Abstract—Predicting network performance for achieving better capacity design in mobile networks is crucial. One of the key challenges is forecasting the assessment of spectral efficiency under high network loads. Based on experience, the primary factor influencing network performance and spectral efficiency is the Channel Quality Indicator (CQI) metric. This article aims to examine the performance of the Hamrah Aval mobile network using CQI prediction through deep learning. As predictive CQI is a key factor for network performance and spectral efficiency, the selection of appropriate features for deep learning networks significantly impacts improving predictive capabilities. Features such as frequency band, Physical Resource Block (PRB) allocation in neighboring cells, the number of neighboring cells within a 2.5-kilometer radius, downlink and uplink transport capacity, and the use of higher-order modulation are chosen as inputs to the deep learning network. Ultimately, the proposed model achieves a 96% accuracy in predicting the CQI for the dataset of Hamrah Aval cells.
Index Terms— Cellular Telecommunications, Capacity, Deep Artificial Intelligence, 3CQI.
I.INTRODUCTION
One challenge for network designers in the past decade has been the escalating traffic in mobile networks. Despite the growth rate of mobile networks being dependent on various factors, it has been claimed that, on average, the traffic has doubled every two years. This rapid development pace adheres to Moore’s Law. Alongside the increasing load, network performance undergoes changes, and in cases where adding capacity to the network is not feasible, it leads to a decline in network performance. Simultaneously, user demands, or more precisely, application-specific demands, have increased for operational capability and reduced latency.
On the other hand, the process of adding capacity to mobile networks is highly time-consuming. Mobile operators typically require six months to add a new layer of 4G or 5G and two years to construct a new base station. In general, significant economic justification is necessary for new investments. Therefore, designing based on predictive modeling becomes essential. Hence, the decision-making process regarding network capacity expansion must rely on accurate predictions of future network performance and the evaluation of various traffic growth scenarios, network performance, and capacity enhancements.
II.The Challenge of Predictive Modeling and User Experience
The challenge of predictive modeling in terms of operational data capacity in 4G and 5G mobile networks, based on Orthogonal Frequency Division Multiple Access (OFDMA) technology, is significant. The frequency spectrum for mobile networks is divided into different frequency bands, and the channel bandwidth in LTE systems is typically 5, 10, 15, or 20 megahertz, while in 5G, it can range from 50 to 100 megahertz in lower frequency bands and up to 400 megahertz in higher frequency bands.
Both LTE and 5G systems have resource blocks (RBs) allocated in channels, where the available spectrum is divided. In LTE, each RB has a size of 180 kilohertz, while in 5G, the size of each RB can vary between 180 kilohertz and 1440 kilohertz based on use cases. The operational capacity of user data in these systems is determined by two main components.
The first component is the number of RBs available for each user, dependent on network density, user density, and allocated capacity. The second crucial component is the achievable operational power in each RB within the system spectrum.
This article focuses on modeling spectral efficiency and analyzing it in an active network with real changes in network load over a specific time interval. It emphasizes the importance of understanding that the patterns of change in each mobile operator are unique and depend on various factors. Some of the key factors include the frequency spectrum holdings, network density, topology, quality of design, and the implemented radio access strategies.
Therefore, instead of finding a one-size-fits-all model for different networks and diverse spectrum holdings, the goal is to define and construct a framework and methodology that can adapt to various spectrum holdings while considering different features. In cellular communication systems, advanced capabilities for performance monitoring have been utilized, enabling the recording and analysis of different network parameters in real-time. Machine learning techniques are then employed to analyze indicators and learn from past data, allowing for predictive modeling of conditions and parameters in the future.
III.A Review of Past Approaches
The presented articles [4,5] delve into various aspects of machine learning for addressing system improvement challenges, including beamforming optimization, multi-user detection, coding/decoding techniques, RF description, anomaly detection, network traffic prediction, resource management, spectrum allocation, and delivery prediction. While many studies discuss capacity design and optimization challenges using deep learning methodologies, a predominant focus is placed on traffic prediction issues.
Some of the recent relevant examples can be found in articles [6-10], where different approaches with high or superior accuracy, depending on the ability to capture various network conditions such as seasonal variations, abnormal events, or changes in network configurations, were explored and tested. However, in the capacity design process, traffic prediction can often be solely represented by the first component, the user experience, discussed in the introduction, which involves the number of RBs available for each user. The second important factor, the spectral efficiency, according to the authors, has not received much attention so far.
In situations where mobile operators are grappling with increasing traffic and a data deluge, precise evaluation and measurement of spectral efficiency and its accurate prediction with artificial intelligence models will alleviate many challenges. The novel idea introduced in this article is the presentation and evaluation of several deep learning models capable of analyzing network effects during increasing traffic conditions and accurately predicting spectral efficiency.
