concave in spectral efficiency, confirming a clear trade-off.Different numbers of active antennas produce different energy–spectral-efficiency curves.
  • Precoding and CSI: At low SNR, MRT has higher energy efficiency than ZF because of its lower complexity; at high spectral efficiency, ZF offers higher spectral efficiency for a given energy efficiency.ZF benefits from intra-cell interference cancellation.
  • Precoding and CSI: Statistical CSI provides more stable and accessible feedback than instantaneous CSI, but at the expense of lower network capacity.This approach is useful when instantaneous CSI is difficult to obtain or maintain.
  • Transmit-power trade-off: With the reported model, energy and spectral efficiency increase together with transmit power until the joint optimum at PTX = 35 dBm.Beyond that point, transmit power becomes the dominant consumption factor and further increases reduce energy efficiency.

B. Multi-cell scenario

Multi-cell mMIMO energy-efficiency analysis remains difficult because large-scale networks combine many parameters, complex protocols, and interference effects. Existing work uses stochastic-geometry and tractable-bound approaches to derive initial insights into capacity, pilot reuse, and optimal antenna and user dimensioning.

  • Challenges: Multi-cell mMIMO is less theoretically understood than single-cell mMIMO because large-scale networks are difficult to model and analyse.The difficulty reflects sophisticated protocols, many tunable parameters, and higher system complexity.
  • Challenges: No general closed-form expressions holistically describe multi-cell mMIMO energy-efficiency bounds and trade-offs.Research remains fragmented across particular system aspects and simplifying assumptions.
  • Analytical approaches: Stochastic geometry has been used to study multi-cell mMIMO capacity scaling, pilot contamination, heterogeneous deployments, and energy efficiency.Analyses include uplink and downlink settings, with QoS or spectral-efficiency constraints and tractable lower bounds for average-user performance.
  • Optimization insights: Energy-efficiency optimization identifies optimal multiplexed-UE and antenna counts while accounting for circuit power, transmit power, and pilot reuse.Reported gains arise from intra-cell interference suppression and sharing circuit-power costs among multiplexed UEs; larger pilot-reuse factors can protect against pilot contamination.
  • AEE–ASE trade-offs: Smaller ASE targets require fewer antennas and multiplexed UEs for optimal AEE, whereas higher ASE requirements reduce energy efficiency.Increasing BS density and adding antennas can improve energy efficiency, but density benefits saturate when circuit power dominates.

3) Downlink mMIMO network deployment perspectives:

Downlink mMIMO studies extend tractable stochastic-geometry analyses to energy-efficiency optimization under spectral-efficiency requirements, but often simplify channel knowledge and interference modeling. Their results support dimensioning mMIMO deployments and using QoS-aware optimization to improve energy efficiency.

  • Downlink modelling: Downlink mMIMO energy-efficiency analysis extends an uplink methodology using a tractable average-UE SINR bound and a BS power-consumption model.The optimization targets downlink energy efficiency while providing minimum spectral efficiency to the average UE.
  • Downlink modelling: Perfect CSI was assumed in the downlink model, so pilot contamination was neglected and model accuracy decreased.This simplification avoids jointly modeling uplink channel estimation and downlink data transmission.
  • Optimization results: The downlink analysis found the same general energy-efficiency conclusions as the uplink analysis.In one example, optimal mMIMO dimensioning used 193 antennas and 21 multiplexed UEs per BS.
  • Implications: QoS-aware mMIMO dimensioning is important because inadequate BS or antenna counts can produce suboptimal energy efficiency.The reviewed downlink results advocate deployment dimensioning and operation optimization according to end-user QoS demands.
  • Lean carrier and symbol shutdown: Lean carrier and symbol-shutdown mechanisms adapt BS hardware activity to short-term traffic variations while preserving service continuity and QoS.The survey covers time-domain mechanisms operating from hundreds of microseconds to hundreds of milliseconds.
  • Lean carrier and symbol shutdown: 3GPP NR permits sleep periods up to 159 ms and a 99.38% sleeping ratio when SS-burst periodicity is 160 ms.With 15 kHz subcarrier spacing and a 20 ms periodicity, sleep can reach 19 ms and 95%.
  • Lean carrier and symbol shutdown: Replacing 3GPP LTE with 3GPP NR and proposed sleep modes can provide up to 80% energy savings under low-to-medium traffic loads.The reported savings remain possible with stringent delay constraints.

