AI-Ready Energy Modelling for 5G RAN:

Quick, Lightweight, Open-Source Simulation

Kishan Sthankiya∗ , Keith Briggs† , Mona Jaber∗ and Richard G. Clegg∗
∗ School of Electronic Engineering and Computer Science
Queen Mary University of London, London, UK
Email: {k.sthankiya, m.jaber, r.clegg}@qmul.ac.uk
† Wireless Research, BT Labs, Martlesham Heath, UK

Email: [email protected]

Abstract—Recent sustainability drives place energy- tier deployments with tens of User Equipment (UE) devices
consumption metrics in centre-stage for the design of future attached to each. For the scenarios presented in this paper, a
radio access networks (RAN). At the same time, optimising single run generally takes less than 3 seconds on consumer
the trade-off between performance and system energy usage
by machine-learning (ML) is a promising approach but may hardware.
requires a large amount of granular RAN data to train Emerging Open RAN [4] architecture positions the RAN
models, and to adapt in near realtime. In this paper, we Intelligence Controller (RIC) as an interface to optimise the
present extensions to the 5G system-level discrete-event AIMM network with AI and ML models. By design, the AIMM
Simulator, generating realistic figures for throughput and energy Simulator offers users an interface to write code in line with
efficiency (EE). We further investigate the trade-off between
maximising either EE or spectrum efficiency (SE). To this the RIC functionality. In this work, we created a function
end, we have run extensive simulations of a typical macrocell to calculate and monitor the power and energy consumption
network deployment under various transmit power-reduction of base stations in an idealised urban 5G deployment and
scenarios. Our results demonstrate that the EE and SE objectives study scenarios of reduced transmit power. Three-dimensional
often require different power settings. Importantly, we also positions (i.e. x, y, and z spatial coordinates) of UEs in the
show that the AIMM Simulator is a powerful tool for quick
prototyping of scalable system models which can interface simulation area and their CQI and resulting throughput are
with ML frameworks, and thus support future research in calculated, allowing a look at realistic estimates for energy
energy-efficient 5G networks. efficiency (EE) and spectral efficiency (SE).
Index Terms—5G, RAN, modelling, power consumption, en- Advancing technology crucially requires dynamic optimi-
ergy efficiency, Discrete event simulation sation for resource allocation in rapidly fluctuating network
and traffic conditions, where traditional static EE and SE
I. I NTRODUCTION points quickly become obsolete. Our work accentuates this
The significance of 5G telecommunications cannot be over- by demonstrating the trade-off between SE and EE and the
stated, with global subscribers approaching 6 billion. Com- possibility of major savings by adapting models to the current
pared to earlier systems, 5G radio access networks (RAN) situation in the network. Minimising the time required to
have more flexibility to become energy efficient using various obtain these results and the potential benefits of increasing EE
techniques [1]. Meanwhile, energy usage remains consider- or SE make dynamic optimisation targets crucial for efficient
able, with up to 80% of this concentrated at the base stations and adaptable network resource utilisation.
(BS) [2]. However, exactly how to continuously optimise The contribution of this paper is presenting the AIMM
networks to achieve a trade-off between energy usage and Simulator and demonstrating its use for measuring energy
performance is an open question. Artificial Intelligence (AI) consumption in different scenarios. The simulator is extremely
and, more specifically, machine learning (ML) techniques are promising for AI and ML use cases as it can quickly produce
promising for this optimisation, but their training cannot take a detailed estimate of BS energy consumption in scenarios
place on a real network. with tens of BSs and several hundred users. This will enable
This paper implements and assesses an extension to estimate the deployment of both traditional optimisation techniques
energy use using a system-level 5G simulator, AIMM Simu- and AI training for optimisation techniques that look at, for
lator [3]. We add major new features for tracking energy use example, varying power levels and sleep modes. We show that
in all network components. With our extensions, the AIMM considerable savings can be made in this way, by adjusting
Simulator meets two key needs for using AI to optimise only three of the nineteen BS in our scenarios we can increase
energy usage: rapid scenario analysis for large numbers of the network energy efficiency by up to 14.8%.
training rounds to run and faithful, real-world, network key
performance indicator (KPI) tracking; e.g. Channel Quality II. BACKGROUND
Index (CQI) and Modulation and Coding Scheme (MCS). The An extensive survey [1] for 5G EE, estimates over 50%
simulator is designed to simulate tens of BSs, including multi- of all network energy usage will be the radio access element
 TABLE I
C OMPARISON OF 5G S IMULATION T OOLS

