The Long Road to Sobriety: Estimating the Operational

Power Consumption of Cellular Base Stations in France

Arsalan Ahmed, Marceau Coupechoux

To cite this version:

Arsalan Ahmed, Marceau Coupechoux. The Long Road to Sobriety: Estimating the Operational
Power Consumption of Cellular Base Stations in France. The International Conference on Information
and Communications Technology for Sustainability (ICT4S), Jun 2023, Rennes, France. �hal-04082598�

HAL Id: hal-04082598

https://telecom-paris.hal.science/hal-04082598v1
Submitted on 26 Apr 2023

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.

Distributed under a Creative Commons Attribution 4.0 International License

 The Long Road to Sobriety: Estimating the
Operational Power Consumption of Cellular Base
Stations in France
Arsalan Ahmed† and Marceau Coupechoux§
† Faculté des Sciences et Technologies, Université de Lorraine, 54506, Vandœuvre-lès-Nancy, France
§ LTCI, Télécom Paris, Institut Polytechnique de Paris, 91120, Palaiseau, France

Email: [email protected], [email protected]

Abstract—As the Information and Communication Technology gains achieved with the arrival of newer cellular generations
(ICT) sector represents 1.8% to 3.9% of the global Green House of 4G and 5G, it is clear that these gains have at least
Gas (GHG) emissions, it is of upmost importance to know how been compensated by the increase in data traffic [4], [5].
much energy is spent annually in mobile networks and how this
consumption is evolving. It is quite likely that the huge energy Between 2002 and 2012, GHG emissions directly related to
efficiency gains achieved by technology evolution have at least ICT, grew twice as fast as the overall global emissions. From
been compensated by the surge in data traffic. Therefore, in this 2012 till 2015, studies lead to contrasting conclusions [3].
paper, we estimate the operational power consumption of cellular However, all authors agree on the fact that this share will not
Base Stations (BSs) deployed in France from 2015 to 2022. significantly reduce if the business keeps running as usual.
However, unfortunately, the lack of openly available data hinders
the estimation process. In order to work around this issue, we As mobile networks constitute a large part of the ICT sector,
rely on a public dataset on radio electric installations, on widely they are subject to ambitious targets of cutting down on GHG
adopted power consumption models and on a set of assumptions emissions, see e.g. the International Telecommunication Union
backed by the scientific literature. We demonstrate that, over (ITU) recommendation [6] which, by 2030, envisions a 45%
the considered period, the numbers of BSs and transceivers have reduction in the emissions of mobile network operators. But
grown at a sustained Compound Annual Growth Rate (CAGR)
of 7.55% and 18.27%, respectively. Within the same period, the the rise in BS installations and deployments are a serious
average BS power consumption has increased at a CAGR of hindrance to this goal as BSs alone are responsible for
9.89% while the total operational power consumption of BSs approximately 80% of the total energy consumption of mobile
has grown at a CAGR of 18.18%. We further show that the networks [7], [8]. Due to such trends, the increased energy
introduction of 5G has accelerated this trend despite the recent consumption of BSs has become a major issue which should
decommissioning of 2G and 3G transceivers. These alarming
figures advocate for proactive digital sobriety policies. gain more attention from the researchers.
Index Terms—Mobile Network, 5G, Base Station, Power Con- The modern end-user devices with newer mobile genera-
sumption, Digital Sobriety, France. tions like 5G are becoming more and more energy-efficient in
terms of energy per transmitted bit. However, these energy
I. I NTRODUCTION gains can easily be nullified due to the so called rebound
In recent years, the widespread proliferation of wireless effect [9], [10]. Direct rebound effect in mobile networks
mobile devices have prompted tremendous growth in the cel- occurs when the changing consumer behaviour increases the
lular communications industry [1]. Such industry expansions usage of energy-efficient end-user devices to the point that
are carefully guided and encouraged by the mobile operators the aggregate usage outweighs the potential energy savings.
who continually strive to provide more capacity through their Hence, it is necessary to evaluate the total energy consumption
networks. This addition of capacity is not only to cater the in J or Wh to be able to quantify the overall impact and not
increasing user demands but it is also an incentive for further only the energy efficiency in J/bit. However, historically,
traffic growth which eventually translates into higher profits. energy consumption at the BSs has been difficult to evaluate
For this purpose, mobile operators deploy additional physical due to the lack of openly available data from the mobile
Base Stations (BSs) or reinforce the current ones. Several operators and ICT equipment providers. On the other hand,
studies have investigated the Information and Communication users are expected to remain aware and conscious of their
Technology (ICT) sector’s global climate impact. ICT rep- mobile usage in light of the growing environmental impact of
resents between 1.8% and 3.9% of the global Green House mobile networks, especially if digital sobriety policies were
Gas (GHG) emissions* [3]. Despite the huge energy efficiency to be implemented. This serious contradiction can partially be
resolved with reliance on models and estimates which have
* There is an inherent complexity in evaluating the carbon footprint of
their own share of advantages and drawbacks.
ICT due to, e.g., the lack of accurate data or the difficult definition of the
sector’s boundary, see [2] for a discussion. The range provided by [3] is rather There exists a number of previous studies that try to estimate
pessimistic among literature’s estimates. the aggregate energy consumption of mobile networks [11],
[12], [13], [14]. The data used in these studies is mostly pro- Support Support
(SUP_ID) (SUP_ID)
vided upon request by telecom network operators. Malmodin
et al. investigate the electricity consumption and operational
carbon emissions of telecom operators from 2010 to 2015 [11]. Station Station
(STA_NM_ANFR) (STA_NM_ANFR)
The same authors in [12], present the operational electric-
ity consumption and GHG emissions for selected European
telecom network operators from 2015 to 2018. A similar Antenna Antenna
study is conducted for Finland where the current and future (AER_ID) (AER_ID)

