Machine Learning enabled Wireless Communication

Network System
Jiajun Chen∗† , Yin Gao∗† , Yingjun Zhou ∗† , Zhuang Liu∗† , Da peng Li∗† and Man Zhang ∗†
∗ State Key Laboratory of Mobile Network and Mobile Multimedia Technology, Shenzhen, 518055, China
† Algorithm Department, Wireless Product R&D Institute, ZTE Corporation, Shanghai, 201203, China

Email:{chen.jiajun1, gao.yin1, Yingjun Zhou, liu.zhuang2, li.dapeng, and zhang.man4}@zte.com.cn

2022 International Wireless Communications and Mobile Computing (IWCMC) | 978-1-6654-6749-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/IWCMC55113.2022.9824835

Abstract—The 5th generation of mobile communication will to improve the network management and user experience is to
support three application scenarios of eMBB, uRLLC and adopt advanced AI/ML technology. The capability of AI/ML
mMTC. To meet the requirements, wireless communication to analyze large quantity of data, and predict the future status
systems needs to continue to develop, with the development of
artificial intelligence(AI). Machine learning (ML) is expected to or decisions has the potential for intelligence in the network
optimize wireless systems by tackling complex problems which optimization deployment and many other aspects.
cannot be solved using traditional mathematical models. AI/ML In this paper, indicative use cases using AI/ML techniques
techniques, which could extract features from raw data in the in conjunction with its benefits, and its potential deployment
wireless access network and predict the future state by the solutions are provided, and even the consideration on the
previous features, are useful to network optimization. In this
paper, indicative use cases using AI/ML techniques in conjunction potential prototype of future wireless communication system
with its benefits, and its potential deployment solutions are is proposed. The main contributions of this work can be
introduced. Furthermore, this paper provides the consideration summarized as follows.
on the potential prototype of future wireless communication • Introduce the several high-level AI/ML based use cases,
system.
e.g., AI/ML based Network Energy Saving, AI/ML based
Index Terms—Machine learning, LSTM, Network Energy
Saving, Load Balancing, Mobility Optimization, load prediction, Load Balancing, and AI/ML based Mobility Optimiza-
trajectory prediction tion, and bring out theirs benefits to the network opti-
mization through extensive simulation experiments.
• In order to deploy the AI/ML techniques into the real
I. I NTRODUCTION
environment, we introduce potential solutions on indica-
With the continuous development of Artificial Intelligence tive AI/ML use cases, including the procedures used for
(AI), AI technology has raised a new round of technological AI.ML training and AI/ML inference.
revolution in all human disciplines. As an important research • Further consideration on future wireless communication
direction of AI technology, machine learning takes advantage network,we propose to integrate AI into every aspect of
of the depth of the neural network non-linear processing wireless communication system to depict vision of the
capability, which successfully solved a series of previously intrinsic AI through intelligent data perception, intelligent
intractable problems, in the image recognition, speech pro- modeling, distributed architecture, and intelligent moni-
cessing, natural language processing, games showed stronger toring.
than the human and the ability of traditional algorithms [1]. The rest of this paper is organized as follows. Section II
Thus, AI technology has achieved increasing attention. describes the high-level AI/ML based use cases in wireless
Wireless communication systems continues to develop, with communication system in conjunction with its benefits and
the development of artificial intelligence algorithms. The 5th the potential solutions in real environment. The consideration
generation of mobile communication will support three ap- on future wireless communication system is provided in Sec-
plication scenarios of eMBB, uRLLC and mMTC. However, tion III, followed by the conclusion in Section IV.
5G network, and even future wireless network system are
II. H IGH - LEVEL AI/ML BASED USE CASES IN WIRELESS
expected to meet the challenges of consistent optimization
COMMUNICATION SYSTEM
of increasing numbers of key performance indicators(KPIs)
including latency, reliability, user experience, density, etc. At A. Network Energy Saving
present, a large amount of human costs need to be invested To meet the 5G network requirements of key performance
in the operating costs of wireless networks, and inaccurate and the demands of the growth of the mobile subscribers,
network optimization decisions hinder the development of millions of the base stations are being constructed. Such
intelligent wireless networks. Operators see traditional human- rapid growth brings the issues of high energy consumption,
machine interaction as slow, error-pone, expensive to handle CO2 emissions and operation expenditures [2] [3]. Therefore,
these issues and requirements. Moreover, the mobile environ- energy saving is an important use case which may involve
ment is becoming much more complex, heterogeneous and different layers of the network, with mechanisms operating at
constantly developing. As a result, a potential tool and solution different time scales.

