IEEE ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY SECTION

Received 5 August 2024, accepted 19 August 2024, date of publication 26 August 2024, date of current version 3 September 2024.
Digital Object Identifier 10.1109/ACCESS.2024.3449700

Energy Efficiency in Open RAN: RF Channel

Reconfiguration Use Case
MARCIN HOFFMANN 1,2 , (Member, IEEE),
AND MARCIN DRYJAŃSKI 1 , (Senior Member, IEEE)
1 Rimedo Labs, 61-131 Poznan, Poland
2 Institute of Radiocommunications, Poznan University of Technology, 60-965 Poznan, Poland
Corresponding author: Marcin Hoffmann ([email protected])
This work was supported in part by the Polish Ministry of Education and Science in terms of Publication Fee and Computer Simulations
under Grant 0312/SBAD/8167, and in part by the Rimedo Labs own budget in terms of algorithm design.

ABSTRACT Recently, energy efficiency (EE) has been pointed out as one of the key requirements within
mobile networks. The development of intelligent algorithms providing Radio Access Networks (RAN) with
EE features is possible when having access to the Key Performance Indicators (KPIs), and proper control
actions, e.g., cell on/off switching, or Radio Frequency (RF) channel reconfiguration. These features together
with a Machine Learning (ML) framework are available in the O-RAN architecture. This paper provides an
overview of the EE framework according to the use cases specified by the O-RAN ALLIANCE. It is followed
by the implementation of the Energy Saving rApp (ES-rApp). The rApp utilizes Deep Q-Learning (DQL) to
increase EE through intelligent RF channel reconfiguration. Simulation results show up to a 24.8% EE gain
over the static RF Channel Configuration (RCC).

INDEX TERMS Open RAN, 5G, energy efficiency, RAN intelligent controller, deep Q-learning, massive
MIMO.

I. INTRODUCTION MNOs with unified ES mechanisms [3]. The most recent

Recently, Energy Efficiency (EE) has become one of the example related to beyond 5G standardization which relates
key topics within RAN [1]. It is currently on the industry significantly to energy consumption is Integrated Sensing
and research agendas. As high user throughput is targeted and Communications (ISAC) [4]. This is motivated by the
within 5G and beyond networks, an increasing number of fact that ISAC utilizes dedicated hardware to perform both
energy-consuming hardware is deployed. Moreover, energy- functionalities, e.g., M-MIMO downlink communications,
efficient RAN must offer a high Quality-of-Service (QoS) and radar [5].
to the network users. The Possibility of dynamically con- Energy savings come at the cost of less available resources,
trolling the used resources and hardware by adapting its which may impact the overall performance of the system.
usage to the actual traffic demand, would bring significant Thus, the optimization goal is typically defined taking both
benefits to the Mobile Network Operators (MNOs). In this of these into account. For this purpose, EE is usually defined
perspective, Energy Saving (ES) mechanisms play a crucial as a ratio between the average user throughput and average
role. They are involved in all layers and domains within power consumption. EE monitoring can help MNOs balance
RAN, from energy-efficient hardware, through the cloud, the user throughput-power consumption tradeoff to lower
software stack, and orchestration [2]. Significant efforts have energy costs or reduce carbon footprint. However, to enable
been put in during the standardization phase of 5G New intelligent and case-dependent EE optimization, algorithms
Radio (NR) by 3GPP within Release-18 onwards to provide need access to network interfaces, both for monitoring and
control purposes. There are various options possible to
The associate editor coordinating the review of this manuscript and improve ES in RAN, e.g., using legacy RAN built on physical
approving it for publication was Adamu Murtala Zungeru . hardware. They are, however, of low efficiency due to the
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 118493
 M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

FIGURE 1. O-RAN use cases aimed at providing energy savings a) COOS b) RCR.

