Abubakar, A. I., Öztürk, M., Rais, R. N. B., Hussain, S. and Imran, M. A.
(2020) Load-Aware Cell Switching in Ultra-Dense Networks: An Artificial
Neural Network Approach. In: 5th International Conference on UK - China
Emerging Technologies (UCET 2020), Glasgow, UK, 20-21 Aug 2020,
ISBN 9781728194882 (doi:10.1109/UCET51115.2020.9205365).
This is the author’s final accepted version.
There may be differences between this version and the published version.
You are advised to consult the publisher’s version if you wish to cite from
it.
http://eprints.gla.ac.uk/222512/
Deposited on: 21 August 2020
Enlighten – Research publications by members of the University of Glasgow
http://eprints.gla.ac.uk
1
Load-Aware Cell Switching in Ultra-Dense
Networks: An Artificial Neural Network Approach
Attai Ibrahim Abubakar
∗
, Metin Ozturk
∗
, Rao Naveed Bin Rais
†
, Sajjad Hussain
∗
, Muhammad Ali Imran
∗
Abstract—Most online cell switching solutions are sub-optimal
because they are computationally demanding, and thus adapt
slowly to a dynamically changing network environments, lead-
ing to quality-of-service (QoS) degradation. This makes such
solutions impractical for ultra-dense networks (UDN) where the
number of base stations (BS) deployed is very large. In this paper,
an artificial neural network (ANN) based cell switching solution
is developed to learn the optimal switching strategy of BSs in
order to minimize the total power consumption of a UDN. The
proposed model is first trained offline, after which the trained
model is plugged into the network for real-time decision making.
Simulation results reveal that the performance of the proposed
solution is very close to the optimal solution in terms of trade-off
between the power consumption and QoS.
I. INTRODUCTION
Base station (BS) switching is one of the most generally
accepted techniques for energy saving in mobile cellular net-
works (MCNs) [1]. It takes advantage of the spatio-temporal
variations in user traffic demands to match power consumption
with traffic demand per time thereby avoiding energy wastage
during period of low or no traffic load. Various BS switching
schemes have been proposed in literature employing analyti-
cal [2], heuristic [3], and machine learning [4], [5] techniques.
Among them, machine learning implementations are be-
ginning to gain more application in wireless communications
because of their ability to learn complex network behaviours
that are hard to be accurately modelled analytically. They also
have the ability to adapt to dynamic network environment,
which is difficult for most hard-coded heuristic algorithms [6].
In addition, machine learning models can be trained offline
and then plugged into the network in order to enhance real-
time decision making by reducing the network delays [7]. One
of such machine learning algorithms is the artificial neural
networks (ANN), which are known as universal approximators
because they are able to find the relationships between com-
plex non-linear functions and are also known for their excellent
generalization ability [8], hence they have found numerous
applications in the field of wireless communications [7].
A few research works have been carried out regarding
the application of ANN for cell switching purposes [4], [9],
[10]. In [4], an ANN algorithm was proposed to determine
the switching pattern that maximizes the energy efficiency
of the network while ensuring that the minimum bit rate
requirements of the users is satisfied. A context-based energy
saving approach for cache enabled BSs using Bayesian neural
networks was proposed in [9]. The authors in [10] applied
∗
James Watt School of Engineering, University of Glasgow, UK.
†
Electrical and Computer Engineering, Ajman University, UAE.
ANN for traffic prediction and cell switching decisions with
two different ANN architectures.
However, the previously mentioned solutions only consider
simplistic network scenarios where very few small cells (SCs)
are deployed, hence, such solutions will not be suitable when
network dimensions becomes very large and complexity in-
creases. In addition, only one type of SC was considered in the
aforementioned works which is not the case in a real network
where different types of SCs (remote radio head (RRH),
micro, pico, and femto cell) are deployed, thus making their
considered scenarios unrealistic.
In this paper, we exploit ANN to determine the optimal cell
switching strategy in an ultra-dense network (UDN). Specifi-
cally, we develop an ANN-based cell switching framework,
which is referred to as offline-trained online cell switch-
ing (OTOcell), to learn the optimal switching strategy for
the SCs in a UDN. The developed model is computationally
efficient since the training is done offline, after which the
trained model is implemented in the network for real-time
decision making. This is particularly important for UDNs,
where the macro cells (MCs) are already over-burdened with
signalling and computational operations, in which case, adding
a cell switching algorithm on top of these would make their
workload more severe. We also consider various types and
number of SCs to validate the robustness and scalability of
the proposed solution.