ns3-o-ran-
Feature AIMM Py5cheSim 5G-LENA Vienna
e2
3D modelling
Free-to-use
Open-source
Low dependencies

by 2025. This projection accounts for advanced sleep modes, sensitivity to these details. In contrast, [10] proposes a model
lean carriers, massive MIMO layer adaption and ML in 5G NR based on data flow for the electric power consumption of the
(see [1, Section 2] for a complete list). At the same time, the active antenna unit in a 5G BS using statistical traffic data and
report [5, Section 5.6], envisages AI as a technology that can data-rate estimations.
anticipate network traffic dynamics and continuously optimise
In lieu of the first requirement for low latency access to
operation. The authors in [6] conclude ML techniques may be
network performance data, described earlier, wireless net-
more promising. However, ML models share issues with all AI
work simulators provide valuable insights. More specifically,
models, in that they cannot be trained or tested in operational
system-level simulators with discrete-event frameworks (see
networks without severe risk of catastrophic outage. Mitigating
Table I) are essential for the dynamic modelling of nodes
with virtual representations [7] is a potential solution, but
– reflecting their changes in state and facilitating runtimes
also faces two key challenges when trying to apply ML to
which are faster than in real-time. However, there is diversity
optimise energy efficiency of the system. Firstly, they require
in the tools available to model 5G systems in the literature. For
low latency access to large amounts of reliable performance
example, Py5cheSim [11] falls short in capturing features such
and energy consumption data, captured with high precision.
as the simulation nodes having x, y, and z spatial coordinates
Secondly, the computational results from the virtual replica
(3D modelling). Use of these co-ordinates are crucial for pre-
must return fast enough to still apply to the physical network.
dicting service interruptions and user experience degradation
For a given network, the energy data captured can be de-
for key enabling technologies in 5G, such as massive MIMO
scribed by an energy consumption model. A common approach
and beamforming operations. Meanwhile, efforts to support
is the analytical model in [2], estimating that power used by a
open innovation can be fostered through transparent practices,
BS, PBS = NTRX (P0 + δp Pout ), where NTRX is the number of
such as open-source and free-to-use software. Most simulators
transceivers, P0 and δp are power consumption parameters de-
share this trait, yet the Vienna [12] project runs in MATLAB
pendent on cell-type and Pout is the transmit power. Assuming
which is proprietary and incurs licensing fees. Tools with
a fixed number of NTRX , we see that power consumption for
many dependencies necessitate a larger codebase that not only
the whole BS is directly proportional to the transmit power.
increase complexity for debugging, but crucially, introduce
Later studies model energy use in systems of high and low
longer runtimes for a virtual replica. Amongst the ns-3 based,
transmit power nodes across a variety of cloud-RAN (C-RAN)
5G-LENA [13] and ns3-o-ran-e2 [14], the former requires four
deployment scenarios [8] and BS sleep modes [9]. Each of
supplementary libraries and the latter requires two. Given that
these works attempts to model energy or power consumption
virtual replicas require the ability to evaluate many scenarios
with a specific goal, such as sleep modes [6], [9], traffic
with incredible speed, applying an upper limit of three external
behaviour [10], or network deployment [8]. However, a lack
libraries, the AIMM Simulator [3] becomes a clear contender,
of realistic parameters is demonstrated by the level of transmit
only requiring Python 3.8+ and two additional libraries with
power required to satisfy user demands. This taken from a path
the entire codebase standing at ⩽ 10MB.
loss model following Shannon’s theorem [9, Eq(2)] and [8,
Eq(6)], which does not account for the required signal strength It is evident that there is a lack of realistic 5G network
at the UE to ensure less than 10% blocker error rate (BER). simulation tools that dynamically model energy efficiency
Moreover, it is not a focus of these studies to model EE in towards facilitating fast optimisation in data-driven virtual
a network simulator to facilitate fast optimisation in a virtual replicas. To address the challenges of providing accurate
replica of an operational network. energy data that mirror real-world conditions at low latencies
In isolation, EC alone does not account for how effectively and support open research efforts, the AIMM Simulator is
the energy is being used. This is better described by energy positioned as prominent solution. These findings indicate a
efficiency (EE), often expressed as the amount of data trans- timely opportunity to expand the capabilities of the AIMM
mitted (in bits) per unit of energy consumed (in Joules). Indeed Simulator, by building an extension for energy modelling. This
EE is a complex behaviour that is affected by the network extension bridges the gaps in the leading 5G network simu-
structure and features (see [1]), the stochastic nature of users’ lation and energy modelling research to enable the potential
spatial and temporal profiles, and environmental conditions. of ML techniques to optimise energy efficiency in operational
In this case, an analytical model would not offer the required networks. It is described in section III-B.
 III. M ETHODS • PRx : Received power of the signal from serving BS,
A. AIMM Simulator • Iinter,i : Sum of interference powers from all BSs in the
set J, excluding the jth BS,
• Inoise : Noise power at the UE.

