Received April 19, 2022, accepted May 6, 2022, date of publication May 11, 2022, date of current version

May 18, 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3174089

Energy Saving Technologies and Best Practices

for 5G Radio Access Network
RUMENG TAN 1, YING SHI 1, YINGYING FAN 2, WENTAO ZHU 3, AND TONG WU 4
1 China Telecom Research Institute, Guangzhou 510630, China
2 Shanghai Research Institute, ZTE Corporation, Shanghai 201203, China
3 China Mobile Group, Design Institute Company Ltd., Beijing 100080, China
4 National Institute of Metrology, Beijing 100029, China

Corresponding author: Tong Wu ([email protected])

ABSTRACT This article identifies energy-saving potential of the fifth generation (5G) Radio Access Net-
work, and describes main energy-saving principles and technologies. It explores how to use network energy
saving technologies, such as carrier shutdown, channel shutdown, and symbol shutdown in 5G network, that
have been inherited from 4G. Some enhanced technologies for 5G like equipment deep sleep and symbol
aggregation have also been introduced in this article. However, it is far from enough and an innovative energy-
saving solution should be considered. To meet the requirements and development of intelligent and self-
adaptive energy-saving solution, Artificial Intelligence (AI) and big data analysis are introduced to form a
more precise energy-saving strategy based on site-specific traffic and site-related conditions, thus improving
the efficiency and reducing the manpower. Finally, two commercial application practices of AI-based energy-
saving solution are elaborated. One is the practice of AI-based service awareness energy saving for 4G/5G
collaborative networks, the energy benefits can be improved up to 20%; The other practice is the adoption
of a new architecture Active Antenna Unit (AAU) with beam pattern optimization, its energy benefits can
be promoted by 30%. These two practices could help mobile network operators (MNOs) to achieve the most
energy-efficient network with good network performance and lower Operating Expense (OPEX).

INDEX TERMS 5G, artificial intelligence, base station, energy efficiency, energy saving, radio access
network.

I. INTRODUCTION some forecast, and 5G network would be a big part of the

First of all, climate change is arguably the biggest challenge increase [4]. And the biggest challenge for mobile network
of our time. Even a Nobel Prize was awarded in 2021 to energy saving is about the base stations (BS), which account
researchers in this field. And many of MNOs have set ambi- for about 73% of total energy usage of a typical mobile
tious targets about carbon neutral and/or net-zero reported network according to GSMA [5]. So the critical thing to do
by GSMA [1], for instance, Orange has set 2040 for net- is to reduce carbon footprint in every possible way, but not at
zero and 2050 for carbon neutral based on 1.5-degree centi- the cost of user experience promised.
grade climate changing. Meanwhile, in the guide for the The good news is we don’t have to re-invent all the wheels
cellular mobile network operators (MNOs) by GSMA about to 5G energy saving. Network energy saving technologies,
climate target, the goal is to reduce carbon emissions by 45% such as carrier shutdown, channel shutdown, symbol shut-
from 2020 to 2030 [2], [3]. down etc., that have been emerged since the 4G era, can be
In order to achieve these targets, we need new ways leveraged to mitigate 5G energy consumption. Furthermore,
of thinking how we do green network, because the pres- enhanced technologies like deep sleep, symbol aggregation
sure would be huge. The total energy consumption by the shutdown etc., have been developing in the 5G era. In this
mobile industry would be tripled in ten years, according to article it aims to detail these fundamentals.
However, it is far away from being enough, an innovative
The associate editor coordinating the review of this manuscript and energy saving solution should be taken into consideration.
approving it for publication was Francisco Rafael Marques Lima . Powered by AI and big data, a more precise energy saving
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 10, 2022 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 51747
 R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

strategy based on site-specific traffic and site-related condi- scenario, shutdown duration, and shutdown depth leads to an
tions, improves the efficiency and reduces the manpower. improvement in an overall energy-saving capability.
It is however not just about the network itself. Because the
other industries can also benefit from a greener network either
with direct reduction of carbon emission, or using 5G as an
underlying technology to make their own businesses more
efficient, more productive and therefore greener.