The remaining parts of the paper is organized as follows:
Section II presents the system model, followed by a description
of the proposed OTOcell framework in Section III. Section IV
evaluates the performance of the model while Section V
concludes the paper.
II. SYSTEM MODEL
A. Network model
A heterogeneous UDN, with separate control and data plane,
comprising both MC and SCs is considered. Four types of
SCs (RRH, micro, pico, and femto cell) are considered in
the work. The MC—which encompasses the footprints of the
SCs—is constantly kept on, and also orchestrates the switching
operation of the SCs via its backhaul connection to them. The
SCs, on the other hand, can be turned ON/OFF based on their
traffic load and are responsible for handling high data traffic
demands. Vertical traffic offloading is considered, such that
the traffic load of the SCs that are switched OFF are offloaded
to the MC to ensure that the quality-of-service (QoS) of the
offloaded users are maintained.
2
B. Power Consumption model
The power consumption model of a BS proposed in [11] is
adopted and is expressed as:
P
BS
=
(
P
c
+ σ
y
P
tx
, if 0 < P
tx
< P
m
P
s
, if P
tx
= 0,
(1)
where P
BS
represents the BS total power consumption, σ
y
denotes the slope of the load dependent components, P
c
is
the constant power consumption component of the BS, and P
s
is the sleep mode power consumption. P
tx
and P
m
denote the
instantaneous and maximum BS transmission power, respec-
tively. The relationship between the traffic load of the BS and
transmission power can be expressed as:
P
tx
= τ P
m
, 0 ≤ τ ≤ 1 (2)
where τ is the normalized traffic load of the BS.
The total power consumption of the UDN comprises the
power consumption of the MC and that of all the SCs deployed
under its coverage. This can be expressed as:
P
T
= P
MC
+
N
X
n=1
P
SC,n
, (3)
where P
T
, P
mc
and P
sc,n
are UDN’s total, MC’s, and n-th
SC’s power consumption, respectively.
C. Problem Formulation
The aim of this research is to select the optimal combination
of SCs to switch OFF, during periods of low or no traffic in
order to minimize the total power consumption of a UDN
while ensuring that the QoS of the users originally connected
to the switched OFF SCs are maintained by the MC.
Hence, the optimization objective can be defined as:
min
φ∈Φ
P
T
(φ)
s.t. ˆτ
MC
≤ 1.
(4)
where φ is the selected SC switching policy, while Φ is the set
of all the possible SC switching combinations. P
T
(φ) is the
expected power consumption of the network with φ switching
policy. ˆτ
MC
is the traffic load of MC after the offloading is
complete, and is given as
ˆτ
MC
= τ
MC
+
N
X
n=1
τ
SC,n
Γ
n
, (5)
where τ
MC
and τ
SC,n
are the original traffic demands (i.e.,
before offloading) of MC and n-th SC, respectively, and N
is the total number of SCs in the network. Γ is a control
parameter, which is responsible for offloading the traffic load
of only the switched OFF SCs, such that
Γ
n
=
(
1, if SC
n
is OFF
0, if SC
n
is ON,
(6)
where SC
n
is the n-th SC.
The constraint in (4) is to ensure that there must be sufficient
capacity in the MC to accommodate both the original traffic
demand of the MC, τ
MC
, and the total traffic demand of all
the SCs that are switched OFF in order to maintain the QoS.
Traffic
Database
Exhaustive
Search
Algorithm
Input/Output
Mapper
Training Data Set Generator
Offline Model Generator
ANN Model
Training
Validation
Testing
Real-Time
Switching Policy
Maker
UDN
Fig. 1. Overview of the proposed OTOcell framework.
III. OTOCELL FRAMEWORK
Most of the cell switching solutions developed using heuris-
tic approaches, such as exhaustive search (ES) or genetic
algorithm, are not suitable for real-time implementation, par-
ticularly in networks with large dimensions because they are
usually computationally demanding. As a result, before these
algorithms decide which set of SCs to switch ON/OFF and
execute the decision, the network state would have changed,
thereby leading to sub-optimal switching decision and delays.