dedicated capacity allocated and powered on all the time, and Cell On/Off Switching (COOS), which is well investigated
limited management capabilities offered by incumbent RAN in the research community, and RF Channel Reconfigu-
vendors. As an alternative, the Open RAN concept fits here ration (RCR) in Massive Multiple-Input Multiple-Output
better with virtual components, open interfaces, and unified (M-MIMO) scenarios, also known as antenna selection (see
native intelligence enabling multi-vendor competition and Fig. 1.)
programmability [6]. A specific and industry-wide adopted The existing RCR algorithms rely widely on the ML
realization of Open RAN is defined by O-RAN ALLIANCE techniques, e.g., [12]. However, these solutions require
and referred to as O-RAN. In the O-RAN standardization detailed Channel State Information (CSI), i.e., radio channel
and development effort, EE is one of the key requirements, coefficients between the user and each of the Base Station’s
namely, the O-RAN ALLIANCE is currently providing (BS’s) antennas. Such physical layer data is not available
means to support ES topics [7], e.g., with the use of according to the 3GPP, and O-RAN specifications, which
centralized ES policies. makes those algorithms hard to implement in practice. On the
The key aspect of O-RAN is to embed native intelligence, other hand, there is an algorithm that utilizes deep learning to
through the use of a RAN Intelligent Controller (RIC), which predict RCC based on the average channel gain and precoder
is standardized in two flavors, namely a Non-Real Time RIC weights [13]. Those weights are hard to be exposed in
(Non-RT RIC) and Near-Real Time RIC (Near-RT RIC), real-time by the BS while considering adaptive beamforming
which operate in timescales of above 1 s, and between 10 ms schemes like Zero Forcing, or Maximum Ratio Combining.
and 1 s respectively [8]. Both RICs serve as platforms for Of course, they can be obtained for the Grid of Beams (GoB)
the so-called RIC applications, which can be developed either beamforming, where a strictly specified set of weights is
by the RIC vendors themselves or by third parties: rApps for used, but as fixed values, they do not contribute to the ML
Non-RT RIC, and xApps for the Near-RT RIC. xApps and model. Finally, RCR can be performed based on the cell
rApps can utilize ML algorithms and fit well for addressing load threshold as proposed in [14], but this approach offers
the ES aspects, e.g., xApps for symbol-level micro sleeps, and neither flexibility nor high performance. Thus, in this paper,
rApps for algorithms that perform hardware reconfiguration, the authors propose an intelligent RCR that not only utilizes
like scaling down the antenna array. ML to optimize EE defined as the ratio between the user
In the literature, there are some works on the O-RAN average throughput and average power consumption, but also
architecture with a focus on the utilization of ML algorithms. takes as input features that are exposed by the BS according to
For example, in [6] a Deep Reinforcement Learning (DRL) the O-RAN ALLIANCE and 3GPP specifications [11], [15].
agent is proposed for resource allocation, and in [9], The aim of this paper is, first, to discuss the role of
ML is utilized to predict Service-Level Agreement violation. O-RAN in providing ES, and improving EE of the 5G
However, both works do not cover the topic of EE. EE, and beyond networks concerning the O-RAN ALLIANCE
in turn, is mentioned in [10], however, not as a main topic, and 3GPP. Secondly, to propose an ES-rApp that utilizes
but as one of the many use cases. In that paper, it is achieved DRL to perform RCR aimed at the maximization of EE.
through the per Resource Block (RB) power allocation, which The proposed ES-rApp utilizes input parameters that are
is not currently planned to be supported according to the aligned with the O-RAN ALLIANCE use-case definition,
O-RAN ALLIANCE specifications. The existing literature and exposed by the BS according to the 3GPP specifications.
lacks an overview of EE use cases that are agreed upon In addition, most of the current implementations for a
in the O-RAN community as possible for implementation particular use case are realized with a single xApp or
according to [11]. These use cases are, among others, rApp. Therefore, the third novelty presented in this paper

