Received 29 December 2023; revised 28 March 2024; accepted 20 May 2024.

Date of publication 3 June 2024; date of current version 18 June 2024.

The associate editor coordinating the review of this article and approving it for publication was Z. Yang.
Digital Object Identifier 10.1109/TMLCN.2024.3407691

Data-Driven Energy Efficiency Modeling in

Large-Scale Networks: An Expert
Knowledge and ML-Based
Approach
DAVID LÓPEZ-PÉREZ1 (Senior Member, IEEE), ANTONIO DE DOMENICO 2 (Member, IEEE),
NICOLA PIOVESAN 2 (Member, IEEE), AND MÉROUANE DEBBAH 3 (Fellow, IEEE)
1 Institute of Telecommunications and Media Applications, Universitat Politècnica de València, 46022 Valencia, Spain
2 Huawei Technologies, 92100 Boulogne-Billancourt, France
3 KU 6G Research Center, Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates

CORRESPONDING AUTHOR: A. DE DOMENICO ([email protected])

This work was supported in part by NSF Project under Grant CNS 1910153; in part by the Generalitat Valenciana through the CIDEGENT
PlaGenT, Project iTENTE, under Grant CIDEXG/2022/17; and in part by the MCIN/AEI/10.13039/501100011033 and the European Union
‘‘NextGenerationEU/PRTR’’ under Grant CNS2023-144333.

ABSTRACT The energy consumption of mobile networks poses a critical challenge. Mitigating this concern
necessitates the deployment and optimization of network energy-saving solutions, such as carrier shutdown,
to dynamically manage network resources. Traditional optimization approaches encounter complexity due
to factors like the large number of cells, stochastic traffic, channel variations, and intricate trade-offs.
This paper introduces the simulated reality of communication networks (SRCON) framework, a novel,
data-driven modeling paradigm that harnesses live network data and employs a blend of machine learning
(ML)- and expert-based models. These mix of models accurately characterizes the functioning of network
components, and predicts network energy efficiency and user equipment (UE) quality of service for any
energy carrier shutdown configuration in a specific network. Distinguishing itself from existing methods,
SRCON eliminates the reliance on expensive expert knowledge, drive testing, or incomplete maps for
predicting network performance. This paper details the pipeline employed by SRCON to decompose the
large network energy efficiency modeling problem into ML- and expert-based submodels. It demonstrates
how, by embracing stochasticity, and carefully crafting the relationship between such submodels, the overall
computational complexity can be reduced and prediction accuracy enhanced. Results derived from real
network data underscore the paradigm shift introduced by SRCON, showcasing significant gains over a state-
of-the-art method used by a operator for network energy efficiency modeling. The reliability of this local,
data-driven modeling of the network proves to be a key asset for network energy-saving optimization.

INDEX TERMS Cellular networks, 5G, 6G, energy efficiency, carrier shutdown, optimization, data,
machine learning (ML).

I. INTRODUCTION bands with wider bandwidths through radios1 equipped with

T HE energy consumption of mobile networks is concern-

ing. Recent studies have shown that, while third genera-
tion partnership project (3GPP) new radio (NR) deployments
additional radio frequency (RF) chains to enhance beam-
forming and multiplexing capabilities [2]. The 4× energy
efficiency increase is far from the 100× target of the inter-
are 4× more energy efficient than 3GPP long term evolu- national telecommunication union (ITU) for fifth generation
tion (LTE) ones, they consume up to 3× more energy [1].
1 A radio denotes a radio unit. A wideband radio can manage multiple
The increased energy consumption is primarily attributed to carriers/cells, and a base station (BS) site may comprise one or multiple
the deployment of denser networks handling new frequency radios.

2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
780 VOLUME 2, 2024
López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

(5G) networks [3]. The 3× larger energy consumption is pos- to privacy concerns, and they usually do it for a short period
ing a threat on the environmental and business sustainability of time.
of cellular networks, where the energy cost of a mobile net- Nevertheless, it is essential to acknowledge that cell-
work already accounts for 23 % of the total operator cost [4]. level statistics, CDRs, DTs, and minimization of DT alone
Since the radio access network (RAN) is the biggest con- are not sufficient for network optimization. These metrics
tributor to energy consumption in a mobile network, with assess network performance only for the specific configu-
an average share of 73 % [5], the mobile industry has been ration —or, at most, a few configurations— set during the
working in solutions to enhance network energy efficiency. measurement campaigns. They do not provide insights into
One key tool to minimize network energy consumption, while network performance across the billions of other potential
meeting the quality of experience of end-users, is to tailor network configurations. Consequently, state of the art trial
network resources to user equipment (UE) requirements [2]. and error optimization approaches based on expert knowl-
To this end, the 3GPP has defined various network energy- edge as well as cell-level statistics, CDRs and DT-based data
saving features to facilitate the implementation of different cannot accommodate for the requirements of today’s real
types of shutdown solutions, i.e. symbol, channel and carrier world 4G/5G large-scale network optimization.
shutdown, which allow the online (de)activation of time,
space and frequency network resources, respectively [6]. 2) SYSTEM-LEVEL SIMULATION
The potential of these network energy-saving solutions is,
System-level simulation is another widely used option for
however, suboptimally exploited in current fourth generation
assisting network optimization [11]. Precise modeling of the
(4G) and 5G network deployments due to the complexity of
propagation environment and protocol stack are crucial to
their optimization. The modelling and further optimization
obtain an accurate evaluation and generalization of network
of network energy-saving solutions are intricate problems,
performance to any network configuration.
largely unsolved in both industry and academia. The main
Available tools for modeling the wireless propagation envi-
challenges emanate from (i) the large number of cells and
ronment range from basic models that solely rely on statistics
parameters per cell to configure, (ii) the stochastic and non-
to advanced ray-tracing systems [12]. However, these precise
stationary nature of end-user traffic demands and the wireless
tools often require hard-to-obtain map information and time-
channel, and (iii) the intricate coupling/trade-offs between
consuming operations, making them expensive. Examples of
energy consumption and network/UE performance (in terms
such radio propagation prediction tools are [13], [14].
of coverage quality, throughput, reliability and latency) [6].
With regard to the protocol stack, frameworks such as
Effectively optimizing network energy-saving solutions,
NS-3 5G-LENA [15] and its evolution [16] have gained
not only enhances network energy efficiency and reduces
popularity due to their capabilities. Nevertheless, they fail at
operator costs, but also significantly contributes to environ-
modelling the complexity of true products. For example, they
mental sustainability.
do not take into account the heterogeneity —and different
performance— of off-the-shelf BS and UE products.
A. LITERATURE REVIEW
Various methodologies have been employed to tackle the 3) THEORETICAL TOOLS
issue of optimizing network energy-saving solutions. None
Theoretical tools have also been developed to aid network
of these techniques, however, have completely overcome the
energy efficiency optimization. Importantly, optimal shut-
challenges at hand. In the following, we review the most
down policies for a single server with bursty traffic were
common approaches to network energy-saving modelling and
derived in [17]. The optimal shutdown policy was shown to
optimization, introducing the role of ML.
be a two-threshold policy, with one threshold to drive the
sleep procedure and a different one to manage the cell wake
1) EXPERT KNOWLEDGE up. Even if there is no theoretical results available for more
Currently, expert knowledge still remains a primary means than one server, the findings of this work are in line with
for optimizing network parameters. Experts rely on cell- today’s carrier shutdown solutions, which implement two
level statistics [7], gathered from BSs, as well as call detail different sets of conditions for shutdown and reactivation (see
records (CDRs) [8] and drive tests (DTs) [9] measurements Section II-B).
to evaluate network performance and UE quality of service When theoretically studying more complex multi-cell
at various locations, and take decisions. However, CDRs and macrocellular networks, we should highlight the coverage,
DTs necessitate large measurement campaigns and special- capacity and energy efficiency trade-offs derived in [18],
ized equipment, which make them economically expensive [19], and [20] for large-scale cellular networks with massive
and diminish their worth. multiple-input multiple-output (mMIMO), carrier aggrega-
To alleviate this cost challenge, the 3GPP introduced min- tion (CA) and joint mMIMO and CA capabilities, respec-
imization of DT [10], which employs UE geolocated radio tively.
measurements to inexpensively assess network performance. In [21], the authors compared the energy efficiency trade-
However, only a few UEs enable this feature nowadays due offs associated with a macrocellular network equipped with

VOLUME 2, 2024 781

shutdown capabilities to those of a small cell network, dis- These ML estimators are designed to mitigate the complexity
cussing the pros and cons of these two deployment strategies inherent in channel detection procedures while minimizing
The further analysis in [22] showed the benefits of deploying performance degradation [33]. In terms of radio resource
advanced sleep modes at the small cells, proposing multiple management functionalities, new deep learning approaches
carrier shutdown strategies, and comparing them in terms of aim at streamlining wireless resource allocation algorithms,
complexity, blocking rate probability, throughput and energy showcasing a significant leap in managing the complexities of
efficiency. medium access control (MAC) layer network operations [34].
From an optimization perspective, the authors of [23] In contrast, only a few studies have explored the use of real-
investigated how to fine-tune network energy efficiency by world network data for modeling and enhancing large-scale
jointly managing the long-term cell activation, UE associ- networks, the subject at hand.
ation and power control in a heterogeneous network with
mMIMO capabilities, and proposed a distributed solution a: TRAFFIC PREDICTION
based on game theory, proven to converge to the Nash Within this domain, substantial strides have been made in
equilibrium. data-driven traffic prediction with real-world data employing
The work in [24] also modelled hardware failure rate due various ML techniques, such as long short term memory
to cell (de)activation, and proposed a heuristic to control the cells (LSTMs), convolutional neural networks (CNNs), and
BS shutdown, which minimises the acceleration over time of graph convolutional networks (GCNs) [35], [36], [37], [38].
the hardware failure rate, while satisfying the UE quality of These methodologies are pivotal for facilitating proactive
service demands. network optimization efforts. Cutting-edge traffic prediction
Although the models produced and insights gained from approaches are increasingly leveraging generative adversarial
these studies are valuable, their reliance on numerous networks (GANs), combined with traffic data and expert
assumptions —stemming from the theoretical nature and knowledge [39], [40]. Integration of urban environment
complexity of large-scale networks—limits their applicabil- knowledge graphs have further enriched traffic prediction
ity. For instance, the carrier shutdown scenarios explored models, improving their accuracy and thus the quality of
in this research significantly differ from those implemented network optimization.
in actual network environments. Therefore, these theoretical
approaches fall short when applied to the optimization of
b: ANOMALY DETECTION
energy-saving solutions in real-world network settings.
The application of data-driven modeling using real-network
data has also been significantly explored for anomaly detec-
4) DATA-DRIVEN MODELLING AND OPTIMIZATION tion and network alarm predictions. Advanced ML models,
Given the limitations noted, there is an increasing interest similar to those used for traffic prediction, have been adapted
in leveraging ML to develop more scalable and flexible to detect anomalies by learning from vast datasets of net-
approaches for network optimization [25]. Recognizing this work activity. These models are trained to recognize patterns
potential, the 3GPP has initiated a shift towards a new indicative of potential issues, enabling network operators
paradigm in data-driven network modeling and optimization. to preemptively address problems before they escalate [41],
This effort focuses on identifying the crucial measurements [42], [43], [44].
and protocols necessary to devise advanced data-driven
strategies for localized network optimization [26]. By apply- c: CELL RATE AND UE SPECTRAL EFFICIENCY
ing ML techniques, the aim is to predict network performance PREDICTIONS
accurately and improve various dimensions of network func- Gijon et al. assessed cell throughput prediction accuracy
tionality, notably energy efficiency [27]. using real network key performance indicator (KPI) coun-
To date, such explorations into data-driven network mod- ters, exploring a variety of ML methodologies such as
eling and optimization have predominantly focused on support vector regression, k-nearest neighbors, and decision
developing ML models that harness data from numeri- trees and artificial neural network (ANN)-based ensemble
cal simulations and test environments. These models are approaches [45]. Advancing this research, Xing et al. intro-
not just academic exercises, but are intended to introduce duced a neural network framework aimed at modeling the
new mechanisms at the physical layer of communication interference and spectral efficiency of downlink (DL) 5G
systems [28], as well as to refine the functionalities at mMIMO transmissions, employing reference signal received
the network’s upper layers [29]. Significant advancements power (RSRP) data from DTs [46]. Their work underscores
in physical layer mechanisms include the design of neu- the potential of ML in accurately forecasting cell throughput
ral receivers specifically for 5G multi-UE multiple-input and enhancing understanding of UE spectral efficiencies.
multiple-output (MIMO) configurations [30], alongside the However, these models lack the capability to dynamically
introduction of deep learning-based synchronization [31] and forecast KPIs in response to changes in network configu-
decoding schemes [32]. Intelligent MIMO channel estima- rations, revealing a significant gap in their adaptability to
tors have also emerged as a notable example application. real-world optimization challenges. Furthermore, they face