Unlike traditional approaches that rely on Shannon’s theory

to estimate the user throughput, AIMM uses pre-computed
tables to look up CQI values based on SINR values at
10% Block Error-Rate thresholds. These are derived from
experimental results based on clustered delay line type-C with
practical channel estimation parameters. Linear scaling of CQI
to MCS index allows lookup of SE from [16, Table 5.1.3.1-1].
It follows that the throughput of the i-th UE attached to the
j-th BS is:

Fig. 1. AIMM Simulator Block Diagram [3] Ti,j = SEi,j × B, (3)

The AIMM Simulator is a fast system-level simulator de- where SEj,i represents the SE, determined by the MCS index
veloped to recreate 5G New Radio (NR) elements as seen in (obtained from the SINR in (2)) and B is the bandwidth.
Fig. 1. Written in Python, the core simulator (Sim) employs B. Extension Implementing Power Consumption Model for
a discrete-event library for fine control over component in- Base Stations
teraction. Physical and network layer concepts (e.g. resource
The power consumption of BSs is estimated based on a
elements, IP packets) are abstracted to reduce runtime. Finally,
model inspired by [2]. The total power consumption of BSj
numerical Python enhances operation speed comparable to
is denoted as PBSj ,
compiled C code.
The flexibility of simulations is first demonstrated within PBSj = NTRX × Nant × (P0 + f (Pj )), (4)
a Sim instance, where a central loop interval orchestrates
downstream events for attached components. BSs (Cell) and and is the product of the number of transceiver chains (NTRX ),
user equipment (UE) are created in a 3D space with the number of antennas (Nant ) and the static power consumption
option to override default parameters (see Table II). A Mobility that is not related to the actual load of the BS (P0 ) plus a
Management Entity (MME) maintains UE attachment and function of the actual transmit power (Pj ) of BSj , where
handover, where strategies and link-level threshold logic can Pj
be defined. More sophisticated control is reserved for the ηPA ·(1−σfeed ) PRF + PBB
f (Pj ) = . (5)
RIC, which provides a ‘clean’ interface to allow integration (1 − σDC ) (1 − σMS ) (1 − σcool )
with ML libraries (e.g. TensorFlow, PyTorch) and development Refer to Table II for the definitions and values of the variables
of rApps and xApps. Dynamic features (e.g. cell transmit in this model.
power) may be customised specifically to an experiment via To add energy modelling capabilities, an extension module
the Scenario class, and polling the state of the system is was created for the AIMM Simulator which can be found on
handled by the Logger. These features highlight the tool’s GitHub [17]. This module includes a system power consump-
versatility and foster a firm grounding for evaluating EE and tion model [2], typical parameters different types of BSs [18],
SE in 5G NR systems. 5G CQI reference tables [16], cell sleep mode, per-user SINR
The AIMM Simulator features standardised pathloss mod- based handover and logging for later analysis. Furthermore,
els [15], namely: Urban Macro (3D-UMa), Urban Micro (3D- we define experimental scenarios to allow tracking of UE
UMi), and Indoor Hotspot (3D-InH) — including both line- movements and changes to transmit power of BSs.
of-sight (LoS) and non-line-of-sight (NLoS) variants. These
models are used to calculate Reference Signal Received Power C. System Model
(RSRP) and Signal-to-Interference-plus-Noise Ratio (SINR): We consider the downlink homogeneous system to explore
RSRPi,j = GMIMO,j + Gant,j + PTx,j − PLj,i (1) the energy efficiency of a 5G macro BS RAN in a 3D plane.
PRx,i,j As seen in Fig. 2, we deploy macro BSs, (j = 1, 2, ..., J), in a
SINRi,j = (2) regular hexagonal grid, with centre frequency ν and bandwidth
Iinter,i + Inoise,i
B. BSs are evenly spaced with an inter-site distance Disd and
where indices i and j refer to the UE and the BS, respectively. height (HBS ). Each BS serves the geographical area of a BS
The following parameters are used: numbered in Fig. 2, composed of transceivers (NTRX ) each
• GMIMO : MIMO (Multiple-Input Multiple-Output) gain, served by a number of antennas (Nant ). At each antenna, we
• Gant : Antenna gain, assume an omnidirectional radiation pattern and a transmit
• PTx : Transmit power, power in Watts (Pj ), where Pj ⩽ Pmax , the maximum transmit
• PL: Pathloss between BS and UE, power. We group BSs by position, described in Table III.
 TABLE III
BS S CENARIO D EFINITIONS ( REFER TO F IG . 2)
16 17 18 Centre Cell
Inner Ring Cells Scenario Description Variable Power BSs (Kv )
12 13 14 15
Outer Ring Cells Scenario 1 Centre 9
7 8 9 10 11 UEs Scenario 2 Inner ring, antipodal 8, 10
Base Stations Scenario 3 Inner ring, alternate 4, 10, 13
3 4 5 6 Simulation Area Scenario 4 Central triad 4, 8, 9
0 1 2