II. ENERGY SAVING DESIGN PRINCIPLES FOR BS

A. DYNAMIC SCALING
The dynamic capability is the capability of scaling power
consumption with real-time changes of service loads. The
dynamic capability is decomposed into the capability of
FIGURE 1. Three dimensions of shutdown capabilities.
scaling resource consumption with changes of service loads,
and the ability of equipment power consumption to scale as
resource consumption changes. Extended shutdown scenarios can be that parameters are
adjusted such that the shutdown can be applied more often.
An improvement of the shutdown duration can be achieved
1) RESOURCE ON-DEMAND SCALING
by reducing the common information overhead in no-load or
Based on the network performance and user experience, light-load scenarios, or to perform aggregation scheduling for
time-, space-, and frequency-based resource occupation and small-packet services. Finally, the shutdown depth is related
air-interface power resource allocation are adjusted. During to how many hardware components that can be de-activated.
low- and medium-load period, reducing resource overhead Furthermore, the three dimensions are not independent,
through proper scheduling is a key path for energy-saving. but are mutually coupled and mutually constrained. When
The evolution of wireless networks is a process of expand- one dimension is strongly expanded, the other two dimen-
ing network capacity and resources to cope with increasing sions may be affected. For example, increasing the shutdown
service requirements. The resources are increased to meet depth increases the energy saving gain at the time when the
the peak capacity of the network. However, traffic is tidal, shutdown takes effect, but increasing the shutdown depth
and the proportion of peak scenarios in a day is not so high. may increase the wakeup duration [6]. The increase of the
It is therefore necessary to adjust the resources to the required wakeup duration affects the application scenarios and the
demand. possibility of entering the sleep state. The overall gain of
the whole day may not be the optimal state. Therefore, these
2) POWER CONSUMPTION SCALES WITH RESOURCES three dimensions need to be considered comprehensively in
When resource consumption decreases, equipment power the design of the solution, and strive to find a good balance
consumption should also decrease. The resource herein may between them.
be a bandwidth, a quantity of carriers, a quantity of channels,
a transmit power, or the like. For example, the transmission C. TRADE-OFFS
power consumption varies with the number of RF (radio Energy-saving involves various trade-offs, as was seen in the
frequency) channels and PA (power amplifier) output power, section on shutdown capabilities. One trade-off is between
while the compute power consumption typically varies with network performance and energy saving. Based on the
the bandwidth and the quantity of channels [6]. impact on performance, many energy-saving are lossy as
they scale resources and the number of resources is closely
B. SHUT-DOWN CAPABILITIES related to performance. Therefore, it is difficult to achieve
In order to get power consumption to scale with resources, the lossless performance. On the other hand, as mentioned in
static power consumption may become an obstacle. In other the preceding analysis, there is a large margin for network
words, when scaling down resources the power consumption experience in off-peak hours, such as night. If the margin
reaches a level (the static power consumption) where it does is properly squeezed out, there is energy-saving to harvest.
not scale any more. In such cases, a shutdown capability of the However, how to squeeze this space will involve another
hardware component(s) is desirable, i.e., a deactivation of the issue, the qualitative and quantitative analysis of the impact
component(s) which decrease the static power consumption of various energy-saving measures on performance. In many
stepwise. A drawback with this is that the component is taken scenarios, this impact cannot be accurately modeled in a
out of service, and it will require a delay to be re-activated. white box manner. Although certain parameter thresholds are
The energy saving gain from shutdown capabilities designed for most features to control the impact on perfor-
depends on three dimensions as illustrated in Fig.1; S: Sce- mance, it is difficult to provide accurate guidance on how to
nario (shutdown scenario) T: time (shutdown duration) adjust parameters. Artificial intelligence (AI) might be a tool
D: Depth (shutdown depth). An improvement in a shutdown to handle this, as will be discussed in section VII.

51748 VOLUME 10, 2022

R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

Another trade-off is between network energy saving and

User Equipment (UE) energy saving. In most cases, net-
work energy saving is decoupled from UE energy saving.
However, trade-off exists in some scenarios. For example,
to save power for UEs, the network performs a lot of adapta- FIGURE 2. Illustration of time-domain energy saving. During symbols
tion in the design of common signaling, and these adjustments with no information to transmit, the PA is de-activated.

usually increase the overheads on the network side. However,

to allow for more time-domain energy saving. The most
an Internet of Things terminal has a strict requirement on low
necessary control signals for access are still broadcasted but
power consumption, which exacerbates this contradiction.
many other signals are only transmitted on-demand from the
The terminal may selectively listen to a paging of a network
accessing user devices. Furthermore, the periodicity of the
for energy saving, but the network cannot arbitrarily reduce
necessary control signals can be configured between 5 ms
its paging density.
and 160 ms. For a coverage-providing carrier the default
The number of UEs in a cell fluctuates. However, the
periodicity is 20 ms [8]. This means that in idle mode a
network needs to implement flexible scaling of common
5G carrier can have approximately 100 times longer silent
signaling overheads to reduce the power consumption of the
periods than a LTE carrier (20 ms vs. 0.2 ms) [9], during
base stations in light-load scenarios. The UE side needs to
which time-domain energy saving can be applied.
be aware of and compatible with the dynamic adjustment
Furthermore, also when the carriers are loaded, time-
of the common signaling on the base station side to reduce
domain energy saving can be applied. Due to the bursty nature
the impact on processes such as initial UE access and cell
of traffic, empty symbols are common also in loaded traffic
handover and re-selection.
scenarios. It is also possible to create empty symbols, time
slots, and even sub-frames, by smart scheduling decisions.
III. BASIC ENERGY-SAVING FUNCTIONS FOR 5G RAN
Hence, time domain energy saving is a very important energy
5G will add many new functionalities to the RAN, and will
saving technology in 5G.
also deploy more resources in the form of carriers, band-
width, and transmit antennas. This not only means increased
B. SPATIAL-DOMAIN ENERGY SAVING
capacity and capabilities of the network, by following the
design principles in the previous section there are also large In spatial domain energy saving, a number of the RF channels
opportunities for energy saving technologies. are de-activated in low-load scenarios [10], [11], see Fig.3.
However, the complexity of the 5G network and the diffi- This is sometimes referred to antenna or multiple input mul-
culty of operation and management will also increase. This tiple output (MIMO) muting, or RF channel shutdown. The
also holds for the management and control of energy sav- idea is that the remaining capacity should be enough to serve
ing technologies, as also other aspects such as coverage, the traffic, but coverage and data rates may be affected since
interference, user QoS (quality of service) and so on need some of the PAs and thereby available output power is taken
to be considered when using the different energy saving out of service. However, there are methods to compensate
technologies. It is difficult to use fixed or simple algorithm this, e.g. by common signal power boosting.
to guarantee optimal performance. Therefore, AI needs to be Since 5G in general, and massive MIMO in particular, uses
introduced and how to use AI technology to realize energy more antennas and RF channels than previous generations,
saving applied to operation and management of wireless spatial domain energy saving is an important energy saving
communication units becomes more and more important. technology in 5G.
Therefore, section V will be dedicated to AI-driven energy
saving, while this section will focus on the energy saving
technologies themselves.