118494 VOLUME 12, 2024

M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

is the interworking of an xApp - rApp tandem. The rApp that define general schemes and signaling flows that the
oversees their operation and controls the xApp’s operation algorithms (deployed as xApps and rApps) shall fit in.
for fine-tuning the performance. In Sect. II authors discuss Fig. 1-a shows the concept of COOS, where the idea is
the role of O-RAN architecture in providing ES at the feature to shut down (or put in sleep mode) the whole cell. In the
level, i.e., COOS, and RCR. It is followed up by an overview figure, pico/small cells serve as the capacity layer and are
of the generalized O-RAN framework for the development subject to switching off or on depending on the traffic needs.
of ES-rApps in Sect. III. Here, the authors also highlight Note that COOS does not only concern pico/small cells - it
the importance of cooperation between rApps and xApps, refers to a ‘‘cell’’ generically. Therefore, it can be utilized in
to support ES mechanisms. Based on the general framework, other scenarios, e.g., switching off/on a regular carrier, or a
in Sect. IV, a detailed description of an O-RAN-compliant Component Carrier in the Carrier Aggregation scenario. The
ES-rApp is provided. It is supported by numerical results cells subject to switching off are capacity cells, while the
showing the benefits of the solution presented in Sect. V. The coverage cell should always be on. The other O-RAN use
paper is summarized in Sect. VI. case, presented in Fig. 1-b is defined as RCR, where instead of
The contribution of the paper can be summarized as switching off/on the whole cell, the algorithm switches off/on
follows: some of the RF channels that correspond to the number of
beams in a M-MIMO cell. If the traffic is low, the number of
• We provide the description and discuss the role of
beams can be decreased, not lowering users’ QoS.
O-RAN in the improvement of EE in the current 5G and
In addition to the definition of use cases, O-RAN provides
future 6G networks based on both O-RAN ALLIANCE
a framework for the operation of various algorithms. This is
and 3GPP specifications.
realized through the separation of concerns using Near-RT
• We propose an ES-rApp that unlike [12] and [13] is
and Non-RT RIC, and standardized interfaces, like, O1, A1,
design in compliance with O-RAN ALLIANCE and
and E2. When working with ES it is important to consider that
3GPP specifications. The proposed ES-rApp utilizes
those algorithms cannot work in a vacuum. The decision to
DRL for RCR, which allows for higher EE gains than
switch off the whole or part of the BS needs to be coordinated
the load-based solution [14].
with other features [18]. Such coordination and a holistic
• We propose a scheme of cooperation between the
approach to ES help to avoid instability in the network.
ES-rApp that performs RCR in a large timescale and
BMM xApp that performs near-real-time user-to-beam III. OPEN RAN rAPPS FOR INCREASING NETWORK
association [8], [16]. ENERGY EFFICIENCY
The traffic load varies in the mobile network over the day,
II. HOW TO INCREASE EE IN O-RAN? or even between different days of the week. It is related
As mentioned in the previous section, O-RAN enables the to the users’ mobility and behavior, e.g., from Monday to
introduction of ES optimization algorithms through Non-RT Friday people work in city centers, thus usually BSs therein
and Near-RT RICs. More specifically, Non-RT RIC, along are highly loaded from morning till afternoon. On the other
with rApps, belongs to the management plane, while Near-RT hand, a lot of people have their homes in the suburbs,
RIC serves as a platform for xApps, abstracting out part of thus BSs deployed therein remain almost empty until late
the RAN and enabling direct control of BS entities. There afternoon. This creates an opportunity to either switch off
are two control interfaces associated with those: A1, which entire cells or perform RCR to increase EE, which in O-RAN
operates between Non-RT RIC and Near-RT RIC, facilitating architecture can be implemented in the form of rApp or
policy management, enrichment information transfer, and xApp. Considering the ES applications, one should note
ML model updates; and E2, which connects Near-RT RIC that the reconfiguration of the O-RAN Radio Unit (O-RU),
with RAN nodes, serving as a control loop responsible for by either switching on/off or changing the RCC requires
executing commands and gathering measurements from BS some transition time, and sometimes even a restart of the
nodes (in O-RAN called E2 Nodes). The O1 interface is device. Thus, we propose to implement ES algorithms in
responsible for configuration and management. It connects the form of an ES-rApp. The generalized framework for the
the Service Management and Orchestration (SMO) system in deployment of the ES-rApp is depicted in Fig. 2. The rApp
which Non-RT RIC resides with RAN nodes [17]. Through is hosted in a Non-RT RIC residing inside the SMO, along
the abovementioned framework and interfaces, O-RAN with a Collection and Control (CC) entity. They are used
enables the development of use case-dependent xApps, and to provide the ES-rApp with data captured from E2 Nodes
rApps aiming at the improvement of EE, e.g., by disabling to perform control actions (e.g., switching off/on of cells,
some parts of RAN hardware during low traffic load. The or reconfiguration of RF channels).
decisions to switch off hardware to save energy are typically
taken based on long-term trends and statistics. Therefore, A. INFORMATION FLOW AND RL FRAMEWORK
ML-based schemes can significantly help in improving ES Fig. 2-a depicts the signaling flow (based on O-RAN
schemes by, e.g., predicting traffic changes before they documentation [11]) between Non-RT RIC, with an ES-
happen. O-RAN specifications [11] provide the ES use cases rApp, and other O-RAN entities. At first, the ES-rApp

VOLUME 12, 2024 118495

 M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

FIGURE 2. Information flow diagram (a) and corresponding RL framework (b) applicable to COOS and RCR O-RAN
use-cases.

obtains information about supported configurations of O-RU, The whole process is cyclic, i.e., the agent recognizes the
e.g., sleeping modes, and RCCs. This is done through the state, takes action and observes the reward. After multiple
hierarchical connection utilizing the O1 interface between iterations, the rApp is trained on how to configure O-RU to
SMO and E2 Node (i.e., a BS node connected through the E2 maximize rewards, e.g., EE.
interface to RIC), and the Open Fronthaul (O-FH) interface
between E2 Node and O-RU. Next, the ES-rApp is configured B. COOPERATION WITH xApps/NEAR RT CONTROL LOOP
to receive from E2 Node and O-RU, data compliant with ES features should interact with other Radio Resource
the 3GPP specification [15]. The ES-rApp can monitor, e.g., Management (RRM) algorithms. The decision to switch off
cell load, power consumption, and user throughputs. The some of the resources needs to be coordinated. Fig. 3 shows
ES-rApp would then utilize the ML to process the collected an example where a Beam Mobility Management xApp
data to dynamically switch off/on cells, or perform RCR. The (BMM-xApp) operates within a near-RT control loop, and
resultant decision made by the ES-rApp is communicated to the ES-rApp operates within the non-RT control loop. The
Non-RT RIC, which utilizes CC to change the configuration ES-rApp aims to intelligently adjust the configuration of
of O-RU via a cascade of O1, and O-FH. Both COOS and the antenna array based on the network state, to maximize
RCR affect the network performance. Thus, in the last stage, EE. The BMM-xApp is responsible for switching the users
the ES-rApp utilizes a previously established O1 connection between beams based on the location information, to reduce
to evaluate network performance and retrain the ML model if
needed.
The performance of the network after COOS or RCR
actions is hard to predict based on analytical calculations.
Thus, the ES-rApp must act in a trial-and-error manner, i.e.,
select O-RU configuration, observe the performance in terms
of users’ QoS or ES, and use this observation to update the
ML model. This approach is known as RL, i.e., learning
through interaction, which is proven to fit the O-RAN
architecture [6]. To fulfill this requirement, the authors
propose to use the RL framework as shown in Fig. 2-b.
In this framework, a so-called agent (ES-rApp) interacts
with the environment (E2 Nodes, O-RUs), by making actions
(decisions on O-RU reconfigurations) based on the state
(KPIs, e.g., cell-load), and observes the reward (KPIs, e.g.,
power consumption, user throughput), as depicted in Fig. 2-b. FIGURE 3. Cooperation between rApp and xApp.