782 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

challenges in accurately predicting UE rates, mainly because This includes considering the complex interplay between
they do not adequately consider crucial factors such as band- energy-saving strategies, mobility management, and other
width allocation per UE, which are profoundly influenced by network dynamics. Such a framework promises to advance
network energy-saving strategies. the current state of network optimization by providing a more
accurate, adaptable, and practical approach to energy-saving
d: ENERGY CONSUMPTION AND CARBON EMISSIONS solutions, thereby addressing the limitations identified in the
ESTIMATIONS existing body of work.
To evaluate the energy-related performance of the RAN,
the mobile industry has defined several measurement meth- B. RESEARCH QUESTIONS AND OBJECTIVES
ods and metrics for different network levels (network [47], In light of the current state of the art, this paper sets forth
site [48], BS [49], and UE [50]) as well for different scenarios a series of research questions and scientific objectives to
(dense urban, urban, and rural coverage) [47], and services delineate the research problem at hand, and move towards the
(enhanced mobile broadband, ultra-reliable low-latency com- creation of such comprehensive end-to-end framework:
munications, and massive machine type communications)
[51]. More recently, Rappaport et al. have introduced the
1) RESEARCH QUESTIONS
Waste Factor [52], a new metric for quantifying energy effi-
• Is it feasible for a data-driven model, reliant solely on
ciency in a wide range of circuits and systems applications,
off-the-shelf network data —no DTs— to deliver accu-
including data centers and RANs.
rate predictions of network energy consumption and user
Li et al.’s study, utilizing network data from Nan-
throughput?
chang, modeled carbon emissions from mobile networks and
• How can we seamlessly blend machine learning algo-
expanded these insights across China [53]. The research
rithms with domain expertise to enhance energy effi-
used energy consumption and network traffic data from
ciency modeling in large-scale networks?
Nanchang, alongside BS and mobile user counts across
• What advancements in network energy efficiency pre-
Chinese provinces, using simple models for extrapolation.
diction can this type of modeling framework introduce,
Additionally, they introduced a reinforcement learning (RL)
and how does it measure up against existing methodolo-
framework designed to shut down network cells as a means to
gies?
prevent capacity over-provisioning. A significant limitation
of the model, however, is its simplistic approach to energy-
saving schemes —cells are shut down as deemed necessary 2) SCIENTIFIC OBJECTIVES
and UE handovers are ideal— coupled with an inability • Establish a Data-Driven Modeling Framework: Our goal
to evaluate UE performance following changes. Maggi et is to devise a holistic framework that can precisely
al. have made a substantial contribution by introducing a forecast network energy efficiency and user through-
sector-specific, closed-loop Bayesian method to optimize put using only inexpensive available network data. This
carrier shutdown thresholds, effectively reducing power con- requires the harmonization of machine learning (ML)
sumption while maintaining network KPIs, particularly UE technologies with expert insights to depict the multi-
rates [54]. This data-driven approach, validated in a live faceted nature of large-scale networks.
4G network, emphasizes the practical application and effec- • Consider Practical Network Energy-Saving Solutions:
tiveness of their model. However, the focus on optimizing The framework aims to furnish practical energy saving
carrier shutdown thresholds, excluding the refinement of han- solutions into the large-scale network modelling for ele-
dover algorithms, and the model’s limitation in simulating vating energy efficiency in telecom networks.
network behavior offline might present challenges in more • Validate the Framework with Real-World Data: We
complex scenarios where Bayesian optimization may fail to intend to prove the framework’s utility by applying it
converge [55]. to actual network scenarios, showcasing its capability
Drawing from the literature, it becomes evident that for precise energy consumption predictions and its value
while significant strides have been made in network opti- in fostering user throughput when fine tuning energy
mization through various methodologies, a crucial gap saving solutions.
remains: a cohesive, data-driven modeling framework. Exist-
ing research, valuable though it is,often compartmentalize
aspects of network performance, neglecting the intricate bal- C. PAPER’S CONTRIBUTIONS
ance required between energy efficiency and maintaining Building upon the research questions and scientific objectives
optimal network KPIs in dynamic real-world scenarios. This outlined previously, this paper introduces a pioneering solu-
highlights the urgent need for a comprehensive framework tion to the challenges of optimizing network energy-saving
that not only bridges these divides, but also utilizes real-world solutions. We present a novel data-driven framework for
data to thoroughly model and optimize the trade-offs between modeling network energy consumption and UE throughput,
energy consumption and key network performance metrics. referred to as SRCON [56], which uses measurement data

VOLUME 2, 2024 783

from live networks to jointly fit (i) ML-based, black-box and handover procedures across a vast, heterogeneous net-
(ii) expert-based, white-box models2 that can work of hundreds of 4G and 5G cells. This eliminates
• accurately characterize the functioning of the most rele- the need for expensive drive testing or ray tracing-based
vant network components, and simulation tools.
• generalise the modelling of network energy consump- • Decomposed Modeling Framework: We introduce a
tion and UE throughput to any possible network energy novel methodology by segmenting the overall mod-
saving solution configuration, even those not observed eling challenge into distinct ML- and expert-based
in the training data. subproblems. Outputs from expert-based models inform
SRCON, akin to digital twins, functions as a virtual repre- ML-based models, enabling precise generalization of
sentation of the physical system; in this instance, the network, network energy consumption and UE throughput pre-
acting as its digital counterpart. Digital twins are increasingly dictions across any carrier shutdown configuration,
viewed as a vital paradigm for monitoring, controlling, and including unencountered scenarios.
• Data and Model Construction Pipeline: This paper pro-
optimizing communication systems [57], [58]. They offer a
sandbox for testing new features, e.g.artificial intelligence vides a detailed account of the necessary data and the
(AI) solutions, potentially reducing the need for field data pipeline used to construct ML-based models for network
collection and algorithm testing. Most digital twins often energy consumption and UE throughput. This method is
rely on Internet of things (IoT) sensors to proactively ana- crucial for accurately capturing the varied performance
lyze system performance in real-time and make decisions characteristics of network products and end-user devices
— an approach not seen in current literature. Each prob-
accordingly [59]. A key challenge in digital twin is ensuring
lem necessitates a unique framework.
that virtual control optimization is safe and reliable, prevent-
• Minimal Assumption Requirement: A distinguishing
ing incorrect decisions due to ‘‘model exploitation’’ [60].
feature of our modeling approach is its minimal depen-
SRCON diverges from typical digital twin frameworks —
dence on the statistical assumptions often mandated by
and attempts to enhance reliability— by not attempting to
methods like Bayesian optimization, which typically
replicate network behaviors on a small timescale, e.g., mil-
assume Gaussian behavior. Our models’ capability to
liseconds —a task that is fundamentally impractical due to
provide reliable predictions without such assumptions
the significant modeling challenges faced in real-world large-
significantly increases their robustness and adaptability
scale networks. Unlike the relatively straightforward sensor
across various network scenarios.
data utilized in many digital twin applications, the network’s
• Innovative Expert-Based ABM Model: Introducing an
‘sensors’ are the UEs themselves, whose interactions and per-
innovative agent-based model (ABM) marks a sig-
formance data are inherently multi-dimensional and complex
nificant advancement. This expert model thoroughly
to analyze. Given this intricacy, SRCON instead is set to
simulates network stochasticity and the dynamics of
achieve a statistically indistinguishable emulation of network
carrier shutdown across wide-area networks with man-
behaviors. We should also underscore that, unlike many digi-
ageable complexity. It includes handover procedures and
tal twins focused on real-time optimization, SRCON’s current
approximates UE transfers. The ABM enables precise
goal is not real-time modeling and optimization but rather to
predictions using established telecom theories, supply-
establish foundational functions for such future optimization
ing inputs to ML models for any configuration of carrier
capabilities. Further details on these aspects are provided
shutdown solutions and handover parameters.
throughout the document.
• Data-Driven Comparative Results: We conclude with
It is important to acknowledge that our current scope
empirical findings that underscore the extensive benefits
focuses on network energy consumption and UE throughput,
of our modeling approach. A comparative evaluation
pivotal KPIs for evaluating the efficiency of energy-saving
against a benchmark tool, utilized by a leading network
solutions. However, we recognize the significance of other
operator for carrier shutdown optimization, illustrates
metrics, such as delay, Diving into the modeling of such
our models’ superior performance and effectiveness.
metrics represents a promising avenue for future research,
The rest of this paper is structured as follows. In Section II,
requiring additional modelling. We leave this to future work.
we introduce the basics of a state of the art carrier shut-
With respect to the available literature earlier reviewed in
down solution, with the corresponding formal definitions.
Section I-A, the novelty of this paper resides in the following
In Section III, we highlight the main challenges to model
aspects:
carrier shutdown performance in a large-scale network,
• Novel Modeling Approach: Unlike existing literature,
and present the main logic behind our proposed modelling
this work uniquely focuses on modeling network energy approach, with a mix of black- and white-box modelling.
consumption and UE throughput within the context of In Section IV, we detail the different types of data, which
a practical carrier shutdown solution and its associated are available to us to carry on our modelling exercise.
2 In contrast to a black-box model, in a white-box model logic, functioning In Section V and VI, we present the energy consumption
and programming steps are transparent and known. As a result, it’s decision and UE rate models developed to investigate the energy effi-
making process is interpretable. ciency of practical networks. These models are constrained

784 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

by —tailored to— the available data. In Section VII, we up decisions are generally assessed by capacity booster cells
present the main contribution of this paper, our expert-based, and coverage cells, respectively, with a frequency in the order
white-box, ABM, which mimics the stochastic behaviour of seconds, while the carrier shutdown duration may well vary
in a large-scale network with carrier shutdown, and allows from tens of seconds to minutes or even hours. The time that it
to generalise the network energy consumption and UE rate takes to shutdown and wake up a carrier is around 3 seconds.
predictions to any carrier shutdown and handover parameter Since carrier shutdown allows deactivating the entire cell,
configuration. The outputs of this model are inputs to the it enables deeper sleeps than symbol and channel shut-
models presented in Section V and Section VI. To demon- down, and in turn, larger energy-savings. However, coverage
strate the generalization capabilities of our framework, and capacity losses can be significant, if carriers are not
in Section VIII, we describe the scenarios and conditions (de)activated in a coordinated manner across the network. For
under which our modelling approach has been tested in a example, coverage holes may appear when shutting down a
real network, and discuss the performance results. Finally, cell in those cases where no other cell can provide coverage
in Section IX, we drawn the conclusions, and highlight new to the former UEs of the shutdown cell. Capacity may also
research directions. be affected, as a shutdown cell does not allow for the spatial
reuse of spectrum, and thus the service quality of the UEs
II. CARRIER SHUTDOWN AND MODELLING of the cells receiving the UEs of the shutdown cell may be
CHALLENGES compromised due to resource sharing. Similarly, the load in
Carrier shutdown allows to fully deactivate the cells/carriers3 the cell receiving the UEs of the shutdown cell may grow after
mounted on a given radio unit by switching off most of its some time, and if the shutdown cell is not reactivated on time,
RF and digital front-end components. The base band unit they may also suffer from service quality degradation. This
and the interface for waking up the shutdown carriers remain gives rise to an intricate trade-off between network energy
active. For more details on carrier shutdown, please refer to consumption and network/UE performance.
reference [6]. The target of this paper is to present a novel, data-driven
Importantly, to enable a dynamic carrier shutdown opera- modelling approach to assess this trade-off, which can be
tion, the 3GPP has specified a number of related features [61], later used to optimize carrier shutdown operations in practical
building on the concepts of capacity booster cell and coverage networks. In the following, we detail the specific carrier
cell,4 among which we should highlight: shutdown logic assumed in this paper, which is in line with
• Cell pairing: Capacity booster cells and coverage cells
that used in real solutions [6].
can be paired for energy efficiency purposes, where an A. CELL PAIRING
LTE or an NR cell can be the coverage cell of a capacity As indicated earlier, for energy efficiency purposes, cells
booster cell. are divided between (higher carrier frequency) capacity and
• Autonomous shutdown: A capacity booster cell can
(lower carrier frequency) coverage cells, and every capacity
take autonomous carrier shutdown decisions based on cell is paired with at least one coverage cell, which should (i)
estimations, not only of its own number of connected provide service to the UEs of the capacity cell, should this
UEs and DL and uplink (UL) cell loads, but also on one shut down, and (ii) assist its wake up process. Thus, the
available load information from paired neighbouring paring process between capacity and coverage cells is capital
coverage cells exchanged through the respective X2/Xn to the overall carrier shutdown performance.
interfaces. The coverage cell paired with a capacity cell should be
• Reactivation: The paired coverage cell owning a capac-
the coverage cell that has the largest coverage overlap with
ity booster cell can autonomously request an inter-BS the capacity cell. Given the difficulty to predict the over-
cell reactivation over the X2/Xn interface based on its laps among cells in a planning stage, these relationships are
own load information. usually computed based on inter-frequency measurements
• Shutdown information sharing: A coverage cell can
periodically reported by the UEs. In a nutshell, the following
inform all its neighboring cells of the (de)activation of conditions should be met to consider a given inter-frequency
one of its capacity booster cells. cell as a pairing candidate:
The 3GPP efforts in this front still continue today [27], [62]. • The number of times that such coverage cell is reported
With respect to the general functioning of carrier shutdown by the UEs of the capacity cell is larger than a threshold,
solutions, it should be noted that carrier shutdown and wake and
3 Note the while cell refers to the area covered by a carrier operating
• the ratio of the number of times that such coverage cell
at a given frequency with a given bandwidth, these two terms are used
is reported to the total number of measurement reports
interchangeably in this paper. is larger than another threshold, and
4 To ease the complexity of carrier shutdown, in a multi-layer network, • among those measurement reports in which such cell
cells are usually divided into two groups: Capacity booster cells are deployed is reported, the fraction of reports with an RSRP larger
for capacity enhancements, typically use a higher carrier frequency, and
can be shut down. Coverage cells, instead, are deployed to provide blanket than an RSRP coverage threshold is larger than another
coverage, usually use a lower carrier frequency, and cannot be shut down. threshold.

VOLUME 2, 2024 785

Although of importance, the modelling of this capacity and The A4 inter-frequency handover is triggered when the
coverage cell pairing process is out of the scope of this paper, biased RSRP, Mn , of a neighbouring cell, n, of cell, i, becomes
freq
and it will considered as optimised and given in the following. better than a threshold, τi A4 , i.e. Mn + Oi,n + Oi,n cell −