A. TIME-DOMAIN ENERGY SAVING

The scheduler allocates a certain number of symbols for
downlink data in accordance with the system load and service
data forecasts and turns off the PAs to save energy when
there is no information being transmitted. Symbol power
saving sometimes refers to as cell discontinuous transmission FIGURE 3. Illustration of spatial-domain energy saving. During low-load
periods, a number of antenna elements and corresponding RF channels
(DTX) [7], as illustrated in Fig.2. are de-activated.
However, even an empty LTE (long term evolution) carrier
transmits the cell-specific reference signals (CRS) approx-
imately every 0.2 ms, which means that only very short C. FREQUENCY-DOMAIN ENERGY SAVING
(2-3 Orthogonal Frequency Division Multiplexing symbols) Frequency domain energy saving relates to reducing the
component de-activations are possible. 5G has addressed this available frequency bandwidth and thereby save energy.
issue by drastically reducing the signal load in idle mode, There are two classes of frequency domain energy saving:

VOLUME 10, 2022 51749

 R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

1. Large-scale, in this case an entire carrier is shutdown, 5G era such as equipment deep sleep, LTE/NR carrier coop-
and consequently it is sometimes referred to as carrier shut- erative shutdown, enhanced channel shutdown and symbol
down. Carriers corresponding to some frequency bands or aggregation shutdown. The rationales of these new developed
frequencies are shut down, as illustrated in Fig.4. When these technologies in 5G are described in [12].
carriers are mapped to independent physical hardware com- On the other hand, the standards organizations are accel-
ponents (e.g. a radio unit), corresponding components can be erating the development of relevant energy-saving standards.
shut down and thereby save energy. The 5G NR standard allows more components to switch off or
2. Small-scale, such as bandwidth shrinking and subcarrier go to sleep when the base station is in idle mode and requires
shutdown. In this case only carrier specifications are adjusted. far fewer transmissions of always-on signaling transmissions.
The entire carrier is still working, and component shutdown In 3GPP Release 18, a new study item - Network Energy
cannot be implemented. In this case, the specifications of Savings [13] - was established as a priority in Dec. 2021,
certain processing units can only be adjusted to reduce power aimed to further explore the energy-saving potentials.
consumption, which means that the dynamic capability is This study item plans to study and develop a network
crucial for the achievable energy saving. energy consumption model especially for the base station,
KPIs (key performance indicators), an evaluation methodol-
ogy and to identify and study network energy savings tech-
niques in targeted deployment scenarios. The study should
investigate how to achieve more efficient operation dynami-
cally and/or semi-statically and finer granularity adaptation
FIGURE 4. Illustration of large-scale frequency-domain energy saving,
of transmissions and/or receptions in one or more of net-
carrier shutdown. One of the two carriers is shutdown, and the related work energy-saving techniques in time, frequency, spatial,
radio unit is de-activated to save energy. and power domains, with potential support/feedback from
UE, potential UE assistance information, and information
With the roll-out of 5G on new frequency bands, in par- exchange/coordination over network interfaces. The study
ticular large-scale frequency-domain energy saving, such as not only evaluate the potential network energy consumption
carrier shutdown, will be an important energy saving technol- gains, but also assess and balance the impact on network
ogy. With low load in the network and overlapping network and user performance, e.g. by looking at KPIs such as spec-
coverage, some carriers can be shutdown to save energy. This tral efficiency, capacity, user perceived throughput, latency,
can also be coordinated over different radio access technolo- UE power consumption, complexity, handover performance,
gies. Taking the overlapping network of NR (New Radio) and call drop rate, initial access performance, service level agree-
LTE as an example, we can think that NR is the capacity ment (SLA) assurance related KPIs, etc. The energy-saving
system and LTE is the basic coverage system. If the NR techniques to be developed should avoid having a large
traffic is low, the NR can be intelligently shut down and the impact to such KPIs.
traffic transferred to LTE at the same time. Once LTE service
exceeds a certain threshold, the NR carrier is re-activated. V. AI-BASED ENERGY-SAVING SOLUTIONS
In this way, the power consumption of the whole network A network intelligent solution based on AI [14] and big
changes with the traffic volume. data [15] is proposed after review of the research and appli-
cation progress of AI in the cellular network field [16]–[19].
IV. DEVELOPMENT OF ENERGY-SAVING TECHNOLOGIES The solution believes that AI and big data can be introduced
FOR 5G into a network on three levels: network element intelligence:
The beginning of network energy saving came with the fact operation and maintenance (O&M) intelligence; and service
that many sites had traffic peaks and troughs, which means intelligence, with the principles of tiered, on-demand and
certain parts of BSs could be shut down to save energy, phased. In this way, ubiquitous intelligence can be achieved.
and these included carrier frequency block, carrier frequency
shutdown, channel shutdown and symbol shutdown. Details A. WHAT IS AI
of these energy-saving technologies and their rationales can Artificial intelligence is the area of computer science focus-
refer to [12]. ing on creating machines that can engage on behaviors that
With the evolution of 5G network, new technologies used humans consider intelligent. It combines computer science,
in 5G provide faster and more services which make 5G physiology and philosophy that is a broad topic consisting of
system more complexity than 4G. However, providing ser- different fields, from machine vision to expert system.
vices with cellular network, transferring signals with multiple The element that the fields of AI have in common is the
channel, etc., 5G and 4G have something in common. Thus creation of machines that can ‘‘think.’’ The ability to cre-
5G energy saving does not have to start from scratch, all ate intelligent machines has intrigued humans since ancient
energy saving technologies of 4G can be leveraged. times, and with the advent of the computer and over 50 years
However, it is far away from being enough, more new tech- research of AI programming techniques, the dream of smart
nologies and enhanced technologies need to be developed in machines is becoming a reality [14].

