Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Alexandria Engineering Journal (2023) 72, 621–634

H O S T E D BY
Alexandria University

Alexandria Engineering Journal


www.elsevier.com/locate/aej
www.sciencedirect.com

ORIGINAL ARTICLE

Adaptive cell selection algorithm for balancing cell


loads in 5G heterogeneous networks
Emre Gures a,*, Ibraheem Shayea a,*, Muntasir Sheikh b, Mustafa Ergen a,
Ayman A. El-Saleh c

a
Department of Electronics and Communication Engineering, Faculty of Electrical and Electronics Engineering, Istanbul
Technical University (ITU), 34467 Istanbul, Turkey
b
Electrical and Computer Engineering Department, College of Engineering, King Abdulaziz University, Jeddah, Saudi Arabia
c
Department of Electronics and Communication Engineering, College of Engineering, A’Sharqiyah University (ASU), Ibra 400,
Oman

Received 9 January 2023; revised 20 March 2023; accepted 7 April 2023


Available online 25 April 2023

KEYWORDS Abstract Heterogeneous networks (HetNets) are a promising solution for managing the exponen-
Load balancing; tial increase in the number of mobile users while maintaining high data rates and coverage. HetNets
Mobility management; consist of various cell types with different cell coverage and system capacities. However, the traffic
Heterogeneous networks; load in HetNets is variable, uneven and random over time, leading to unequal cell loads. Some cells
5G; have excessive user presence where the competition for system resources is high, while other cells
Millimetre-wave; have low user presence where system resources are not fully utilised. This paper proposes a mobility
Handover load balancing algorithm that prioritises millimetre-wave (mmWave) cells in target cell selection
and user association to provide load balancing in 5G HetNets and improve overall system perfor-
mance. A two-step target cell selection method that considers the load level of cells and reference
signal received power (RSRP) is proposed. This method prioritises mmWave cells that meet cell
load level and RSRP conditions in target cell selection, taking full advantage of the unique prop-
erties of mmWave cells (enhanced network throughput, spectral efficiency and enormous band-
width) to balance traffic loads. The HO triggering and decision-making process is independently
performed for each user. In case the serving cell is overloaded, different HO procedures are applied
for load balancing depending on the target cell type (macro base station (BS) or mmWave BS). The
scope of the study is further expanded by applying different HO procedures according to the serving
and target cell types to maintain mobility robustness in case the serving cell is not overloaded. This
research proposes various system scenarios with fixed HO margin (HOM) values and fixed time-to-
trigger (TTT) durations to examine the effects of HCP settings on the proposed algorithm’s perfor-
mance in terms of average load level of the serving cell, throughput, spectral efficiency and call
dropping ratio (CDR). The system that provides the highest overall performance is applied to
the proposed algorithm. To further assess and validate the performance of the proposed algorithm,
it is compared with other load balancing algorithms from the literature. The simulation results

* Corresponding authors.
E-mail addresses: gures.emre@gmail.com (E. Gures), shayea@itu.edu.tr (I. Shayea).
Peer review under responsibility of Faculty of Engineering, Alexandria University.
https://doi.org/10.1016/j.aej.2023.04.012
1110-0168 Ó 2023 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
622 E. Gures et al.

reveal that the proposed algorithm does provide noticeable enhancements in network performance
in terms of load level, throughput, spectral efficiency and CDR for various mobile speed scenarios
as compared to existing load balancing algorithms.
Ó 2023 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria
University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction balancing purposes, mmWave BSs will be more popular in


future cellular networks since they significantly increase net-
HetNets are a remarkable solution for managing the explosive work throughput, the available bandwidth and spectral
growth of mobile data traffic. HetNets are based on the inte- efficiency.
gration of multiple radio access technologies, small cells and Handover (HO) is the procedure that changes the serving
distributed antenna systems. This integration provides users cell of a mobile user equipment (UE) in radio resource control
with enhanced quality of service (QoS) in terms of data rate (RRC) CONNECTED mode [18]. HOM and TTT are two
and coverage [1]. Although this solution optimises the perfor- HCPs that have a significant impact on HO performance
mance of network operators and subscribers in fifth-generation and must be carefully adjusted to ensure efficient HO proce-
(5G) HetNets, numerous challenges are present [2–5]. HetNets dures [19]. Since the movement speed of the UE is an impor-
consist of small cells placed in macro cell coverage areas that tant factor that influences the optimum TTT duration and
improve network coverage and network capacity at irregular HOM value, it should be considered when setting HCPs.
user traffic distribution areas or hotspots. Small cells (such The TTT parameter is defined as the time interval during
as picocells, femtocells, microcells and relays) are the cellular which the RSRP value measured throughout the HO proce-
coverage offered by low-power BSs compared to macro-BSs. dure must satisfy certain signal strength criteria for the HO
They are the key element for offloading mobile traffic from decision to be made. A low TTT setting allows the traffic of
macro cells. Due to the transmit power difference between cell serving cells to be offloaded to neighbouring cells in a shorter
types, users tend to associate with macro cells that provide the amount of time in case of load imbalance within the network.
strongest downlink signal. As a result, the number of user clus- In deployment scenarios where the mmWave is used, dynami-
ters associated with small cells is significantly lower than that cally reducing TTT in cases where the target cell is a small BS
of macro cells. Load imbalance in HetNets leads to unfair data will enable users to associate with these cells in a shorter time.
rate distribution among users as well as inconsistency in the The features of mmWave can achieve performance enhance-
quality of experience (QoE) and quality of service (QoS) [6]. ments in throughput and spectral efficiency. However, low
Load balancing mechanisms are needed to ensure the efficient TTT settings can cause an increase in HOP and HOPP, leading
use of system resources. These mechanisms should be designed to the undesired signalling overhead between BSs and UEs. A
to fairly distribute the traffic load among BSs while providing high TTT setting can alternatively lower the HOP and HOPP
satisfactory QoS for all users, especially mobile devices at the probability. However, these settings will delay corrective (load
cell edge. balancing) actions in cases of load imbalance and may increase
The introduction of new radio (NR) bands will have a sub- RLF.
stantial impact on communication performance in 5G systems HOM is the margin value that controls the HO decision.
and future HetNets [7,8]. The 5G NR can operate in low and For instance, when the serving cell is overloaded, instead of
high frequency bands ranging from 4.5 GHz to 6 GHz (sub- increasing the transmit power of the target neighbouring cell,
6 GHz) and 24.25 to 52.6 GHz (mmWave) [9–11]. MmWave the HOM value is added to the RSRP of the target cell. Thus,
provides high data rates due to enormous bandwidth and the target cell is made more attractive in user association.
low-order modulation. However, mmWave connections are Users in the overloaded cell can be offloaded to neighbouring
very sensitive to rapid channel changes and suffer from high cells even if the RSRP value provided to the UEs from the
path loss, severe channel interruptions (due to increased block- serving cell is higher. UEs in an overloaded cell can also be off-
ing and shadowing effects) and obstacles caused by various loaded to neighbouring cells if the received signal strength
structures, such as buildings [12,13]. To overcome these chal- (RSS) from the serving cell is higher than that of neighbouring
lenges and provide users with an acceptable service quality cells. However, care should be taken in determining the HOM
level, mmWave BSs are the solution for ultra-dense deploy- values. A low HOM value may prevent users in overloaded
ments in future 5G networks. However, smaller footprints cells from offloading to neighbouring target cells. This causes
and the ultra-dense deployment of mmWave BSs will signifi- the load level of serving cells to remain high, and their
cantly increase HO probability (HOP) and lead to an upsurge resources cannot be efficiently used. This may decrease both
in the number of mobility issues, such as high radio link failure throughput and spectral efficiency. Conversely, if the HOM
(RLF) and HO ping-pong (HOPP) [14]. This makes HO man- value is set too high, more users may be associated with the
agement increasingly challenging since it will result in higher neighbouring target cell than it can serve. This leads to reduced
signal overhead with lower QoS and QoE [15]. When deter- throughput and spectral efficiency of users associated with the
mining the optimal BS for HO, it is necessary to seek solutions target cell. This setting can further increase the HOP and
that consider the maximum instantaneous RSRP or signal to HOPP probabilities. Therefore, an efficient solution will
interference plus noise ratio (SINR), the network behaviour require user-specific HCP settings.
as well as all relevant information, such as channel conditions, The key contributions of this paper can be summarised as
load levels of BSs, bandwidth and throughput [16,17]. For load follows:
Adaptive cell selection algorithm for balancing cell loads 623