Fig. 2. Topology of our system with nineteen BSs in a regular grid and UEs In each of the four scenarios in Table III, we study the
deployed over this grid. effects of reducing the transmit power (Pj ), from Pmax to
0 W (Psleep ) in steps of 0.5 mW on multiple metrics. We
UEs (i = 1, 2, . . . , NUE ) are distributed with a homogeneous first capture the throughput as defined in (3) and power
Poisson point process with density λUE UEs per km2 . UE consumption as defined in (4) for each of the J BSs. The
height is defined as HUE and is the same for all. UEs mean throughput of BSj is:
are assumed to be outdoor and stationary, with a 3D-UMa PNUE
Ti,j × Xi,j
NLoS pathloss model with serving and neighbouring BSs. T̄j = i=1PNUE , (7)
| i=1 Xi,j |
The operating mode is frequency-division duplexing (FDD),
and UEs attach each to a maximum of one BS and select where Xi,j indicates if a given UEi is connected to BSj and
the BS with the highest SINR based on (2). The user-to-BS Ti,j is defined in (3). Based on the BS-based metrics defined
connectivity is represented by defining in Xi,j for each UEi in (3), (4), and (7), we formulate four set-based metrics, where
and each BSj : a set S could be Kv , Ks , or Kv ∪ Ks .
( • Mean throughput (Mb/s) per set (Tk ) as in (8),
1 if UEi is connected to BSj • Mean power consumption (kW) per set (PCk ) as in (9),
Xi,j = (6)
0 otherwise • Mean spectrum efficiency (Mb/s/Hz): SEk = Tk /B where