51750 VOLUME 10, 2022

R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

According to the GSMA technical report [16], artificial even 15% energy consumption has been saved in the entire
intelligence is made up of 3 principal branches, big data, network [12], [19].
automation and artificial intelligence. Big data gathers large Some leader operators have introduced intelligent capabil-
data sets on which analytics are applied to gain insights and ities in their network progress, like network operation and
enhanced decision making. Automation is where machines maintenance, network planning and optimization. According
follow pre-programmed rules to run processes, generally used to the forecast of Analysys Mason, 80% operators hope that
for repetitive tasks. The final area is most advanced – Artifi- 40% network intelligence will be realized, while one third of
cial intelligence where machines perform cognitive functions them expect the rate will exceed 80% by 2025 [16].
similar to those attributed to humans. In response to the requirement of an intelligent and self-
AI is the ability of a computer or machine to emulate adaptive energy saving solution, AI and big data technologies
human tasks through learning and automation, generally are accelerating to be exploited and introduced into network
understood to be the simulation of the higher order functions energy saving for improving the efficiency and reducing the
of intelligent beings in areas such as visual processing, speech manpower required [17], [19].
processing and analytics.
AI algorithms take decisions as a consequence of the appli- C. AI-BASED ENERGY SAVING FOR 5G
cation of advanced analytical techniques and may be applied
Nowadays the 5G network deployment is on the fast track
in combination with automated advanced feedback loops to
around the world. Many MNOs are currently running 2G, 3G,
solve problems.
4G and 5G networks at the same time. The time distribution of
In general, a more detailed AI should seek to emu-
cellular network traffic has often obvious peaks and troughs,
late or simulate higher-order biological system functions
basic functions applied to the entire cellular network is not a
including visual processing, speech/natural language pro-
site-specific strategy, resulting in less efficient due to ignoring
cessing, outcomes prediction, objects or data categorization,
the traffic and neighboring sites patterns varied from site to
and problem-solving. However, it should be noted, AI must
site, especially in more complexity network.
exclude software systems based on traditional rule-based and
That is why AI and big data technology are introduced to
determined algorithms, for example, where a specific process
form a more precise energy saving strategy based on specific
or algorithm is designed/programmed by one or more people,
site traffic and other site-related conditions. The AI-based
which are not based on AI tools or techniques. This is because
network energy saving solution can forecast the traffic load
AI should include a significant element of learning from or
of base stations based on historical traffic load, service type,
adapting to data either as the whole process or an identifiable
site coverage and user behaviors. Energy saving strategy can
part of the process.
automatic configuration based on coverage identification and
Artificial intelligence can be further defined by the appli-
configuration identification by AI technology. The suitable
cation of learning that may be undertaken; machine learning
energy saving strategy combined with different energy sav-
and deep learning.
ing functions, including an initial relative threshold to the
Machine learning uses statistical techniques to perform
scenario and executable energy saving time schedule, will
specific tasks, often requiring a smaller amount of data.
be enabled for the sites that are expected to have energy
In doing so, machine learning can be conducted by low-end
saving effects. Meanwhile, AI-based network energy saving
systems though usually need labelling and features extrac-
solution can also ensure the balance between network power
tion to perform problem/task breakdown. This means that
consumption and network performance based on sufficient
machine learning applications are faster to train, but testing
model training.
may be slower to ensure the validity of results. However, these
The basic functions of an intelligent energy-saving solu-
are more readily explainable as the process is understood.
tion driven by AI should include but not limited to the fol-
Deep learning, on the other hand, uses artificial neural
lowing: scene identification, traffic forecast, multi-network
networks which require a more substantial amount of data
coordinate control, online iteration and optimization based on
to train models. This, in turn, requires high-performance
network performance, etc. [12]
GPUs (graphics processing units) but allows deep learning to
process unlabeled data and solve end-to-end problems. As a
result of its reliance upon large data sets, deep learning often D. ENERGY SAVING SCHEME ARCHITECTURE
is slower to train, however, faster to test, the biggest drawback BASED ON AI
here is what is referred to as the ‘‘black box’’ – while the Fig.5 shows the overall architecture of the AI-based energy-
inputs and outputs may be understood, the steps taken may saving network [12], [20]. Based on the intelligent application
not be. platform, the artificial intelligence algorithm is adopted to
achieve the maximum balance between the system perfor-
B. AI TECHNOLOGY IN NETWORK ENERGY SAVING mance and the energy-saving effect, so as to achieve the
Major vendors are currently offering AI-based energy saving network energy-saving and consumption reduction.
solutions, and the commercial application results showed By shutting down the resources that are not utilized and
that 80% activation time has been increased and over 10% keeping network capacity adequate yet minimum, network

VOLUME 10, 2022 51751

 R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

and dynamic policy generation will come on the next

stage.