1. A two-step target cell selection method that considers our knowledge, no research has examined the effect of dif-
resource availability of cells and RSRP is proposed. This ferent HCP settings on load balancing optimisation. There-
method prioritises mmWave cells that satisfy cell load level fore, this study fills a gap in the literature.
and RSRP conditions in target cell selection, taking full 4. The performance of the proposed algorithm is analysed,
advantage of the unique properties of mmWave cells (en- compared and validated with other load balancing algo-
hanced network throughput, spectral efficiency and enor- rithms selected from the literature. The simulation results
mous bandwidth) to provide load balance. In the first demonstrate that the proposed algorithm outperforms
step, a restricted NCL is generated from cells with load other algorithms in terms of load balancing, throughput,
levels below the threshold load level. In the second step, spectral efficiency and CDR.
if the mmWave cell in NCL has an RSRP value higher than
the threshold RSRP value, this cell is selected as the target This paper is organised as follows: Section 2 reviews the
cell; otherwise, the macro cell with the highest RSRP in current research on load balancing in HetNets. Section 3 high-
NCL is selected as the target cell. Overall system perfor- lights the proposed load balancing algorithm. Section 4 pre-
mance is enhanced by increasing the adaptability and sents the system and simulation models. Section 5 briefly
usability of mmWave cells. By including the load level of introduces the load balancing performance metrics. Section 6
cells in target cell selection, load balance can be achieved provides the simulation results, and Section 7 concludes this
in the network. paper.
2. Different HO procedures are proposed for load balancing
depending on the target cell type (mmWave and macro) 2. Background and related works
in case the serving cell is overloaded. Applying conven-
tional HO procedures to the proposed algorithm is ineffi- This section presents the research background and related
cient due to the following: a) the target cell selection works that mainly focus on cell load balance in 5G HetNets.
method is implemented by waiving max-RSRP to increase The first subsection discusses the research background, while
adaptability and usability of mmWave cells in 5G HetNets, the second subsection provides a brief summary of several
b) the transmit power is disproportion between mmWave related works selected from the literature.
and macro BSs and c) mmWave channels are unpredictable
(rapid channel changes, high path loss and severe channel 2.1. Research background
interruptions). The scope of this study is further expanded
by applying different HO procedures according to the serv-
ing and target cell types in order to maintain mobility HetNets, formed by the integration of mmWave BSs with tra-
robustness in cases where the serving cell is not overloaded. ditional sub-6 GHz macro-BSs, have emerged as a promising
The proposed model prioritises mmWave cells that satisfy solution for increasing user capacity and data rates. HetNets
the load level and RSRP conditions in user association. It are considered to be one of the fundamental architectures of
ensures that users remain associated with these cells as long 5G systems (see Fig. 1). However, the inherent propagation
as conditions are satisfied, otherwise, it associates users characteristics of mmWave frequencies may present new chal-
with the best performing macro BS. lenges such as higher path loss, atmospheric absorption and
3. Various system scenarios with fixed HOM values and fixed rain attenuation. Such challenges increase the outage probabil-
TTT durations are suggested to investigate the effects of ity, thereby reducing overall system performance. When reli-
several HCP settings on the performance of the proposed able mmWave communication is unlikely, mmWave BSs
algorithm. This study facilitates a better understanding of must coexist with conventional sub-6 GHz macro-BSs to pro-
the effect of HCP on load balancing performance in 5G vide users with seamless connection [20].
HetNets. Several studies have examined the effect of differ- HetNets present new challenges for cellular resource man-
ent HCP settings on mobility robustness optimisation agement. Traditional max-SINR or max-RSRP user associa-
throughout the literature [14,19]. However, to the best of tion algorithms cause serious load imbalance problems in

Fig. 1 Loads are balanced between cells according to reSource availability. Cell edge UEs in high-traffic macro cells are offloaded into
small low-loaded mmWave cells.
624 E. Gures et al.