This power consumption model provides an estimation of the B is the bandwidth,

• Mean energy efficiency (Mb/J): Tk × τ /PCk , where τ is
energy consumed by the BSs, taking into account various
components and efficiency factors. the duration of one run in seconds where,
P
T̄k
TABLE II Tk = ∀ BSj ∈ S, (8)
|S|
D EFINITION OF VARIABLES IN THE POWER CONSUMPTION MODEL P
PBSj
Variable Description Value PCk = ∀ BSj ∈ S. (9)
ν Carrier frequency 3.5 GHz
|S|
B BS bandwidth 10 MHz Each of the four scenarios in Table III is repeated for 100
J No. of sites 19
HBS BS height 25 m
different seed values and the metrics are averaged over all
Pmax Max transmit power 20 W runs per scenario. In each scenario, the performance metrics
Disd BS Inter-site distance 500 m for three sets are calculated Kv , Ks , and Kv ∪ K s .
λUE UE density 400
HUE UE height 1.5 m IV. R ESULTS AND D ISCUSSION
NTRX No. of transceivers in the base station 6
Nant No. antennas 1 In this section, we present the results of implementing the
Pj Output power 0–20 W scenarios in Table III, as shown in Fig. 3. For each scenario,
P0 Static power 130 W
ηPA Power amplifier efficiency 0.311 we show the mean network throughput (as in (8), plot (a) and
PRF RF processing power 12.9 W mean network power consumption (as in (9), plot (b). Next,
PBB Baseband processing power 29.6 W the mean values of EE and SE in the whole network for each
σfeed Feeder loss 0.5
σDC DC-to-AC losses 0.075 scenario are presented in (c) and (d). In the following sections,
σMS Mains supply losses 0.09 we discuss the impact of power setting in each scenario based
σcool Cooling losses 0.10 on these four metrics.
A. Impact on throughput and power consumption
D. Experiment Setup The network throughput in Fig. 3(a) is affected by power Pj ,
We begin by varying transmit power (Pj ) of different with lower output power resulting in lower throughput. How-
BSs within the network in Fig. 2, implementing scenarios as ever, the peak throughput for each scenario occurs at different
defined in Table III. In each scenario, the BSs in Fig. 2 are power settings. For example, reducing the transmit power of
separated into two sets: Kv which includes the selected BSs the Inner Ring, alternate has a greater impact on throughput
for varying power level and Ks which includes the other BSs, than Central Triad, despite similar power consumption profiles
such as | Kv ∪ Ks |= J. and the same number of cells. It follows that optimising the
 19.8

Mean Network Energy Efficiency (Mbps/J)

11.0
19.6
Throughput (Mbps)

10.5
19.4
10.0

19.2
9.5

19.0
9.0

-inf 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 -inf 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43
(a) (c)

Mean Network Spectral Efficiency (bps/Hz)

2.05
2.10
Power Consumption (kW)

2.00
2.05

2.00 1.95

1.95
1.90
1.90

1.85 Centre Inner Ring, antipodal 1.85

Central triad Inner Ring, alternate
1.80
1.80
-inf 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 -inf 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43
(b) (d)

Transmit power of all cells in Kv (dBm)

Fig. 3. Comparison of Mean Network Throughput (a), Mean Network Power Consumption (b), Mean Network Energy Efficiency (c) and Mean Network
Spectral Efficiency (d). Each subplot x-axis represents the Pj ∀ Kv , ranging from sleep mode (−∞ dBm) to 43 dBm. All plots illustrate the effect of
reducing transmit power of BSs in (Kv ), and the impact that has on the network mean (i.e., across all cells).