energy consumption trends for mobile networks are discussed

[13]. Golard et al. present a method to evaluate and project
the total energy consumption of radio access networks using Transceiver Transceiver Transceiver
(EMR_ID) (EMR_ID) (EMR_ID)
on-site measurements provided by operators. This method is
applied to come up with 4G and 5G BS models which are
then used to predict the energy footprints for Belgium [14]. Band Band Band Band Band
However, since most of these works use collective values from (BAN_ID) (BAN_ID) (BAN_ID) (BAN_ID) (BAN_ID)

telecom operators, the break down of energy consumption is

not quite clear and visible especially in terms of location, Fig. 1: Hierarchy in the ANFR dataset.
BSs, operators and mobile generations. Furthermore, the data
collection methodology, calculation details and information
about continuous historical data remain unknown. II. ANFR DATASET AND DATA P ROCESSING
Hence, this work attempts to estimate the national opera- A. ANFR Dataset
tional power consumption of cellular BSs in France between
April 2015 and September 2022 using a public dataset on This study is based on the public dataset on radioelectric
radioelectric installations, some specific power consumption installations above 5 W [15] from the Agence Nationale des
models from the literature and accompanying assumptions FRéquences (ANFR) which is a governmental agency man-
related to the BS equipment. The key findings of this study aging all radio frequencies in France (https://www.anfr.fr/en).
are: The ANFR is tasked with the authorization of all transmission
site deployments above 5 Watts in mainland France as well
• The number of BSs has increased at a Compound Annual as in the overseas French territories. This publicly accessi-
Growth Rate (CAGR) of 7.55% while the number of ble dataset provides a unique opportunity for researchers to
transceivers has grown at a CAGR of 18.27%. Further- conduct studies on a national scale. The dataset contains data
more, existing BS sites are reused for newer genera- from all public and private operators of radio installations. It
tions instead of dedicated BS deployments. In addition, has been maintained from April 2015 and is updated on a
since the introduction of 5G in November 2020, the monthly basis. Each of the data files can be considered as a
decommissioning of several 2G and 3G transceivers has table containing some specific information. The dataset gathers
compensated the inclusion of 5G transceivers. information about different elements of the radio installations,
• Existing BSs are continually being reinforced with newer namely physical supports, BSs, antennas, transceivers and
cellular technologies, leading to multi-generational BSs frequency bands.
with an increased number of frequencies. Consequently, There exists a hierarchy in the dataset that defines the
the average BS power consumption has increased at a relationship between these elements as depicted in Fig. 1. It
CAGR of 9.89%. can be seen that a support (also known as a site) is at the
• The total BS power consumption has increased at a top of the hierarchy. Each support or site consists of several
CAGR of 18.18%. Since the introduction of 5G, this BSs which can belong to the same or different operators. In
increase has accelerated with a high CAGR of 19.75%. some rare cases, a single BS is attached to multiple supports.
This highlights the existence of a direct rebound effect For example, when a BS uses two antennas that are more
where the energy efficiency gains are negated due to the than 90 meters apart, then the BS is said to be supported by
rise in BS deployments. two different physical supports. A BS is owned by a single
• We show that our estimates are coherent with other data operator and utilizes one or more antennas. It is also possible
sources, despite the uncertainty in the estimation process. that several BSs share the same antenna. This is due to the
This validates our approach. sharing of telecom infrastructure where competitors become
partners to lower costs and investments. In France, leading
We provide a brief description of the dataset and data mobile operators have signed mutual strategic agreements to
processing in Section II. Section III presents the BS power share infrastructure among each other [16]. From the dataset,
consumption models. The numerical results are discussed in we observe that as of September 2022, around 17.86% of
Section IV while Section V is for discussion. We conclude the the antennas are shared between BSs belonging to several
paper in Section VI. different operators. An antenna is associated with one or more
transceivers. However, one transceiver can only use a single
antenna. Lastly, a transceiver can operate multiple frequency

Active Cooling
Pin or PBS

Mains Supply
bands but those bands belong to only a single type of wireless RF Pout

DC - DC
PA
communications standard/technology. BB Feeder
RF PA

B. Data Processing
We use python as the programming language for dataset
processing. Specifically, python’s Pandas library is mainly
(a) Integrated BS Model.
employed because of its support for cleaning, manipulating
and analyzing tabular data.
RRH
First, we preprocess the dataset before performing the Pin or PBS Optical Fibers

Mains Supply
analysis and computations related to power consumption. This RF Pout