118496 VOLUME 12, 2024

M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

the number of beam reselections and radio link failures that, selects one of the RCCs (action a) and measures its
[8], [16]. performance in terms of EE (reward r). We propose the
The ES-rApp and BMM-xApp cooperate hierarchically. following detailed definitions of RL framework components:
The former performs the reconfiguration of the M-MIMO
• An Environment is a single cell of an O-RAN
BS within relatively long intervals. BMM-xApp provides fast
compliant network equipped with an M-MIMO antenna
decisions on user-to-beam association and relies on the Radio
array.
Environment Map (REM), representing the spatial distribu-
• An Agent is the ES-rApp placed in the Non-RT RIC
tion of received power associated with each beam. When the
aiming at EE maximization through dynamic RCR. The
ES-rApp changes the array configuration, the number and
agent both takes actions and monitors their performance
shape of beams also change. Thus, the BMM-xApp must
with the use of the CC mechanism that provides the
be equipped with different sets of REMs proper for each
termination of O1 and O-FH.
available array configuration. When the ES-rApp decides to
• An Action is defined as one of the possible RCCs
change the array configuration, an appropriate notification is
at O-RU and is enforced by the agent via O1 and
sent to the BMM-xApp.
O-FH. We assumed a discrete action space A =
One should note that when there are multiple Apps in a
{a1 , a2 , a3 }, where actions a1 , a2 , and a3 corresponds
single system, conflicts may occur. Beyond obvious ones, like
with 100%, 50%, and 25% of Active RF Channels
two Apps concurring to change the same parameters, others,
(ARCs) respectively. To compensate for beamforming
like different objectives set for two Apps, can be challenging,
gain variations related to scaling up/down the number
e.g., one App maximizes saved energy, while another
of active antenna elements, a constant Tx power for the
maximizes user performance. If both work concurrently,
whole array is assumed.
they can constantly modify the network behavior, causing
• A State s is constituted as a tuple containing one-hot
instabilities and making it difficult for both of them to
encoded current RCC (oneHotRCR) and several KPIs
converge. Such conflicts should then be considered on
defined according to the 3GPP specifications [15]:
multiple levels, between two rApps, between two xApps, and
mean PRB utilization over all M-MIMO layers
between rApp and xApp [19]. One approach is to jointly
within the whole cell (RRU.PrbTotSdmDl, scaled
design a set of Apps, so that they do not come into conflict
to range 0 − 1), means power consumption per
by design, but rather work in a coordinated manner - which
carrier (PEE.AvgPower, divided by the maximum
is the case here. The other is to design conflict resolution or
power consumption measured under full cell load),
mitigation functions, which are out of the scope of this paper.
average downlink (DL) user throughput per cell
(DRB.UEThpDl, divided by 107 ), the total number
IV. IMPLEMENTATION OF ML-BASED ES-rApp
of users connected to the cell (RRC.ConnMean),
In the previous section, the general information flow and RL
and the number of users associated with each beam
framework were introduced to provide an overview of the
(L1M.SSBBeamRelatedUeNbr.<beam index>) reflect-
environment where the ES-rApp is deployed. In this section,
ing their spatial distribution (in case of reduced number
we propose an implementation of an ES-rApp aiming at the
of beams caused by RCR values are zero-padded).
maximization of EE through dynamic RCR, according to the
In detail, s = (oneHotRCR, RRU.PrbTotSdmDl,
input/output data, and architecture requirements defined in
PEE.AvgPower, DRB.UEThpDl, RRC.ConnMean,
the O-RAN ALLIANCE use-case [11].
L1M.SSBBeamRelatedUeNbr.<min beam index>, · · ·
L1M.SSBBeamRelatedUeNbr.<max beam index> )
A. DESIGN OF THE ES-rApp
• A Reward is defined as EE, being the ratio between
As per Fig. 2, ES-rApp is deployed in Non-RT RIC. Its
average DL user throughput, and average power con-
main goal is to perform an O-RU antenna array selection
sumption per carrier measured over an observation
through the cascade of O1 and O-FH interfaces to max-
period, i.e., r = DRB.UEThpDl
PEE.AvgPower . The KPIs related to the
imize EE. The algorithm aims to adjust the number of
computation of the reward are obtained through O1 and
ARCs, taking into account the dependencies between O-
are compliant with [15].
RU configuration, power consumption, user throughput, PRB
utilization, and distribution of users over beams, provided The above definition of the RL framework components is
to ES-rApp through the O1 interface. This approach allows based on the KPIs defined in 3GPP specifications, that are
the reconfiguration of O-RU to enable both high ES during available in an O-RAN M-MIMO network through the O1
low-load hours and preserve users’ QoS during peak-load interface. This allows ES-rApp to be independent of any
hours. The decisions of RCR may result in modifications specific data models, e.g., power consumption. On the other
of the maximum number of spatial streams, or the O-RU hand, the proper KPIs can be generated through computer
antenna Tx power, thus they should be made within the simulations to enable evaluation of the proposed ES-rApp.
non-RT control loop. The ES-rApp fits in the RL framework To implement an RL agent inside ES-rApp, we utilize
discussed in Sec. III-A, i.e., it performs the following steps Deep Q-Network (DQN), which is adequate for dealing with
periodically: observes multiple KPIs (state s), based on discrete and finite action space. Therefore, DQN fits well