choice of cells to reduce power and the setting of Pj is crucial However, when all active BSs operate at maximum capacity,
for achieving the highest network throughput with the lowest less power is available for data transmission per unit of time,
energy consumption. resulting in decreased SE. For instance, the Central Triad
In plot (a), it is evident that reducing Pj by 6 dBm leads to scenario in our analysis experienced a 3.9% reduction in SE
an increase in throughput, except in the Centre scenario. This performance. On the other hand, the Centre scenario had only
increase in throughput is explained by (2), where a reduction in a minimal impact on SE, causing a 1.5% reduction. Moreover,
unwanted signals from nearby BSs leads to an increase in both the Inner Ring, alternate scenario had better SE performance
SINR and CQI. Therefore, this change in output power proves in the range of Pmax to 22 dBm. In contrast, the Inner Ring,
to be an effective solution for mitigating interference and antipodal performed better at lower Pj levels. Therefore,
improving the system’s overall performance in most scenarios. reducing BS Pj can have varying impacts on SE, depending
on the specific scenario. The error bars in Fig.3(c) and (d)
B. Energy and Spectrum Efficiency represent one standard deviation from the mean, primarily
This section evaluates the network’s EE across different due to the static placement algorithm and the influence of UE
scenarios, as shown in Fig. 3(c). The EE enhancements range placement on both EE and SE.
from 34-43 dBm, with the most significant gains observed in
the Inner Ring, alternate setup, resulting in a 0.768 Mbps/J
C. Lessons learnt for future network design
boost. However, the EE decreases slightly when the base
station (BS) power drops below these levels in all scenarios. The findings of this study underscore the intricate nature
Notably, the EE significantly improves when the BSs in Kv of jointly optimising energy and spectrum efficiency, necessi-
are in sleep mode (-inf). The Inner Ring, alternate setting tating meticulous selection of BSs and precise configuration
yields the most substantial EE improvements of 14.8%, while of power levels. For instance, in the context of pursuing pure
the Centre scenario shows the lowest improvement at 3.79%. EE objectives, the Inner Ring, alternate scenario emerges as
These highlight the importance of optimal strategy selection the optimal choice when three BSs in Kv are in sleep mode.
for reducing network power and its implementation timing. Conversely, when the objective shifts to pure SE, the same
Reducing the transmit power of a BS (Pj ) can decrease scenario achieves peak performance with an output power
the network’s overall SE. In Fig.3(d), we observed a slight of Pj = 37 dBm. These instances exemplify the nuanced
increase in SE when Pj was reduced from Pmax to 37 dBm. trade-offs between EE and SE goals, necessitating a judicious
balance in optimisation based on network metrics and user [4] M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia,
requirements. “Understanding O-RAN: Architecture, interfaces, algorithms, security,
and research challenges,” IEEE Commun. Surveys Tuts., vol. 25, no. 2,
The research further highlights the AIMM Simulator’s po- pp. 1376–1411, 2023. [Online]. Available: https://ieeexplore.ieee.org/
tential to model interactions between EE and SE and its document/10024837/
capacity for realistic throughput, in contrast to Shannon-based [5] ITU, “Future technology trends of terrestrial IMT systems towards 2030
and beyond,” International Telecommunications Union (BR), Tech. Rep.
estimations. The operational mode of FDD supports energy- M.2516-0, 2022.
efficient carrier shutdown methods, with strategies such as [6] N. Piovesan, D. Lopez-Perez, A. De Domenico, X. Geng, H. Bao,
prolonged periods of deep sleep lasting minutes or hours. and M. Debbah, “Machine learning and analytical power consumption
models for 5G base stations,” IEEE Commun. Mag., vol. 60, no. 10,
Strategies like cell zooming enable a macro cell’s coverage pp. 56–62, Oct. 2022. [Online]. Available: https://ieeexplore.ieee.org/
expansion while allowing others to remain in a deeper sleep for document/9928089/
extended periods [1], ensuring efficient energy usage without [7] G. Nardini and G. Stea, “Using network simulators as digital twins of
5G/B5G mobile networks,” in 2022 IEEE 23rd Int. Symp. on a World
compromising network performance. of Wireless, Mobile and Multimedia Networks (WoWMoM), Jun. 2022,
Additionally, the study underscores the practical appli- pp. 584–589.