DC - DC
PA
BB
is to filter out irrelevant information from it and reduce the RF PA
data size. For this preprocessing stage, we only consider the
cellular technologies/standards. These include GSM 900, GSM
1800, GSM 900/1800, UMTS 900, UMTS 2100, UMTS
2100/900, LTE 700, LTE 800, LTE 1800, LTE 2100, (b) RRH based BS Model.
LTE 2600, NR 700, NR 2100 and NR 3500. Overall, all
the mobile communication generations are taken into account Fig. 2: Two types of BS models along with their sub-units.
while excluding their experimental bands such as LTE 1400
Expe. Furthermore, we only consider the four leading French
operators as they hold the majority of the market share and number of transceiver chains present in the BS. The power
hence, their data samples are representative of the entire consumption of the PA is calculated as follows:
dataset. In addition, only the in service BSs are taken into con-
sideration, i.e., authorized BSs not yet in service are excluded Pout
PP A = (2)
from the analysis. We then utilize the compact preprocessed ηP A × (1 − σf eed )
dataset for faster data analysis and power consumption related
calculations. The complete source code of the project with where Pout is the output RF transmit power, ηP A is the power
further details on the dataset and data processing is publicly efficiency of the PA and σf eed is the feeder/cable loss arising
available on GitHub [17]. due to the physical distance between the BS and the antenna.
Note that Pout is generally an increasing function of the cell
III. BS P OWER C ONSUMPTION M ODELLING load and of the signal bandwidth.
We rely on accepted models from the literature for perform- We consider two types of BS models namely, the integrated
ing our power consumption estimations. BS model and the Remote Radio Head (RRH) based BS
model [19]. These models are shown in Fig. 2. An integrated
A. Power Consumption Models BS exists as a single unit placed at a distance from the
1) EARTH Model: The Energy Aware Radio and Network antennas. The PAs require active cooling within the BS and
Technologies (EARTH) project defines a widely used BS are connected to the antennas via long feeder cables. This
power consumption model [18]. The EARTH model assumes results in power losses of σcool and σf eed which are already
that all BSs can be generalized such that they consist of taken into account by Equation 1. On the other hand, in RRH
multiple transceivers where each transceiver serves a single based BSs, a RRH is introduced for hosting the PAs at the
antenna element. Each transceiver chain includes a power same physical location as that of the antennas. The RRH is
amplifier (PA), a RF small-signal transceiver module, baseband connected to the BS using optical fibers while the antennas
boards (BB), a DC-DC power supply, an active cooling system are connected to the RRH through small flexible jumpers. As
and an AC-DC unit (mains supply) for connection to the a consequence, the feeder loss is reduced to a lower value.
electrical power grid [18]. The mathematical formulation of The cooling loss is completely avoided as the PAs inside the
this model is the following: RRH are now cooled through natural air circulation.
2) 5G NR 3500 specific Model: The EARTH model is
PP A + PRF + PBB
PBS = NT RX × (1) applicable to a wide range of BS types. However, it does not
(1 − σDC )(1 − σM S )(1 − σcool ) accurately model 5G BSs operating at the 3500 MHz band.
where PBS is the total BS power consumption, PP A is This is because at this frequency, Massive MIMO is employed
the PA power consumption, PRF is the power consumption which requires the presence of several active antenna elements
of the transceiver module and PBB is the baseband power inside an Active Antenna Unit (AAU). Hence, we select a
consumption. In addition the σDC , σM S and σcool are the specific power consumption model that models 5G BSs con-
power losses in the DC-DC power supply, mains supply and sidering the impact of this massive beam-forming functionality
active cooling system, respectively. Furthermore, NT RX is the [20]. This model estimates the total BS power consumption as
follows: TABLE I: Assumption set for the EARTH model.
 