VII. CARRIER-DOMAIN (CARRIER SHUTDOWN) ENERGY SAVING-BASED SOLUTIONS

Carrier-domain mechanisms address long-term traffic variation by switching carriers or cells into deep dormancy. They enable deeper sleeps than symbol shutdown, but operate on larger timescales and are not primarily designed to preserve cell service continuity.

  • Mechanism scope: Carrier shutdown mechanisms adapt network configuration to long-term traffic variations, including regular daily and weekly load patterns.Because networks are sized for peak traffic, low- and medium-load periods can waste energy.
  • Mechanism scope: Deep-dormancy schemes deactivate BS hardware for minutes or hours and allow deeper sleeps than symbol-shutdown mechanisms.They target long-term load changes rather than short-term traffic fluctuations.
  • Operational boundary: Carrier-domain mechanisms switch off carriers rather than preserving continuous cell service and therefore can operate over the top of the lean-carrier signalling framework.3GPP NR work is still needed to coordinate carrier activation and deactivation.
  • Approaches: The 3GPP considers intra-BS, inter-BS, and inter-RAT approaches to carrier-domain energy saving.These approaches differ in their degree of BS coordination and the types of BS involved.

A. Intra-BS energy saving mechanisms

Intra-BS mechanisms reduce energy use by adapting active carriers, bandwidth parts, and radio resources to current UE requirements. Their benefits must be evaluated network-wide because local shutdown decisions can alter coverage, load, interference, QoS, and hardware lifetime.

  • Intra-BS mechanisms: Dynamic control of active component carriers is presented as more feasible than switching off BS sectors when coverage risks are significant.Carrier aggregation and NR dormant states support faster adaptation to UE requirements.
  • Intra-BS mechanisms: 3GPP NR bandwidth parts restrict transmission and signalling to a configured spectrum portion, reducing BS and UE power consumption at low cell load.BWPs can be activated or deactivated through timers, DCI, or RRC signalling.
  • Network effects: BS-centric shutdown can leave UEs without coverage, increase neighbouring loads, and change inter-cell interference, rank, and MCS selection.These effects can influence packet success rate and broader network performance.
  • Network effects: Network-wide energy optimization therefore must account for coverage, load, interference, and QoS effects when adjusting individual BS configurations.The survey emphasizes that local energy decisions should not be evaluated independently of neighbouring cells.
  • Optimization results: A joint BS activation, UE association, and power-control solution achieved 20% energy savings while increasing blocking rate by only 0.1%.The approach was evaluated in a HetNet with mMIMO capabilities using mixed-integer programming and complexity-reduction methods.
  • Inter-BS coordination: Inter-BS distributed optimization offers lower energy savings and UE performance than centralized optimization but is proven to converge to a Nash equilibrium.The centralized scheme relaxes integer variables and iteratively solves activation, association, and power-control problems.
  • Hardware lifetime: Frequent deep transients between BS states increase hardware failure rates through temperature gradients and can raise maintenance costs.One LTE approach achieved around 30% power savings while keeping the acceleration factor close to one.

C. Inter-RAT energy saving mechanisms

Inter-RAT energy-saving mechanisms target low-demand 5G capacity cells, coordinating 3GPP NR with underlying LTE coverage to reduce consumption while preserving service requirements. Antenna-selection studies show that energy savings depend on circuit-power costs, traffic load, QoS, and coordination constraints.