FIGURE 5. Architecture diagram of AI-based energy-saving network.

energy consumption can be optimized in line with the net-

work traffic load forecast by AI-based.
For 5G era, enhanced AI-based energy-saving solutions
can take account of the different efficiency levels of fre-
quency bands and factors in that the power efficiency of
different networks can vary. By directing users from their
less power-efficient spectrum band(s) to other band(s) that FIGURE 6. Flowchart of 5G energy-saving based on manual configuration
are more power-efficient, more radio resources can be shut mode.

down to lower network energy consumption.

Alongside energy-saving potential, AI-based solution also B. AI-BASED MACHINE LEARNING MODE
can constantly monitor customer experience, network avail- Site-specific configuration, historical traffic and network
ability and data traffic to ensure there is no impact on network KPIs will be collected. Combined with time series prediction
performance. algorithm like LSTM (long short-term memory), intelligent
optimization algorithm like ant colony optimization and clus-
tering algorithm like K-Means, the network coverage, user
VI. APPLICATIONS FOR 5G ENERGY-SAVING SOLUTIONS
distribution and initial threshold will be output. With traffic
The applications of 5G energy-saving solutions mostly sepa-
forecast and evaluation radio resource utilization in real-time,
rate into two stages respectively with two modes: traditional
cell-specific energy saving strategy will be applied to the
manual mode and AI-based machine learning mode.
network to ensure the balance between network performance
and energy consumption.
A. TRADITIONAL MANUAL MODE Fig. 7 illustrates a typical flowchart of AI-based 5G
Based on basic analysis of coverage, traffic, the traditional energy-saving (ES) application procedure. Based on the RAN
energy saving solution is applied for simple scenarios with intelligent application platform, the AI algorithm is used
unified strategy, for instance, channel shutdown will be acti- to achieve the maximum balance between system perfor-
vated when the PRB utilization is below 10% during 10pm to mance and energy-saving effect to realize the optimal net-
6am. The strategy is configured on OMS (Operation Manage- work energy-saving and consumption reduction by using the
ment System) by manpower, and the energy saving efficiency basic data such as configuration, performance statistics and
will be calculated with performance counters for energy con-
sumption, while the KPIs are also monitored by manpower
to ensure the network performance. Fig.6 illustrated a step-
by-step procedure of 5G energy-saving based on manual
configuration mode.
In the early commercial operation stage, the energy-saving
effect of manual mode is quite good notwithstanding the
lack of historical data. But as the network users and traffic-
loads increase, the risk of degradation of network KPIs will
increase.
With the continuous development of 5G networks and
users, traditional manual mode that mainly rely on engi-
neers’ experience can no longer meet the users’ and verticals
requirements.
Thus, the second mode, AI-based technology with high- FIGURE 7. Flowchart of 5G energy-saving based on AI configuration
computational data analysis, cross-domain feature mining, mode.

51752 VOLUME 10, 2022

R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

measurement report or call detail trace (MR/CDT) of the 1) Idle or low traffic period based on historical traffic
existing network. analysis
2) Energy saving threshold based on network traffic load
VII. BEST PRACTICES OF AI-BASED ENERGY-SAVING 3) Energy saving activation time based on threshold
SOLUTIONS FOR 5G 4) Energy saving function combination
A. THE PRACTICE OF SERVICE AWARENESS ENERGY
SAVING FOR 5G 2) TRAFFIC LOAD FORECAST
AI-based energy saving solution with traffic forecast does The intra-week sub-sequence split prediction method,
improve the energy saving efficiency of basic functions in which each of the seven days in a week needs to be
[16]–[19]. But in most multi-mode and multi-frequency cel- put into the series of its responding days of all the weeks,
lular network, it still has some limitation that service effi- is combined with cell types, holiday factors and forecast of
ciency varies from mode to mode, and/or band to band. network traffic load. After putting all the algorithm candi-
And if all services/users are concentrated in part of the dates (linear regressive, ARIMA (autoregressive integrated
network/band, more energy consumption could be saved after moving average model) and LSTM, second-order exponential
idle network/band shut down or deep sleep. smoothing, etc.) into tests, the one with the best results is
AI-based service awareness in 5G network should also selected.
be taken into consideration, which exploits the differences The result of commercial application cases showed that the
in energy efficiency of different types of services to deliver prediction accuracy exceeds up to 90% for uplink/ downlink
certain services to the most energy-efficient network, helping PRB utilization and RRC connected users. Fig. 9 illustrates
achieve the most efficient energy usage without impact on the prediction accuracy of the commercial application case,
user experience [18]. compared the predicted value with the real traffic in DL/UL
Fig.8-Fig.11 illustrates an AI-based pilot solution based on PRB usage and RRC users every 15 minutes. The predicted
service-awareness energy-saving. The solution involves pro- value matches well with the actual value in normal scenarios.
viding automation capabilities for the service management
layer and resource management layer. Based on network-
level AI-based intelligent energy saving policy management
and site energy saving scheduling control, the mobile net-
work energy saving solution implements network scene adap-
tion, one site one policy, and multi-network collaboration
for intelligent base station energy saving management. This
maximizes network energy saving benefits while ensuring
stable network performance and achieves the optimal balance
between energy consumption and KPIs.

FIGURE 9. The commercial application case and the prediction accuracy.