Pout Integrated BS RRH based BS
PBS = NT RX × + NT XRU PC + PBase (3)
ηP A Tech. 2G, 3G and 2G/3G 4G, 5G, 3G/4G, 4G/5G, 2G/4G,
2G/5G, 3G/5G, 2G/3G/4G,
where Pout is the output RF transmit power of the AAU, ηP A 3G/4G/5G, 2G/4G/5G,
is the power efficiency of the PA, NT XRU is the number of 2G/3G/5G and 2G/3G/4G/5G
transceiver units inside the AAU, PC represents the additional MIMO Single transmitter For 4G and 5G transceivers only.
• 2T2R MIMO for lower bands
digital and RF processing needed for each antenna branch including LTE 700, LTE 800 and
and PBase is the baseline power consumption. Furthermore, NR 700.
NT RX corresponds to the number of complete transceiver • 4T4R MIMO for higher bands
including LTE 1800, LTE 2100,
chains utilizing a single AAU present in the BS. In practice, LTE 2600 and NR 2100.
NT RX is equal to the number of sectors at the BS (typically 3). σf eed 0.5 0.2
It is to be noted that only a few models take Massive MIMO σcool 0.1 0
σDC 0.075
into account. Another one is described in [21]. However, nu- σM S 0.09
merical values of [20] provide better estimates when compared PRF 12.9 W
to publicly available data, like in [22]. PBB 29.6 W
3) Dependence on Load: The output power Pout is an in- ηP A 0.311
Pmax 4 W per downlink
√ MHz.
creasing linear function of the load, i.e., the average proportion LPRF 2
of radio resources used for the transmission. Assuming that LPBB 2
the BS has a maximum transmit power Pmax , we compute LηP A 1.02
the output power as follows: Pout = ρ × Pmax , where ρ is the dLT 8 years
ρ 0.3
BS load. This linear model is sufficiently accurate to be well
accepted by the literature, see e.g. [18], [23], [21].
B. Performance Improvements Note that we apply these performance improvements to the
The variable values in equations (1), (2) and (3) do not EARTH model only and not to the 5G NR 3500 specific
remain fixed but are updated with time in order to take into model. This is because the 5G NR is a relatively new tech-
account the continuous technological improvements resulting nology and its models do not yet require time based scaling.
in increased BS power efficiencies. In general, the analog
RF circuitry and the digital baseband circuitry are subject IV. N UMERICAL R ESULTS
to technology scaling on the basis of Moore’s law [24]. A. Numerical Assumptions
Moreover, a reduction of 2% per year is observed in the power
In order to obtain the variable values of the BS con-
lost inside a PA [24]. Therefore, the values of PRF , PBB and
sumption models, we make appropriate assumptions since
ηP A are updated as follows:
the information regarding BS equipment is proprietary and
n PRF kept confidential by the operators as well as the equipment
PRF = ⌊n⌋
(4)
LPRF
2 manufacturers.
The information coming directly from the preprocessed
n PBB dataset includes NT RX that is the number of complete
PBB = ⌊n⌋
(5)
LPBB
2
transceiver chains in each BS, total bandwidth associated with
each transceiver, the duplex mode of each transceiver and the
ηPn A = ηP A × LnηP A (6)
year in which each BS becomes in service. In order to get
n the value of NT RX at a BS, we group together and count
where PRF is PRF at year n, PRF is at year 0 and LPRF is
n all transceivers with the same characteristics. In addition, we
the constant scaling factor for PRF . Similarly, PBB is PBB
at year n, PBB is at year 0 and LPBB is the constant scaling calculate the total bandwidth of each transceiver by using the
factor for PBB . Lastly, ηPn A is ηP A at year n, ηP A is at year starting and ending frequency fields provided in the dataset.
0 and LηP A is the constant per year scaling factor for ηP A . Furthermore, for determining the duplex mode, we observe
In (4) and (5), n2 is subject to integer division since Moore’s their associated frequency bands and the related regulation.
law is applied every two years. Lastly, for each BS, a date of service field is provided in the
Equations (4), (5) and (6) are applied for state-of-the-art dataset which we use for getting the in service date.
BSs at year n. We assume that, according to the dataset, when Tab. I presents all the assumed values for each transceiver
a new BS is put in service at year n, such a state-of-the-art signal chain for power consumption related calculations with
equipment is installed. This particular BS will keep the same the EARTH model. We use the values measured (in 2011) and
parameters until year n + dLT , where dLT is the average BS not estimated (for 2020) by the EARTH model but then apply
lifetime duration in years. At year n+dLT , the values of PRF , the improvement procedure as described in Section III-B. The
n+dLT n+dLT
PBB and ηP A are updated to PRF , PBB and ηPn+d
A
LT
improvement figures for LPRF , LPBB and LηP A are taken
respectively, as if a current state-of-the-art BS were deployed. from [24]. From Tab. I, it can be seen that the BSs are
 TABLE II: Assumption set for the 5G NR 3500 model.
100000
All Generations
2G
MIMO 64T64R MIMO with the use of an AAU. 3G
PBase 260 W 2G/3G
80000 2G/3G/4G
ηP A 0.25 2G/3G/4G/5G