HetNets since users will tend to associate with macro BSs that explores the effect of biases in terms of cell load and UE
have the highest signal strength [3]. The competition for system throughput to determine the optimal biases that maximise
resources in macro-BSs is therefore high due to the excessive SINR coverage or throughput.
number of associated users, while the system resources of In [25], user association was formulated as a local optimisa-
mmWave BSs are not fully utilised due to the limited number tion problem in macro BS. A limited CSI feedback user asso-
of associated users. Load imbalance between cell types leads to ciation scheme was proposed to develop a low-complexity
inequitable data rate distribution among users within the net- successive offloading scheme. However, user association was
work as well as inconsistencies in QoE and QoS. Cell associa- examined in a static environment without considering user
tion schemes that are cognizant of both mmWave and sub- mobility.
6 GHz systems are needed for balancing loads and improving The popularity of load balancing models that involve
system performance in HetNets. The full utilisation of system machine learning (ML) has been steadily rising in recent years.
resources is required for enhancing user experience. The vast data source that exists from mobile networks can help
in the efficient management of HetNets, allowing more
informed decisions. Meaningful information can be learned
2.2. Related works using ML approaches and large datasets. Related tasks can
be automatically optimised without human intervention. ML
Several load balancing methods in HetNets are present algorithms are basically divided into three main categories:
throughout the literature. supervised learning, unsupervised learning and reinforcement
In [21], a cell selection algorithm was proposed for balanc- learning (RL). In supervised learning, with the help of a learn-
ing loads in HetNets. The suggested method consists of two ing agent, sample inputs form the labelled dataset and its
SINR thresholds used to guarantee the throughput require- desired outputs. A general rule is learned to map the inputs
ment of users. When the SINR received is below the first SINR to outputs. In unsupervised learning, since the expected output
threshold, a candidate list is prepared based on the RSS values is unknown, the learning agent tries to extract useful informa-
of candidate femtocells. The user then updates the candidate tion from the training data by analysing the data structure.
list by extracting several femtocells according to the predicted Unlike other ML techniques, the agent in RL discovers which
RSS values to create a second list. The target femtocell is deter- actions to take by constantly interacting with an environment
mined according to the number of free resource blocks and and attempting the action that will maximise a reward.
pilot signal power. The physical resource blocks (PRB) with In [26], an online RL-based load balancing algorithm was
higher SINRs are allocated to the user. The algorithm forces proposed to exploit the spatio-temporal regularity of data traf-
users to associate with the macro cell if the femtocell is fic in vehicle networks due to the seasonality of urban traffic
unavailable. The simulation results indicate that the proposed flow. The proposed algorithm consists of two stages: initial
algorithm outperforms the conventional Max-SINR method in RL and historically based RL. The initial RL method is used
terms of user throughput. However, the proposed algorithm to exploit spatio-temporal regularities in vehicle networks.
does not perform well in high mobility scenarios. The past relationships provided by the initial RL are consid-
In [22], a new parameter (the load balancing efficiency fac- ered as reference. The historical association models use the
tor) was suggested along with a utility-based load balancing similarities between the past state and future traffic flows. As
algorithm that simultaneously considers operator utility and the current network continues to change, the BS employs a
user utility for each under-loaded neighbouring cell. The load greedy method to learn association actions from historical
balancing efficiency factor orders overloaded cells based on the association models.
remaining capacity of neighbouring cells and the estimated Ref. [27] proposed an algorithm that learns the optimum
load of edge-UEs in neighbouring cells. The centralised con- HOM between the source cell and all neighbouring cells from
troller periodically collects load information of small cells previous experience using the Q-learning algorithm. Deep
and uses it to optimise HCPs. Each neighbouring cell’s aggre- learning (DL) automatically infers distinctive features that
gated utility is calculated, and load balance is accomplished by help make sense of data through multiple layers that represent
transferring each edge-UE candidate from the overloaded cell data abstractions [28]. The DL approach learns to represent
to the best neighbouring cell (with the highest utility). data through a nested hierarchy of concepts, providing
In [23], a cluster-based load balancing algorithm for Het- tremendous power and flexibility in complex real-world sce-
Nets was proposed. A central controller collects information narios [29]. The curse of dimensionality, caused by an expo-
of edge-UEs and their best neighbouring cells to form clusters. nential increase in the number of state-action pair values
Through this information, the network is modelled as a routed that must be stored in RL, makes model adaptation to com-
multigraph. The algorithm calculates an adaptive threshold plex network scenarios exceedingly difficult [30]. Deep rein-
value from the load information of cells in the graph and forcement learning (DRL), which is a combination of DL
detects overloaded cells that exceed this threshold value. Clus- and RL, is anticipated to become more popular in performing
ters are dynamically created from suitable neighbouring cells adaptive and autonomous large-scale load balancing.
selected from an overloaded cell and its neighbours. To bal- Ref. [31] presented a DRL-based load balancing algorithm
ance loads, the HOMs of cells are adjusted to the neighbouring that adjusts the HOM values of BSs. It is assumed that the net-
cells in the cluster and the UEs are offloaded from the over- work is equipped with a central agent capable of monitoring
loaded cells to these cells. network performance indicators and controlling the HOM val-
In [24], a two-step load balancing mechanism based on two ues of cells. The action is the HOM values of BSs, while the
biases was proposed to adjust the selection of layer and radio state is the input layer of the neural network which consists
access technologies (RAT) in multilayer HetNets of LTE-A of performance indicators. Three hidden layers are added to
macro cells and mmWave small cells. The SINR distribution determine the dependencies of actions on states, and each layer
Adaptive cell selection algorithm for balancing cell loads 625

has a corrected linear ReLU activation function. The simula- appears primitive and ineffective for addressing the unique fea-
tion results revealed an increase in overall throughput as com- tures of 5G networks due to the disproportionate transmit
pared to the model without the application of HOM values. power of cells in 5G HetNets (the transmit powers of LTE-A
Drone BSs, as opposed to terrestrial BSs, offer numerous and 5G cells in this study are 46 dBm and 30 dBm, respec-
advantages in wireless communication networks due to their tively). This decreases the overall system performance by
deployment flexibility and mobility [32]. Drone BSs provide reducing the adaptability and usability of mmWave cells. More
extra capacity in hotspot areas where traffic rapidly increases, specifically, making target cell selection decisions that are
thereby reducing heavy traffic loads and efficiently using sys- solely based on the RSRP (or SINR) criterion can lead to
tem resources. However, drone BSs have limited battery life, unbalanced network loads, inefficient HOs and service
giving them limited airtime. The mobile nature of drone BSs interruptions.
also causes fluctuations in the channel quality between these This study proposes a two-step target cell selection that
BSs and users, which may result in HO failures and/or packet considers the load level of cells and RSRP and prioritises
loss [33]. Therefore, efficient HO mechanisms are needed in mmWave cells in target cell selection to enhance system perfor-
cells where drone BSs are integrated. The challenge of balanc- mance and provide load balance. In the first step, the load level
ing loads in such networks is determining the locations of BSs of cells is simultaneously and periodically measured, and a
according to the load status of cells. ML-based solutions that restricted neighbouring cell list (NCL) is created from cells
address the predictability of drone mobility are likely to be whose load levels are below the threshold level (TLthr ¼ 0:5).
more prominent in future. In [33], the Q-learning algorithm This step avoids associating a user with an already congested
was used to learn the optimum bias value of each BS according BS. It positively affects the overall load distribution of the net-
to the loads of static BSs. In [34], the number of future users work by including the load level criterion of cells in the target
within cells was estimated from past data using hybrid cell selection. In the second step, if an mmWave cell in NCL
ARIMA-XGBoost algorithms to balance loads. Accordingly, has a higher RSRP value than the threshold RSRP value
drone BSs were dynamically deployed. (RSRPthr ¼ 80dBm), that cell is directly selected as the target
The studies presented in this section have the following cell. Otherwise, the macro cell with the highest RSRP in
limitations: NCL is selected as the target cell. The goal of this step is to
enhance overall system performance by increasing the adapt-
1. Most existing works were developed for the 4G technology, ability and usability of mmWave cells with a high RSRP level
which may not be efficient for the 5G network that has dif- and sufficient resources. This is accomplished by providing a
ferent features and requirements. Further investigations are margin of protection against the erratic nature of the mmWave
needed to effectively implement these algorithms designed channel (rapid channel changes, high path loss and severe
for the 4G technology in 5G networks with different mobil- channel interruptions). Stretching the load level threshold
ity and distribution scenarios. and the threshold of RSRP values can increase or decrease
2. Most studies were designed to operate according to a cen- the availability and usability of mmWave cells.
tral optimisation model. This means that the optimisation The second phase of the proposed mobility load balancing
process is performed based on the performance of the entire model is the HO triggering and decision-making process. As
network, not the individual user experience. Centralised previously mentioned, the HO triggering and decision-
optimisation may cause increased HO issues for some users. making process is independently performed for each user.
Several users may require optimisation at time T , while Conventional HO decision is based on a comparison of RSRP
others may need optimisation at a different time. from both the serving and neighbouring BSs. In its simple
3. The most recent studies perform target cell selection deci- form, the HO decision is initiated if the following condition
sions based on the strongest RSRP (or SINR). However, prevails for a certain period of time:
the resource availability of cells, which directly affects net-
RSRPt > RSRPs þ HOM ð1Þ
work performance metrics (such as load level of serving
cells, throughput and CDR), was not included in the target where RSRPs and RSRPt represent the RSRP of the serv-
cell selection. ing cell and the target cell, respectively. Load balancing opti-
4. Some existing algorithms do not optimise all HCPs. For misation (when the serving cell is overloaded) allows users to
example, some algorithms found in [27,31,33] only optimise offload user traffic to underloaded target cells at the expense
one HOM. Using a static TTT can cause other HO issues, of lower RSRP by setting the HOM to negative values. In this
such as low throughput, CDR and RLF. study, cells within the network are categorised into four groups
according to their load levels: low (load  25 %), medium
(25 % < load  45 %), high (45 % < load  65 %) and over-
3. Proposed load balancing model loaded (load  65). However, user traffic may not be offloaded
from overloaded serving macro cells with strong RSRP to
underloaded mmWave cells if the conventional HO decision
This study proposes a mobility load balancing algorithm that
mechanism is applied in the proposed algorithm. This is
prioritises mmWave cells in target cell selection and user asso-
because in the target cell selection phase, mmWave cells that
ciation to provide load balance and enhance overall system
satisfy the RSRP and load level conditions are prioritised by
performance in 5G HetNets. This algorithm conducts target
sacrificing the highest RSRP, unlike the traditional target cell
cell selection, HO triggering and decision-making processes
selection method. Therefore, some adjustments are needed
for each UE according to their own experience.
for user association.
Traditional target cell selection is performed depending on
In case the cell that the user is associated with (serving cell)
the maximum RSRP (or max-SINR). Such cell selection
is overloaded, different HO procedures are applied in this
626 E. Gures et al.