978-1-6654-6749-0/22/$31.00 ©2022 IEEE 1285

Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on May 17,2023 at 04:26:14 UTC from IEEE Xplore. Restrictions apply.
 Figure 3. Overall structure of AI/ML based energy saving

formance on energy saving. In addition, the actual energy

Figure 1. Statistics of the Switch-off Duration (Hours) saving of each cell per day is also significantly increased.
Fig. 2 shows that the power saving used AI/ML techniques
could reach up to 1.24 KW·H, and no matter what switch-
off strategy used, AI/ML based ES shows the best results for
power saving. Based on all the simulations and analysis given
above, compared with the conventional ES, AI based ES shows
better performance, less impact on system KPIs. Hence, how
to deploy the AI/ML architecture into the current NG-RAN
structure and interfaces should be considered.
The structure, as shown in Fig. 3 is composed of data
collection , tool box, energy saving decision and energy-
saving execution. Data collection is mainly responsible for
the data collection, and required data is used to do the
load prediction, scenario classification through ML algorithms.
Energy saving decision making may be based on the various
information including predicted information. There are many
kinds of predicted information that can be used by operators to
make energy-saving decisions, such as energy-saving scenarios
Figure 2. Statistics of the energy consumption (KW·H) classification, traffic load prediction and trajectory prediction.
The more accurate the predicted information is, the better the
energy-saving decisions based on the predicted information
Switching on/off cells is a kind of ES methods [4] [5] results will be.
in the spatial domain that exploits the traffic offloading in
a layered structure to reduce the energy consumption of the
whole wireless access network. When the traffic load predicted B. Load Balancing
is lower than the fixed threshold, the cells can be switched off The rapid growth of commercial network traffic and the
and the UEs may be offloaded to the new target cell. Machine application of multi-band make the balanced distribution of
learning algorithms can be used to predict the state of the next traffic become a challenge. With the distribution and the
period, specially the traffic load. movement of users in radio access networks, it often happens
The main aim of the cell-switching on/off scheme [6] is to when some of the base stations burdens a extremely high
save the energy consumption in the wireless access network. amount of load while the others is almost free of load. This is
The switch-off strategies used in the BSs include symbol the so called load imbalance which often happens in current
switch-off, channel switch-off and carrier switch-off. We eval- networks, and it brings more attention to load balancing. The
uate the application of ML techniques with real scenarios. goal of which is to distribute traffic load to different nodes of
After the AI energy saving is enabled, due to the dynamic the network for the maximum utilization of network resources.
calculation and adjustment strategy of each cell, the switch-off Base stations could decide to offload some load to other base
duration per cell per day is significantly improved compared stations, based on some load information. With the help of load
to the conventional energy saving methods, as shown in Fig. 1. prediction through AI/ML techniques, the base stations could
AI/ML based energy saving predicts load accurately and get more accurate decisions of load balancing to enhance the
switch off the cell in time, which achieve the better per- load balancing performance.

1286
Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on May 17,2023 at 04:26:14 UTC from IEEE Xplore. Restrictions apply.
 Figure 4. Flow chart of load prediction results

Figure 6. The predicted latitude of the UE

could request input information for load balancing training and

inference from neighbouring RAN nodes and UE, e.g., own
current or predicted load information, current or predicted load
information from neighbouring RAN nodes, UE radio mea-
surements (RSRP, RSRQ, SINR,etc). After model inference,
feedback needs be exchange between NG-RAN nodes in order
to evaluate model performance and decide whether to trigger
model update/retraining. Load balancing decision could be the
recommended target cell to be offloaded for UE.

C. Mobility Optimization
With the development of wireless communication technolo-
gies and explosive increases number of UEs, data traffic grows
rapidly so that much denser deployment is essential. More
frequent handover cause higher latency and throughput reduc-
tion, which has a negative impact on the network performance
Figure 5. Potential solution to AI/ML based load balancing and users’ satisfaction. UEs are handed over frequently, which
may cause handover anomalies: e.g., short stay, ping-pong
handover, early handover, late handover. For the applications
Regarding load prediction, which is important part of of 5G network including energy saving, load balance, resource
AI/ML based load balancing, linear ensemble model composed allocation and mobility management, it is essential to predict
of three sub-models could predict the resource status in terms the positions of mobile users in the future so as to make
of time, space and historical pattern respectively [7]. Different preparation in advance.
methods such as time series analysis, linear regression and Mobility management is the scheme to guarantee the
regression tree are applied to the sub-models, after which ad- service-continuity during the mobility by minimizing the call
justed weights are calculated and allocated to each of the three drops, RLFs, unnecessary handovers, and ping-pong. UE tra-
sub-models to create the ensemble model. Fig. 4 shows the jectory prediction is one of most important AI/ML techniques
simulated performance comparison of load prediction based on used for mobility optimization [8] [9]. Capability of predicting
the various ML algorithms. The evaluation index is ANRMSE mobile users’ next-period positions is not only of great benefits
(Average Normalized Root-mean Square Error). With higher to network resource allocation but also improve its intelligence
accuracy of load prediction, load balancing decisions relying and effectiveness [10] [11]. Accurate next-location prediction
on current, past, and future cell resource status are sufficient, enable the network to obtain users’ information in advance,
which could avoid ping-pong handover between different cells, and target NG-RAN node can assign the resource and opti-
cell overload and degradation of user service quality. mize network configuration for the incoming user to increase
Considering AI/ML could help network to make better load users’ quality of services and the utilization rate of network
balancing decision, the potential solution to AI/ML based resources.
load prediction over current NG-RAN structure and potential AI algorithms, Long short-term memory (LSTM) [12] is
interfaces is provided, as shown in Fig. 5. NG-RAN node selected as the location prediction models. The models exploit