Number of In Service BSs

Pmax 240 W 3G/4G
NT XRU 64 3G/4G/5G
PC 1W 60000
ρ 0.3

40000

classified as either integrated or RRH based BSs. This is

determined based on the general industry practice where BSs 20000
belonging to the recent cellular generations of 4G and 5G
are using RRHs. Similarly, the Multiple-Input Multiple-Output
02015 2016 2017 2018 2019 2020 2021 2022 2023
(MIMO) configurations are also assumed. Mainly, the lower
bands of 4G and 5G employ 2T2R MIMO while the higher Fig. 3: Number of in service BSs by technology in France (the
bands up to 2.6 GHz use 4T4R MIMO. Furthermore, the vertical line indicates the introduction of 5G in France).
values for σf eed , σcool , σDC , σM S , PRF , PBB and ηP A are
adopted from the original work of the EARTH model [18].
Specifically, the value of σf eed is reduced from 0.5 to 0.2 for transceivers in November 2020 as indicated by the vertical
RRH based BSs in comparison with integrated BSs. This is line, 2G/3G/4G/5G and 3G/4G/5G BSs have started to
because the general feeder loss value for integrated BSs is 3 appear which are now quickly increasing in number. We
dB while for RRH based BSs is 1 dB [25]. In addition, the determine that between April 2015 and September 2022, the
maximum output power Pmax depends on the bandwidth of number of BSs has grown at a sustained CAGR of 7.55%.
the transceiver. We follow here the simple 4 W/MHz rule as Even if only the period after the introduction of 5G BSs is
in [14], where the considered bandwidth is only for downlink. considered, the CAGR value does not deviate much and is
For Frequency Division Duplexing (FDD), this is exactly the 7.2%. From this, we deduce that the newer generations reuse
transceiver bandwidth. For Time Division Duplexing (TDD), existing infrastructure instead of dedicated BS deployments.
the considered bandwidth is 2/3 of the transceiver bandwidth. Hence, BSs are now supporting an increased number of
We assume the average BS lifetime duration to be 8 years after cellular standards and frequency bands. This trend is also
which it is replaced with up-to-date equipment. This value is exemplified by the fact that in 2020, 97% of the 5G towers in
based on the fact that the lifetime of a typical BS is between 5 China were constructed on existing sites [27].
to 10 years [26]. We also assume a realistic load of ρ = 30% Fig. 4 presents the number of transceivers by cellular
which is close to the average load in busy hours [14]. technology. Currently, the total number of transceivers exceeds
Tab. II presents all the assumed values for the 5G NR more than 1, 600, 000. As expected, we observe that the
3500 specific model. The values in Tab. II are largely taken growth of 2G and 3G transceivers has stagnated in the recent
from [20]. The Pmax is now significantly higher at a value of past whereas 4G and 5G transceivers have been rising. A
240 W as compared to the EARTH model due to the presence CAGR of 18.27% is calculated for transceivers from April
of a 5G NR 3500 AAU in the signal chain [22]. 2015 to September 2022. Since the introduction of 5G, the
transceivers’ CAGR for only the next 12 months is found to
B. Equipment Count be 21.65%. However, when we consider the complete duration
In this section, we discuss the insights extracted from the since the introduction of 5G, the CAGR drops to 17.09%. This
preprocessed dataset. These insights do not depend upon the is due to the increase in decommissioning of several 2G and
selected power consumption models nor on our assumptions. 3G transceivers in 2022.
Fig. 3 shows the total count of in service BSs with respect
to the cellular technology in France from 2015 to 2022. C. Operational Power Consumption
From the most recent data point, it is revealed that there Fig. 5 depicts the calculated total BS operational power
are over 100, 000 in service cellular BSs in the country. consumption according to the cellular technology. As BSs
Furthermore, it is observed that the majority of the BSs generally include several generations of the technology, we
support multiple cellular generations. Since 2G and 3G mobile only represent the most significant combinations. It can be
networks are gradually becoming obsolete, a sharp downward seen that there is an almost 3.5-fold increase in the aggregate
trend is observed in the numbers of only 2G, only 3G and BS power consumption over the span of 7 years. Currently, the
2G/3G BSs. On the contrary, during the same time period, total BS power consumption crosses 350 MW which translates
BSs of the type 2G/3G/4G and 3G/4G have been increasing into 3 TWh energy consumption per year (approximately 0.7%
due to the addition of 4G LTE transceivers to the BS sites. of the total electricity consumption in 2021 in France [28]).
Furthermore, with the authorization and commissioning of 5G Since the introduction of 5G transceivers with high power
 All Generations All Generations
1600000 2G 350 2G
3G 3G
1400000 4G 2G/3G
5G 300

Total BS Power Consumption (MW)

2G/3G/4G
2G/3G/4G/5G
Number of Transceivers

1200000
250 3G/4G
3G/4G/5G
1000000
200
800000

600000 150

400000 100

200000
50
02015 2016 2017 2018 2019 2020 2021 2022 2023
02015 2016 2017 2018 2019 2020 2021 2022 2023
Fig. 4: Number of transceivers by technology in France.
Fig. 5: Total BS power consumption in France.

demand at the end of 2020, there has been a slight acceleration 7.51 7.59
in energy consumption. We compute that between April 2015 7 6.74
and September 2022, the total BS power consumption has

Average BS Power Consumption (KW)

increased at a sustained CAGR of 18.18%. Moreover, since 6 5.88
the introduction of 5G BSs, we found the power consumption
to have grown at a higher CAGR of 19.75%. This exhibits 5
the presence of a direct rebound effect where increasing BS 4.55
installations nullify the energy efficiency gains attributed with 4
newer cellular generations such as 5G. 3.54