VOLUME 12, 2024 118497

 M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

for selecting one of several RCCs by ES-rApp. It estimates TABLE 1. Simulation parameters.
the expected reward (EE) associated with each action (i.e.,
possible RCCs) in the observed state (see Fig. 4). Usually,
consecutive states in RL affect each other. However, in the
case of RCR, ES-rApp can select any possible configuration
independently of the previous decisions, i.e., only based on
observed KPIs (state). Thus, it is enough for ES-rApp to be
focused on the maximization of EE separately in each state.
This approach is known as myopic DQL. DQN is built of an
input layer, two hidden layers and an output layer. The input
layer takes the features according to the state definition, i.e.,
there are 7 + <number of beams> inputs. Each hidden layer
is the so-called dense layer built of 256 neurons (the output
of a single neuron from the former layer is provided as input V. SIMULATION RESULTS
to every neuron in the following layer) and is followed by the The proposed ES-rApp was verified in a python-based sim-
Rectified Linear Unit. The output layer has a size equal to the ulation environment developed by the authors, considering
number of elements in the set representing the action space. DL in an Urban Macrocell (UMa) of an O-RAN network
In this case, |A| = 3), and maps directly on EE expected to be according to the scenario definition and BS configuration
achieved while selecting a particular action a in a given state in [3] and [21]. The UMa has a radius of 500 m and operates
s. Based on [20] the computational complexity of the utilized in the n78 band, utilizing 100 MHz bandwidth split into 273
dense layer is O(n · m), where m is the layer’s input size, and RBs, at the carrier frequency of 3.5 GHz. BS is installed
n is the layer’s output size. In order to ensure stable network at a height of 25 m and transmits with a total Tx power
operation, the RCR is expected to be done by rApp within the of 55 dBm, it utilizes the GoB scheme, i.e., a fixed set of
timescale of minutes. This enables rApp to control multiple up to 32 predefined beams. The user is associated with a
cells. However, evaluation of such scenarios is considered for beam that provides the highest Rx power. The beam’s Rx
future work. power is calculated based on the UMa path-loss model and
array radiation pattern from the 3GPP specification [21].
In addition, a spatially correlated shadowing model is used
B. CHALLENGES
independently for each beam, which models the presence of
During the design of the ES-rApp, several challenges need to
obstacles [22]). Within a single beam, radio resources are
be taken into account:
split equally between users, with a limit of 30 RBs per user.
We assume that users do not have specified QoS requirements
1) SWITCHING OFF RF CHANNELS
for radio resources. The number of ARCs can be scaled down
changes the properties of the antenna array that enforces to 8 × 4 (32 elements), or 8 × 2 (16 elements), which supports
whole O-RU reconfiguration, e.g., the number of supported 16, and 8 beams respectively. According to the initial studies,
beams, or the number of possible layers. in [11], the 64 antenna O-RU during full load consumes the
power of 1120 W. This value changes with the traffic load and
2) THE RL ALGORITHM can be scaled down by reducing the number of ARCs. Based
must be supported with mechanisms that prevent testing on the 3GPP power consumption model described in [3] the
of obviously wrong solutions, e.g., during busy hours, the power consumption for such a 64 antenna O-RU is given by:
configuration with 20% of ARCs should never be tested.
Pdl [W] = 630 · sa + 360 · sa · sf + 220 + Ptx , (1)
3) THE rApp
must strictly cooperate with other xApps, e.g., with BMM- where sa is the fraction of active antennas, sf is the cell
xApp. Changing the O-RU configuration usually requires its load, and Ptx is the Tx power. Simulation parameters are
restart, and the users should be temporarily offloaded to other summarized in Table 1. The simulations are arranged in so-
beams to ensure QoS fulfillment. called episodes. An episode corresponds to one day and
consists of 144 steps, i.e., the decision on RCR is made by
4) SWITCHING OFF RF CHANNELS the ES-rApp once every 10 minutes. Within a single episode
results in fewer beams, thus the overall Tx power is the number of users in the cell changes according to the
decreased, which affects the coverage. Thus, the power of sinusoidal traffic profile, from 5 at midnight to 150 at noon.
the remaining beams needs adjustment to compensate. One At every step, a number of users are randomly placed over
approach is to establish a constant total power for the whole the cell area, to reflect the real network where the spatial
antenna array, but in-depth studies are needed related to the placement of users in a cell varies between days. Next the
power amplifiers’ operation points and potential non-linear ES-rApp selects the RCC and observes the reward. Finally,
distortion. a state-action-reward tuple is put into the database.