  • Inter-RAT operation: 5G NR booster cells can be switched off under low traffic demand, requiring tight inter-working with the underlying LTE network.NR base stations consume more power than LTE because of wider bandwidth, more complex hardware, and larger mMIMO antenna arrays.
  • Inter-RAT operation: Release 16 enhancements let an LTE coverage cell request NR booster-cell reactivation over the S1/NG interface using cell-load information.The NR capacity booster cell can also autonomously switch off based on its own load.
  • Antenna selection: When circuit power per antenna is small, activating all BS antennas can minimize power; otherwise, deactivating a subset saves energy.The multi-cell case adds pilot-contamination interference that limits SINR gains and makes more antennas not necessarily optimal for total BS power.
  • Antenna selection: A geometric-programming relaxation can efficiently optimize antenna activation and transmit powers to deliver requested traffic with minimal power.The cited work nevertheless lacks detailed system-level analysis.
  • Antenna selection: Fine-grained antenna or transceiver deactivation can match activation to QoS demands, while longer sleeps may increase activation delay and latency.Practical schemes seek deeper sleeps and greater savings, creating a trade-off between energy reduction and responsiveness.

A. ML for Traffic Prediction

Accurate traffic forecasting is needed to support energy-efficient RAN decisions, but network demand varies across time and space and depends on complex mobility and environmental factors. The survey contrasts statistical and machine-learning approaches, emphasizing models that capture nonlinear and spatio-temporal dependencies.

  • Forecasting requirements: Traffic forecasting supports energy-efficient decisions such as carrier shutdown, but prediction must reflect changing QoS demands across time and space.UE mobility and urban factors create dependencies among neighboring and distant cells.
  • Forecasting requirements: Prediction time scale should match the energy-saving decision period: hourly carrier adjustment requires hourly forecasts, whereas daily operation requires 24-hour predictions.Longer-horizon prediction is more challenging.
  • Statistical methods: Statistical methods such as ARIMA rely on lagged time-series values but generally struggle with rapid variations, nonlinear behavior, and seasonal traffic patterns.SARIMA extends ARIMA to address seasonality.
  • Machine-learning methods: Machine-learning methods can model nonlinearities using the large data volumes collected by base stations, although traditional methods require tuning and have limited memory.These limitations can reduce prediction accuracy for complex traffic behavior.
  • Deep sequence models: LSTM networks address vanishing gradients and learn long-term time-series dependencies through input, forget, and output gates.Their recurrent architecture uses prior inputs, memory, and outputs to model sequence behavior.
  • Spatio-temporal modelling: LSTM-only methods may omit spatial dependencies, while incorporating cross-domain data or spatial information can improve prediction accuracy.One approach achieved similar performance with temporal models as with spatio-temporal models at the cost of a small training-time increase.

B. ML for 5G Energy Efficiency Optimisation

Machine learning and reinforcement learning are surveyed as tools for adapting 5G energy-saving decisions to complex, partially observed, and dynamic network conditions. Reported studies combine sleep-mode control, buffering, coordination, and learned policies to trade energy savings against QoS and latency.

  • Reinforcement-learning foundations: Wireless-network optimization is difficult because traffic, mobility, interference, and channel conditions vary dynamically and may be only partially observed.Delayed rewards further complicate learning by making action effects difficult to evaluate.
  • Reinforcement-learning foundations: Reinforcement learning maps network states to actions through interaction, maximizing reward while balancing exploration against exploitation.The agent receives new environment states and numerical rewards after acting.
  • Reinforcement-learning limitations: RL faces a curse of dimensionality because computational requirements grow exponentially with state and action-space size.Function approximation with linear functions or neural networks is used when spaces are large or continuous.
  • Distributed optimization: Fuzzy Q-learning jointly controls cell and backhaul-node DTX while accounting for buffer state, capacity, QoS, and interference.The distributed scheme coordinates neighboring BS activation and achieved up to 38% energy savings versus buffering-free baseline DTX.
  • Deep reinforcement learning: Action-wise experience replay and adaptive reward scaling improved learning stability and adaptability, with gains over baseline Q-learning in energy saving and UE QoS.These techniques address practical learning behavior in dynamic environments.
  • Deep reinforcement learning: Deep Q-learning reduced cumulative network cost by up to 30% versus Q-learning, improved stability under non-stationary traffic, and reduced convergence time.The reported framework used an action-exploration method to improve training behavior.