We also observe that, from April 2015 to September 2022, 3

2.51
the average BS power consumption went from 1.75 kW to
3.53 kW. This translates into a sustained CAGR of 9.89%. 2
1.6
Considering only the period after the introduction 5G, the 1.0
1
average BS power consumption CAGR has escalated to 11.7%. 0.6
This is caused by the continuous reinforcement of current 0.3
0 1 2 3 4 5 6 7 8 9 10 11
BSs with new cellular technologies. Fig. 6 demonstrates how
Number of Frequencies per BS
the average BS power consumption rises with the number
of supported frequency systems as of September 2022. As Fig. 6: Average BS power consumption at 30% by the number
expected, the lower the number of frequency systems, the of frequencies (France, Sept. 2022).
lower is the average power consumption at the BS.
Fig. 7 shows the average BS power consumption with
respect to the highest technology present at the BS as of A. Coherence with other sources
September 2022. Here, a highest BS technology type includes Since we don’t have access to private data of the operators,
all systems before it. For example, 3G BSs include 2G, 3G it is difficult to verify the accuracy of the obtained results
and 2G/3G BSs. We make a distinction between the lower over the years. We thus crosscheck our estimates with other
bands and the 3500 MHz band of 5G in order to clearly sources of data. Lundén et al. present the operational electricity
demonstrate the effect of using 5G AAUs on the average BS consumption for European telecom network operators between
power consumption. The general trend is that the average 2015 to 2018 [12]. In this study, values for mobile access
BS power consumption rises with the existence of higher networks are provided for a number of mobile subscriptions
cellular generations at the BS. Specifically, BSs with the 5G ranging between 255 millions to 265 millions. In the same
3500 MHz transceivers consume the highest power due to the period, the number of mobile subscriptions in France went
presence of Massive MIMO supporting AAUs. from 66.68 million to 70.42 million [29]. We thus scale down
(French subscriptions represent 26%) the values of [12] and
V. D ISCUSSION provide the comparison in Tab. III. From the second column
of the table, it can be seen that our estimate accounts for
In this section, we check the coherence of our results with 40% to 70% of Lundén et al. evaluation. The difference can
other sources and perform a sensitivity analysis related to some be explained by the scope of [12] which includes the whole
model parameters. mobile network while we are focusing on BSs. BSs are indeed
 6.47 TABLE IV: Comparison with ARCEP collected data (2016-
2020) [30].
6
Average BS Power Consumption (KW)
Year ARCEP (fixed and ARCEP×60%×80% Our Work
5 mobile networks) [MW] (BSs) [MW]
[MW]
2016 340 163 126
4 3.91 2017 353 169 160
3.53 2018 387 186 197
2019 408 196 234
3 2020 434 208 250

2 1.96

we are able to estimate an average power consumption of

1
26.6 MW in 2017. This shows that our estimate is very close
0.33 0.5
to the reported value.
0 All Generations 2G 3G 4G 5G 5G According to [27], a 64T64R three-sectorized 5G macro BS
(700 & (3500 MHz)
2100 MHz) consumes approximately 3−4 kW which is 2 to 3 times higher
Highest BS Technology
than 4G equipment. The factor of 3 between 4G and 5G BSs
Fig. 7: Average BS power consumption at 30% load by highest is also given by an Executive Vice-President of China Mobile
BS technology (France, Sept. 2022). in [32]. The president of EJL Wireless Research mentioned
in [32] that a typical 5G site consumes 10 kW. The ITU-T
TABLE III: Comparison with Lundén et al.’s average annual Recommendation [33] reports a typical consumption of 8 kW
power consumption (2015-2018) [12]. for a 2G/3G/4G/5G BS. In [22], a 64T64R 5G NR BS with
transmit power of 240 W operating on 100 MHz bandwidth
Year Lundén et al. scaled Lundén et al. Our Work is said to consume 4297 W of power (about four times the
(whole mobile net- scaled×80% [MW] (BSs) [MW]
work) [MW]
consumption of a 8T8R 4G BS with 40 W on 20 MHz). These
2015 261 208 107 sources do not precise the corresponding load values. Based on
2016 267 213 126 our results shown in Fig. 7, the average power consumption
2017 273 218 160 of a 4G (highest technology) BS is 1.96 kW. Furthermore,
2018 282 225 197
on average, BSs equipped with only the lower bands of 5G
consume 3.91 kW while BSs running the 3500 MHz band
consume 6.47 kW at 30% load. Our results verify that 5G BSs
known to represent approximately 80% of the consumption consume 2 to 3 times more power than 4G BSs. In addition, for
of mobile networks [7]. If we apply this factor to the scaled 64T64R 5G BSs, our value of average power consumption
values, we obtain the third column of the table. With these lies between the reported values of 4.23 − 10 kW.
rough approximations, our approach provides values at 48%
and 41% from the estimation in 2015 and 2016, respectively,
but at only 26% and 12% in 2017 and 2018. Note that the B. Sensitivity analysis
linear scaling of the European values to just France, ignores
the country-wise dynamics that define this contribution to the Given the nature of our employed methodology to calculate
total consumption power consumption, it is possible that small changes in some
Another important source of comparison is a recent report model parameters may cause a large deviation in the final
from the French telecom regulator ARCEP on the data col- results. Hence, we present here a brief sensitivity analysis by
lected in 2020 [30], see Tab. IV. This data is related to both varying two parameters one by one. It is to be noted that
fixed and mobile networks of the operators (all equipment in except for the varied parameter, all other values are the same
the access, back-haul and core networks). If we apply a “rule as before. Fig. 8 depicts the total BS power consumption in
of thumb” of 60% for the mobile networks [31] and 80% for France at loads varying from 10% to 40% with a step of
the BSs [7], we obtain the third column of the table. Again, 5%. Similarly, Fig. 9 shows the total BS power consumption
we obtain the same order of magnitude without relying on curve at several different average BS lifetime (LT) duration.
proprietary data: differences of 29%, 6%, 6%, 16% and 17% These curves suggest that the assumption on the average BS
can be observed from 2016 to 2020. lifetime duration does not significantly impact the results. On
In [31], it is mentioned that the French mobile operator the contrary, the average load has a determining influence on
“Free” consumed 456 GWh of electricity in 2017, out of which the computed total power consumption. This means that more
60% was for its mobile network. Taking into account a 80% accurate load estimations, taking into account the space and
share only for BSs [7], leads to an installed power of 24 MW time dynamics, are required to continue the work presented in
in 2017. By filtering the BSs deployed by “Free” in the dataset, this paper.
 500 comparisons with recent publications show that we obtain the
All Generations - Load = 40 %
All Generations - Load = 35 % right order of magnitude. Certainly, our estimates are intended
All Generations - Load = 30 % to be preliminary in nature with the major goal of starting
400 All Generations - Load = 25 %
Total BS Power Consumption (MW)

All Generations - Load = 20 % a discourse around the topic. These estimates can be further
All Generations - Load = 15 % refined in the future to make them more accurate. For example,
All Generations - Load = 10 %
the space and time heterogeneity of the traffic could be taken
300
into account in place of an average load per BS, or a model
for sleep modes in 5G could be introduced. However, if our
200 estimates are confirmed, the alarming figures we show here
would call for proactive digital sobriety policies.