Hi A4 > τi . Specifically, according to this expression,

A4

B. REFERENCE CARRIER SHUTDOWN (DE)ACTIVATION it should be noted that, in the A4 inter-frequency handover
LOGIC entry condition, a hysteresis, Hi A4 , a cell-individual offset,
In the reference implementation used in this paper, assuming cell , and a cell-specific frequency offset, O freq , are used to
Oi,n i,n
that a capacity cell, c, is paired with only one coverage avoid ping-pongs, and prioritize a given neighboring cell or a
cell, b(c), and to maintain the quality of service (QoS) of given frequency, respectively.
UEs affected by the transfer, carrier shutdown can only be Accordingly, the optimization of the A4 inter-frequency
autonomously activated in capacity cell, c, when handover parameters also plays a role in network energy
• the number, UcUE , of radio resource control (RRC) con- efficiency, and are hence part of our framework.
nected UEs in the capacity cell, c, is smaller than an entry
threshold, χcUE , and
III. NETWORK ENERGY EFFICIENCY MODELLING
• DL = 1 DL + 1 DL , of used DL physical
the sum, 1c,b(c) c b(c) As can be inferred from previous sections, the network energy
resource blocks (PRBs) in the capacity cell, c, and the
efficiency modelling problem deals with multiple objectives
paired coverage cell, b(c), is smaller than another entry
(e.g. network energy consumption and UE throughput) and
threshold, χc,b(c)
DL , and
optimization variables (e.g. carrier shutdown and A4 han-
• the sum, 1c,b(c)
UL = 1 UL + 1 UL , of used UL PRBs in
c b(c) dover parameters), and is particularly challenging due to
the capacity cell, c, and the paired coverage cell, b(c), its large-scale, stochasticity and non-stationarity as well as
is smaller than another entry threshold, χc,b(c)
UL .
the complex coupling between cells and the intricate trade-
In contrast, a coverage cell, b(c), may wake up its paired offs between energy consumption and UE performance. For
capacity cell, c, in carrier shutdown when instance, the shutdown of a cell may save energy, but impacts,
• UE , of RRC connected UEs in the paired
the number, Ub(c) not only the performance of the UEs connected to such cell,
coverage cell, b(c), is larger than a leaving threshold, but also the overall network coverage and the performance of
9c,b(c)
UE , or those UEs connected to the neighboring cells as well as the
• the number, 1b(c) DL , of DL PRBs in the paired cover- possibility of nearby capacity cells to shutdown. In practice,
age cell, b(c), is larger than another leaving threshold, these challenges make the utilization of precise networking
9c,b(c)
DL , or models essential to perform a rigorous network energy effi-
ciency optimization.
• the number, 1b(c)
UL , of used UL PRBs in the paired cov-
Fortunately, recent advancements in big data acquisition
erage cell, b(c), is larger than another leaving threshold,
and processing have made possible to efficiently store —and
9c,b(c)
UL .
subsequently process— the large amount of radio measure-
Time windows are used to average and smooth these statis- ments to which BSs have access, opening the door to new,
tics. As one can imagine, the optimization of these thresholds data-driven modelling and optimization paradigms [25], [27].
plays a major role on energy-savings, and are thus part of our The main idea behind solving the network energy effi-
model. ciency optimization problem in a data-driven manner is
The careful consideration of these algorithms ensures that to leverage such easily accessible measurements to gener-
the paired coverage cells can seamlessly support the commu- ate accurate network-specific models, without the need of
nications of all UEs in the capacity cell that is shutting down, resorting to expensive expert knowledge, DT-based data or
in addition to those of its already connected UEs, thereby incomplete three-dimensional (3D) high-definition maps for
maintaining QoS without disruption. A bias could be added to ray-tracing purposes.
the corresponding thresholds to ensure that a larger number of Data-driven network modelling helps addressing most of
PRBs are reserved for the UEs being transferred. By ensuring the challenges of state of the art approaches surveyed in
that the paired coverage cells can accommodate the PRBs of Section I, but it also brings its own. Among the challenges
the transferred UEs, we uphold the QoS. introduced by data-driven network modelling, it is worth
highlighting the following two:
C. REFERENCE UE TRANSFER LOGIC
In the reference implementation used in this paper, once the 1) Massive data sets: A typical cellular network in a
capacity cell, c, decides to shutdown, it instructs its RRC metropolis has around 50 thousands BSs, each of them
connected UEs to perform an A4 inter-frequency handover generating nearly 3000 KPIs per hour [63]. When
to the frequency of its paired coverage cell, b(c). If such UEs including UE measurement reports, this results in 1 ter-
are able to handover within a given predefined time frame abyte of data per hour [56]. This overwhelming amount
(usually of tens of seconds), the capacity cell, c, shuts down. of information needs to be stored and processed in time
Otherwise, it abandons its intention to shutdown. for its productive utilization.

786 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

FIGURE 1. SRCON framework for network energy efficiency.

2) Generalization: This is arguably the most relevant issue words, said white-box model takes as input, among others,
in data-driven modelling, and pertains to the inference the carrier shutdown and A4 handover parameter configu-
of the network performance when applying the model ration, and provides as output the inputs of our black-box
in a scenario –or using a combination of parameters– models. In this manner, we can drive universal ML-based
never measured before and thus not observable in the BS energy consumption and UE throughput predictions, and
training data. further derive network energy efficiency.
To address these challenges, SRCON —our proposed mod- Fig. 1 present this framework, which is further detailed in
elling approach— combines a variety of wireless as well the following sections.
as data-driven and ML concepts. Rather than attempting For the sake of space, note that, in general, we do not
to replicate network behaviours at a time-scale of millisec- mention UL related statistics in the rest of the paper, but
onds, which is fundamentally infeasible, considering the whenever we refer to a DL statistic or process, the analogous
aforementioned modelling challenges in practical networks, UL one is generally implied.
SRCON emulates network behaviours in a statistically indis-
tinguishable manner. This statistical indistinguishability lays IV. MODELLING OBJECTIVE AND DATA AVAILABLE
the ground for a practical and accurate overall assessment To estimate the goodness of a network optimization cam-
of UE performance, allowing the use of sufficient statistics paign, the performance of the network is usually measured
—instead of all available samples— through an efficient before and after a change of network parameters. Embracing
processing of massive data sets. To enable an accurate gen- this methodology, at least two measurement campaigns are
eralization, and deal with the heterogeneity and complexity usually conducted to collect data from the network when
of devices, SRCON uses measurement data from live net- carrying a network optimization exercise. During the first
works to jointly fit ML-based, black-box and expert-based, measurement campaign, before the optimization, all energy
white-box models, as hinted earlier. For our network energy saving solutions, including carrier shutdown, are deactivated
efficiency modelling problem, in more details, we propose to estimate the baseline network energy efficiency and per-
to model BS energy consumption and UE throughput statis- formance. The parameters of the network, also refereed to
tics using customised ML-based modeling approaches to the as engineering parameters, are recorded too. In the context
available data to capture the particularities of the specific BS of this paper, we will refer to the data collected in this
products and off-the-shelf UEs in the area of study (e.g. not phase as unbiased data, as it is not affected/biased by any
all BS product versions have the same energy consumption energy-saving policy. During the second measurement cam-
characteristics, not all UEs have the same decoding capabili- paign, after the optimization, new measurements are collected
ties). A systematic feature importance analysis was used to to estimate the network energy efficiency and performance
identify the most relevant input features of such two ML- resulting from the activation of carrier shutdown when using
based models, e.g. the amount of time for which capacity cells the optimized carrier shutdown and A4 handover parameter
will be in carrier shutdown, the number of UEs transferred to configurations. The new engineering parameters of the net-
each neighbouring cell when carrier shutdown takes place, work are also recorded.
the resulting PRB load in those neighbouring cells. To allow To develop our new, data-driven modelling framework, we
generalisation, a customised expert-based model is used to take advantage of several such campaigns in different cities
derive the inputs of the ML-based models for any carrier and at different times of the year. In a nutshell, our mod-
shutdown and A4 handover parameter configuration. In other elling objective is to predict, using the unbiased data of the

VOLUME 2, 2024 787

first measurement campaign, the network energy efficiency TABLE 1. Engineering parameters.
and performance with the minimum possible error under the
carrier shutdown and A4 handover parameter configuration
in the second measurement campaign. The corresponding
metrics to assess the quality of the prediction are generally
the mean absolute error (MAE) and mean absolute percentage
error (MAPE). Importantly, it should be noted that the quality
of the model is impacted by the amount of data features
collected, the amount of samples collected per feature and
their time granularity.
Given the use of ML modeling tools in our work, the data
of the second measurement campaign is divided into two sets,
one used for training and validating such ML models, as it
will be shown in Sections V and VI, and the other for test the
overall model, as it will be discussed in Section VII.
In the rest of this section, we describe the data sets and
features available and used in our modelling exercise, which
can be categorised into three types, i.e. engineering param-
eters, cell-level KPIs and UE measurement reports. We also
formally present the notation used in this paper. Importantly,
to make sure that this modelling framework can be widely
used, we target at solely using data readily available to any
operator.

A. ENGINEERING PARAMETERS
The engineering parameters data sets, d EP , describe the
network configuration. We should distinguish 3 types of engi-
neering parameter data sets:
• Network parameters data set, d NEP : Information related
to the configuration of each BS, radio unit and cell
(e.g. type of radio unit, location, number of RF chains,
number of supported and configured carriers per radio
unit, as well as frequency, bandwidth, bearing, tilt and
other information per cell);
• Mobility parameters data set, d M EP : Information related
• Traffic statistics data set, d TKPI : Information on the ser-
to the configuration of handover procedures of every cell
viced traffic per cell (e.g., average number of active UEs
(e.g. A4 handover parameters);
EP : Information per transmission time interval (TTI), average number of
• Energy-saving parameters data set, d ES
used PRBs per TTI, sum traffic volume);
related to the configuration of energy-saving solutions of KPI : Information on
• Energy-saving statistics data set, d ES
every cell (e.g. capacity and coverage cell pairing, carrier
the activated energy-saving modes per cell (e.g., dura-
shutdown thresholds).
tion of the carrier shutdown activation);
Energy consumption statistics data set, d EC KPI : Informa-
Table 1 details the most relevant engineering parameters •
used in our modelling framework, and provides a formal tion on the energy consumed by each radio unit.
definition. Table 2 details the most relevant cell-level KPIs used in our
modelling framework, and provides a formal definition. Note
B. CELL-LEVEL KPIs that the level of aggregation was 60 minutes in our cell-level
The cell-level KPIs data set, d KPI , describes the performance KPIs data sets, and thus we have one entry in each cell-level
of each cell in the network [7]. Importantly, it should be KPIs data set per cell, i, and hour, h.
noted that, to conserve memory at the BS, this cell-level KPIs
information is typically aggregated over configurable periods C. UE MEASUREMENT REPORTS
of 5, 15, 30, or 60 minutes using sums or averages. As a result, The UE can be directed to perform a variety of measurements,
it does not provide an accurate understanding of the network including intra-frequency, inter-frequency, and inter-radio
behavior at the subframe or slot level. We can distinguish access technology (RAT) measurements, in accordance with
among three types of cell-level KPIs data sets: the measurement configuration provided by the network [64].

788 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

TABLE 2. Cell-level KPIs. TABLE 3. UE measurement reports.

modeling. Accurately approximating RAN energy consump-

tion involves summing the energy expended by its radio and
base band units. In this section, we outline the framework
underpinning our radio unit energy consumption model, cen-
tered around ML. Notably, an ANN architecture emerged as
our choice to model the energy consumption, Er,h RU , per radio

unit, r, at hour, h, after rigorous exploration of various ML

models due to its robust performance and broad applicability.
While the selection process is not presented here for brevity,
a comprehensive account can be found in [65] for further
insights into this radio unit energy consumption model and
our choices.

A. FEATURE IMPORTANCE ANALYSIS

To discern the most impactful factors in estimating radio
This valuable information gathered at the UE is transmitted unit energy consumption, a feature importance analysis was
to the serving BS through UE measurement reports. Each UE conducted using data obtained over a span of 12 days from an
measurement report is uniquely identified by the identifier extensive deployment featuring 7500 4G/5G radio units. This
(ID) of the UE that conducted the measurement and the time encompassed 24 distinct types of radio units (commercial
at which the measurement was taken. Among other statistics, products). The data sets utilized in this analysis included
these reports include the RSRP measured by the UE from engineering parameters, d EP , and cell-level KPIs, d KPI ,
its serving cell, as well as those from a limited number of from the measurement campaign. These data sets comprised
neighboring cells. 150 features per cell, with the most pertinent ones highlighted
Our data set of measurement reports, denoted as d MR , in Table 1 and 2. It is worth noting that during the data
comprises the compilation of all measurement reports collection, energy-saving measures such as symbol, channel,
received by all cells in the network from their respective UEs and carrier shutdown were active. Additionally, as indicated
during specific hours of the day. Table 3 provides a compre- in Section IV-B, the cell-level KPIs statistics were recorded
hensive breakdown of the most pertinent features present in a on an hourly basis.
UE measurement report as used in our modeling framework, The feature importance analysis was conducted in two
accompanied by a formal definition. This data set contains phases: i) Gradient boosting models were trained with vary-
one entry per measurement report. ing sets of input features; ii) Then, an examination of the
Considering that measurement reports can be transmitted shapley additive explanations (SHAP) values for each feature
by UEs in connected mode as frequently as every 5 mil- was undertaken using these models. To elaborate, the SHAP
liseconds, the measurement report data set, d MR , occupies value assigned to each feature signifies the alteration in the
significantly more memory than the engineering parameters, model’s anticipated prediction when that specific feature is
d EP , and the cell-level KPIs, d KPI , data sets. As a result, taken into account [66].
measurement reports are usually collected and stored for Fig. 2 illustrates the SHAP values associated with the key
shorter durations, often just a few hours during specific time numerical features in the data set. Specifically, the diagram
periods. illustrates both the magnitude and direction of each feature’s
impact on the model output in relation to the average model
V. ML-BASED RADIO UNIT ENERGY CONSUMPTION prediction. The right-side y-axis denotes the corresponding
MODEL feature value, displayed in a color gradient from blue (low
Understanding and optimizing energy-saving aspects neces- values) to red (high values). Each individual scatter dot cor-
sitates a thorough grasp of RAN energy consumption responds to a data instance.

VOLUME 2, 2024 789

 depend on the carrier shutdown and A4 handover parameter
settings. Therefore, they need to be modeled as functions
of these parameters to accurately predict network energy
consumption for different configurations, enabling effective
optimization.
Section VII presents the white-box model that allows such
generalization.