X. OPEN RESEARCH DIRECTIONS

The survey identifies unresolved theoretical, simulation, and planning challenges that limit energy-efficiency understanding in realistic large-scale 5G and future RANs. Future work must better capture heterogeneous deployments, complex channels, multi-cell interactions, and accurate power consumption.

  • Modelling gaps: Current energy-efficiency understanding is largely based on simplified average-user, uniform-network, channel, operational, and BS power-consumption models.These assumptions do not capture all relevant features of large-scale multi-cell RANs.
  • Modelling gaps: New theoretical analyses are needed for non-uniform BS and UE distributions, local performance distributions, and advanced features such as mMIMO.Existing frameworks for these settings remain in an early stage and have mostly addressed simpler single-antenna small-cell networks.
  • Channel models: Highly directional channels weaken the usual channel-hardening behavior, requiring new capacity and energy-efficiency bounds for realistic RAN configurations.Existing mMIMO bounds work best when useful-signal coefficients have non-zero means and small variance.
  • Simulation and analysis: Sophisticated numerical and system-level tools are needed to analyze large-scale multi-cell interactions among mMIMO, carrier aggregation, coordinated transmission, and power consumption.These tools can provide deployment guidance and motivate new theoretical research.
  • Network planning: RAN planning tools remain mainly capacity-driven rather than designed to derive energy-efficient deployments.Planning must account for topology, site deployments, UE and traffic distributions, and accurate BS power models.
  • Network planning: Energy-efficient planning must balance modelling accuracy against complexity and use flexible propagation, system-level simulation, and tailored optimization methods.Such methods should support practical choices between fewer sites with larger mMIMO arrays and more sites with smaller arrays.

C. Multi-carrier and Heterogeneous Network Analysis

5G heterogeneous deployments must coordinate NR and LTE activation, scheduling, and sleep modes according to traffic load and cell characteristics. Machine learning can support energy-aware optimisation, but practical deployment remains constrained by biased measurements, prediction uncertainty, data requirements, and computational cost.

  • Multi-carrier coexistence: NR and LTE may share spectrum, requiring dynamic time scheduling to enable their coexistence.NR and LTE can also be deployed at different frequencies, depending on spectrum availability.
  • Practical optimisation: Energy-efficiency optimisation must account for distinct NR and LTE coverage, bandwidth, antenna, interference, deployment, and resource-use characteristics.The literature lacks sufficient studies of practical technology inter-working from an energy-efficiency perspective.
  • Load-aware inter-working: Coordinated NR/LTE operation should adapt to traffic load: LTE may cover low-load periods, NR may serve medium loads, and both may aggregate at high loads.Downlink/uplink splitting can use NR for downlink and LTE for uplink, but only when energy-efficient and necessary.
  • Heterogeneous networks: Small and millimetre-wave cells can be activated on demand for capacity, while coordinated sleep modes and macrocell control-plane coverage support energy savings.Separating control and data planes allows macrocells to maintain connectivity while capacity cells provide locally needed high-rate transmissions.
  • Machine-learning limitations: Prediction accuracy should be evaluated in energy-savings terms because biased measurements, forecasting errors, and optimisation-induced load shifts can undermine energy-saving decisions.Large RAN datasets are difficult to acquire, processing is energy demanding, and ML computational costs can be substantial.
  • Machine-learning optimisation: Traffic prediction can guide reinforcement learning to decide when to activate or deactivate base-station functionality, potentially improving policy convergence and exploration.Joint supervised-learning and reinforcement-learning approaches can provide multi-step forecasts for network optimisation.
Loading 2101.11246v2…