study depending on the target cell type (macro BS or mmWave


a (continued)
BS) for load balancing purposes. Accordingly, if the target cell
is the mmWave cell, HO is directly triggered regardless of the Algorithm 1: Proposed Mobility Load Balancing Algorithm
RSRP value of the serving cell (Algorithm 2). It should be 1 begin
noted that the mmWave cell selected as the target cell satisfies
20 if Conditions are met
the load level and RSRP conditions, which are sufficient con- 21 then HODecision False
ditions (types of prioritisation) to trigger HO. If the target cell 22 Run Trigger Timer
is a macro cell, HOM is added to the target cell RSRP value 23 else if Conditions are not met
(i.e., the transmit power is increased in a virtual way) to make 24 Run Algorithm 3
HO to the target cell more attractive to the user (Algorithm 3). 25 end if
In this way, HO of the user from the high-loaded serving cell to 26 end if
the low-loaded target cell is accomplished. Similarly, it should 27 else if Serving BS is ‘‘mmWave BS”
be noted that the load level of the macro cell selected as the tar- 28 Check conditions
29 if Conditions are met
get cell is less than the threshold load level and it is the cell with
30 then HODecision False
the highest RSRP level in NCL. The pseudocode of the pro-
31 Run Trigger Timer
posed algorithm is presented in Algorithm 1. 32 else if Conditions are not met
Although the proposed algorithm is a load balancing algo- 33 Check Target BS type
rithm, different HO procedures are applied to maintain mobil- 34 if Target BS is ‘‘macro-BS”
ity robustness in cases where the serving cell is not overloaded. 35 Run Algorithm 2
If the serving cell is the mmWave cell that is not overloaded, it 36 else if Target BS is ‘‘mmWave BS”
remains as a serving cell as long as it satisfies the load level and 37 Run Algorithm 3
RSRP conditions. The mmWave cell with a high RSRP level 38 end if
and sufficient resources is prioritised in user association. 39 end if
40 end if
Otherwise (if the mmWave cell does not satisfy the RSRP con-
41 end if
dition), HO decision is made to prevent possible RLF that may
42 end if
occur due to the decrease in signal quality. If the serving cell is 43 end
a macro cell that is not overloaded, different HO procedures
are applied depending on the target cell type, as in the load bal-
ancing part. In this case, if the target cell is the mmWave cell,
HO is directly triggered (Algorithm 2). Otherwise (if the target
cell is a macro cell), Algorithm 3 is applied. For this case,
HOM would be set to zero. It should be noted that in all
HO procedures applied, the load level of the target cell is less Algorithm 3: HO Trigger and Decision (macro-BS)
than 50 %. To conclude, the proposed algorithm prioritises 1 if then RSRPTBS þ HOM > RSRPSBS *
mmWave cells that satisfy the load level and RSRP conditions 2 if Triggertimer  TTT then
in target cell selection and user association. The proposed algo- 3 HODecision True
rithm takes full advantage of the unique features of mmWave 4 Send HO request
cells (enhanced network throughput, spectral efficiency and 5 else if
enormous bandwidth) to provide load balance. 6 then HODecision False
7 Run Trigger Timer
8 end if
9 end if
Algorithm 1: Proposed Mobility Load Balancing Algorithm
RSRPTBS = RSRP of target BS,
1 begin RSRPSBS = RSRP of serving BS,
2 Set up HOM and TTT values * HOM is zero if the serving cell is not overloaded.
3 if Simulation time t = 1 then
4 HODecision False
5 else
6 if Serving BS is ‘‘overloaded”
7 Check Target BS type
8 if Target BS is ‘‘macro-BS”
9 Run Algorithm 2 Algorithm 2: HO Trigger and Decision (mmWave BS)
10 else if Target BS is ‘‘mmWave BS” 1 if Target BS is ‘‘mmWave BS”
11 Run Algorithm 3
2 if Triggertimer  TTT then
12 end if
3 HODecision True
13 else if Serving BS is not ‘‘overloaded”
4 Send HO request
14 if Serving BS is ‘‘macro-BS”
5 else if
15 Check Target BS type
6 then HODecision False
16 if Target BS is ‘‘macro-BS”
7 Run Trigger Timer
17 Run Algorithm 2
8 end if
18 else if Target BS is ‘‘mmWave BS”
9 end if
19 Check conditions
Adaptive cell selection algorithm for balancing cell loads 627