1287
Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on May 17,2023 at 04:26:14 UTC from IEEE Xplore. Restrictions apply.
 Figure 7. The predicted longitude of the UE
Distance Error (Unit:M)

Figure 9. Potential solution to AI/ML based mobility optimization

and potential interfaces is provided, as shown in Fig. 9. It has

the similar understanding as the solution to AI/ML based load
balancing. NG-RAN node could request input information for
Location
Location Orientation RSRP_S RSRP_S+RSRP_N RSRP_S+2RSRP_N
load balancing training and inference from neighbouring RAN
nodes and UE, e.g., UE historical location information, UE
Figure 8. Average distance error of LSTM based location and various features Radio measurements, moving velocity, etc. The decision of
AI/ML based mobility optimization includes predicted target
handover, estimated arrival probability,etc.
the historical UE geographic location (e.g. latitude, longitude)
over a period of time, current UE orientation and RSRP III. F UTURE CONSIDERATION
information for serving and neighbour cells to predict the 5G communication system is a service-based architecture,
future position, so that in this case, historical geographic providing a modular framework for meeting stringent latency
location, orientation and RSRP information are taken into and reliability requirements. Barely introducing AI technology
consideration when constructing the AI model [13]. The in wireless network to solve a certain network optimization
simulation results of the location prediction is shown in the problem does not enable network to be intelligent. Contin-
Fig. 6 and Fig. 7. In addition, prediction performance of LSTM uously changing radio environment requires to retrain and
with different features are evaluated in Fig. 8. It is obvious update the fixed ML models resulting in the repetitive work,
that introducing the orientation information into the LSTM and hindering the intelligence of the wireless system. Future
network is of benefit to the location prediction. In addition, communication system is not only considered to apply AI
LSTM using RSRP information from the serving base station to enhance the certain function, e.g. energy saving, mobility
exhibits apparently better performance than the LSTM only management, etc, but also is designed to integrate AI into
using the geographic location, while RSRP values from two every aspect of wireless communication system to depict
neighbour base stations can enhance the ability of LSTM vision of the intrinsic AI.
to predict the next location. Therefore, it is concluded that Therefore, to bring autonomous learning, autonomous
the features, such as the orientation, measurement RSRP, are decision-making, self-optimization and self-evolution, AI-
helpful to the location prediction which also helps to make Native radio network will be a intelligent loop, in which is
suitable mobility optimization decision. intelligent data perception, intelligent modeling, distributed
Based on the discussion and simulation given above, AI architecture, and intelligent monitor.
based trajectory prediction is of benefits to optimize the
network. Accurate prediction helps RAN to make effective A. Intelligent data perception
mobility optimization management. The potential solution to Large quantity of data transportation will bring burdens to
AI/ML based load prediction over current NG-RAN structure the current interface. On the other hand, data sensed from the