cability and potential advantages of integrating the AIMM [8] A. Israr, Q. Yang, and A. Israr, “Power consumption analysis of access
network in 5G mobile communication infrastructures – An analytical
Simulator into AI training and real-world deployments. Anal- quantification model,” Pervasive Mob. Comput., vol. 80, 2022.
ysis of execution times across various scenarios reveals a [9] S. K. G. Peesapati, M. Olsson, M. Masoudi, S. Andersson, and C. Cav-
remarkable mean execution time of 1.625 seconds ± 0.974 (2 dar, “An analytical energy performance evaluation methodology for 5G
base stations,” in 17th Int. Conf. on Wireless and Mobile Computing,
s.d.), achieved without specialised hardware. This outstanding Networking and Communications (WiMob), Oct. 2021, pp. 169–174.
speed positions the AIMM Simulator as a promising tool for [10] Y. Zou et al., “Electric load profile of 5G base station in distribution
evaluating AI applications in the context of a 5G digital twin systems based on data flow analysis,” IEEE Transactions on Smart
Grid, vol. 13, no. 3, pp. 2452–2466, May 2022. [Online]. Available:
and within operational live networks. https://ieeexplore.ieee.org/document/9708427/
[11] G. Pereyra, C. Rattaro, and P. Belzarena, “Py5cheSim: a 5G Multi-Slice
V. C ONCLUSIONS Cell Capacity Simulator,” in 2021 XLVII Latin American Computing
Conference (CLEI), Oct. 2021, pp. 1–8.
In this work, we present a study on 5G RAN efficiency [12] M. K. Müller et al., “Flexible multi-node simulation of cellular mobile
under different scenarios of BS power settings. To this end, communications: the Vienna 5G system level simulator,” EURASIP
we first describe the new energy consumption model and its Journal on Wireless Communications and Networking, vol. 2018,
no. 1, Dec. 2018. [Online]. Available: https://jwcn-eurasipjournals.
use in the AIMM Simulator. We define a system of 19 BSs and springeropen.com/articles/10.1186/s13638-018-1238-7
formulate four scenarios for manipulating the transmit power [13] M. Mezzavilla et al., “End-to-end simulation of 5G mmwave networks,”
levels. Our results shed light on the interdependence between IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 2237–2263, 2018.
[Online]. Available: https://ieeexplore.ieee.org/document/8344116/
energy and spectral efficiency in a realistic representation of [14] A. Lacava et al., “Programmable and customized intelligence for
throughput estimation. Our findings emphasise the practical traffic steering in 5G networks using Open RAN architectures,”
application and benefits of utilising the AIMM Simulator IEEE Trans. Mobile Comput., pp. 1–16, 2023. [Online]. Available:
https://ieeexplore.ieee.org/document/10102369/
for fast and lightweight simulations. Furthermore, given the [15] 3GPP, “Study on 3D channel model for LTE (Release 12),” 3rd
performance, we believe the AIMM Simulator is an effective Generation Partnership Project (3GPP), Technical Report (TR) 36.873,
AI-ready tool for practical 5G system modelling, yielding fast Jan. 2018, version 12.7.0.
[16] ——, “NR; Physical layer procedures for data (Release 17),” 3rd
results as an enabler for future mobile networks. Generation Partnership Project (3GPP), Technical Specification (TS)
38.214, Dec. 2022, version 17.0.0.
ACKNOWLEDGMENT [17] Kishan Sthankiya, “AIMM Simulator Energy Modelling Exten-
Keith Briggs developed the original AIMM Simulator code sion,” Oct. 2023. [Online]. Available: https://github.com/apw804/
EnergyModels
as part of the AIMM project1 . The full source code is avail- [18] H. Holtkamp, G. Auer, V. Giannini, and H. Haas, “A parameterized
able at https://github.com/keithbriggs/AIMM-simulator. Kis- base station power model,” IEEE Communications Letters, vol. 17,
han Sthankiya developed extensions for energy modelling. no. 11, pp. 2033–2035, Nov. 2013. [Online]. Available: http:
//ieeexplore.ieee.org/document/6600717/
R EFERENCES
[1] D. Lopez-Perez et al., “A survey on 5G radio access network energy
efficiency: Massive MIMO, lean carrier design, sleep modes, and
machine learning,” IEEE Commun. Surveys Tuts., vol. 24, no. 1,
pp. 653–697, 2022. [Online]. Available: https://ieeexplore.ieee.org/
document/9678321/
[2] G. Auer et al., “How much energy is needed to run a wireless network?”
IEEE Wireless Commun. Mag., vol. 18, no. 5, pp. 40–49, Oct. 2011.
[Online]. Available: http://ieeexplore.ieee.org/document/6056691/
[3] AIMM Project, “AIMM 5G System Simulator,” Dec. 2022. [Online].
Available: https://pypi.org/project/AIMM-simulator/

1 https://www.celticnext.eu/project-aimm/