118498 VOLUME 12, 2024

M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

FIGURE 4. Input and output features of DQN utilized by the ES-rApp.

FIGURE 6. Energy Efficiency averaged over 50 days.

FIGURE 5. Offline training loss for ES-rApp, both raw and processed
through a 20-point moving average.

a comparison between EE averaged over 50 days achieved

First, we captured the training data by exploring different by each of the evaluated algorithms. All algorithms provide
RCCs over 15 days (episodes), which resulted in 2160 train- EE gains over Full-Array. A simple PRB threshold-based
ing samples. We used them to train the DQL agent offline, Ref-Load provides only a 10.6% EE gain. Thus, it is more
using batch size equal to 4, and Mean Absolute Error loss beneficial to utilize ML-based methods that achieve 19.3%
function. The training loss (both raw, and smothen with a 20- and 24.8% EE gains for Ref-ML and ES-rApp respectively.
point moving average) is depicted in Fig. 5. It can be seen that Here the set of input features plays a crucial role. The Ref-
after 20000 training steps (iterations of stochastic gradient ML utilizes precoder weights as one of those. However,
descent), the mean absolute error of EE prediction stabilizes under GoB they are constant and do not contribute much
at the level of about 1 kbit/J. to the model accuracy. On the other hand, the proposed
After the training phase, we compared the proposed ES-rApp utilizes a set of input features that is proper for GoB
ES-rApp against two state-of-the-art algorithms: Ref-Load beamforming, e.g., the number of users per beam. As a result,
algorithm [14], which scales down the number of ARCs when it can improve EE by 24.8% compared to the Full-Array and
PRB utilization is below 10%, and Ref-ML algorithm [13] performs close to the upper bound defined by the Ex-Search,
selecting RCC based on the average channel gain and i.e., the absolute difference in EE is only about 50 bit/J.
precoder weights. As Ref-ML originally has a constraint To better understand the resultant average EE, in Fig. 7,
on the minimum user throughput, for a fair comparison, we compared its components: the power consumption and
we set its goal to maximize EE, which already cares about average user throughput achieved by the tested algorithms
users’ QoS. In addition, we have evaluated a Full-Array over the day. Ref-ML is characterized by the lowest power
scenario with 64 ARCs, serving as a reference level for EE consumption of up to about 1.1 kW, compared to almost
gains, and Exhaustive Search Ex-Search that evaluates each 1.4 kW consumed by Full-Array. However, it comes at
possible RCC under identical placement of users, to select the cost of poorer throughput during peak load hours.
one providing the highest EE. The Ex-Search stands for On the other hand, the Ref-Load algorithm preserves users’
the upper bound on EE gains. Before capturing the results, throughput but acts cautiously, and does not fully utilize the
we trained both ES-rApp and Ref-ML with the use of the data opportunity to increase EE. The proposed ES-rApp provides
captured during a 15-day-long simulation. Fig. 6 presents both high ES during low-load hours and high throughput

VOLUME 12, 2024 118499

 M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

FIGURE 7. Traffic profile, power consumption (upper subplot), and average user throughput
(lower subplot) achieved over the daytime. Results are averaged over 50 days.

the proposed ES-rApp achieves EE gains of 8% over the

Full-Array also under busy hour load. Only under the medium
load, EE obtained by the ES-rApp is lower than the one
associated with an Ex-Search, i.e., 20% EE gain over the Full
Array for ES-rApp versus 22% for the Ex-Search.