4. System and simulation models called Directional Mobility Model, was chosen because it
allows the consideration of various HO.
The simulation model of this study is designed to represent a scenarios as well as additional possibilities. Each simulation
real 5G environment. The network is modelled on the LTE- cycle independently measures the performance of users,
Advanced Pro 3GPP Rel.16 specifications. Simulations are thereby increasing the accuracy of results. User movements
performed using the MATLAB software. We designed the sim- are periodically updated at 40 ms intervals, and the movement
ulation environment as a HetNet of 39 macro-BSs (hexagons) distance is periodically matched. All considered measurement
and 117 mmWave BSs (circles), as shown in Fig. 2. Each results made throughout the study are the average perfor-
macro-BS includes three sector antennas equipped with an mance values obtained from these 15 users.
omnidirectional antenna to connect each user to the service The Euclidean distances of each user to the BSs are calcu-
network and improve system performance. The mmWave BS lated in the distance matrix. In multipath scenarios, the dis-
is placed in the centre of each hexagonal macro-BS sector. tance matrix estimates the signal path loss in addition to the
The radii of the macro-BS and mmWave BS are 500 m and Rayleigh fading and log-normal shadowing. The RSRP and
200 m, respectively. These cells operate at different carrier fre- SINR values are then computed over each carrier signal. Dur-
quencies to avoid interference between macro-BS and ing simulation, each BS within the network updates its
mmWave BS (macro-BS’s carrier frequency = 2.1 GHz and throughput, spectral efficiency and CDR reports. The BS also
mmWave BS’s carrier frequency = 28 GHz). updates the load status and sends load information to another
A number of mobile users are created in each cell at ran- BS. Each user measures the average SINR and RSRP, which is
dom coordinates according to the cell type. A random number then forwarded to the respective BS for optimisation. Modula-
of mobile users are assigned between 0 and 100 for each tion and Coding Scheme (MCS) selection is accomplished by
macro-BS sector, and between 0 and 750 for mmWave BSs. the BS according to the received SINR and RSRP data [40,41].
The blue and pink dots represent users associated with
macro-BSs and mmWave BSs, respectively. The number of 5. Key performance indication (Kpi) metrics
users in each BS is randomly and periodically changed
throughout the simulation cycles to represent a real network Load is the ratio of current users in a cell to the maximum
environment. The developed model randomly replicates gener- number of users that the same cell can serve.
ated load traffic throughout the simulation and completely Numberofuserinthecell
enables appropriate control functions in the target cell during Load ¼ ð2Þ
Maxnumberofuserinthecell:
user mobility.
In the simulation, six different user speeds are considered to In this work, the maximum UE number that each sector of
evaluate the impact of users at different speeds (40, 60, 80, 100, macro-BS and mmWave BS can serve is 100 and 750,
120 and 140 kmph) on network performance. These speeds are respectively.
assigned for theoretical investigations and represent vehicle Throughput is the volume of data that is moved from one
speeds in urban and suburban areas. Table 1 provides a sum- place to another in a specific amount of time. It is a parameter
mary of the main parameters used in the simulation. that depends on the availability and efficient use of network
In this research, 15 users are selected to evaluate the perfor- resources. In networks with balanced load distribution, high
mance of the proposed algorithm. The 15 users are represented throughput is achieved due to the efficient use of resources.
by the cross in Fig. 1 and positioned at random coordinates in In overloaded cells, fluctuations in data rates can occur due
Cell 1 for the first simulation cycle. Since user mobility is in to lack of resources, and these fluctuations affect users’ QoE.
one direction, each user moves in different parallel paths from Throughput (bps) is defined as follows:
the other selected users. The 15 measured users all move in a  
NUE symbs  NRE mSymb
RB
straight line parallel to each other. This user mobility model, RBs NSCs NSC RB
bits
Rbits ¼
UE
CR ð3Þ
TSF
where:
RUE
bits : Total Number of bits per second received to one UE
NUE
CCs : Number of CCs connected to each UE
NUE
RBs : Number of PRB assigned to each UE
NRB
SCs : Number of subcarriers in each PRB
NSC
symbs : OFDMA symbols in time domain
NRB
RE : Number of resource elements
mSymb
bits : Number of bits in one time-slot (modulation
symbols)
CR: Code Rate
TSF : Sub-frame interval of time
Spectral Efficiency is the data rate that is normalised with
the cell bandwidth. The mmWave spectrum bands can provide
additional data rates, resulting in increased spectral efficiency.
Spectral efficiency (bps/Hz) can be formulated as follows:
Fig. 2 Simulation environment.
628 E. Gures et al.

Table 1 System and simulation parameters [18,35–37].


Parameters Values
LTE-A cell 5G cell
Transmit Power [dBm] 46 30
Carrier Frequency [GHz] 2.1 28
System Bandwidth [MHz] 20 500
Number of BS 39 117
Shadowing Standard Deviation [dB] 6 7.8
Cell Radius [m] 500 200
BS Antenna Height 25 15
Number of PRBs 100 per sector 2500
Shadow Fading Model A Gaussian-distributed random variable with zero mean and rdB standard deviation in dB [37,38]
Resource Distribution Evenly distributed across all active UEs
Fast Fading Model Rayleigh fading model [39]
BS Noise Figure [dB] 5
UE Noise Figure [dB] 9
UE Height [m] 1.5
UE Antenna Gain [dBi] 0
UE Antenna 1, Omni-directional
Tested UE number 15 UEs randomly distributed
Mobility Model DMM
UE Speed [km/hour] 40,60,80,100,120,140
White Noise Power Density [dBm/Hz] 174
Thermal Noise Power [Np] Np = Nt + 10 log (BW  106)
 
Path Loss Model [dB] 40x 1  4x103 xDhb xlog10 ðRÞ  18xlog10 ðDhb Þ þ 21xlog10 f þ 80
Modulation Scheme AMC
Q_rxlevmin [dBm] 101.5
T311 Interval [s] 10
HO preparation time [ms] 50
HO execution time [ms] 40
Dhb = Antenna height, R = Distance, Q_rxlevmin: Minimum required RX level

 
Symbs  NRS mSymb
NUE 6.1. Fixed HCP values
X NUE
CCs
RBs
CCn
NRB
sc NSC RB
bits
gUE ¼ CR ð4Þ
n¼1
TSF xUE
BW
This section examines and validates the effect of various fixed
Call Dropping Ratio is the call admission control (CAC) HCP settings on the load balancing performance of the pro-
unit responsible for handling incoming calls. Thus, it prevents posed model. Performance represents the average values com-
the HO of users to an already congested cell, thereby guaran- puted for the 15 users throughout the simulation cycles and
teeing the QoS of both the users in the candidate cell and the under various UE speeds. Load balancing performance is
calling users. CDR refers to the ratio of calls dropped due to assessed using various HOM and TTT settings, as shown in
lack of resources in the serving cell to the total accepted calls. Table 2. The TTT and HOM values are selected as low, med-
Mathematically, it is expressed as: ium and high. The performance of different HCP settings is
Ndropped analysed using four KPIs: the load level of the serving cell,
CDR ¼ ð5Þ throughput, spectral efficiency and CDR.
Naccepted
Fig. 3 displays the average serving cell load levels for all
selected HCPs at various mobile speeds. The HCP2 system
6. Simulation and performance analyses with HOM and TTT values of 16 dB and 100 ms, respectively,
achieved the best performance with an average serving cell
This section presents the findings from the simulation study. load level of 35.80 %. In contrast, the HCP6 system with the
The performance of the proposed algorithm is thoroughly
assessed throughout various conditions. Six different mobile
speed scenarios are applied in six systems with several HOM Table 2 HCP settings.
values and TTT durations, as shown in Table 2. These mobile System no HOM (dB) TTT (ms)
speeds have been chosen for theoretical review since they rep-
resent vehicle speeds in urban and suburban areas. All pre- HCP1 0 0
HCP2 16 100
sented results in this section are the average measurement
HCP3 4 320
values from the 15 users initially determined. Performance is HCP4 10 320
independently measured for each user at each simulation step HCP5 16 320
(40 ms). Next, the average is obtained for all users. Each sce- HCP6 16 5120
nario has been simulated for over 2000 steps.
Adaptive cell selection algorithm for balancing cell loads 629