3) SERVICE AWARENESS
Based on energy saving functions and AI-driven traffic load
prediction, this pilot solution is an industry’s trial project
to introduce AI-powered service-awareness energy saving.
By identifying service types and their energy efficiency dif-
ferences, this solution can evaluate service requirements in
real time and support the service with networks of higher
FIGURE 8. AI-based network energy-saving accurate deployment. energy efficiency to maximize energy efficiency in the entire
network.
To improve the network energy efficiency based on user
1) CELL-SPECIFIC INITIAL STRATEGY redistribution, there are three main steps: target network/band
Initial energy saving strategy can automatic configuration selection, suitable user selection and consequent user direc-
based on coverage identification and configuration identifi- tion. Service efficiency varies from network to network,
cation by AI technology. The suitable energy saving strategy and/or band to band, the most energy-efficient network/band
combined with different energy saving functions, including will be selected as target and the most suitable users will be
an initial relative threshold to the scenario and executable selected. Also, energy efficiency pattern may change after
energy saving time schedule, will be enabled for the sites that user direction, leading a new round of optimization. Thus the
are expected to have energy saving effects. three steps together form a closed loop.

VOLUME 10, 2022 51753

 R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

It is estimated that more than 5 million kilo-watt-hours will

be saved annually per thousand stations and at least 2,500 tons
of carbon emissions will be reduced if the pilot solution is
deployed in the whole 4G/5G network in Chengdu.
And according to typical network configuration calcula-
tions, the energy saved by this pilot solution is twice as much
as that of the conventional AI-based energy saving solutions,
and it can save up to 20% of energy in a multi-mode network,
FIGURE 10. Service-awareness energy saving.
thereby effectively reducing the operational expenditure.

B. THE PRACTICE OF NEW ARCHITECTURE AAU WITH

4) ONLINE ITERATION AND OPTIMIZATION OF THE BEAM PATTERN OPTIMIZATION
THRESHOLD The cooperation and optimization of beams between cell and
In order to improve the energy saving efficiency, online iter- UE could avoid aimless radiation of base station. The practice
ation and optimization of the threshold can be used instead of this type of energy saving are still on trial and ongoing
of the traditional ways which does not take site variations research. Here is a description of the early practice results.
into consideration and results in safe but inefficiency energy For antenna array of AAU, the larger size of the antenna
saving settings. array, the narrower beam of the channel, thus, the more
Clustering algorithm is used to find out the optimal energy focused channel energy, and the longer distance of the net-
saving threshold settings with the best efficiency and shortest work coverage. With the introduction of an ultra-large-scale
time. In any case when the KPI baseline is compromised, the antenna array, the uplink and downlink coverage can be
threshold can be rolled back. increased synchronously without increasing the transmission
power. Therefore, the use of ultra-large-scale antenna arrays
5) APPLICATION AND PERFORMANCE is an important innovation direction to improve coverage
This solution is put into commercial use in Chengdu city, and reduce energy consumption. Through the integration and
to verify this commercial energy saving solution with ser- innovation of new architectures and new algorithms, both
vice pilot. The results with Chengdu networks show that the network performance and energy saving can be promoted,
over 35% network energy consumption of 4G/5G can be helping MNOs to build better and greener 5G networks.
reduced without impact on the network performance or user A new pilot practice with such ultra-large-scale antenna
experience. arrays of AAU are deployed in Xiamen city, which doubled
This commercial trial in Chengdu 4G/5G network involves the size of the antenna array (elements from 192 to 384) and
three phases. With only the basic energy saving function, manufactured with new materials and architectures. It adopts
about 9 kWh energy is saved daily per site; when the adaptive beam optimization and intelligently adapts the user’s
AI-driven traffic forecast is enabled simultaneously, approxi- wireless channel changes to improve the utilization efficiency
mately 12 kWh energy is saved daily per site. After AI-based of air interface resources; and adopts high-resolution beam
service pilot is enabled, up to 14 kWh energy can be saved domain noise reduction to improve the efficiency of multi-
daily per site. The proportion of energy saving is increased user pairing. This enables the new AAU to achieve ‘‘accurate-
from 16.6% to 24.5% (as illustrated in Fig. 11). During the alignment,’’ ‘‘fast-following,’’ and ‘‘good-pairing,’’ which
trial, there is less impact to network performance, both the can greatly improve user experience and cell capacity. The
4G network and 5G network. comparison and performance gains of this new architecture
AAU are illustrated in Figure 12 and Table 1.

FIGURE 12. Comparison of Traditional AAU and New Architecture AAU

(number of antenna elements is from 192 to 384).

Test results show that compared to the traditional solu-

FIGURE 11. Energy saving per site daily (kWh and Proportion). tion, the new AAU can increase the cell’s 30% uplink and