However, these heuristic approaches can be combined with
ANN to accelerate the computation of the optimal cell switch-
ing strategy. Our proposed framework is built upon two basic
observations: 1) cell switching can be considered to be a
problem of deciding the mapping between the traffic demand
and optimal switching pattern; 2) ANN are popularly referred
as universal function approximators, implying that they can
learn the mapping between almost any input and output [8].
Based on these observations, we propose an OTOcell frame-
work to determine the optimal switching strategy that maps the
traffic demand of the BSs to the optimal cell switching pattern.
The proposed OTOcell framework is summarized in Fig. 1.
The traffic loads of the MC and all the SCs associated with
it are collected and stored in the Traffic Load Database. The
traffic loads are then passed to the Training Data Set Generator
which consists of the ES algorithm, and Input/Output Mapper.
The ES algorithm uses the optimization function in (4) to
decide the optimal set of SCs to switch ON/OFF per time while
the Input/Output Mapper prepares the training data set—which
includes the traffic loads, and optimal switching combinations.
The training data set is then transferred to the Offline Model
Generator for ANN model training, validation and testing.
The ANN model utilized is a feed-forward architecture
comprising one input layer, three hidden layers (HLs), and
one output layer (OL). The number of neurons in the input
layer is determined by the input features of the training data
set (number of SCs and MC), the number of neurons in the
HL are selected empirically by trying different combinations,
and the OL neurons is given by 2
N
, where N is the number
of SCs. The activation function for the HL neurons is ReLU
while that of the output neurons is softmax (since the cell
switching problem is a multi-classification problem).
3
TABLE I
PARAMETERS FOR THE DEVELOPED ANN MODEL
Parameter Scenario 1 Scenario 2
HLs, Neuron size 3, 128 × 128 × 128 3, 32 × 32 × 32
OL neuron size 16 4096
Learning rate 0.0001 0.001
Batch size 30 50
Epochs 1000 1000
HL Activation Function
ReLU
OL Activation Function
Sofmax
Loss function Categorical-crossentropy
Optimizer Adam
The function of the ANN is to learn the mapping between
the SC traffic demands and the optimal switching pattern
through training. The training process involves adjusting the
ANN parameters, using gradient-descent algorithms, such
that the difference (error) between the expected output (i.e.,
predicted output) and the actual output (label) is as close
to zero as possible. This error is usually estimated using a
loss function and in our case the categorical cross-entropy
function is employed. The trained cell switching model is then
transferred to the Real-time Switching Policy Maker for real-
time SC switching.
The justification for using the proposed framework is
twofold: 1) once the ANN model is fully trained, the optimal
cell switching pattern can be obtained in real-time, that is,
whenever the network status changes, without resorting to
computing the optimization objective afresh; 2) both training
data set generation and the ANN training stage, which are the
computationally intensive processes, can be done offline, thus
enabling the trained model to be implemented for real-time
cell switching without additional computational overhead and
delays to the network.
IV. PERFORMANCE EVALUATION
A. Simulation Scenario and data-set generation
Two simulation scenarios, Scenario-A and Scenario-B, with
different number of SCs are considered to test the performance
of the proposed model on varying network sizes. Both sce-
narios consists of 1 MC, but Scenario-A has 4 SCs (1 of
each type of SC), while Scenario-B has 12 SCs (2 RRH, 3
micro, 4 pico, and 3 femto cells)
1
. The traffic load of both
MC and SCs are generated using uniform random distribution
model, such that τ
MC
∈ [0, m
MC
] and τ
SC
∈ [0, m
SC
] where
m
MC
, m
SC
are the maximum normalized loads of the MC and
SCs, respectively. In Scenario-A, BS switching pattern were
generated for 7 days with one-minute resolution using ES
amounting to about 10,000 observations, while Scenario-B was
for 35 days
2
resulting in about 50,000 observations. For the
remaining simulation parameters, we adopted values in [11].
1
The number of each type of SC was selected randomly.
2
More data set is generated in Scenario-B because the increase in network
dimension and complexity makes the training process more difficult.