highest HOM and TTT setting is the worst-performing system 5120 ms and 16 dB, respectively. It is clear that the average
with an average serving cell load level of 56.40 %. The HCP1 throughput is higher for HCPs with low TTT settings. The
system with zero HOM and TTT values exhibited the second- HCP2 system (TTT = 100 ms) outperforms HCP5
best performance with an average cell load level of 37.60 %. (TTT = 320 ms) and HCP6 systems (TTT = 5120 ms) by
When the performances of HCP2, HCP5 and HCP6 systems 38.81 % and 1063 %, respectively. In HCP3, HCP4 and
with fixed HOM values and different TTT durations are collec- HCP5 with similar TTT settings, changes in HOM do not have
tively examined, the performance gains of the HCP2 system a noticeable effect on the average throughput. According to
(TTT = 100 ms) compared to the HCP5 (TTT = 320 ms) the proposed algorithm, HOM is a parameter applied when
and HCP6 systems (TTT = 5120) are 16.19 % and 36.52 %, users HO to the macro BS. This result clearly shows that in
respectively. For HCP3, HCP4 and HCP5 systems with the terms of throughput, actual performance gains are achieved
same TTT duration and various HOM values, the load levels when users HO to mmWave cells that satisfy the required con-
decrease with increasing HOM values. As can be seen from ditions (load level and RSRP).
Fig. 3, the serving cell load level for all HCPs is less than Fig. 5 illustrates the average spectral efficiency performance
65 % (overloaded threshold) in most cases. This proves that for all examined HCPs at various speed scenarios. With zero
the load level condition used in the target cell selection is an TTT and HOM settings, HCP1 has the highest average spec-
effective solution in terms of load balancing. tral efficiency of 2.49. In contrast, the lowest average spectral
Fig. 4 presents the average throughput performance for dif- efficiency is seen in HCP6, which has the highest HOM and
ferent HCP settings in various mobile speed scenarios. The TTT settings. The average spectral efficiency in HCPs with a
highest average throughput is 2.48 Mbps for HCP1, which fixed HOM of 16 dB and TTT duration of 100, 320 and
has zero HOM and TTT time. The second-highest average 5120 ms are 2.27, 2.00 and 0.69 bit/s/Hz, respectively. The
throughput is 2.16 Mbps for the HCP2 system, where the results reveal an increase in spectral efficiency with reduced
HOM and TTT values are 100 ms and 16 dB, respectively. TTT. When the spectral efficiency performances of HCP3,
In contrast, HCP6 achieved the lowest average throughput HCP4 and HCP5 with similar TTT durations are examined,
of 0.21 Mpbs with very high TTT and HOM values of the effect of changes in HOM are more clearly seen in low
mobile speed scenarios. However, fluctuations steadily con-
tinue in faster mobile speed scenarios.
Fig. 6 presents the average CDR performance for all
selected HCPs in various mobile speed scenarios. The HCP1
system with zero TTT and HOM settings performs best with
an average CDR of zero. The highest CDR is seen in the
HCP6 system which has the highest TTT and HOM settings
with an average of 2.60 %. The average CDR values for
HCP3, HCP4 and HCP5 systems with the same TTT duration
and various HOM values are 0.45 %, 0.38 % and 0.35 %,
respectively. The results indicate that the TTT duration has a
more significant effect on the CDR performance. HCPs with
low TTT duration have better CDR performance since they
can trigger HO in a shorter time against call drops that may
occur due to lack of resources and poor signal quality. In addi-
tion, the average CDR of all selected HCPs (excluding HCP6)
is less than 0.50 %, proving that the load level and RSRP con-
ditions used in the proposed mobility load balancing algorithm
Fig. 3 The average load of serving cells versus simulation cycles. do enhance the CDR performance.

Fig. 4 The average throughput versus mobile speed scenarios.


630 E. Gures et al.

Fig. 5 Average spectral efficiency versus mobile speed scenarios.

Fig. 6 Average CDR probability versus mobile speed scenarios.

Table 3 Comparison of HO performance for all HCPs.


HCP1 HCP2 HCP3 HCP4 HCP5 HCP6 Average
Load level (%) 37.60 35.80 45.20 43.53 42.72 56.40 43.5417
Throughput (Mbps) 2.4847 2.1697 1.7929 1.7917 1.7866 0.2135 1.7065
Spectral efficiency (bit/s/Hz) 2.4905 2.2705 2.0429 2.0362 2.0080 0.6953 1.9239
CDR (%) 0 0.0178 0.4506 0.3822 0.3544 2.6006 0.6343

Table 3 presents the average HO performance of the exam- 49.34 %, respectively. This result indicates that low TTT set-
ined system settings (from HCP1 to HCP6) for all measured ting allows the proposed algorithm to conduct load balancing
users (15 users) and all mobile speed scenarios (from 40 to actions at a shorter time, resulting in improved performance in
140 km/h) throughout the entire simulation time (2000 simula- load balancing related metrics such as load level, throughput,
tion cycles). Table 4 displays improvement in the average HO spectral efficiency and CDR. When examining the effect of the
performance for all HCPs compared to the best system perfor- HOM setting on the overall performance of the proposed algo-
mance. As seen from Table 4, HCP1 (TTT = 0 ms and rithm, it can be seen that the HCP3, HCP4 and HCP5 systems
HOM = 0 dB) achieves the best performance among the anal- generally achieve similar outcomes to each other (83.23 %,
ysed systems with an overall outcome of 98.80 %. The overall 82.71 % and 83.93 %, respectively). Analysing the effect of
outcomes for HCP2 (TTT = 100 ms and HOM = 16 dB), HOM is limited to scenarios where the target cell is a macro
HCP5 (TTT = 320 ms and HOM = 16 dB) and HCP6 cell when the serving cell is overloaded. This is noteworthy
(TTT = 5120 ms and HOM = 16 dB), which have the same since it was proven that users can experience greater enhance-
HOM and different TTT settings, are 94.61 %, 83.93 % and ments in overall system performance if associated with
Adaptive cell selection algorithm for balancing cell loads 631

Table 4 Average HO performance gain for all HCPs as compared to the HCP with the best system performance.
HCP1 HCP2 HCP3 HCP4 HCP5 HCP6 Average
Load level (%) 95.2128 100 79.2035 82.2421 83.8015 63.4752 83.9892
Throughput (%) 100 87.3224 72.1576 72.1093 71.9041 8.5926 68.6810
Spectral efficiency (%) 100 91.1664 82.0277 81.7587 80.6264 27.9181 77.2495
CDR (%) 100 99.9822 99.5494 99.6178 99.6456 97.3994 99.3657
Overall outcome (%) 98.8032 94.6178 83.2345 82.7150 83.9320 49.3463 82.3213

mmWave cells. Table 5 presents the abbreviations and corre-


sponding explanations used in this paper.