B. INPUTS OF THE MODEL

Each of the input features listed in Table 4 underwent pre-
FIGURE 2. Energy consumption SHAP analysis performed on the processing based on its type to eliminate outliers, and was
most important numerical features in the collected then fed into the ANN. The numerical features were normal-
measurements data.
ized prior to entering the model, while the categorical features
were encoded using one-hot encoding.
TABLE 4. Energy consumption model input parameters. To ensure maximum generality and flexibility, our ANN
model takes input data from C MAX carriers, where C MAX rep-
resents the highest number of carriers that the most capable
radio unit, can manage. In our data set, C MAX is six. When a
radio unit handles fewer carriers, C < C MAX , the input neu-
rons corresponding to the remaining C MAX − C carriers are
assigned zero values. This universal model approach enables
the implementation of a unique ANN model with a fixed num-
ber of input neurons. It can be trained using data from all radio
units, ∀r ∈ RU, in the data set, regardless of their number of
configured carriers, |Cr |, while resulting in minimal loss in
accuracy, as it will be discussed in Section V-F.
Our feature importance analysis revealed that the DL PRB
1 DL
load, represented as i,h , holds the highest significance in C. OUTPUTS OF THE MODEL
Ni DL
modeling radio unit energy consumption. Additionally, the The analysis of the collected data revealed instances where
different energy consumption values, Er,h RU , are reported
second most crucial feature is the maximum transmit power
denoted by Pimax . These two features enable the model to for the same input feature values. This variability can be
effectively capture the power transmitted, PiTX , by cell, i, attributed to several factors, including:
across various DL PRB load levels. The duration of car- 1) Features that have a slight impact on energy consump-
rier shutdown activation, ti,h CS , also emerges as significant, tion but are not included in our data sets.
offering insights into the sleep behavior of the cells. The esti- 2) Potential errors in measurements or data collection.
mation of radio unit energy consumption is also influenced by 3) Tolerances of hardware components, which influence
factors such as the radio unit type, the number of transceivers their energy consumption behavior.
(TRXs), carrier transmission mode, frequency, and band- For the sake of presentation, we use the following change of
width. These parameters define the radio unit’s hardware RU , in this section.
notation, ȳ = Er,h
and capabilities, making them crucial for accurate modeling. To characterize this noise, we define the measured energy
A notable finding from our analysis is the strong correlation consumption, ȳ, as ȳ = y + n, where y represents the energy
between the modulation and coding scheme (MCS) and the consumption for a given input configuration, and n accounts
number of MIMO layers used per DL PRB and the DL PRB for the noise arising from the aforementioned factors. Based
load. This implies that modeling the DL PRB load across all on our data analysis, we deduced that the noise term n can
cells operated by a radio unit may be sufficient to capture their be approximated as a normally distributed variable with a
average energy consumption behavior. mean of zero and a standard deviation of σ . Consequently,
Table 4 provides a comprehensive list of the 10 selected the measured energy consumption ȳ follows a normal distri-
features following the feature importance analysis. Features bution with a mean of µ = E[ȳ] and a standard deviation
that have minimal impact on energy consumption or are of σ .
highly correlated with the chosen ones, thus offering limited To account for this noise, our ANN model was designed
additional information, have been excluded. to estimate and output both of these parameters, µ and σ ,
Remark 1: While the maximum transmit power, Pimax , for a given input, x. Importantly, the outputs of these two
of cell, i, and other selected inputs in Table 4 are static parameters enable the computation of a confidence interval
variables provided by the engineering parameters, the DL for each energy consumption estimate, enhancing the overall
1i,h
DL
CS , of carrier shutdown
PRB load, , and the duration, ti,h trustworthiness of the estimation process.
Ni DL

790 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

To align with optimization methods that aim to minimize,

we adopted the following loss function for training the ANN
model:
(ȳ − µ)2
l(ȳ, µ, σ ) = − log (P(ȳ|µ, σ )) = log(σ ) + . (2)
2σ 2
This loss function serves the dual purpose of minimizing
both prediction error and associated uncertainty. The first
term is minimized when the standard deviation, σ , is low,
indicating high confidence in the estimation. The second
term is minimized when the prediction error, ȳ − µ, is
reduced.
FIGURE 3. Structure of our proposed ANN, highlighting the
Regarding the training and testing data sets, we utilized the
selected input variables, the organization and interconnection same data set as employed in the feature importance analysis
of the two hidden layers, and the configuration of the two (Section V-A). The data samples were chronologically sorted
output neurons which represent the mean and standard and divided into a training set (80 % of samples, the first
deviation of the power consumption estimates.
10 days) and a testing set (the remaining 20 %, the last
2 days). For training the ANN model, 80 % of the train-
D. ARCHITECTURE OF THE MODEL ing set samples were randomly chosen, with the remaining
The fundamental architecture chosen for our proposed ANN 20 % serving for model validation during training. The Adam
model is the multilayer perceptron, which comprises several version of the gradient descent algorithm was employed for
fully connected layers of neurons [67]. Fig. 3 offers a com- model training [67], incorporating an early stopping mecha-
prehensive visualization of the overall ANN employed for nism to halt training after 200 epochs without validation loss
estimating power consumption in this paper. improvement.
Following from the figure, the input layer comprised n I =
NRU + 10 ∗ C MAX neurons, where NRU represents the number F. PERFORMANCE OF THE MODEL
of distinct radio unit types modeled by the ANN. In our spe-
To evaluate performance, we conducted a comparison
cific data set, NRU equated to 24, and the number of features
between the estimated energy consumption and the actual
considered per radio type was 10, resulting in an input layer
measurements from the test set, utilizing the MAE and MAPE
of nI = 84 neurons. This architecture configuration allowed
as evaluation metrics.
for effective representation of input data.
It is important to mention that the training of the ANN
Two hidden layers followed the input layer, with n H,1 =
model took approximately 75 minutes in a machine powered
40 and n H,2 = 15 neurons, respectively. These dimensions
by a Intel® CoreTM i7-9700 CPU @ 3.00 GHz with 32 GB
were determined through an optimization process aimed at
of Random Access Memory (RAM), encompassing 1086 iter-
maximizing model accuracy.
ations with a learning rate of 0.001.
The output layer consisted of n O = 2 neurons, captur-
From the results, we can observe that our energy consump-
ing the mean µ and standard deviation σ of the radio unit
RU for radio unit, r, at hour, h, tion model achieved a MAE of 10.94 W, and a remarkably
energy consumption ȳ = Er,h
low MAPE of 5.87 %, when estimating the energy consumed
as explained earlier. Given that both metrics are inherently
by each radio unit across all hours of the test period. As an
positive, a sigmoid activation function was employed at the
example, Fig. 4 illustrates the comparison between actual and
output layer.
estimated normalized energy consumption for various radio
units of the same type. It is worth noting that energy con-
E. TRAINING AND TESTING OF THE MODEL
sumption follows a linear relationship with the DL PRB load,
The objective of the model optimization process was to 1i,h
DL

simultaneously minimize prediction error and uncertainty. , and the presence of three distinct slopes is attributed
Ni DL
Specifically, during the training phase, the aim was to to different configurations of the maximum transmit power,
enhance the probability that energy consumption samples, ȳ, Pimax , within the data set. The proposed ANN model adeptly
for a given input, x, fall within the estimated distribution, captures the energy consumption characteristics for each of
N (µ, σ ). This approach ensures that the statistical distribu- these configurations.
tion of energy measurements output by the model aligns with To complement these results, it should also be noted that
the distribution of energy measurements in the data. the accuracy loss incurred by our universal ANN modelling
Given that energy consumption, ȳ, follows a normal distri- approach (i.e. one model for all radio units) with respect to
bution, this probability was computed as one in which an ANN model is trained per group of radio units
2
supporting the same number of carriers is equal to 1.86 %.
1 − (ȳ−µ)
P (ȳ|µ, σ ) = √ e 2σ 2 . (1) This shows that the flexibility of the universal model comes
σ 2π at the expense of a reduced accuracy loss.

VOLUME 2, 2024 791

 FIGURE 5. Example of UE rate SHAP analysis performed on the
most important numerical features in the collected
FIGURE 4. True and estimated normalized energy consumption measurements data in a given cell.
vs DL PRB load for multiple BSs of a given type.

avg
• As per [7], the average UE rate, Ri,h , for each cell,
VI. ML-BASED DL UE RATE MODEL i, at hour, h, is calculated by subtracting the number
In the context of optimizing a large-scale network for minimal of DL bits, Vi,hDL- , successfully transmitted during the
energy consumption, maintaining an acceptable UE rate is last time slots when the DL buffer becomes empty in
crucial to avoid compromising network performance. In this cell, i, at hour, h, from the total number of DL bits,
section, we present a summary of the proposed ML model for DL , transmitted at the RLC layer at the same cell and
Vi,h
estimating the UE rate, while considering the diversity of end- hour. This value is then divided by the duration, Ti,hDL- ,
user devices in the network. It’s important to emphasize that during which cell, i, was transmitting DL bits at hour, h,
while we selected an ANN architecture to model the energy excluding the last time slots when the DL buffer became
consumption, Er,h RU , per radio unit, r, at hour, h, we opted for DL −V DL-
avg Vi,h i,h
a gradient boosting architecture to model the average, Ri,h ,
avg empty. Mathematically, Ri,h = DL- .
Ti,h
ce On the other hand, the 5 %-tile UE rate, Ri,h ce , for each
and the 5%-tile, Ri,h , UE rate per cell, i, at hour, h. This choice •
was based on considerations of both accuracy and complexity. cell, i, at hour, h, can be estimated from the counters
DL in cell, i, at hour, h. These counters indicate the
RNg,i,h
Unlike energy consumption, which remains relatively con-
sistent for the same radio unit operating under similar number of samples falling within a predefined UE rate
conditions, the UE rate is sensitive to the geographical char- range, g.
acteristics and channel conditions of a cell’s deployment
First, we conducted a SHAP analysis for each cell in our
area. Therefore, two cells with identical configurations and
data set to determine the most relevant features for modeling
traffic loads but deployed in different locations could exhibit
the UE rate per cell and hour [66]. An example of the SHAP
vastly different average and 5%-tile UE throughput. To cap-
values of the ten most significant features for a specific cell is
ture these nuances effectively, our experiments showed that
depicted in Fig. 5. In this instance, the DL PRB load emerges
using separate models for individual cells is more appropri-
as the most influential feature on average. As expected, the
ate than using a universal model for all cells. Furthermore,
SHAP analysis indicates that a higher DL PRB load cor-
when pursuing this approach, gradient boosting not only yield
responds to a lower estimated UE rate. This connection is
improved accuracy, but also reduced complexity in terms
logical, given that a higher DL PRB load generally implies a
of training time compared to an ANN architecture or other
greater number of connected UEs, resulting in less bandwidth
alternatives we tested.
per UE.
To provide a comprehensive analysis across all cells, we
A. FEATURE IMPORTANCE ANALYSIS identified the five most contributing features in each cell
Similar to the approach described in Section V-A, we con- based on their SHAP values and ranked them according to
ducted a feature importance analysis on the collected data their frequency of occurrence. Fig. 6 illustrates how fre-
sets to determine the key features for estimating average and quently a feature is included in the five most influential
5%-tile UE throughput. For consistency, we also utilized the features across all cells and UE rate models. This graph
same data sets as in Section V-A. However, it is important to underscores the significance of the DL PRB load, DL traffic
note that our cell-level KPIs data sets do not directly provide volume, number of RRC connected UEs, and the frac-
avg ce ,
features corresponding to the average, Ri,h , or 5%-tile, Ri,h tion of RSRP samples below −100 dBm in predicting UE
UE throughput for each cell, i, at hour, h: throughput.

792 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

FIGURE 6. Feature importance analysis for the UE rate model.

Consequently, given the enhanced importance of these

features, as an ideal baseline estimator, we consider the
scenario where all four features are available and utilized
to formulate the UE rate model for each cell. However,
in practical implementation, due to modelling complexity
issue and the correlations among some of these features,
we construct a model using only two out of the four: the
DL PRB load and the number of RRC connected UEs. This
decision aims to simplify the complexity of our expert-based,
white-box ABM, which will be elaborated on in Section VII.
We will refer to the model that employs only these two
aforementioned features as the ‘‘ABM-friendly estimator’’.
This section will include a comparison between the ideal
baseline estimator and the ABM-friendly estimator.
1 DL UE
Remark 2: The DL PRB load, N i,hDL , and the number, Ui,h ,
FIGURE 7. Average UE rate for a cell mostly experiencing low
i loads.
of RRC connected UEs depend on the carrier shutdown and
A4 handover parameter settings. Therefore, they need to be
modeled as functions of these parameters to accurately pre-
dict network energy consumption for different configurations, dependent on the specific traffic characteristics, such as file
enabling effective optimization. sizes. For instance, if there are only enough bits to occupy
25 % or 75 % of the transmission slot’s bandwidth, the result-
ing UE rate would be 25 % or 75 % of the cell’s maximum
B. INPUTS AND OUTPUT OF THE MODEL capacity. As accurately predicting traffic nature and file sizes
Before constructing our gradient boosting models, we con- based on available data is challenging, we decided to model
ducted pre-processing on the selected input features as average and 5%-tile UE throughput using our gradient boost-
discussed in Section VI-A. An important insight from our ing approach, focusing solely on input data corresponding to
data analysis was that cells with high DL PRB load levels 1 DL
DL PRB load smaller or equal than 0.1, i.e. i,h ≥ 0.1. For
tend to have more accurately predicted average and 5%-tile Ni DL
UE throughput. lower loads, they were modeled using a random variable with
In Fig. 7, we observe the average UE rate of a cell with a its probability mass function (PMF) learned from the data.
low number of UEs and consistently low loads, while Fig. 8
depicts a cell with higher UE count and larger average loads. C. ARCHITECTURE AND TRAINING OF THE MODEL
The former graph highlights the considerable variance in The gradient boosting for regression, as developed by [68],
1 DL was employed within this framework. This method constructs
average UE throughput when i,h < 0.1, which impacts the
Ni DL
an additive model using a forward stage-wise approach.
accuracy of UE rate estimation using available data.
1 DL At each stage, a regression tree is fitted to the negative
In intuitive terms, when the DL PRB load, i,h , is low, gradient of the specified loss function.
Ni DL
the UEs sharing the transmission slot might have lim- Given our objective of minimizing prediction errors for a
ited payload, potentially leading to underutilized bandwidth. single output, we opted for the mean squared error as the
avg
Consequently, the average UE rate, Ri,h , becomes highly chosen loss function.