51754 VOLUME 10, 2022

R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

TABLE 1. Performance gains for new architecture AAU compared to 2) Impact the consumer product such as cell phone and
traditional AAU.
cell phone software to more intelligent and corporate with the
network in power and signal strength usage.
3) Vendor equipment design to be more integrated with AI
to reduce power usage.
With the fact that 5G network deployment and its operat-
ing time is relatively short, compared to existing 2G/3G/4G
networks with abundant data accumulation, how to set initial
power saving strategy of 5G power saving and its reference
parameters and how to perform the power saving collabo-
ration between 4G and 5G are all undergoing studies and
will be optimized along the way when more experience is
accumulated after the solution been widely deployed.
Nevertheless, with the continuous innovation and evolution
of 5G energy-saving technologies based on digital technolo-
gies we have sound reasons to believe that the energy effi-
downlink coverage, and the average experience rate gains for
ciency of future networks will become more efficient, and
edge users can be increased by 25%. If the cell coverage
the OPEX of MNOs will become much lower to achieve the
index on edge area remains unchanged, the base station can be
NetZero goals [1], [2] for ICT industry early.
configured with lower transmission power, thereby to reduce
the energy consumption of the base station. Compared with
REFERENCES
the traditional AAU, the energy consumption can be reduced
[1] Mobile Net Zero: State of the Industry on Climate Action 2021, GSMA,
by about 30%. London, U.K., 2021.
[2] Setting Climate Targets: A Step by Step Guide for Mobile Network Opera-
tors to Set Science Based Targets, GSMA, London, U.K., 2020.
VIII. CONCLUSION [3] GHG Emissions Trajectories for the ICT Sector Compatible With the
In this article, we discussed the energy saving principles UNFCCC Paris Agreement, document ITU-T L.1470 (01/2020), Recom-
and the practical application technologies such as the carrier mendation ITU-T, 2020.
[4] A. S. G. Andrae, ‘‘New perspectives on internet electricity use in 2030,’’
shutdown, channel shutdown and symbol shutdown emerged Eng. Appl. Sci. Lett., vol. 3, pp. 19–31, Jun. 2020.
since 4G era for mobile communication networks, which [5] Going Green: Benchmarking the Energy Efficiency of Mobile, GSMA,
can also be leveraged to mitigate 5G energy consumption. London, U.K., 2021.
We also elaborated some enhanced technologies, e.g., deep [6] B. Debaillie, C. Desset, and F. Louagie, ‘‘A flexible and future-
proof power model for cellular base stations,’’ in Proc. IEEE
sleep and symbol aggregation shutdown, which have been 81st Veh. Technol. Conf. (VTC Spring), May 2015, pp. 1–7,
developing in 5G era. Meanwhile, with the rapid develop- doi: 10.1109/VTCSpring.2015.7145603.
ment of AI and other emerging technologies, energy-saving [7] P. Frenger, P. Moberg, J. Malmodin, Y. Jading, and I. Godor, ‘‘Reduc-
ing energy consumption in LTE with cell DTX,’’ in Proc. IEEE
solutions incorporating AI can speed the realization of effi- 73rd Veh. Technol. Conf. (VTC Spring), May 2011, pp. 1–5, doi:
cient collaboration of multi-networks and can be introduced 10.1109/VETECS.2011.5956235.
into all aspects of communication network operation and [8] Physical Layer Procedures for Control (Release 16), v16.0.0, docu-
ment TSG RAN (2019). TS 38.213; NR; 3GPP, 2019.
maintenance. As an application illustration, this article pre- [9] P. Frenger and R. Tano, ‘‘More capacity and less power: How
sented two best practices for AI-based smart energy saving 5G NR can reduce network energy consumption,’’ in Proc.
solutions. One is the practice on AI-based service awareness IEEE 89th Veh. Technol. Conf. (VTC-Spring), Apr. 2019, pp. 1–5,
doi: 10.1109/VTCSpring.2019.8746600.
energy saving for 4G/5G collaborative network, the other one
[10] P. Skillermark and P. Frenger, ‘‘Enhancing energy efficiency in LTE with
is adoption of new architecture AAU with AI-based beam antenna muting,’’ in Proc. IEEE 75th Veh. Technol. Conf. (VTC Spring),
optimization. Both of them could help to achieve the most May 2012, pp. 1–5, doi: 10.1109/VETECS.2012.6240248.
energy-efficient network with good performance and lower [11] H. Halbauer, A. Weber, D. Wiegner, and T. Wild, ‘‘Energy efficient massive
MIMO array configurations,’’ in Proc. IEEE Globecom Workshops (GC
OPEX. Wkshps), Dec. 2018, pp. 1–6, doi: 10.1109/GLOCOMW.2018.8644331.
However, network-wide intelligence is difficult to be [12] Smart Energy Saving of 5G Base Station: Based on AI and Other Emerging
achieved overnight, a long-term development is required. Technologies to Forecast and Optimize the Management of 5G Wireless
Network Energy Consumption, document FG-AI4EE D.WG3-02, ITU-T
With the continuous accumulation of data from commer- Technical Report, 2021.
cial networks, the machine-learning algorithm will enable [13] (Dec. 2021). RP-213554. New SI: Study on Network Energy Saving
AI based power saving solution to be perfect gradually. for NR. 3GPP TSG RAN Meeting #94e. [Online]. Available:
https://www.3gpp.org/ftp/TSG_RAN/TSG_RAN/TSGR_94e/Docs/RP-
The AI algorithm itself will be evolved iteratively to realize 213554.zip
higher efficiency and accurate strategy adapting when net- [14] H. H. de Garis, ‘‘The artilect war—Cosmists vs. terrans—A bitter contro-
work topology and traffic model changes. versy concerning whether humanity should build godlike massively intel-
Here is the thought of widen the impact: ligent machines,’’ Etc Pubns., Pittsburgh, PA, USA, Tech. Rep., Feb. 2005.
[15] V. Mayer-Schönberger and K. Cukier, Big Data: A Revolution That Will
1) Feedback to vendor industry to design more intelligent Transform how we Live, Work and Think. Boston, MA, USA: Houghton
equipment to be more adaptive. Mifflin Harcourt, 2013.