VI. CONCLUSION
The O-RAN architecture acts as a key enabler for the
implementation of ES features in 5G and beyond networks
by providing a unified environment for the deployment of
ES-oriented rApps and xApps. The environment includes
FIGURE 8. Relation between cell load and Energy Efficiency, based on the
ETSI traffic profile [23]. Results are averaged over 50 days. interfaces for data CC, and ML features for model train-
ing, deployment, and inference. The O-RAN specifications
provide a detailed description of ES use cases, includ-
ing message flow between engaged entities and an ML
during peak load. Moreover, its RCR decisions are almost the framework. Based on all these, the authors have proposed
same as the optimal ones derived by Ex-Search. Compared to an ES-rApp that performs RCR, using DQL to maximize
Ref-Load, it is due to the utilization of ML, while compared EE. Due to the utilization of RL, ES-rApp can follow the
to Ref-ML, due to the better selection of input features. hidden dependencies between O-RU configuration, power
While, in Fig. 7 the network traffic is represented by the consumption, throughput, PRB utilization, and distribution of
number of users it is beneficial to investigate the explicit users over beams to intelligently scale up/down the number of
relation between obtained EE and the cell load. In Fig. 8 ARCs. Such an intelligent reconfiguration of O-RU enables
there is a relation between achieved EE, and cell load both high ES during low-load hours and preserving users’
for the proposed ES-rApp and reference algorithms. There QoS during peak-load hours. The simulation results show
are 3 load scenarios considered, according to the ETSI up to 24.8% improvement in achieved EE while utilizing
traffic profile [23]: low load, medium load, and busy hour the proposed ES-rApp compared to the Full-Array. It also
that corresponds to the 10%, 30% and 50% of cell load outperforms state-of-the-art algorithms: Ref-Load [14], and
respectively. The load is calculated for the Full-Array case Ref-ML [13]. The future work aims at addressing aspects
in terms of mean PRB utilization over all M-MIMO layers like scaling with a larger simulation environment (dozens
within the whole cell (RRU.PrbTotSdmDl [15]). The results of BSs); integration of the algorithm within a realistic setup
confirm the previous observations. The Ref-Load provides using a commercial-grade RIC and RAN emulator; varying
EE gains over Full-Array only under low load scenarios users’ QoS profiles; or utilizing real-life traffic patterns and
(about 36%) due to its cautious behavior. The Ref-ML traffic demands. The other topic is to analyze multiple Apps
provides EE gains over Full-Array compared to the proposed working together to achieve energy saving in harmony within
ES-rApp under low and medium load (about 65% and the hierarchy - i.e., a joint rApp/xApp operation. Finally,
20% respectively), while under busy hour tries to reduce as the RIC framework should support various Apps from
the number of active antennas too much resulting in about various vendors that achieve different goals, the authors see
15% EE degradation compared to the Full-Array. Finally, it as important to concern conflict mitigation.