VOLUME 2, 2024 793

 The results, presented in Table 5, reveal that, as anticipated,
the ideal baseline estimator exhibits the highest performance
when evaluating the average and 5%-tile percentile UE
throughput. Specifically, it achieves a MAE of 2.24 Mbps and
0.14 Mbps, alongside a MAPE of 10.60 % and 27.94 % for the
average and 5%-tile percentile UE throughput, respectively.
The 5%-tile percentile UE throughput, characterized by lower
values than the average, incurs larger MAPE values.
In contrast, the mean estimator demonstrates the weakest
performance, yielding a MAE of 4.27 Mbps and 0.14 Mbps,
along with a MAPE of 21.52 % and 35.9 %, for the average
and 5%-tile percentile UE throughput, respectively. This rep-
resents a 2× and a 28 % increase in MAPE when estimating
the mean and the 5%-tile UE rate, respectively.
FIGURE 8. Average UE rate for a cell experiencing high and low Significantly, the ABM-friendly estimator strikes a com-
loads. mendable balance between complexity and performance.
Specifically, it achieves a MAE of 3.34 Mbps and a MAPE
TABLE 5. Accuracy of the UE rate model. of 15.74 % for forecasting the average UE rate, and records
a MAE of 0.14,Mbps and a MAPE of 28.14 % for predicting
the 5%-tile percentile UE rate. These figures represent a 48 %
increase in MAPE (1 MBps in MAE) compared to the ideal
baseline estimator for the average UE rate, and a negligible
increase for the 5%-tile UE rate. This suggests that the inclu-
sion of DL traffic volume and the fraction of RSRP samples
below 100 dB offers limited accuracy improvement for this
Regarding the training and testing phases, we utilized the process at the cell-edge.
same data set as employed in the feature importance analysis
(refer to Section VI-A). Notably, this data set also aligns with VII. AGENT-BASED STOCHASTIC CARRIER SHUTDOWN
the one used throughout Section V. We applied the same MODEL
training and testing methodology as described in Section V-E, Drawing inspiration from concepts in the Monte Carlo
thereby yielding consistent training and testing sets. method [70] and agent-based modeling [71], we introduce
The process of model training was executed through the a novel, expert-based, white-box model in this section. This
adoption of the Friedman version of the gradient descent model replicates the behavior of a large-scale network when
algorithm, as documented in [69]. Given the manageable time a carrier shutdown solution, as outlined in Section II, is
frame allocated for training, the technique of early stopping implemented. With the aid of this model, we can extend the
was intentionally omitted. network’s energy consumption and UE rate predictions of the
previous two sections to encompass any carrier shutdown and
D. PERFORMANCE OF THE MODEL A4 handover parameter configurations.
To evaluate performance, we conducted a comparison As identified in such Sections V and VI, it is essential to
between the estimated average and 5%-tile percentile UE model the following parameters as functions of the carrier
throughput using both the ideal baseline and ABM-friendly shutdown and A4 handover parameter setup in order to enable
estimators, against the UE throughput obtained from mea- this generalization:
surements in the test set. This assessment employed the • The number, Ui,h UE , of RRC connected UEs.
formulation introduced in Section VI-A. For this comparison, 1i,h
DL

we employed the MAE and MAPE metrics. Additionally, • The DL PRB load, .
Ni DL
we employed the mean estimator as a further benchmark, • The duration, CS ,
ti,h of carrier shutdown for each cell, i,
which characterizes the DL UE rate by computing the mean and hour, h.
values of the related data observed in the training set. 1i,h
DL

Regarding complexity, it is worth noting that our gradi- With respect to the DL PRB load, , and given that
Ni DL
ent boosting model required less than 1 minute per cell for the number, Ni DL , of available DL PRBs for each cell, i,
training the ideal baseline or the ABM-friendly estimator in is provided by the engineering parameters, our focus lies in
a machine powered by a Intel® CoreTM i7-9700 CPU @ modeling the number, 1i,h DL , of utilized DL PRB for every

3.00 GHz with 32 GB of Random Access Memory. The cell, i, during each hour, h.
hyperparameters of the gradient boosting model were con- It is crucial to note that this expert-based, white-box mod-
figured as follows: 500 estimators, a maximum depth of 4, eling solely relies on unbiased information. As discussed in
and a learning rate of 0.01. Section IV, this information is collected during an initial

794 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

measurement campaign conducted prior to optimization. Let Xi,h MC-IN represent the set of inputs utilized to seed
′ ,r
During this campaign, remind that all energy-saving solu- each Monte Carlo run, r, where this set includes the random
tions, including carrier shutdown, are deactivated to establish realizations, Ũi,h UE and 1̃ DL , drawn from the unbiased
′ ,r i,h′ ,r
a baseline network energy efficiency and performance. distributions, fU UE and f1 DL within Xi,h IN .
i,h i,h
The anticipated network behavior at each hour, h′ , namely
A. PROBLEM STATEMENT ˆf UE , fˆ DL , and fˆ CS within X OUT (as defined in Defini-
Ui,h′ 1i,h′ ti,h′ i,h′
Using the unbiased cell-level KPIs data set, d KPI (refer to tion 2), is subsequently derived from statistics obtained across
Section IV-B), and given the extensive data collected over all Monte Carlo runs r ∈ {1, · · · , R}, as it will be discussed
multiple days, we can characterize the distributions, fU UE and later in this section.
i,h
UE , of RRC connected UEs and the
f1 DL , of the number, Ui,h Definition 4 (Outputs of Each Monte Carlo Run, r): Let
i,h
Xi,hMC-OUT symbolize the set of outputs for each Monte Carlo
number, 1i,hDL , of utilized DL PRBs for each cell, i, at each ′ ,r
run, r, wherein this set encompasses the predicted number,
hour, h, under the condition that the carrier shutdown solution UE , of RRC connected UEs, the predicted number, 1̂ DL ,
Ûi,h ′ ,r i,h′ ,r
is not activated. Our analysis suggests that the probability CS
density functions (PDFs), fU UE and f1 DL , can be approxi- of used DL PRBs, and the predicted duration, t̂i,h′ ,r , of the
i,h i,h carrier shutdown for each cell, i, hour, h′ , and Monte Carlo
mated by Gaussian distributions.
run, r.
Definition 1 (Inputs to Our Expert-Based, White-Box
It is worth noting that our Monte Carlo method incorpo-
Model):
rates a rolling concept from hour to hour. This implies that
We denote the set of stochastic inputs to our expert-based,
IN , wherein this set encompasses, not only the unbiased data distributions but also the predicted
white-box model as Xi,h
statistics at hour, h′ , if available, serve as inputs to our expert-
among other elements, the PDFs, fU UE = N (µU UE , σU UE )
i,h i,h i,h based, white-box model to predict network behavior at the
and f1 DL = N (µ1 DL , σ1 DL ), representing the number, Ui,h
UE ,
subsequent hour h′ + 1. This concept will also be elaborated
i,h i,h i,h
of RRC connected UEs and the number, 1i,h DL , of utilized DL on further in this section.
PRBs for each cell, i, at each hour, h.
It is important to note that our focus is on working days C. DEALING WITH CARRIER SHUTDOWN -
(Monday to Friday) and the typical 24-hour day. Additionally, AGENT-BASED MODELLING
due to the granularity of the cell-level KPI data set, d KPI , we To model the outputs, Xi,h MC-OUT , of a given Monte Carlo
′ ,r
have a single data sample per working day for each cell, i, and run, r, in relation to any carrier shutdown and A4 handover
typical hour, h, to estimate such PDFs. The PDFs are built parameter configuration, we have developed a customized
using the many days of data. Agent-Based Model (ABM).
Given the aforementioned unbiased distributions, the net- ABMs have found extensive application in economics
work’s engineering parameters (see Section IV-A), and the for simulating macroeconomic structures emerging from the
carrier shutdown logic detailed in Section II, our objective is repeated local interactions among socioeconomic agents.
to predict the network’s behavior when the carrier shutdown The fundamental concept involves constructing a virtual
solution is activated. environment and populating it with agents, each endowed
Definition 2 (Outputs of Our Expert-Based, White-Box with distinct attributes. These agents adhere to fundamen-
Model): tal guidelines governing their interactions with both each
We denote the set of stochastic outputs of our expert-based, other and their surroundings. Typically, these guidelines are
white-box model as Xi,h OUT , wherein this set includes the
′ grounded in insights about behavior and the local environ-
predicted distributions, fˆU UE′ , fˆ1 DL′ , and fˆt CS′ , of the predicted ment. Consequently, ABMs are dynamic, stochastic systems,
i,h i,h i,h
number Ûi,h UE of RRC connected UEs, the predicted number,
′
usually executed on computers, evolving over time through
iterative processes or algorithms. During these processes,
1̂i,h′ , of utilized DL PRBs, and the predicted duration, t̂i,h
DL CS ,
′
′ agents are adjusted based on established rules. Often, ABMs
of carrier shutdown, for each cell, i, and hour, h , respectively.
incorporate randomness, where agents select various behav-
ioral options randomly. Consequently, Markov chain theory is
B. DEALING WITH STOCHASTICITY - MONTE CARLO well-suited for the mathematical formalization of ABMs [72].
METHOD In this context, we propose an ABM defined by a set of C
To account for the stochastic nature of traffic and channel agents, one for each capacity cell, c ∈ CC , with each agent
conditions, and to forecast performance at each hour, h′ , we characterized by individual attributes drawn from a finite
employed the Monte Carlo method [70]. list of possibilities —specifically, active or shutdown.
Our Monte Carlo method encompasses R runs or snap- We denote the set of possible attributes as S = {active =
shots. Each individual Monte Carlo run, denoted as r, is 1, shutdown = 0}, referring to the solution space as
initialized with data drawn from the unbiased distributions, 6 and an agent configuration as x ∈ 6, with x =
IN (as defined in Definition 1).
fU UE and f1 DL within Xi,h
i,h i,h
(x1 , · · · , xc , · · · , xC ). Consequently, the cardinality of the
Definition 3 (Inputs to Each Monte Carlo Run, r): solution space is 26 . The process of updating agent attributes

VOLUME 2, 2024 795

at each time step, t, comprises two parts. First, a random 2) AGENT SELECTION
subset of agents is chosen based on a probability distribution, The agent selection at each time step, t, of Monte Carlo run,
ω. Then, the agents’ attributes are updated according to a rule, r, at hour, h′ , is founded on our novel concept of ‘‘load dis-
u, determined by the subset of agents chosen at that time. tance to threshold’’. This concept is grounded in the following
Both of these processes will be detailed in Subsections VII- logic: Among the capacity cells in the set, C →sd (cells that are
C1 and VII-C2. active and can be shut down), and considering a given shut-
It is important to note that, due to our implementation via down entry condition DL PRB threshold, χ DL , the capacity
a sequential update scheme, only one agent —equivalent to cell, c, with the smallest combined load —combining its load
one capacity cell, c ∈ CC — can alter its attribute at a given and that of its paired coverage cell, 1c,b(c) DL — is more likely
time. This signifies that transitions can only occur between to be shut down first or more frequently compared to other
agent configurations that vary in at most one bit, following a cells. Here, b(c) refers to the paired coverage cell of capacity
bit representation. The agent selection and updating processes cell, c. Similarly, within the set, C →ac (cells that are shut
are sequentially iterated for a set number of time steps until a down and can reactivate), and considering a shutdown leaving
stable agent configuration, x, is reached. An agent configura- condition DL PRB threshold, 9 DL , the capacity cell, c, with
tion, x, is considered stable if no agent can change its attribute the highest load on its paired coverage cell, 1b(c) DL , is more
within a time step, t, i.e. no cell can be shutdown or activated likely to reactivate first or more frequently compared to other
anymore at this hour. To prevent ping-pong effects or loops, cells. Additional details on the shutdown entry and leaving
the maximum number of time steps, t, within a Monte Carlo conditions can be found in Section II-B.
run, r, is constrained by tmax . With this specification, a Building on these observations, we begin by computing
Monte Carlo run, r, can be perceived as a random walk across the ‘‘load distance’’ in terms of the number of DL PRBs to
the solution space, 6. It is important to note that, at the start of the shutdown entry condition DL PRB threshold for each
the first run in the initial hour, i.e. h′ = 1, r = 1, and t = 1, capacity cell, c, in the set, C →sd , denoted as dc,h′ ,r = χc,b(c)
DL −
we assume that all capacity cells are active, i.e. xc = 1 for all 1̃c,b(c),h′ ,r . Similarly, for the set, C
DL →ac , we compute the
c ∈ CC . ‘‘load distance’’ to the shutdown leaving condition DL PRB
1) UPDATING RULE CHECK
threshold as dc,h′ ,r = 1̃b(c),h DL
′ ,r − 9c,b(c) . Subsequently,
DL

we normalize each of these distances by dividing it by the sum

To expedite the processing of the ABM, we implemented a N
of distances across all cells in both sets, resulting in dc,h ′ ,r =
preliminary updating rule check. dc,h′ ,r
Following the principles outlined in Section II-B, we define P
dc,h′ ,r
. The engineering parameters provide the
C →sd ∪C →sd
the set, C →sd , as the set of capacity cells active at hour, h′ , thresholds, χc,b(c)
DL and 9 DL , while the quantities of used
c,b(c)
Monte Carlo run, r, and time step, t, that can alter their DL PRBs, 1̃c,b(c),h ′ ,r and 1̃b(c),h′ ,r , are initially drawn
DL DL from
attribute —specifically, cells that can be shut down. This set is the unbiased distributions, f1 DL , of the corresponding cells,
identified based on a rule, u sd , comprising three conditions: i,h
c and b(c), at the beginning of Monte Carlo run, r (as per
UE < χ UE , and
1) Ũc,h ′ ,r c,h′ Definition 3). These quantities are then updated as described
2) 1̃c,b(c),h′ ,r = 1̃c,h
DL DL + 1̃ DL
′ ,r b(c),h′ ,r < χc,b(c),h′ , and
DL
in Section VII-C4 when t > 1.
N
3) 1̃c,b(c),h′ ,r = 1̃c,h′ ,r + 1̃b(c),h′ ,r < χc,b(c),h′ .
UL UL UL UL Using these normalized estimations, dc,h ′ ,r for all c ∈

In a similar manner, we identify the set, C →ac , as the set C →sd ∪ C →sd , we construct the PMF, ω, and select a
of capacity cells that are shut down at hour, h′ , Monte Carlo capacity cell, c a , from this distribution. If the chosen
run, r, and time step, t, and have the potential to change capacity cell, c a , is active, it initiates the shutdown pro-
their attribute —these cells can be reactivated. This set is cess, and assesses whether its connected UEs, Uc a =
determined using a rule, u ac , comprising three conditions: {1, · · · , u, · · · , Uc a }, can be handed over to neighboring cells
through an A4 inter-frequency handover. Conversely, if the
′ ,r > 9c,b(c),h′ , or
UE
1) Ub(c),h UE
selected capacity cell, c a , is shut down, its paired coverage
2) 1b(c),h
DL
′ ,r > 9c,b(c),h′ , or
DL
cell, b(c a ), starts the reactivation process. The UEs origi-
3) 1b(c),h′ ,r > 9c,b(c),h
UL UL
′. nally connected to this capacity cell, c a , as determined by
If either the updating rule, u sd or u ac , is satisfied, specific the unbiased input data drawn at the beginning of Monte
capacity cells can either initiate the shutdown process or the Carlo run, r (as defined in 3), are returned to it. Addi-
reactivation process, respectively, depending on their state. tional information regarding UE handovers is available in
If no capacity cell meets the conditions of either updating Subsection VII-C3.
rule, u sd or u ac , meaning, C →sd ∪ C →sd = ∅, the current The distance to threshold metric that we have designed
agent configuration, x, becomes stable, and the Monte Carlo effectively implements the earlier-discussed logic, as it
run, r, terminates. The ABM then proceeds to the next run or ensures that the capacity cell with the greatest distance to
advances to the next hour (if the maximum number of runs, R, threshold is most likely to be selected and consequently take
has been reached), or concludes (if it was already processing action first. The stochastic nature of this process should also
the last hour). be noted.