B. ANN Training and Testing
For Scenario-A, two data sets—each comprising about
10,000 traffic load samples of the BSs and their corresponding
optimal switching patterns—were used for training and testing
the proposed model. For Scenario-B, one data set comprising
about 50,000 traffic load samples and their corresponding
optimal switching patterns was utilized, out of which 80%
was used for training and 20% for testing. The training of the
model in both scenarios is carried out using the Adam opti-
mization algorithm [12]. Table I summarizes the parameters
of both models. Upon successful training of both models in
each scenario, the trained models are then applied to the test
data set in order to evaluate the performance of the trained
models.
C. Results and discussion
Figs. 2 and 3 presents a comparison of the total power
consumption of the UDN, for both scenarios of OTOcell
versus two benchmark approaches: 1) All-ON, which is the
conventional approach where no switching is implemented,
that is, all the SCs and MC are constantly kept on; and 2)
ES approach, which tries to find the best switching policy
by considering all the possible switching combinations and
selecting the one that results in the least power consumption
while the constraint in (4) is satisfied. The ES approach is
guaranteed to always return the optimal policy, and hence
the goal of any switching technique is to produce the closest
approximation of this approach.
In Fig. 2, it can be observed that the performance of the
proposed OTOcell is the same as that of ES most of the
time but shows slight variations at some time instances due to
wrong cell switching prediction from the OTOcell. Compared
to the All-ON, it can be observed that the OTOcell shows a
reduction in power consumption, however, we notice that due
to the few number of SCs, the reduction in power consumption
is not significant most of the time as the SCs have fewer
opportunities to sleep.
In Fig. 3, where the number of SCs is increased from 4 to
12, the OTOcell shows a slightly lesser performance compared
to that of Fig. 2 as the deviation from the optimal ES is
more pronounced. This can be traced to the fact that the
network dimensions and complexity is increased in Scenario-B
compared to Scenario-A, and as a result the OTOcell is prone
to more prediction errors. However, compared to the All-ON
method, the proposed method shows a significant reduction in
power consumption at all time instances owing to the fact the
number of SCs has tripled, hence there are more opportunities
to switch OFF more SCs.
The QoS evaluation of the proposed framework is also
carried out using the throughput metric to ascertain the impact
of the OTOcell framework on QoS of the network. Here, the
network throughput is considered to be the traffic demand that
is supported by all the active BSs in a given time interval.
Table II presents the throughput per time slot of the UDN
using the OTOcell framework and the two benchmarks. The
throughput per time slot is the total network throughput for
a given time slot, i.e., every hour and TP avg. is the average
4
network throughput. It should be noted that not all the time
slots are shown in Table II for conciseness purpose.
1 2 3 4 5 6 7 8 9 10 11 12
43.1
43.2
43.3
43.4
43.5
43.6
43.7
43.8
43.9
44
Time slot, [hr]
Power Consumption, P[kW]
All-ON
ES
OTOcell
Fig. 2. Power consumption of OTOcell and benchmarks when N = 4.
1 2 3 4 5 6 7 8 9 10 11 12
74.5
75
75.5
76
76.5
77
Time slot, [hr]
Power Consumption, P[kW]
All-ON
ES
OTOcell
Fig. 3. Power consumption of OTOcell and benchmarks when N = 12.
From Table II, when N is 4, it can be observed that the
throughput of OTOcell is the same as that of ES, and All-
ON approaches, which means than it is able to guarantee
the QoS of the network. However, when N increases to 12,
a slight decrease in network throughput is observed with
OTOcell compared to ES at certain time instances. This can be
traced to inappropriate switch ON/OFF decisions due to wrong
predictions from the proposed framework occasioned by the
increase in network dimension and complexity which makes
it more difficult to accurately train the ANN model. Hence,
the QoS of the network is slightly reduced when the network
dimension increases, but the overall effect on the network is
very minimal as revealed in the average throughput (TP avg.)
values.
V. CONCLUSION
In this paper, a UDN with four types of SCs and two
deployment scenarios is considered. A cell switching mecha-
nism based on ANN is proposed to determine the optimal cell
switching pattern per time instance based on the traffic loads
of both the MC and SCs. The simulation results reveals that a
significant amount of power savings can be achieved with the
proposed model. In addition, the performance of the proposed
OTOcell framework is very close to the optimal ES-based
solution in terms of power consumption minimization with
TABLE II
THROUGHPUT OF OTOCELL AND BENCHMARKS
N Method
Throughput per time slot [Mbps]
4 8 12 16 20 24 TP avg.