VOLUME 10, 2022 51755

 R. Tan et al.: Energy Saving Technologies and Best Practices for 5G Radio Access Network

[16] (2019). AI and Automation: An Overview. London: GSM Associ- YINGYING FAN was born in Shanghai, China,
ation. (Mar. 10, 2022). [Online]. Available: https://www.gsma.com/ in 1982. She received the B.S. and M.S. degrees
futurenetworks/wiki/ai-automation-an-overview/ in communications and signal processing from
[17] AI in Network: Intelligent Autonomous Network Case Report, GSMA, the Nanjing University of Aeronautics and Astro-
London, U.K., 2019. nautics, Nanjing, China, in 2004 and 2007,
[18] (2020). 5G Energy Efficiencies: Green is the new Black. GSMA Intelli- respectively.
gence. (Mar. 10, 2022). [Online]. Available: https://data.gsmaintelligence. From 2007 to 2015, she was a Software Testing
com/api-web/v2/research-file-download?id=54165956&file=241120-5G-
Engineer with the Shanghai Research Institute,
energy.pdf
ZTE Corporation, Shanghai, who was major in ser-
[19] (2019). 5G-Era Mobile Network Cost Evolution. London: GSM Asso-
ciation. (Mar. 9, 2022). [Online]. Available: https://www.gsma.com/ vice procedure of GSM/UMTS/LTE system. Since
futurenetworks/wiki/5g-era-mobile-network-cost-evolution/ 2015, she has been the RAN Production Planning Director of the Shanghai
[20] T. Rumeng, W. Tong, S. Ying, and H. Yanpu, ‘‘Intelligent energy saving Research Institute, ZTE Corporation, who was major in energy saving of
solution of 5G base station based on artificial intelligence technologies,’’ in radio access network and network sharing. She is an active participant in
Proc. IEEE Int. Joint EMC/SI/PI EMC Eur. Symp., Jul. 2021, pp. 739–742, the development of ITU-T recommendations, working within ITU-T SG5
doi: 10.1109/EMC/SI/PI/EMCEurope52599.2021.9559261. for researching environmental efficiency for artificial intelligence and other
emerging technologies.

RUMENG TAN was born in Shucheng, Lu’an,

Anhui, China, in 1973. He received the B.S. degree
in electromechanical engineering from the Univer-
sity of Petroleum, Beijing, China, in 1995, and WENTAO ZHU was born in Changge, Henan,
the M.S. degree in electromechanical engineer- China, in 1986. He received the B.S. degree
ing from Southeast University, Nanjing, China, in optical information science and technology
in 2003. and the M.S. degree in electromagnetic field
From 2003 to 2021, he was a Researcher with and microwave technology from Beijing Jiaotong
the Mobile Communication Institute of China University, Beijing, China, in 2007 and 2009,
Telecom Research Institute, Guangzhou, China. respectively.
He is an active participant in the development of ITU-T Recommendations From 2009 to 2021, he was a Researcher with
and Chinese ICT industrial standards, working within ITU-T SG5 and CCSA China Mobile Group Design Institute Company
TC9, which creates and manages the EMC/EMF standards for ICT industry. Ltd., Beijing. He is an active participant in the
He has led the drafting of ITU-T recommendations K.121, K.141, and development of ITU-T recommendations and Chinese ICT industrial stan-
L.Suppl.43, and published over 15 Chinese communication industrial stan- dards working. He has participated in the drafting of recommendation ITU-T
dards. His research interests include the evolution of mobile communication K.149 and participated in the application of about 20 patents in the antenna
networks, radio frequency electromagnetic fields, and energy saving and and wireless fields. He has published over ten Chinese communication
carbon reduction technologies and solutions in ICT industry. standards. His research interests include antenna system test and analysis,
Mr. Tan also serves as the Vice-Chair for the CCSA TC9 WG3 (Electro- electromagnetic radiation test, antenna diagnosis based on big data, and
magnetic Radiation and Safety). massive MIMO technologies.

YING SHI was born in Wuhan, Hubei, China,

in 1974. She received the B.S. degree in electri-
cal engineering from Wuhan University, Wuhan, TONG WU worked with the EMC Certifica-
in 1995, the M.S. degree in electrical engineering tion Center, Radio Department, Department of
from the Guangdong University of Technology, Bioenergy and Environment, and Department of
Guangzhou, China, in 1998, and the M.S. degree Information and Electronics, National Institute of
in telecommunication management from the Uni- Metrology, China. He also worked as the Secretary
versity of Strathclyde, U.K., in 2007. of the National Technical Working Group on Stan-
From 1998 to 2008, she was a Researcher dardization of Assessment Methods in the fields
with the Metrology Testing Center, China Telecom of electricity, magnetics and electromagnetism for
Research Institute, Guangzhou. From 2009 to 2021, she was a Manager and human irradiated bodies, undertook the technical
a Senior Expert of the Research and Development Management Department, counterpart work of IEC/TC 106. He is currently
China Telecom Research Institute, Guangzhou. She is an active participant an Associate Researcher and the Deputy Director. He is mainly responsible
in the development of ITU-T recommendations and Chinese ICT industrial for the measurement standards of electromagnetic environment, data cen-
standards. She has published over ten Chinese communication standards. ter electromagnetic environment, and interference testing research. He is
Her research interests include environmental engineering, electromagnetic familiar with measurement technology, research method and standardization
protection, and energy saving and carbon reduction technologies in ICT of electromagnetic environment, and established EMC capacity verification
industry. system.
Ms. Shi currently serves as the Vice-Chair for CCSA TC9.

51756 VOLUME 10, 2022