1288
Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on May 17,2023 at 04:26:14 UTC from IEEE Xplore. Restrictions apply.
 radio environment sometimes don’t have the corresponding IV. C ONCLUSION
labels. Now with the Generative Adversarial networks (GAN), In this paper, three high-level AI/ML based use cases,
it will avoid transferring large amount of data between various e.g., AI/ML based Network Energy Saving, AI/ML based
nodes in the network, and protect the data privacy. GAN have Load Balancing, and AI/ML based Mobility Optimization, and
the ability to generate the required data to simulate real data, bring out theirs benefits to the network optimization through
and improve the performance of model to a certain extent. extensive simulation experiments. And its corresponding so-
lutions in real environment to these AI/ML based use cases
B. Intelligent modeling are provided. Moreover, the considerations on future wireless
communication system are proposed including autonomous
The implementation of machine learning usually requires a learning, autonomous decision-making, self-optimization and
lot of manual intervention, such as data pre-processing, feature self-evolution. More investigations on AI/ML are required
selection, model selection, hyper-parameter and adjustment. to realize intelligent RAN systems. Given these research
Each individual ML model or algorithm has a specific struc- directions, a more rapid development of RAN intelligence will
ture and usually comes with a set of strategies or rules for be achieved.
model construction. The purpose of intelligent modeling is
to reduce manual intervention so that the radio network can R EFERENCES
automatically model and train to achieve autonomous learning [1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
capabilities. Currently, AutoML technology can automatically with deep convolutional neural networks,” in Advances in neural infor-
mation processing systems, 2012, pp. 1097–1105.
select a machine learning model on a given data and task, [2] R. ElKazzi and A. Khalil, “Energy-saving solution for future cellular
and automatically select an optimization algorithm, so that systems,” in 2018 4th International Conference on Renewable Energies
the model has the characteristics of high performance and low for Developing Countries (REDEC). IEEE, 2018, pp. 1–6.
[3] R. Tafazolli et al., “emobility mobile and wireless communications
computational complexity on the task. technology platform: Strategic applications research agenda,” NetWorks
European Technology Platform, 2010.
[4] T. En, W. Ye, D. Fei, P. Zhiwen, and Y. Xiaohu, “A practical enb
C. Distributed architecture off/on based energy saving scheme for real lte networks,” in 2015
17th International Conference on Advanced Communication Technology
The centralized AI server collects the data of each network (ICACT). IEEE, 2015, pp. 12–17.
element and each node in the network, which will cause prob- [5] T. Han and N. Ansari, “On greening cellular networks via multicell
cooperation,” IEEE Wireless Communications, vol. 20, no. 1, pp. 82–
lems such as time delay and large data transmission volume. 89, 2013.
The distributed AI server architecture can effectively solve this [6] Y. Gao, J. Chen, Z. Liu, B. Zhang, Y. Ke, and R. Liu, “Machine
problem. The AI units are distributed on each network nodes learning based energy saving scheme in wireless access networks,” in
2020 International Wireless Communications and Mobile Computing
to jointly perform the calculation tasks of the same set of AI (IWCMC). IEEE, 2020, pp. 1573–1578.
models. Distributed AI network architecture will be the trend [7] Y. Gao, M. Zhang, J. Chen, J. Han, D. Li, and R. Qiu, “Accurate load
of intelligent AI network architecture in the future. Distributed prediction algorithms assisted with machine learning for network traffic,”
in 2021 International Wireless Communications and Mobile Computing
AI network architecture, where each network node can be (IWCMC). IEEE, 2021, pp. 1683–1688.
used as a part of AI training/execution, and a large number of [8] M. Morzy, “Mining frequent trajectories of moving objects for location
related devices can jointly build a common model based on prediction,” in International workshop on machine learning and data
mining in pattern recognition. Springer, 2007, pp. 667–680.
locally collected data sets. It will reduce data transmission load [9] T. V. T. Duong and D. Q. Tran, “A fusion of data mining techniques for
and data privacy protection issues on the radio interface, and predicting movement of mobile users,” journal of communications and
also improve model performance and delay issues . Distributed networks, vol. 17, no. 6, pp. 568–581, 2015.
[10] C. Wang, Z. Zhao, Q. Sun, and H. Zhang, “Deep learning-based
AI will play an important role in the subsequent evolution of intelligent dual connectivity for mobility management in dense network,”
the network architecture. in 2018 IEEE 88th Vehicular Technology Conference (VTC-Fall), 2018,
pp. 1–5.
[11] C. Wang, L. Ma, R. Li, T. S. Durrani, and H. Zhang, “Exploring
D. Intelligent monitor trajectory prediction through machine learning methods,” IEEE Access,
vol. 7, pp. 101 441–101 452, 2019.
Intelligent monitor is the introduction of human control into [12] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
the decision-making process of the network to improve the [13] Y. Gao, J. Chen, Z. Liu, L. Liu, and N. Hu, “Deep learning based location
decision-making of AI algorithms and help the machine better prediction with multiple features in communication network,” in 2021
understand user preferences and make more user-preferred IEEE Wireless Communications and Networking Conference (WCNC).
IEEE, 2021, pp. 1–5.
decisions. For example, when the AI model itself can not
make the correct decision or the cost of making the wrong
decision is high, the AI algorithm can make a decision with
human intelligence. With the reinforcement learning, the agent
obtains reward through interaction with the environment or
feedback from users, learns the characteristics of the external
environment, and improves decision-making strategies to adapt
to the external environment changes.

1289
Authorized licensed use limited to: INDIAN INST OF INFO TECH AND MANAGEMENT. Downloaded on May 17,2023 at 04:26:14 UTC from IEEE Xplore. Restrictions apply.