796 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

3) USER TRANSFER 1̂cDL

a ,h′ ,r = 0. In line with the granularity of 1 hour that is
When capacity cell, c a,
starts its shutdown process, it being employed, the anticipated duration, t̂i,h CS , of the carrier
′ ,r
commands its set of connected UEs, Uc a , to perform inter- a
shutdown for capacity cell, c , during the specified hour, h′ ,
frequency handovers to the frequency of its paired coverage and Monte Carlo run, r, is set to 60 minutes.
cell, b(c a ), through the A4 handover entry condition (see Importantly, our ABM stores all these movements of UEs
Section II-C). We denote by Ic a the set of cells in such and PRBs from the capacity cell, c a , to each of its neigh-
frequency. bouring cells, j ∈ Ic a , at hour, h′ , Monte Carlo run, r, and
In this work, we use the stochastic model presented in [73] time step, t, in a replay memory where all relevant actions
to estimate the probability, pcHO a ,j,h′ , of one such UE being are recorded.
handed by capacity cell, c a , over a neighboring cell, j ∈ Ic a , In the latter case, if the capacity cell, c a , is waking up, the
at hour, h′ . Note the algorithm presented in [73] makes use of agent configuration, x = (x1 , · · · , xc , · · · , xC ), is modified,
the engineering parameters, d EP , and the measurement report with xc a flipping from 0 to 1. We then retrieve, from the
(MR), d MR , data sets (see Section IV). previously mentioned replay memory, the UEs connected
Definition 5 (Transfer Probability, pcHO a ,j,h′ ): Let us denote to —and the used DL PRBs in— capacity cell, c a , at the
by pcHOa ,j,h′ the probability of a UE transferring from capacity beginning of this Monte Carlo run, r, i.e. at t = 1, at this
cell, c a , to a neighboring cell, j ∈ Ic a , when performing hour, h′ , and hand them back to it, updating the set of out-
puts, Xi,h MC-OUT , accordingly. Note that the replay memory is
an inter-frequency handover through the A4 handover entry ′ ,r

condition, at hour, h′ . particularly useful to identify in which neighbouring cells, the

To decide to which specific neighbouring cell, j ∈ Ic a , original UEs of capacity cell, c a , are, and subtract them and
each UE, u, of capacity cell, ca , is handed over at hour, h′ , their numbers of PRBs from their statistics. Note that in this
CS , of the carrier shutdown of
we build an empirical PMF using all values, pcHO a ,j,h′ , from all case, the predicted duration, t̂i,h′ ,r
a ′
capacity cell, c , at hour, h , and Monte Carlo run, r, is set to
neighbouring cells, j ∈ Ic a , and draw a target neighbouring
cell, j t , from it. 0 minutes.
Importantly, if UEs cannot be handed over any neighbour-
ing cell due to, e.g. the existence of coverage holes —no 5) TERMINATION AND EXPERT-BASED, WHITE-BOX
suitable neighbour exits—, i.e. pcHO a ,j,h′ = 0, ∀j ∈ Ic a , MODEL STATISTICS
the capacity cell, c a , aborts its shutdown process, and the Once all the output variables within the set, Xi,h MC-OUT , have
′ ,r
sequential process of the Monte Carlo run moves to the next been updated for a particular time step, t, the ABM proceeds
time step, t + 1. to the subsequent time step, t + 1. At this juncture, a fresh
evaluation of updating rules, agent selection, UE transfers,
and the subsequent update and storage of statistics is carried
4) UPDATING AND STORING OF MONTE CARLO RUN out, as delineated in the preceding subsections. This iterative
STATISTICS progression persists until the current Monte Carlo run, r,
At this juncture, the capacity cell, c a , can be in one of two attains stability, meaning no agent can modify its attributes
operational states: either it is in the process of shutting down due to the inability of any candidate to satisfy the updating
or it is undergoing reactivation. rules, or until the maximum allowed number of time steps,
In the former scenario, if all UEs, Uc a , within the capacity tmax , within a Monte Carlo run is reached.
cell, c a , can be seamlessly handed over to a neighboring cell, Upon reaching either of these conditions, the ABM pro-
j ∈ Ic a , meaning that the sum P of the related handover prob- ceeds to the subsequent Monte Carlo run, r+1, at which point
ability is larger than zero, i.e. j pcHO a ,j,h′ > 0, the capacity the entire process described above is replicated. However,
cell, c a , proceeds to shut down. As a consequence, the agent MC-IN
in this instance, a fresh set of random input variables, Xi,h′ ,r
configuration vector, x = (x1 , · · · , xc , · · · , xC ), is updated, (as defined in Definition 3), is drawn to initialize this new
specifically flipping the value of xc a from 1 to 0. In addition, Monte Carlo run.
the set of outputs denoted as Xi,h MC-OUT (as described in
′ ,r In the event that the maximum specified number of runs, R,
Definition 4) is revised as follows: For each UE, u, within is reached, the ABM moves on to the next hour, h′ + 1, unless
set, Uc a , that transfers to neighboring cell, j t , the predicted it is already processing the final hour. If this is the case, the
number, Ûj UE t
t ,h′ ,r , of RRC connected UEs in cell, j , for that simulation concludes.
′
particular hour, h , and Monte Carlo run, r, is incremented A crucial point to emphasize is that, at the conclusion of
by one unit. Simultaneously, the counter associated with the each hour, h′ , the ABM computes the set, Xi,h OUT , of stochas-
′
capacity cell, c a , is set to 0, i.e. ÛcUEa ,h′ ,r = 0. Furthermore, tic outputs generated by our expert-based, white-box model
the projected number, 1̂j t ,h′ ,r , of utilized DL PRBs in cell, j t ,
DL for that hour. This calculation is performed using the set,
Xi,hMC-OUT , encompassing the outputs from each Monte Carlo
at the specified hour, h′ , and Monte Carlo run, r, is augmented ′ ,r
1̂ DL
a ′ run, r ∈ {1, · · · , R}. For example, the anticipated average
by cU ,ha ,r , with the assumption that all UEs, u, within set, CS , of carrier shutdowns for each capacity cell, c,
duration, t̂i,h
c ′
Uc a , on average employ the same number of PRBs. Similarly, P CS
CS = {1,··· ,R} t̂i,h′ ,r
the counter tied to the capacity cell, c a , is set to 0, i.e. during hour, h′ , can be computed as t̂i,h′ R .

VOLUME 2, 2024 797

This same reasoning can be applied to other variables as test set and calculate the ‘ground truth’, vital for validating
well. Importantly, by utilizing the results from the R Monte our model results. Ground truth denotes the actual measured
Carlo runs for the hour, the ABM is capable of providing truth, as measured directly from our dataset. It is important
not only average statistics but also distributions, as previously to note that each dataset, and thus our ground truth, has been
discussed. aggregated on an hourly basis, providing 24 values per day
for each cell. These values offer insights into traffic, energy
D. ROLLING FROM HOUR TO HOUR consumption, and measurement report statistics (see Table 2
At the onset of the first Monte Carlo run of the initial hour, and 3).
specifically when h′ = 1, r = 1, and t = 1, we made We benchmark our framework’s performance with respect
the assumption that all capacity cells are in an active state, to an expert-based approach, currently utilized by a leading
i.e xc = 1, c ∈ CC . Consequently, by the end of this first network provider, which is presented in the following.
hour, we achieve R stable Monte Carlo runs each with a
corresponding stable agent configuration, that is, xh′ =1,r =
A. BENCHMARK ALGORITHM
(x1 , · · · , xc , · · · , xC ) for all r ∈ {1, · · · , R}.
To benchmark the performance of our modeling framework,
Upon transitioning from the present hour, h′ , to the subse-
we employ an approach used by experts to predict carrier
quent hour, h′ + 1, two scenarios may unfold: i) All capacity
shutdown network performance in the field. It is crucial to
cells are activated anew. This scenario allows for the indepen-
note that this scheme is deterministic. In more detail, it first
dent processing of different hours. ii) A subset of capacity
assumes that all capacity cells are active at the beginning of
cells may be deactivated based on outcomes from the prior
every hour, i.e. xc = 1 for all c ∈ CC . Then, a sequential
hour, h′ . In this case, the hours are no longer independent,
modeling approach is adopted, where the capacity cells are
and a rolling process must be implemented.
shut down one by one in a given order. The average numbers,
In our implemented approach, during the transition from UE and 1 DL , of RRC-connected UEs to and used DL
Ui,h
one hour, h′ , to the next, h′ + 1, a starting agent configura- i,h
PRBs by every cell, i, at each hour, h, are used to check
tion for the new hour is randomly selected from the replay
the shutdown condition rule, u sd (see Section VII-C1), and
memory. This selection is made among the R stable agent
decide the deterministic order of shutdown. Once this order
configurations from the preceding hour, h′ . It is worth noting
MC-IN , in the subsequent hour, h′ + 1, is fixed, the capacity cells go through the UE transfer scheme
that the set of inputs, Xi,h ′ +1,r
presented in Section VII-C3. After processing all capacity
differs from that in the previous hour, h′ , due to unbiased
cells that meet the shutdown condition rule, u sd , the statistics
traffic statistics evolving from hour to hour, i.e. fU UE ̸ = fU UE
i,h i,h+1 for the hour are derived. These statistics include the predicted
and f1 DL ̸ = f1 DL . Consequently, to calculate the statistics UE and 1̂ DL , of RRC-connected UEs to
i,h i,h+1 average numbers, Ûi,h ′ i,h′
for the chosen starting agent configuration, a simulation of the and used DL PRBs by every cell, i, at each hour, h′ , as well
carrier shutdown process from a state with all capacity cells CS , of the carrier shutdown. These
as the predicted duration, t̂i,h′
active to the selected starting agent configuration is needed. values are feed into the ML model presented in Section V to
This simulation involves the requisite UE transfers as well provide an estimate of cell power consumption. As for the UE
as updates and storage of relevant statistics. Importantly, this rate prediction, the mean estimator presented in Section VI
simulation does not necessitate a review of updating rules or a is used. It’s important to note that this method only allows
stochastic agent selection, as the desired agent configuration obtaining one estimate for each target cell KPI in each hour,
is already known. Once the chosen agent configuration is not a distribution, as the model is not stochastic, in contrast
attained, the Monte Carlo run can be initiated. with the proposed framework.
Algorithm 1 summarised this fused Monte Carlo and ABM
framework.
B. CONVERGENCE ANALYSIS OF ABM ESTIMATION
VIII. EXPERIMENTS AND ANALYSIS ERRORS
In this section, we assess the performance of the proposed To validate the capabilities of the proposed model, this section
energy-saving modeling framework in estimating network analyzes the convergence of estimation errors as we increase
performance during carrier shutdown activation. the maximum number of time steps allowed in each Monte
To evaluate the framework introduced in Section VII, Carlo run of the ABM. Fig. 9 demonstrates that the average
we employ datasets from a real network in a metropolitan MAE of the estimates produced by the ABM is significantly
area of China. Our focus is on a set denoted as C, com- influenced by this parameter. Notably, when this threshold is
prising 657 cells categorized into (i) 375 capacity booster extended beyond 150, both types of errors rapidly decrease,
cells (constituting the set C C ) and (ii) 282 coverage cells achieving convergence once the maximum number of steps
(comprising the set C B ). These datasets spanned 12 days per Monte Carlo run reaches 400. By setting the maxi-
in April 2022 and were crucial for constructing the models mum number of steps per Monte Carlo run to this level,
detailed in Sections V, VI, VII, and VIII-A. Additionally, data ABM requires approximately 2 minutes to simulate network
from a 3-day period in May 2023 were utilized to create the performance over a 24-hour period. This outcome lays the

798 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

Algorithm 1 Carrier Shutdown, Expert-Based, White-Box Model

IN ∀i ∈ C, h ∈ {1, · · · , h, · · · , 24}
input : Engineering parameters data sets, d EP , Measurement report data set, d MR , and Unbiased distributions, f UE and f DL in Xi,h
Ui,h 1i,h
(see Definition 1), extracted from cell-level KPIs data set, d KPI
output: Predicted distributions, fˆ UE , fˆ DL and fˆ CS in X OUT′ ∀i ∈ C, h ∈ {1, · · · , h , · · · , 24} (see Definition 2)
′ ′
U ′ 1 ′ t ′ i,h
i,h i,h i,h
for h′ ∈ {1, · · · , h′ , · · · , 24} do
for r ∈ {1, · · · , r, · · · , R} do
X MC-IN ← Ũ UE ← f UE and X MC-IN ← 1̃ DL′ ← f DL // Draw the set of inputs, X MC-IN , of this Monte
i,h′ ,r i,h′ ,r U
i,h i,h′ ,r i,h ,r 1 i,h i,h′ ,r
Carlo run (see Definition3)
if h′ == 1 and r == 1 then
xc = 1, ∀c ∈ CC // Initialize the agent configuration, x = {1, · · · , xc , · · · , XC }
else
/* Use the rolling process from hour to hour to get the starting agent configuration */
end
/* Start random walk within the Monte Carlo run */
for t ∈ {1, · · · , t, · · · , tmax } do
/* Updating rule check */
for c ∈ CC do
if xc == 1 and u sd == 1 then
C →sd ← c // Identify the capacity cells that can be shut down (rule, u sd )
else
C →ac ← c // Identify the capacity cells that can be reactivated (rule, u ac )
end
end
/* Check if a stable agent configuration is found -- Termination of Monte Carlo run */
if C →sd ∪ C →sd then
break;
end
/* Agent selection */
for c ∈ C →sd do
dc,h′ ,r = χc,b(c)
DL − 1̃ DL
c,b(c),h′ ,r
// Compute distance to threshold for cells in set, C →sd
end
for c ∈ C →ac do
DL // Compute distance to threshold for cells in set, C →ac
dc,h′ ,r = 1̃ DL ′ − 9c,b(c)
b(c),h ,r
end
dc,h′ ,r
d N′ = P // Normalise distance to threshold
c,h ,r d
C →sd ∪C →sd c,h′ ,r
N
ω = PMF(d ′ )∀c ∈ C →sd ∪ C →sd // Create PMF, from which a capacity cell will be selected at
c,h ,r
random to change its attribute
ca ←ω// Select at random the capacity cell to change its attribute
/* User transfer */
if xc a == 1 then
Derive Ic a and calculate p HO ∀j ∈ Ic a // Derive the statistics necessary to drive the UE transfer
c a ,j,h′
of the selected capacity cell, using information of data sets, d EP and d MR
if p HO
a ′ == 0∀j ∈ Ic a then
c ,j,h
continue // If a UE cannot be transferred, the shutdown process is abandoned
else
xc a = 0 // Shut down the selected capacity cell
/* Updating and storing of statistics */
Û UE = 0 and 1̂ DL = 0 and t̂ CS = 60 min;
c a ,h′ ,r c a ,h′ ,r c a ,h′ ,r
for u ∈ Uc a do
j t ← PMF(p HO a ′ )∀j ∈ Ic a ;
c ,j,h
1̂ DL
c a ,h′ ,r
Û UE = Û UE + 1 and 1̂ DL = 1̂ DL + ;
j t ,h′ ,r j t ,h′ ,r j t ,h′ ,r j t ,h′ ,r Uc a
end
end
else
xc a = 1 // Reactivate the selected capacity cell
/* Updating and storing of statistics */
Û UE = U UE and 1̂ DL = U UE and t̂ CS = 0 min;
c a ,h′ ,r c a ,h′ ,r c a ,h′ ,r c a ,h′ ,r c a ,h′ ,r
In addition, discount number of UEs and used DL PRBs from the respective counters of the respective cells hosting the UEs of the awaking capacity
cell, c ; a
end
/* Updating of replay memory */
Save the agent configuration and related statistic of this hour, h′ , Monte Carlo run, r, and time step, t, in the replay memory
end
X MC-OUT ← Û UE , 1̂ DL′ and t̂ CS′ ∀i ∈ C // Update set, X MC-OUT , of this Monte Carlo run (see
i,h′ ,r i,h′ ,r i,h ,r i,h ,r i,h′ ,r
Definition 4). Note that the last entry to the replay memory belongs to the stable agent
configuration and statistics of this Monte Carlo run, r, at this hour, h′
end
X OUT ← X MC-OUT ∀i ∈ C // Obtain the set, X OUT , of outputs of our expert-based, white-box model for
i,h′ i,h′ ,r i,h′
this hour, h′ , from the set of outputs, Xi,hMC-OUT
′ ,r , of its R Monte Carlo runs
end