6.2. Proposed algorithm

The proposed algorithm is compared with the algorithms of


Han [25], Ghatak [24] and Aghazadeh [21] to further assess
and validate its performance. For this comparison, the
HCP1 system (TTT = 0 ms, HOM = 0 dB), which achieved
the highest overall performance, is applied to the proposed
algorithm.
Fig. 7 displays the average serving cell load level of the
selected load balancing algorithms and the proposed algorithm
for the entire simulation cycle. The proposed algorithm out-
performs the selected algorithms in terms of load balancing.
The average serving cell load level for the proposed algorithm Fig. 7 The average load of serving cells versus simulation cycles.
is 35.80 %, while the average cell load levels for the algorithms
of Han, Ghatak and Aghazadeh are 58.62 %, 58.66 % and
51.25 %, respectively. This proves that the proposed algorithm
is successful in offloading user traffic from congested cells to
low-loaded cells. The inclusion of the cell load level as a
Table 5 Summary of abbreviations.
parameter in the target cell selection allows users to associate
Acronyms Full Name of abbreviations cells with low load levels.
5G Fifth Generation Fig. 8 presents the average throughput over all mobile
BS Base Station speeds and simulation cycles. The average throughput is
CAC Call Admission Control derived from the overall simulation cycles for each UE speed.
CDR Call Dropping Ratio The results indicate that the proposed algorithm provides a
DL Deep Learning significant increase in throughput compared to other algo-
DRL Deep Reinforcement Learning rithms under all mobile speed scenarios. The proposed algo-
HCP Handover Control Parameters
rithm outperforms the Han algorithm (which attained the
HetNet Heterogeneous Network
HO Handover second-highest performance) by 305 %, 324 %, 447 %,
HOM Handover Margin 309 %, 363 % and 603 % throughout all selected mobile
HOP Handover Probability speeds, respectively. The proposed algorithm achieved the best
HOPP Handover Ping-Pong performance, which is a noteworthy accomplishment. The pro-
KPI Key Performance Indicators posed algorithm’s strategy to prioritise mmWave cells in target
MCS Modulation Coding Scheme cell selection and user association is clearly the contributor to
ML Machine Learning this result. It provides higher throughput by associating users
mmWave Millimetre Wave with mmWave cells that offer more resource blocks at the
NCL Neighbouring Cell List expense of lower RSRP (threshold RSRP kept RSRP at a high
NR New radio
level).
PRB Physical ReSource Blocks
QoE Quality of Experience Fig. 9 shows the average spectral efficiency for all algo-
QoS Quality of Service rithms at different mobile speeds. The figure demonstrates
RL Reinforcement Learning how the proposed mobility load balancing algorithm signifi-
RLF Radio Link Failure cantly increases the average spectral efficiency (especially dur-
RRC Radio ReSource Control ing high mobile speed scenarios) as compared to the
RSRP Reference Signal Received Power algorithms of Han, Ghatak and Aghazadeh. The proposed
RSS Received Signal Strength algorithm achieved performance gains of 1.4 %, 8.7 %,
SINR Signal to Interference plus Noise Ratio 18.9 %, 21.3 %, 34.5 % and 42.9 % compared to the Han
TTT Time to Trigger algorithm, which had the second-highest spectral efficiency
UE User Equipment
throughout all UE speeds. Prioritising mmWave cells increases
632 E. Gures et al.

Fig. 8 Average throughput versus mobile speed scenarios.

Fig. 9 Average spectral efficiency versus mobile speed scenarios.

the adaptability and usability of these cells and improves spec- sistent CDR of 0 % for the entire simulation cycle. The Han,
tral efficiency thanks to the enormous bandwidth. Ghatak and Aghazadeh algorithms attained an average CDR
Fig. 10 presents the average CDR of the proposed and of 1.90 %, 3.65 % and 0.86 %, respectively. In addition to
selected algorithms for all users and mobile speed scenarios applying the HCP1 setting (TTT = 0 ms), the proposed algo-
throughout the simulation cycle. In the proposed load balanc- rithm also considers the load level and RSRP during cell selec-
ing algorithm, the application of HCP1 settings yielded a con- tion, thereby minimising the CDR that may occur due to lack
of resources or reduced signal quality.

7. Conclusion

This paper proposed an algorithm that prioritises mmWave


cells to satisfy load level and RSRP conditions in target cell
selection and user association, taking full advantage of the
unique properties of mmWave cells (enhanced network
throughput, spectral efficiency and enormous bandwidth) to
provide load balance. This study offered a two-step target cell
selection that considers the load level of cells and RSRP. The
mmWave cells that satisfy cell load level and RSRP conditions
are prioritised in target cell selection. The proposed mobility
management scheme balances traffic load and improves overall
system performance. This is accomplished by prioritising
mmWave cells in user association through different HO proce-
Fig. 10 Average CDR probability versus mobile speed dures designated by cell types according to the load status of
scenarios. the serving cell (load balancing optimisation or mobility
Adaptive cell selection algorithm for balancing cell loads 633

robustness optimisation). Various system scenarios with fixed [4] M.A. Tunc, E. Gures, I. Shayea, A survey on IoT smart
HOM values and fixed TTT durations were examined to deter- healthcare: emerging technologies, applications, challenges, and
mine the effects of several HCP settings on the performance of future trends, arXiv preprint arXiv:2109.02042, 2021.
the proposed algorithm. This study contributes to the under- [5] I. Shayea, M. Ergen, M.H. Azmi, S.A. Çolak, R. Nordin, Y.I.
Daradkeh, Key challenges, drivers and solutions for mobility
standing of the effectiveness of HCP on load balancing perfor-
management in 5g networks: A survey, IEEE Access 8 (2020)
mance in 5G HetNets. The system setting that provides the 172534–172552.
highest overall network performance was applied to the pro- [6] E. Gures, I. Shayea, M. Ergen, A.A. El-Saleh, Fuzzy logic-based
posed algorithm. The proposed algorithm was then compared load balancing algorithm in heterogeneous networks, Workshop
with the algorithms of Han, Ghatak, and Aghazadeh to assess Microwave Theory Tech. Wireless Commun. (MTTW) 2022
and validate its performance. A comprehensive analysis was (2022) 210–215.
made by examining the load balancing performance criteria [7] ‘‘3GPP, NR and NG-RAN overall description; stage 2, 2018,
such as load level of the serving cell, throughput, spectral effi- Tech. Rep. TS 38.300, 2018.,” ed.
ciency and CDR. The average serving cell load level for the [8] A. Alhammadi, W.H. Hassan, A.A. El-Saleh, I. Shayea, H.
proposed algorithm is 35.80 %, while the average cell load Mohamad, W.K. Saad, Intelligent coordinated self-optimizing
handover scheme for 4G/5G heterogeneous networks, ICT
levels for the algorithms of Han, Ghatak and Aghazadeh are
Express, 2022.
58.62 %, 58.66 % and 51.25 %, respectively. The proposed [9] ‘‘3GPP, Study on channel model for frequencies from 0.5 to 100
algorithm outperforms the second-highest performing algo- GHz, Tech. Rep. TR 38.901, 2019.,” ed.
rithm (Han) in terms of throughput at all selected mobile [10] ‘‘3GPP, NR base station radio transmission and reception,
speeds by 305 %, 324 %, 447 %, 309 %, 363 % and 603 %, Tech. Rep. TS 38.104, 2018.,” ed.
respectively. Similarly, the proposed algorithm achieves per- [11] T.K. Geok, F. Hossain, S.K.A. Rahim, O. Elijah, A.A. Eteng,
formance gains of 1.4 %, 8.7 %, 18.9 %, 21.3 %, 34.5 % C.T. Loh, et al, 3D RT adaptive path sensing Method: RSSI
and 42.9 % compared to the Han algorithm, which has the modelling validation at 4.5 GHz, 28 GHz, and 38 GHz,
second-highest spectral efficiency throughout all UE speeds. Alexandria Eng. J. 61 (2022) 11041–11061.
The proposed algorithm further considers the load level of cells [12] A.A. Budalal, I. Shayea, M.R. Islam, M.H. Azmi, H.
Mohamad, S.A. Saad, et al, Millimetre-wave propagation
and RSRP in the target cell selection, minimising the CDR
channel based on NYUSIM channel model with consideration
that may occur due to lack of resources or poor signal quality.
of rain fade in tropical climates, IEEE Access 10 (2021) 1990–
For future research, optimal HCP settings should be deter- 2005.
mined using ML algorithms or automatic self-optimisation [13] I. Shayea, T.A. Rahman, M.H. Azmi, M.R. Islam, Real
algorithms while considering individual user experience. The measurement study for rain rate and rain attenuation
algorithm’s performance should be further examined by conducted over 26 GHz microwave 5G link system in
including KPIs related to mobility robustness such as RLF, Malaysia, IEEE Access 6 (2018) 19044–19064.
HOPP and HOP. [14] W.K. Saad, I. Shayea, B.J. Hamza, H. Mohamad, Y.I.
Daradkeh, W.A. Jabbar, Handover parameters optimisation
techniques in 5G networks, Sensors 21 (2021) 5202.
Declaration of Competing Interest
[15] M.S. Mollel, A.I. Abubakar, M. Ozturk, S. Kaijage, M.
Kisangiri, A. Zoha, et al, Intelligent handover decision scheme
The authors declare that they have no known competing using double deep reinforcement learning, Phys. Commun. 42
financial interests or personal relationships that could have (2020) 101133.
appeared to influence the work reported in this paper. [16] M.S. Mollel, S.F. Kaijage, K. Michael, Deep reinforcement
learning based handover management for millimeter wave
Acknowledgement communication, 2021.
[17] M. Mezzavilla, S. Goyal, S. Panwar, S. Rangan, M. Zorzi, An
This research work was funded by Institutional Fund Projects MDP model for optimal handover decisions in mmWave
cellular networks, Eur. Conf. Networks Commun. (EuCNC)
under grant no. (IFPIP: 1037 -135-1443). The authors grate-
2016 (2016) 100–105.
fully acknowledge technical and financial support provided [18] ‘‘Radio ReSource Control (RRC); Protocol Specification
by the Ministry of Education and King Abdulaziz University, (Release 15), document TS 36.300 V15.3.0, 3GPP, Valbonne,
DSR, Jeddah, Saudi Arabia. France, 2018.,” ed.
[19] S. Alraih, R. Nordin, I. Shayea, N.F. Abdullah, A. Abu-Samah,
References A. Alhammadi, Effectiveness of Handover Control Parameters
on Handover Performance in 5G and beyond Mobile Networks,
[1] F. Jiang, C. Feng, H. Zhang, A heterogenous network selection Wireless Commun. Mobile Comput. 2022 (2022).
algorithm for internet of vehicles based on comprehensive [20] F.B. Mismar, B.L. Evans, Partially blind handovers for
weight, Alexandria Eng. J. 60 (2021) 4677–4688. mmWave new radio aided by sub-6 GHz LTE signaling, IEEE
[2] E. Gures, I. Shayea, A. Alhammadi, M. Ergen, H. Mohamad, A Int. Conf. Commun. Workshops (ICC Workshops) 2018 (2018)
comprehensive survey on mobility management in 5G 1–5.
heterogeneous networks: Architectures, challenges and [21] Y. Aghazadeh, H. Kalbkhani, M.G. Shayesteh, V. Solouk, Cell
solutions, IEEE Access 8 (2020) 195883–195913. selection for load balancing in heterogeneous networks, Wireless
[3] E. Gures, I. Shayea, M. Ergen, M.H. Azmi, A.A. El-Saleh, Personal Commun. 101 (2018) 305–323.
Machine learning based load balancing algorithms in future [22] K.M. Addali, S.Y.B. Melhem, Y. Khamayseh, Z. Zhang, M.
heterogeneous networks: a survey, IEEE Access, 2022. Kadoch, Dynamic mobility load balancing for 5G small-cell
634 E. Gures et al.