118500 VOLUME 12, 2024

M. Hoffmann, M. Dryjański: Energy Efficiency in Open RAN: RF Channel Reconfiguration Use Case

ACKNOWLEDGMENT [18] M. Dryjanski and M. Szydelko, ‘‘A unified traffic steering framework for
The research work related to Algorithm Design was founded LTE radio access network coordination,’’ IEEE Commun. Mag., vol. 54,
no. 7, pp. 84–92, Jul. 2016.
by Rimedo Laboratories. [19] C. Adamczyk and A. Kliks, ‘‘Conflict mitigation framework and conflict
detection in O-RAN near-RT RIC,’’ IEEE Commun. Mag., vol. 61, no. 12,
REFERENCES pp. 199–205, Dec. 2023, doi: 10.1109/MCOM.018.2200752.
[20] B. Shah and H. Bhavsar, ‘‘Time complexity in deep learning models,’’
[1] S. Han, T. Xie, and I. Chih-Lin, ‘‘Greener physical layer technologies
Proc. Comput. Sci., vol. 215, pp. 202–210, Dec. 2022.
for 6G mobile communications,’’ IEEE Commun. Mag., vol. 59, no. 4,
[21] Study on Channel Model for Frequencies From 0.5 to 100 GHz (Release
pp. 68–74, Apr. 2021.
17), Standard TS 38.901 v.17.0.0, 3GPP, Mar. 2022.
[2] D. López-Pérez, A. De Domenico, N. Piovesan, G. Xinli, H. Bao, S. Qitao,
[22] C. Zhang, X. Chen, H. Yin, and G. Wei, ‘‘Two-dimensional shadow fading
and M. Debbah, ‘‘A survey on 5G radio access network energy efficiency:
modeling on system level,’’ in Proc. IEEE 23rd Int. Symp. Pers., Indoor
Massive MIMO, lean carrier design, sleep modes, and machine learning,’’
Mobile Radio Commun. (PIMRC), Sep. 2012, pp. 1671–1676.
IEEE Commun. Surveys Tuts., vol. 24, no. 1, pp. 653–697, 1st Quart., 2022.
[23] Environmental Engineering (EE); Metrics and Measurement Method for
[3] Study on Network Energy Savings for NR (Release 18), Standard TR 38.864
Energy Efficiency of Wireless Access Network Equipment; Part 1: Power
v.18.1.0, 3GPP, Mar. 2023.
Consumption—Static Measurement Method, Standard ETSI, ES 202 706-
[4] A. Kaushik, R. Singh, S. Dayarathna, R. Senanayake, M. Di Renzo,
1, v1.7.1, ETSI, Aug. 2022.
M. Dajer, H. Ji, Y. Kim, V. Sciancalepore, A. Zappone, and W. Shin,
‘‘Toward integrated sensing and communications for 6G: Key enabling
technologies, standardization, and challenges,’’ IEEE Commun. Standards
Mag., vol. 8, no. 2, pp. 52–59, Jun. 2024.
[5] A. Bazzi and M. Chafii, ‘‘On outage-based beamforming design for
dual-functional radar-communication 6G systems,’’ IEEE Trans. Wireless
Commun., vol. 22, no. 8, pp. 5598–5612, Jun. 2023.
[6] L. Bonati, S. D’Oro, M. Polese, S. Basagni, and T. Melodia, ‘‘Intelligence
and learning in O-RAN for data-driven NextG cellular networks,’’ IEEE MARCIN HOFFMANN (Member, IEEE) received
Commun. Mag., vol. 59, no. 10, pp. 21–27, Oct. 2021. the M.Sc. degree (Hons.) in electronics and
[7] ‘‘Open RAN technical priorities, focus on energy efficiency,’’ Deutsche telecommunication from Pozna University of
Telekom, Orange, Telefonica, TIM, Vodafone, Tech. Document Technology, in 2019, where he is currently
O-RAN MoU, Jun. 2021. [Online]. Available: https://cdn.prod.
pursuing the Ph.D. degree with the Institute
website-files.com/60b1962ffda0a42f779c765b/60cd8e6f8dd92c298b891
of Radiocommunications. He is currently a
216_Open%2BRAN%2BTechnical%2BPriority%2BDocument%2B-
%2BFinal%2BVersion.xlsx Senior Research and Development Engineer with
[8] M. Hoffmann, S. Janji, A. Samorzewski, Ł. Kułacz, C. Adamczyk, the Rimedo Laboratories working on O-RAN
M. Dryjański, P. Kryszkiewicz, A. Kliks, and H. Bogucka, ‘‘Open RAN software development solutions and spectrum
xApps design and evaluation: Lessons learnt and identified challenges,’’ sharing-related projects. He is gaining scientific
IEEE J. Sel. Areas Commun., vol. 42, no. 2, pp. 473–486, Feb. 2024. experience by being involved in both national and international research
[9] J. Thaliath, S. Niknam, S. Singh, R. Banerji, N. Saxena, H. S. Dhillon, projects. In addition to that, he works on massive MIMO and advanced
J. H. Reed, A. K. Bashir, A. Bhat, and A. Roy, ‘‘Predictive closed-loop beamforming techniques. His scientific articles are published in top
service automation in O-RAN based network slicing,’’ IEEE Commun. journals like IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, IEEE
Standards Mag., vol. 6, no. 3, pp. 8–14, Sep. 2022. TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, and IEEE ACCESS.
[10] A. Giannopoulos, S. Spantideas, N. Kapsalis, P. Gkonis, L. Sarakis, His research interests include the utilization of machine learning and
C. Capsalis, M. Vecchio, and P. Trakadas, ‘‘Supporting intelligence in location-dependent information for the purpose of network management.
disaggregated open radio access networks: Architectural principles, AI/ML
workflow, and use cases,’’ IEEE Access, vol. 10, pp. 39580–39595, 2022.
[11] O-RAN Network Energy Savings Use Cases Technical Report 1.0,
document TR O-RAN ALLIANCE WG1, Mar. 2023.
[12] B. Lin, F. Gao, S. Zhang, T. Zhou, and A. Alkhateeb, ‘‘Deep learning-based
antenna selection and CSI extrapolation in massive MIMO systems,’’ IEEE
Trans. Wireless Commun., vol. 20, no. 11, pp. 7669–7681, Nov. 2021. MARCIN DRYJAŃSKI (Senior Member, IEEE)
[13] N. Rajapaksha, J. Mohammadi, S. Wesemann, T. Wild, and N. Rajatheva, received the Ph.D. degree (Hons.) from Poznan
‘‘Minimizing energy consumption in MU-MIMO via antenna muting by University of Technology, in September 2019.
neural networks with asymmetric loss,’’ IEEE Trans. Veh. Technol., vol. 73, Over the past 15 years, he was a Research and
no. 5, pp. 6600–6613, May 2024. Development Engineer, a Consultant, a Tech-
[14] P. Frenger and K. W. Helmersson, ‘‘Massive MIMO muting using dual-
nical Trainer, the Technical Leader, a Board
polarized and array-size invariant beamforming,’’ in Proc. IEEE 93rd Veh.
Member, and an Advisor. He has been involved
Technol. Conf. (VTC-Spring), Apr. 2021, pp. 1–6.
[15] Management and Orchestration; 5G Performance Measurements, Stan-
in 5G design, since 2012, where he was the
dard 3GPP, TS 28.552 v.18.2.0, Mar. 2023. Work-Package Leader of the FP7 5GNOW Project.
[16] M. Hoffmann and P. Kryszkiewicz, ‘‘Beam management driven by radio Currently, he is the CEO and a Principal Consul-
environment maps in O-RAN architecture,’’ in Proc. IEEE Int. Conf. tant with the Rimedo Laboratories. He is the co-author of many articles on
Commun. Workshops (ICC Workshops), Rome, Italy, 2023, pp. 54–59, doi: 5G, LTE-Advanced Pro and Open RAN and the co-author of the book From
10.1109/ICCWorkshops57953.2023.10283527. LTE to LTE-Advanced Pro and 5G (M. Rahnema, M. Dryjanski, Artech
[17] O-RAN Architecture Description, v.8.0, document O-RAN ALLIANCE, House, 2017).
TS, O-RAN ALLIANCE WG1, Mar. 2023.

VOLUME 12, 2024 118501