4
All-ON 2.55 2.49 2.49 2.48 2.53 2.45 2.48
ES 2.55 2.49 2.49 2.48 2.53 2.45 2.48
OTOcell 2.55 2.49 2.49 2.48 2.53 2.45 2.48
12
All-ON 6.45 6.53 6.57 6.68 6.46 6.41 6.55
ES 6.45 6.53 6.57 6.68 6.46 6.41 6.55
OTOcell 6.44 6.51 6.55 6.68 6.44 6.40 6.54
very minimal effect on the QoS. Future work will consider
using more realistic traffic model, carry out complexity, and
error probability analysis on the proposed model in order to
further ascertain its suitability for UDNs.
VI. ACKNOWLEDGEMENT
This work was supported partly by the EPSRC (GCRF)
funds (Grant no. EP/P028764/1) and Ajman University (Grant
no. 2019-IRG-ENIT-8). The first author was supported by the
Nigerian Tertiary Education Trust Fund (TETfund).
REFERENCES
[1] M. Feng, S. Mao, and T. Jiang, “Base Station ON-OFF Switching in
5G Wireless Networks: Approaches and Challenges,” IEEE Wireless
Communications, vol. 24, no. 4, pp. 46–54, Aug 2017.
[2] A. Shojaeifard, K. Wong, K. A. Hamdi, E. Alsusa, D. K. C. So,
and J. Tang, “Stochastic Geometric Analysis of Energy-Efficient Dense
Cellular Networks,” IEEE Access, vol. 5, pp. 455–469, 2017.
[3] J. Wu, S. Jin, L. Jiang, and G. Wang, “Dynamic switching off algorithms
for pico base stations in heterogeneous cellular networks,” EURASIP
Journal on Wireless Communications and Networking, vol. 2015, no. 1,
p. 117, 2015.
[4] Ruhong Zeng, Shixiang Zhu, Hongwen Yang, and Jiaxin Zhu, “An
artificial neural network based cell switch-off algorithm in cellular
system,” in 2016 2nd IEEE International Conference on Computer and
Communications (ICCC), Oct 2016, pp. 1434–1439.
[5] A. I. Abubakar, M. Ozturk, S. Hussain, and M. A. Imran, “Q-Learning
Assisted Energy-Aware Traffic Offloading and Cell Switching in Het-
erogeneous Networks,” in 2019 IEEE 24th International Workshop on
Computer Aided Modeling and Design of Communication Links and
Networks (CAMAD), Sep. 2019, pp. 1–6.
[6] N. Bui, M. Cesana, S. A. Hosseini, Q. Liao, I. Malanchini, and J. Wid-
mer, “A Survey of Anticipatory Mobile Networking: Context-Based
Classification, Prediction Methodologies, and Optimization Techniques,”
IEEE Communications Surveys Tutorials, vol. 19, no. 3, pp. 1790–1821,
thirdquarter 2017.
[7] F. Hussain, S. A. Hassan, R. Hussain, and E. Hossain, “Machine
Learning for Resource Management in Cellular and IoT Networks:
Potentials, Current Solutions, and Open Challenges,” arXiv preprint
arXiv:1907.08965, 2019.
[8] A. Zappone, M. Di Renzo, and M. Debbah, “Wireless networks design
in the era of deep learning: Model-based, AI-based, or both?” arXiv
preprint arXiv:1902.02647, 2019.
[9] L. Wang, S. Chen, and M. Pedram, “Context-driven power management
in cache-enabled base stations using a Bayesian neural network,” in
2017 Eighth International Green and Sustainable Computing Conference
(IGSC), Oct 2017, pp. 1–8.
[10] I. Donevski, G. Vallero, and M. A. Marsan, “Neural Networks for
Cellular Base Station Switching,” in IEEE INFOCOM 2019 - IEEE
Conference on Computer Communications Workshops (INFOCOM WK-
SHPS), April 2019, pp. 738–743.
[11] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson,
M. A. Imran, D. Sabella, M. J. Gonzalez, O. Blume, and A. Fehske,
“How much energy is needed to run a wireless network?” IEEE Wireless
Communications, vol. 18, no. 5, pp. 40–49, October 2011.
[12] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
CoRR, vol. abs/1412.6980, 2014.