VOLUME 2, 2024 799

FIGURE 9. Average MAEs with respect to the maximum number FIGURE 10. Daily profile of the mean shutting down time of
of time steps per ABM iteration. capacity booster cells; ground truth vs the Benchmark
algorithm and ABM estimations.

groundwork for discussing SRCON modelling performance

in subsequent sections. TABLE 6. Estimation accuracy of the carrier shutdown time of
the capacity booster cells.
C. PREDICTING THE SHUTTING DOWN TIME OF
CAPACITY CELLS
In this section, we evaluate the performance of the proposed
framework in estimating the carrier shutdown duration of the
capacity booster cells.
As explained in Section II, this process is stochastic, influ-
enced by the number of UEs associated with the shutting
down capacity cell and its paired coverage cell, as well as their D. PREDICTING THE LOAD OF NETWORK CELLS
DL and UL PRB loads. Accurately estimating the capacity In this section, we assess the performance of the proposed
cell shutdown duration is a prerequisite to assess network load framework in estimating the cell load.
and UE performance. With respect to energy-savings, it holds At the network level, load variation is a well-known
significance for a precise evaluation of network energy con- phenomenon, decreasing at night and increasing in the morn-
sumption, and enables an understanding of the effectiveness ing until reaching a peak in the late evening. At the cell
of the deployed energy-saving policy. level, in the absence of energy-saving features, load vari-
Fig. 10 illustrates the daily profile of the mean shutdown ation primarily depends on user mobility. However, the
time of the capacity cells in the network under consideration activation of carrier shutdown introduces a new dynamic,
in the test dataset. As expected, the shutdown time increases as lightly loaded capacity booster cells attempt to han-
during the night until a peak, after which the network load dover their UEs (as explained in Sec. II-C) to switch off
begins to rise, causing the mean shutdown time to decrease and reduce network energy consumption. Consequently, car-
accordingly. In the evening, as the network load reduces, rier shutdown also impacts the cell load distribution in the
additional capacity cells can shut down. network.
Additionally, it is evident from Fig. 10 that the proposed Fig. 11 illustrates the daily profile of the average cell load
framework accurately characterizes the average shutdown in the considered network. As expected, its pattern aligns
time during each hour of the day by adapting its estimate with the typical network load profile described in [74]. This
to the variation in network load. In contrast, the benchmark plot demonstrates how both the estimates of the proposed
algorithm consistently overestimates the shutdown time and framework and the benchmark follow the trend of the network
fails to adjust its estimate according to the network load, this load during the day. Our experiments reveal that the ABM
remains rather flat throughout the day. estimate adeptly follows drastic variations in cell load pro-
Table 6 describes the average MAE, and MAPE achieved files, providing more precise estimations than the benchmark
by the investigated algorithms and highlights the accu- algorithm. Specifically, as summarized in Table 7, the ABM
racy improvement of ABM with respect to the Benchmark improves the accuracy of mean cell load predictions with
algorithm. Specifically, over 24 hours, we observe accuracy respect to the benchmark algorithm, achieving a 57.89 % and
gain of 48.53%, and 87.28%, in terms of MAE, and MAPE, 50.74 % reduction in terms of average MAE and MAPE,
respectively. respectively.

800 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

TABLE 8. Estimation accuracy of the energy consumption of the

capacity booster cells.

when it reaches its minimum value. In contrast, the power

consumption profile of the capacity cells is notably flatter
when the energy-saving policy is not active, i.e. without
carrier shutdown, the energy consumption does not scale well
with the actual network load. From Fig.12, it is observed that,
on average, adjusting the network capacity to the variations
FIGURE 11. Daily profile of the average cell load; ground truth vs of network load through energy-saving policies results in an
the Benchmark algorithm and ABM estimations. energy consumption saving of 50 % at the capacity cells.
The power consumption estimations computed through the
TABLE 7. Estimation accuracy of the cell load. power consumption model presented in Section V, using
the estimated shutdown times and loads of the cells in the
network based on the proposed ABM model and the bench-
mark scheme, are also reported in the figure. Notably, both
algorithms successfully capture the general trend of energy
consumption. However, the estimation based on the proposed
ABM model is much closer to reality, owing to its superior
capability in accurately estimating the mentioned loads and
shutdown times.
Table 8 details the average MAE and MAPE achieved
by the investigated algorithms, showcasing the accuracy
improvement of ABM compared to the benchmark algorithm.
Specifically, over 24 hours, we observe a significant accuracy
gain of 62.08 % and 62.06 % in terms of the average MAE and
MAPE, respectively.

F. PREDICTING THE UE DL RATE

To conclude our analysis, we discuss the capabilities of our
framework in characterizing the mean UE DL rate, as illus-
trated in Fig. 13. As expected, this parameter exhibits an
opposite trend to the cell load, detailed in Fig. 11. During the
FIGURE 12. Daily profile of the energy consumption of capacity
night, when the cell load decreases, more resources become
booster cells; ground truth vs the Benchmark algorithm and available in each cell for active UEs, leading to a better end-
ABM estimations. user experience. In contrast, as the cell load increases during
the day, the corresponding user rate decreases due to the
limited resources per active UE.
E. PREDICTING THE ENERGY CONSUMPTION OF From these results, we can observe that the proposed
CAPACITY CELLS UE rate estimator based on gradient boosting presented in
In this section, we assess the performance of the proposed Section VI tends to underestimate the true data rate during
ABM in estimating the energy consumption of the capac- nighttime and slightly overestimate it during the daytime.
ity cells. Fig. 12 depicts the ground-truth average energy In contrast, the benchmark estimator consistently overesti-
consumed by all the capacity cells at each hour of the day, mates the ground truth, resulting in larger errors compared to
both with and without activating the energy-saving policy, i.e. the proposed model. The proposed framework, building on
in the test and training datasets. It is important to note that the our ABM, can account for the impact of the energy-saving
values have been normalized for privacy reasons. scheme on the cell-level KPIs, which ultimately determines
When carrier shutdown is active, the power consumption the UE rate.
of the capacity cells is high during the daytime, reaching its As summarized in Table 9, our solution significantly
peak at 8 pm, and then decreases during the night until 4 am improves the accuracy of the benchmark when estimating the

VOLUME 2, 2024 801

 IX. CONCLUSION
In conclusion, this paper not only introduced the SRCON
framework, but also rigorously evaluated its capabilities in
the context of the research questions and scientific objectives
laid out at the outset. Through this comprehensive data-driven
approach, aimed at modeling the energy efficiency of cellular
networks with a focus on carrier shutdown policies, we have
made significant strides towards addressing the pressing chal-
lenges identified in our research.

A. FEASIBILITY OF DATA-DRIVEN MODELING

Our findings affirmatively answer the first research question,
demonstrating that a data-driven model, relying exclusively
on readily accessible network data and excluding costly DT,
FIGURE 13. Daily profile of the average user rate; ground truth vs can indeed deliver accurate predictions of network energy
the Benchmark algorithm and ABM estimations. consumption and UE throughput. The SRCON framework,
by leveraging advanced ML- and expert-based models, has
accurately estimated key network performance parameters in
TABLE 9. Estimation accuracy of the user rate.
response to diverse carrier shutdown configurations, thereby
showcasing remarkable accuracy.

B. BLENDING MACHINE LEARNING WITH DOMAIN

EXPERTISE
In addressing our second research question, the SRCON
user rate, achieving a nearly 50 % reduction in mean MAE framework exemplifies the seamless integration of ML algo-
and around a 57 % reduction in mean MAPE. rithms with domain expertise. Our innovative, expert-driven
ABM incorporates expert knowledge by implementing prod-
uct algorithms, specifically focusing on carrier shutdown
G. DISCUSSION ON GENERALIZATION TO OTHER
and HO mechanisms. It also effectively interacts with our
SCENARIOS
novel ML-based models for energy consumption and UE
It is imperative to note that the precise calibration of our throughput to ensure precise predictions. This integration has
model is fundamentally intertwined with the specificities of resulted in significant accuracy improvements, underscoring
the vendor’s equipment and the deployment characteristics the efficacy of our approach in advancing energy efficiency
it was originally designed for. Consequently, the model pre- modeling within large-scale networks.
sented in this paper exhibits an optimized performance for
metropolitan networks that deploy analogous technological
frameworks than those used in training scenario, specifically C. ADVANCEMENTS AND COMPARATIVE EVALUATION
4G and 5G networks. Regarding our third question, the SRCON framework has
The necessity for comprehensive retraining of the ML introduced notable advancements in network energy effi-
models arises when attempting to transpose our energy- ciency prediction. When compared to existing methodolo-
saving solutions and performance prediction models to gies, notably a state-of-the-art approach used by a network
alternative network setups or deployment scenarios. This operator, SRCON achieved a 62.08 % reduction in MAE
retraining process is vital for accommodating the varying and a 62.06 % reduction in MAPE for energy consump-
power consumption profiles and environmental conditions tion modeling. For UE throughput estimation, significant
unique to different equipment or scenarios. improvements were observed with approximately a 50 %
Furthermore, deviations in the carrier shutdown and han- reduction in MAE and a 57 % reduction in MAPE, under-
dover (HO) algorithms, from the framework specified, neces- scoring its superior predictive performance.
sitate a thoughtful redesign of the corresponding expert-based In summary, the SRCON framework signifies a paradigm
models within our ABM. Such adaptations underscore our shift towards practical, data-driven network modeling and
commitment to providing precise, localized predictions. optimization. It leverages existing network data to accurately
This deliberation accentuates the importance of a thor- predict resource utilization, energy consumption, and UE
ough evaluation of network infrastructure, vendor-specific throughput, offering a viable solution for modeling and opti-
technologies, and the nuanced characteristics of the network mizing wide-area networks and fulfilling both the research
environment prior to the deployment of our proposed model questions posed and the scientific objectives set forth at the
in real-world settings. beginning of this study.

802 VOLUME 2, 2024

López-Pérez et al.: Data-Driven Energy Efficiency Modeling in Large-Scale Networks