networks based on utility functions, IEEE Access 7 (2019) Consumer Communications & Networking Conference
126998–127011. (CCNC), 2020, pp. 1-6.
[23] M.M. Hasan, S. Kwon, Cluster-based load balancing algorithm [32] Z. Wang, M. Wen, S. Dang, L. Yu, Y. Wang, Trajectory design
for ultra-dense heterogeneous networks, IEEE Access 8 (2019) and reSource allocation for UAV energy minimization in a
2153–2162. rotary-wing UAV-enabled WPCN, Alexandria Eng. J. 60 (2021)
[24] G. Ghatak, A. De Domenico, M. Coupechoux, Coverage 1787–1796.
analysis and load balancing in HetNets with millimeter wave [33] A. Madelkhanova, Z. Becvar, Optimization of cell individual
multi-RAT small cells, IEEE Trans. Wireless Commun. 17 offset for handover of flying base station, in: 2021 IEEE 93rd
(2018) 3154–3169. Vehicular Technology Conference (VTC2021-Spring), 2021, pp.
[25] P. Han, Z. Zhou, Z. Wang, User association for load balance in 1-7.
heterogeneous networks with limited CSI feedback, IEEE [34] J. Hu, H. Zhang, Y. Liu, X. Li, H. Ji, An intelligent uav
Commun. Lett. 24 (2020) 1095–1099. deployment scheme for load balance in small cell networks using
[26] Z. Li, C. Wang, C.-J. Jiang, User association for load balancing machine learning, IEEE Wireless Commun. Network. Conf.
in vehicular networks: An online reinforcement learning (WCNC) 2019 (2019) 1–6.
approach, IEEE Trans. Intell. Transport. Syst. 18 (2017) [35] ‘‘Radio Frequency (RF) System Scenarios (Release 15),
2217–2228. document TR 25.942 V15.0.0, 3GPP, Valbonne, France,
[27] S.S. Mwanje, A. Mitschele-Thiel, A q-learning strategy for lte 2018.,” ed.
mobility load balancing, in: 2013 IEEE 24th Annual [36] Radio ReSource Control (RRC); Protocol Specification
International Symposium on Personal, Indoor, and Mobile (Release 15), document TS 36.331 V15.3.0, 3GPP, 2018.,‘‘ ed.
Radio Communications (PIMRC), 2013, pp. 2154-2158. [37] A. Goldsmith, Wireless communications, Cambridge University
[28] S. Pouyanfar, S. Sadiq, Y. Yan, H. Tian, Y. Tao, M.P. Reyes, Press, 2005.
et al, A survey on deep learning: Algorithms, techniques, and [38] M. Gudmundson, Correlation model for shadow fading in
applications, ACM Comput. Surv. (CSUR) 51 (2018) 1–36. mobile radio systems, Electron. Lett. 27 (1991) 2145–2146.
[29] J. Moysen, L. Giupponi, From 4G to 5G: Self-organized [39] B. Sklar, Rayleigh fading channels in mobile digital
network management meets machine learning, Comput. communication systems. I. Characterization, IEEE Commun.
Commun. 129 (2018) 248–268. Magaz. 35 (1997) 90–100.
[30] B.R. Kiran, I. Sobh, V. Talpaert, P. Mannion, A.A. Al Sallab, S. [40] P. Channels, ‘‘Modulation(Release 12), document TS 36.211
Yogamani, et al, Deep reinforcement learning for autonomous V16. 1.0, 3GPP, Valbonne, France, 2020,” ed.
driving: A survey, IEEE Trans. Intell. Transport. Syst. (2021). [41] I. Shayea, M. Ergen, A. Azizan, M. Ismail, Y.I. Daradkeh,
[31] K. Attiah, K. Banawan, A. Gaber, A. Elezabi, K. Seddik, Y. Individualistic dynamic handover parameter self-optimization
Gadallah, et al., Load balancing in cellular networks: A algorithm for 5G networks based on automatic weight function,
reinforcement learning approach, in: 2020 IEEE 17th Annual IEEE Access 8 (2020) 214392–214412.

You might also like