100 R EFERENCES
[1] “Ericsson Mobility Report,” Ericsson, Tech. Rep., 2021.
[2] A. Stephens, C. Tremlett-Williams, L. Fitzpatrick, L. Acerini, M. An-
02015 2016 2017 2018 2019 2020 2021 2022 2023 derson, and N. Crabbendam, “Carbon impact of video streaming,” 2021.
[3] C. Freitag, M. Berners-Lee, K. Widdicks, B. Knowles, G. S. Blair, and
Fig. 8: Total BS power consumption with varying the average A. Friday, “The real climate and transformative impact of ICT: A critique
of estimates, trends, and regulations,” Patterns, vol. 2, no. 9, 2021.
load (ρ) in France. [4] S. Lange, J. Pohl, and T. Santarius, “Digitalization and energy con-
sumption. Does ICT reduce energy demand?” Ecological Economics,
vol. 176, 2020.
All Generations - BS LT = 10 years [5] D. Bol, T. Pirson, and R. Dekimpe, “Moore’s Law and ICT Innovation
350
All Generations - BS LT = 9 years in the Anthropocene,” in 2021 Design, Automation & Test in Europe
All Generations - BS LT = 8 years Conference & Exhibition (DATE). IEEE, 2021, pp. 19–24.
All Generations - BS LT = 7 years
Total BS Power Consumption (MW)