REFERENCES [23] M. Feng, S. Mao, and T. Jiang, ‘‘BOOST: Base station on-off switching
strategy for green massive MIMO HetNets,’’ IEEE Trans. Wireless Com-
[1] Huawei Technologies. (Aug. 2020). Green 5G: Building a
mun., vol. 16, no. 11, pp. 7319–7332, Nov. 2017.
Sustainable World. Huawei Technol. [Online]. Available:
https://www.huawei.com/en/public-policy/green-5g-building-a- [24] L. Chiaraviglio, F. Cuomo, M. Listanti, E. Manzia, and M. Santucci,
sustainable-world ‘‘Fatigue-aware management of cellular networks infrastructure with sleep
[2] D. López-Pérez et al., ‘‘A survey on 5G radio access network energy modes,’’ IEEE Trans. Mobile Comput., vol. 16, no. 11, pp. 3028–3041,
efficiency: Massive MIMO, lean carrier design, sleep modes, and machine Nov. 2017.
learning,’’ IEEE Commun. Surveys Tuts., vol. 24, no. 1, pp. 653–697, [25] H.-N. Dai, R. C.-W. Wong, H. Wang, Z. Zheng, and A. V. Vasilakos,
1st Quart., 2022. ‘‘Big data analytics for large-scale wireless networks: Challenges and
[3] Framework and Overall Objectives of the Future Development of IMT for opportunities,’’ ACM Comput. Surveys, vol. 52, no. 5, pp. 1–36, Sep. 2019.
2020 and Beyond, ITU-R, document M.2083-0, Sep. 2015. [26] Study on enhancement for Data Collection for NR and EN-DC, docu-
[4] GSMA. (2020). 5G Energy Efficiencies: Green is the ment TR 37.817, V17.0.0, 3GPP, Apr. 2022.
New Black. Accessed: Aug. 19, 2022. [Online]. Available: [27] Study on RAN-centric Data Collection and Utilization for LTE and NR,
https://data.gsmaintelligence.com/api-web/v2/research-file- document TR 37.816, V16.0.0, 3GPP TSG RAN, Jul. 2019.
download?id=54165956&file=241120-5G-energy.pdf [28] C. Zhang, P. Patras, and H. Haddadi, ‘‘Deep learning in mobile and wireless
[5] GSMA. (2022). Mobile Net Zero: State of the Industry on Climate networking: A survey,’’ IEEE Commun. Surveys Tuts., vol. 21, no. 3,
Action 2022. Accessed: Aug. 19, 2022. [Online]. Available: pp. 2224–2287, 3rd Quart., 2019.
https://www.gsma.com/betterfuture/resources/mobile-net-zero-state- [29] N. C. Luong et al., ‘‘Applications of deep reinforcement learning in com-
of-the-industry-on-climate-action-2022-report.pdf munications and networking: A survey,’’ IEEE Commun. Surveys Tuts.,
[6] Network Energy Efficiency (Phase 2), NGMN, Düsseldorf, Germany, vol. 21, no. 4, pp. 3133–3174, 4th Quart., 2019.
Feb. 2023. [30] S. Cammerer et al., ‘‘A neural receiver for 5G NR multi-user MIMO,’’ in
[7] Telecommunication management; Key Performance Indicators (KPI) for Proc. IEEE Globecom, Dec. 2023, pp. 329–334.
Evolved Universal Terrestrial Radio Access Network (E-UTRAN): Defini- [31] F. A. Aoudia, J. Hoydis, S. Cammerer, M. Van Keirsbilck, and A. Keller,
tions, document TS 32.450, V17.0.0, 3GPP, Apr. 2022. ‘‘Deep learning-based synchronization for uplink NB-IoT,’’ in Proc. IEEE
[8] S. Zhang, D. Yin, Y. Zhang, and W. Zhou, ‘‘Computing on base station Global Commun. Conf., Dec. 2022, pp. 1478–1483.
behavior using Erlang measurement and call detail record,’’ IEEE Trans. [32] S. Cammerer, J. Hoydis, F. A. Aoudia, and A. Keller, ‘‘Graph neural
Emerg. Topics Comput., vol. 3, no. 3, pp. 444–453, Sep. 2015. networks for channel decoding,’’ in Proc. IEEE Globecom Workshops,
[9] Umlaut, The Great 2022 Mobile Network Test, Special Mobile Dec. 2022, pp. 486–491.
Netw., Jan. 2022. [Online]. Available: https://www.accenture.com/us- [33] H. Huo, J. Xu, G. Su, W. Xu, and N. Wang, ‘‘Intelligent MIMO detection
en/services/industry-x/umlaut using meta learning,’’ IEEE Wireless Commun. Lett., vol. 11, no. 10,
[10] W. A. Hapsari, A. Umesh, M. Iwamura, M. Tomala, B. Gyula, and pp. 2205–2209, Oct. 2022.
B. Sebire, ‘‘Minimization of drive tests solution in 3GPP,’’ IEEE Commun. [34] Y. Qian, J. Xu, S. Zhu, W. Xu, L. Fan, and G. K. Karagiannidis, ‘‘Learn-
Mag., vol. 50, no. 6, pp. 28–36, Jun. 2012. ing to optimize resource assignment for task offloading in mobile edge
[11] Huawei Technologies. (2018). Huawei 5G Wireless Network Planning computing,’’ IEEE Wireless Commun. Lett., vol. 26, no. 6, pp. 1303–1307,
Solution White Paper. Huawei Technol. Accessed: Feb. 18, 2023. Jun. 2022.
[Online]. Available: https://www.huawei.com/fr/huaweitech/industry- [35] J. Feng, X. Chen, R. Gao, M. Zeng, and Y. Li, ‘‘DeepTP: An end-to-end
insights/outlook/mobile-broadband/insights-reports/5g-wireless-network- neural network for mobile cellular traffic prediction,’’ IEEE Netw., vol. 32,
plan-solution-whitepaper no. 6, pp. 108–115, Nov. 2018.
[12] D. He, B. Ai, K. Guan, L. Wang, Z. Zhong, and T. Kürner, ‘‘The design and [36] C. Zhang, H. Zhang, J. Qiao, D. Yuan, and M. Zhang, ‘‘Deep transfer
applications of high-performance ray-tracing simulation platform for 5G learning for intelligent cellular traffic prediction based on cross-domain
and beyond wireless communications: A tutorial,’’ IEEE Commun. Surveys big data,’’ IEEE J. Sel. Areas Commun., vol. 37, no. 6, pp. 1389–1401,
Tuts., vol. 21, no. 1, pp. 10–27, 1st Quart., 2019. Jun. 2019.
[13] Forsk Atoll. Accessed: Jun. 11, 2024. [Online]. Available: [37] Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, ‘‘Connecting
https://www.forsk.com/atoll-overview the dots: Multivariate time series forecasting with graph neural networks,’’
[14] Ranplan Wireless. Accessed: Jun. 11, 2024. [Online]. Available: in Proc. 26th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining,
https://www.ranplanwireless.com/es/ Aug. 2020, pp. 753–763.
[15] K. Koutlia, B. Bojovic, Z. Ali, and S. Lagén, ‘‘Calibration of the 5G-LENA [38] N. Piovesan et al., ‘‘Mobile traffic forecasting for green 5G networks,’’ in
system level simulator in 3GPP reference scenarios,’’ Simul. Model. Pract. Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2021, pp. 1–6.
Theory, vol. 119, Sep. 2022, Art. no. 102580. [39] S. Zhang et al., ‘‘Deep transfer learning for city-scale cellular traffic
[16] M. Mezzavilla et al., ‘‘End-to-end simulation of 5G mmWave networks,’’ generation through urban knowledge graph,’’ in Proc. 29th ACM SIGKDD
IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 2237–2263, 3rd Quart., Conf. Knowl. Discovery Data Mining, Aug. 2023, p. 4842.
2018. [40] S. Hui et al., ‘‘Large-scale urban cellular traffic generation via
[17] B. Leng, X. Guo, X. Zheng, B. Krishnamachari, and Z. Niu, ‘‘A wait-and- knowledge-enhanced GANs with multi-periodic patterns,’’ in Proc. 29th
see two-threshold optimal sleeping policy for a single server with bursty ACM SIGKDD Conf. Knowl. Discovery Data Mining, Aug. 2023,
traffic,’’ IEEE Trans. Green Commun. Netw., vol. 1, no. 4, pp. 528–540, pp. 4195–4206.
Dec. 2017. [41] A. Zoha, A. Saeed, A. Imran, M. A. Imran, and A. Abu-Dayya, ‘‘Data-
[18] E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, ‘‘Optimal design driven analytics for automated cell outage detection in self-organizing
of energy-efficient multi-user MIMO systems: Is massive MIMO the networks,’’ in Proc. 11th Int. Conf. Design Rel. Commun. Netw. (DRCN),
answer?’’ IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3059–3075, 2015, pp. 203–210.
Jun. 2015. [42] S. Jaffry, S. T. Shah, and S. F. Hasan, ‘‘Data-driven semi-supervised
[19] G. Yu, Q. Chen, R. Yin, H. Zhang, and G. Y. Li, ‘‘Joint downlink and uplink anomaly detection using real-world call data record,’’ in Proc. IEEE Wire-
resource allocation for energy-efficient carrier aggregation,’’ IEEE Trans. less Commun. Netw. Conf. Workshops (WCNCW), Apr. 2020, pp. 1–6.
Wireless Commun., vol. 14, no. 6, pp. 3207–3218, Jun. 2015. [43] S. Sevgican, M. Turan, K. Gökarslan, H. B. Yilmaz, and T. Tugcu, ‘‘Intelli-
[20] D. López-Pérez et al., ‘‘Energy efficiency of multi-carrier massive MIMO gent network data analytics function in 5G cellular networks using machine
networks: Massive MIMO meets carrier aggregation,’’ in Proc. IEEE learning,’’ J. Commun. Netw., vol. 22, no. 3, pp. 269–280, Jun. 2020.
Globecom, Dec. 2021, pp. 1–7. [44] A. Maatouk, F. Ayed, S. Biao, W. Li, H. Bao, and E. Zio, ‘‘A framework for
[21] Y. S. Soh, T. Q. S. Quek, M. Kountouris, and H. Shin, ‘‘Energy efficient the evaluation of network reliability under periodic demand,’’ IEEE/ACM
heterogeneous cellular networks,’’ IEEE J. Sel. Areas Commun., vol. 31, Trans. Netw, Jan. 2023.
no. 5, pp. 840–850, May 2013. [45] C. Gijón, M. Toril, and S. Luna-Ramírez, ‘‘Data-driven estimation
[22] H. Çelebi, Y. Yapici, I. Güvenç, and H. Schulzrinne, ‘‘Load-based on/off of throughput performance in sliced radio access networks via super-
scheduling for energy-efficient delay-tolerant 5G networks,’’ IEEE Trans. vised learning,’’ IEEE Trans. Netw. Service Manag., vol. 20, no. 2,
Green Commun. Netw., vol. 3, no. 4, pp. 955–970, Dec. 2019. pp. 1008–1023, Jun. 2023.

VOLUME 2, 2024 803

[46] Z. Xing et al., ‘‘Spectrum efficiency prediction for real-world 5G networks [60] C. Ruah, O. Simeone, and B. M. Al-Hashimi, ‘‘A Bayesian framework
based on drive testing data,’’ in Proc. IEEE Wireless Commun. Netw. Conf. for digital twin-based control, monitoring, and data collection in wireless
(WCNC), Apr. 2022, pp. 2136–2141. systems,’’ IEEE J. Sel. Areas Commun., vol. 41, no. 10, pp. 3146–3160,
[47] ETSI Environmental Engineering. (Oct. 2020). Es2023228: Oct. 2023.
Assessment of Mobile Network Energy Efficiency. [Online]. Available: [61] NR and NG-RAN Overall Description, document TS 38.300, V16.6.0,
https://www.etsi.org/deliver/etsi_es/203200_203299/203228/01.03.01_60 3GPP TSG RAN, Sep. 2019.
/es_203228v010301p.pdf [62] Study on Network Energy Savings for NR, document TR 38.864, V18.0.0,
[48] ITU. (2016). Energy Efficiency Measurement for Telecommunication 3GPP TSG RAN, Jan. 2023.
Equipment. [Online]. Available: https://www.itu.int/rec/T-REC-L.1350- [63] Z. Ghadialy. (Jul. 2023). Over 21 Trillion KPIs Per Week. [Online].
201610-I/en Available: https://www.operatorwatch.com/2023/07/over-21-trillion-kpis-
[49] ETSI Environmental Engineering. (Dec. 2020). TS103786: per-week.html
Measurement Method for Energy Efficiency of Wireless [64] Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Resource
Access Network Equipment; Dynamic Energy Efficiency Control (RRC); Protocol Specification, document TS 36.331, V17.3.0,
Measurement Method of 5G Base Station (BS) Dynamic Energy 3GPP TSG RAN, Jan. 2023.
Efficiency Measurement Method of 5G BS. [Online]. Available: [65] N. Piovesan et al., ‘‘Machine learning and analytical power consumption
https://www.etsi.org/deliver/etsi_ts/103700_103799/103786/01.01.01_60 models for 5G base stations,’’ IEEE Commun. Mag., vol. 60, no. 10,
/ts_103786v010101p.pdf pp. 56–62, Oct. 2022.
[50] Technical Specification Group Radio Access Network; NR; Study on User [66] S. M. Lundberg and S. Lee, ‘‘A unified approach to interpreting model
Equipment (UE) Power Saving in NR (Release 16), document TR 38.840, predictions,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 30, I. Guyon,
V16.0.0 3GPP, Jun. 2019. U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan,
[51] Technical Specification Group Services and System Aspects; Management and R. Garnett, Eds. Curran Associates, 2017. [Online]. Available:
and Orchestration; Study on New Aspects of Energy Efficiency (EE) for 5G https://proceedings.neurips.cc/paper_files/paper/2017/file/8a20a8621978
(Release 17), document TR 28.813, V17.0.0, 3GPP, 2021. 632d76c43dfd28b67767-Paper.pdf
[52] T. S. Rappaport, M. Ying, N. Piovesan, A. D. Domenico, and D. Shakya, [67] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning.
‘‘Waste factor and waste figure: A unified theory for modeling and analyz- Cambridge, MA, USA: MIT Press, 2016. [Online]. Available:
ing wasted power in radio access networks for improved sustainability,’’ http://www.deeplearningbook.org
2024, arXiv:2405.07710. [68] F. Pedregosa et al., ‘‘Scikit-learn: Machine learning in
[53] T. Li et al., ‘‘Carbon emissions of 5G mobile networks in China,’’ Nature Python,’’ J. Mach. Learn. Res., vol. 12, pp. 2825–2830,
Sustainability, vol. 6, no. 12, pp. 1620–1631, Aug. 2023. Nov. 2011.
[54] L. Maggi et al., ‘‘Energy savings under performance constraints via carrier [69] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Sta-
shutdown with Bayesian learning,’’ in Proc. Joint Eur. Conf. Netw. Com- tistical Learning: Data Mining, Inference and Prediction, 2nd ed.
mun. 6G Summit (EuCNC/6G Summit), Jun. 2023, pp. 1–6. Berlin, Germany: Springer, 2009. [Online]. Available: http://www-
[55] R. Moriconi, M. Deisenroth, and K. Sesh Kumar, ‘‘High-dimensional stat.stanford.edu/~tibs/ElemStatLearn/
Bayesian optimization using low-dimensional feature spaces,’’ in Proc. [70] D. P. Kroese, T. Brereton, T. Taimre, and Z. I. Botev, ‘‘Why the Monte
4th Workshop Bayesian Deep Learn. (NeurIPS), vol. 109, Vancouver, BC, Carlo method is so important today,’’ WIREs Comput. Statist., vol. 6, no. 6,
Canada, 2020, pp. 1925–1943. pp. 386–392, Nov. 2014.
[56] Z.-Q. Luo et al., ‘‘SRCON: A data-driven network performance simulator [71] R. Axelrod, The Complexity of Cooperation: Agent-Based Models of Com-
for real-world wireless networks,’’ IEEE Commun. Mag., vol. 61, no. 6, petition and Collaboration. Princeton, NJ, USA: Princeton Univ. Press,
pp. 96–102, Jun. 2023. 1997.
[57] L. U. Khan, W. Saad, D. Niyato, Z. Han, and C. S. Hong, ‘‘Digital-twin- [72] S. Banisch, Markov Chain Aggregation for Agent-Based Models. Berlin,
enabled 6G: Vision, architectural trends, and future directions,’’ IEEE Germany: Springer, 2015.
Commun. Mag., vol. 60, no. 1, pp. 74–80, Jan. 2022. [73] A. De Domenico, D. López-Pérez, W. Li, N. Piovesan, H. Bao, and
[58] L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong, X. Geng, ‘‘Modelling user transfer during dynamic carrier shutdown in
‘‘Digital twin of wireless systems: Overview, taxonomy, challenges, green 5G networks,’’ IEEE Trans. Wireless Commun, vol. 22, no. 8,
and opportunities,’’ IEEE Commun. Surveys Tuts., vol. 24, no. 4, pp. 5536–5549, Aug. 2023, doi: 10.1109/TWC.2023.3235112.
pp. 2230–2254, 4th Quart., 2022. [74] G. Auer et al., ‘‘How much energy is needed to run a wire-
[59] Y. Wu, K. Zhang, and Y. Zhang, ‘‘Digital twin networks: A survey,’’ IEEE less network?’’ IEEE Wireless Commun., vol. 18, no. 5, pp. 40–49,
Internet Things J., vol. 8, no. 18, pp. 13789–13804, Sep. 2021. Oct. 2011.

804 VOLUME 2, 2024