All Generations - BS LT = 6 years [6] ITU-T, “Recommendation L.1470: Greenhouse gas emissions trajecto-
300
All Generations - BS LT = 5 years ries for the information and communication technology sector compat-
ible with the UNFCCC Paris Agreement,” International Telecommuni-
250 cation Union (ITU), Tech. Rep., Jan. 2020.
[7] Nokia, “White Paper: 5G network energy efficiency,” Tech. Rep., 2016.
200 [8] C. Gabriel, H. Andy, and A. Chern, “Green 5G: Building a Sustainable
World,” Huawei Technologies Co. Ltd., Tech. Rep., 2020, Available
Online at: https://www.huawei.com/en/public-policy/green-5g-building-
150 a-sustainable-world.
[9] J. Bieser, B. Salieri, R. Hischier, and L. Hilty, “Next generation mobile
100 networks: Problem or opportunity for climate protection?” 2020.
[10] V. C. Coroamă and F. Mattern, “Digital rebound–why digitalization will
50 not redeem us our environmental sins,” in Proceedings 6th international
conference on ICT for sustainability. Lappeenranta. http://ceur-ws. org,
vol. 2382, 2019.
02015 2016 2017 2018 2019 2020 2021 2022 2023 [11] J. Malmodin and D. Lundén, “The electricity consumption and oper-
ational carbon emissions of ICT network operators 2010-2015,” KTH
Fig. 9: Total BS power consumption with varying the average Centre for Sustainable Communications, Tech. Rep., 2018.
BS lifetime (LT) duration in France. [12] D. Lundén, J. Malmodin, P. Bergmark, and N. Lövehagen, “Electricity
consumption and operational carbon emissions of European telecom
network operators,” Sustainability, vol. 14, no. 5, p. 2637, 2022.
[13] H. Pihkola, M. Hongisto, O. Apilo, M. Lasanen, and S. Vatanen,
VI. C ONCLUSION “Energy consumption of mobile data transfer-Increasing or decreasing?
Evaluating the impact of technology development & user behavior.” in
In this paper, we propose an estimation of the operational ICT4S, May 2018.
[14] L. Golard, J. Louveaux, and D. Bol, “Evaluation and projection of 4G
power consumption of cellular BSs in France from 2015 and 5G RAN energy footprints: The case of Belgium for 2020–2025,”
to 2022. We base our methodology on a public dataset on Annals of Telecommunications, Nov. 2022.
radioelectric installations, widely adopted BS power consump- [15] “Données sur les installations radioélectriques de plus de
5 watts,” ANFR, [Dataset], 2022. Available Online at:
tion models and realistic assumptions from the literature. https://www.data.gouv.fr/en/datasets/donnees-sur-les-installations-
We highlight the fact that the number of BSs, the number radioelectriques-de-plus-de-5-watts-1/.
of transceivers, the average BS power consumption and the [16] “Bouygues Telecom and SFR conclude a strategic agreement to
national power consumption are growing at sustained CAGRs share a part of their mobile access networks,” Bouygues Telecom,
Tech. Rep., 2014, Available Online at: https://www.bouygues.com/wp-
of 7.55%, 18.27%, 9.89% and 18.18% respectively. So far, content/uploads/2014/01/14 0019 20140131 CP ByT SFR Mutualisa-
the introduction of 5G has not significantly modified the tion-vd%C3%A9f va.pdf.
rate of increase in the numbers of BSs and transceivers [17] A. Ahmed and M. Coupechoux. (2019)
ANFR Dataset Processing Codes - GitHub Repo. [On-
thanks to the practice of co-siting and the decommissioning line]. Available: https://github.com/arsalanqasimahmed/ANFR Data-
of 2G/3G transceivers, respectively. However, it has caused set Processing Codes.git
an acceleration in growth rates of the average and the national [18] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson,
M. A. Imran, D. Sabella, M. J. Gonzalez, O. Blume et al., “How
BS power consumption. Without access to proprietary data, it much energy is needed to run a wireless network?” IEEE wireless
is difficult to check the accuracy of our results. However, the communications, vol. 18, no. 5, pp. 40–49, 2011.
[19] ETSI, “TS 102 706-2 V1.5.1 - Environmental Engineering (EE); Metrics
and Measurement Method for Energy Efficiency of Wireless Access
Network Equipment; Part 2: Energy Efficiency - dynamic measurement
method,” European Telecommunications Standards Institute (ETSI),
Tech. Rep., 2018.
[20] S. Tombaz, P. Frenger, F. Athley, E. Semaan, C. Tidestav, and A. Fu-
ruskar, “Energy performance of 5G-NX wireless access utilizing massive
beamforming and an ultra-lean system design,” in 2015 IEEE Global
Communications Conference (GLOBECOM), 2015.
[21] D. López-Pérez, A. D. Domenico, N. Piovesan, H. Baohongqiang,
X. Geng, Q. Song, and M. Debbah, “A survey on 5g energy efficiency:
Massive mimo, lean carrier design, sleep modes, and machine
learning,” CoRR, vol. abs/2101.11246, 2021. [Online]. Available:
https://arxiv.org/abs/2101.11246
[22] C. L. I, S. Han, and S. Bian, “Energy-efficient 5G for a greener future,”
Nature Electronics, vol. 3, pp. 182–184, 2020.
[23] J. Lorincz, T. Matijevic, and G. Petrovic, “On interdependence
among transmit and consumed power of macro base
station technologies,” Computer Communications, vol. 50,
pp. 10–28, 2014, green Networking. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0140366414000632
[24] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof
power model for cellular base stations,” in 2015 IEEE 81st Vehicular
Technology Conference (VTC Spring), 2015.
[25] G. Micallef, P. Mogensen, and H.-O. Scheck, “Mobile operators have set
ambitious targets–is it possible to boost network capacity while reducing
its energy consumption?” EURASIP Journal on Wireless Communica-
tions and Networking, vol. 2012, no. 1, pp. 1–12, 2012.
[26] Nokia, “Sustainability and the life of a base station,” Nokia, Avail-
able Online at: https://www.nokia.com/networks/insights/5g-future/base-
station/: :text=A%20base%20station%20spends%20its,handle%20a%20
Petabyte%20of%20data.
[27] Y. Ding, H. Duan, M. Xie, R. Mao, J. Wang, and W. Zhang, “Carbon
emissions and mitigation potentials of 5G base station in China,”
Resources, Conservation and Recycling, vol. 182, p. 106339, 7 2022.
[28] “Bilan énergétique de la France en 2021,”
Available: https://www.statistiques.developpement-
durable.gouv.fr/sites/default/files/2022-04/datalab essentiel 275 bi-
lan energetique provisoire 2021 avril2022.pdf, Ministère de la
Transition Énergétique, Tech. Rep., Apr. 2022.
[29] ITU-T, “Mobile Cellular Subscriptions 2000 - 2020,”
https://www.itu.int/en/ITU-D/Statistics/Documents/statistics/2021/Dec-
ember/MobileCellularSubscriptions 2000-2020.xlsx, Tech. Rep., 2020,
accessed: August 2022.
[30] “Enquête annuelle “Pour un numérique soutenable”; Édition 2022 -
Collecte des données 2020.” Available: https://www.arcep.fr/cartes-
et-donnees/nos-publications-chiffrees/impact-environnemental/enquete-
annuelle-pour-un-numerique-soutenable-edition-2022.html, ARCEP,
Tech. Rep., Apr. 2022.
[31] Universfreebox. (2018) Découvrez combien consomment en
électricité les réseaux fixe et mobile de Free. [Online].
Available: https://www.universfreebox.com/article/44272/Decouvrez-
combien-consomment-en-electricite-les-reseaux-fixe-et-mobile-de-Free
[32] L. Dan Jones. (2019) Power Consumption: 5G Bases-
tations Are Hungry, Hungry Hippos. [Online]. Avail-
able: https://www.lightreading.com/mobile/5g/power-consumption-5g-
basestations-are-hungry-hungry-hippos/d/d-id/749979
[33] ITU-T, “Recommendation L.1210: Environment and ICTS, Climate
Change, e-Waste, Energy Efficiency; Construction, Installation and Pro-
tection of Cables and Other Elements of Outside Plant,” International
Telecommunication Union (ITU), Tech. Rep., Dec. 2019.