IV.Theory
Given Cellular communication systems, also referred to as cellular networks, are networks constructed with a large number of base stations, each covering specific geographical areas. These areas are typically divided into three cells. In remote communications, the efficiency and radio channel capacity are fundamentally expressed based on the Signal-to-Noise and Interference Ratio (SINR). The signal level is determined by the antenna power of each base station, and the signal loss in the path to the mobile phone. Noise is thermal noise defined by the channel bandwidth, and interference refers to any unwanted signal received in the channel. Interference can be internal, originating from a similar system, or external, entering the system from other networks. In the downlink performance of fourth and fifth-generation mobile networks, interference is primarily from intra-cellular sources. Both LTE and 5G NR systems are based on Orthogonal Frequency Division Multiple Access (OFDMA), allowing users in a cell to use a portion of the frequency band unless multiple-input multiple-output (MIMO) techniques are implemented. The impact of MIMO interference on system performance is limited. The main source of interference is the received signal assigned to other users; predominantly intra-cellular interference imposed by neighboring cells.
Radio channel quality in mobile communication systems varies dynamically from the nearest point to the antenna to the cell’s edge. Therefore, LTE and 5G NR networks must adjust connection parameters to create the best possible user experience. This process is referred to as link adaptation. Link adaptation is based on certain parameters set by user equipment (UE) devices, as defined by the 3GPP standard in the Channel State Information (CSI) report. These reports have three main components: Channel Quality Indicator (CQI), Precoding Matrix Indicator (PMI), and Rank Indicator (RI).
From the link adaptation perspective, CQI is the most crucial component, as it implicitly indicates the downlink transmission rate. The CQI range has discrete values between 0 and 15, where 15 represents the best channel quality, and 0 indicates the weakest channel quality. According to the 3GPP standard, the mapping of CQI for different modulations is provided in [10] (scaled by a factor of 1024).
Rank Indicator (RI) is another important factor for predicting spectral efficiency, especially in the context of MIMO functionality. It reflects the nearly parallel performance of MIMO in different spatial scenarios. Some initial performance analysis results in this area suggest better spectral efficiency in 5G systems compared to 4G. This improvement is mainly attributed to the use of higher rank indicators and better MIMO performance in terms of spatial multiplexing. The most significant reason for this is the increase in the number of antenna elements in 5G radio systems.
Fig. 1. Logs in CQI cellular
V.Problem Definition and Research Method
Based on the information provided, solving the spectrum efficiency prediction problem can have a significant impact with a good prediction model involving CQI (Channel Quality Indicator) and the rank indicator. In a previous study [reference 11], authors investigated the correlation between CQI and MIMO (Multiple Input Multiple Output), while this article focuses on modeling CQI. Most research in this field concentrates on network architectures and performance models, which are generally comprehensive, with some assumptions that may limit the model accuracy. Additionally, these studies usually focus on system design and algorithms, rarely addressing network design issues. This is mainly because academics and researchers are not often directly involved in industry work and have limited access to real network data.
There is limited quantitative research on CQI that combines empirical and theoretical approaches. Furthermore, in the process of evaluating the performance of global mobile communication systems, from 3G to LTE and up to 5G-NR systems, the network layout has undergone minimal changes. Key aspects of network design, such as antenna height and azimuth, remain relatively constant. Therefore, many performance modeling methods can be applied based on past data, traffic pattern recognition, and the effects of the network.
The authors’ approach to the problem of performance prediction in network efficiency relies on commercial mobile data. Handling extensive datasets, constructing algorithms, and models that can work in any mobile cellular network with different features and impacts are integral to their methodology.
Data collection in mobile networks is performed through Performance Management (PM) systems. The 3GPP standard outlines how vendors of Radio Access Network (RAN) gather network performance data. Although there are slight differences between RAN methods in designing PM systems, the core principles remain similar. The key performance indicators used in this article are reports of the CQI indicator, analyzed as a Cumulative Distribution Function (CDF) in different periods, such as hours and days, in each cell. The evaluation of the CQI reporting process involves tracking and analyzing changes in network load, considering factors such as Resource Block (RB) consumption within the serving cell and neighboring cells.
A.Selecting network features
The objective of this article is to design a model based on deep learning networks capable of predicting the CQI (Channel Quality Indicator) in network load conditions. The main ideas revolve around the subdivision of cell clusters based on crucial and influential factors affecting system performance. The following five features, utilized to enhance the accuracy of the predictive model, include:
- Frequency band
- Physical Resource Block (PRB) utilization in neighboring cells
- Number of neighboring cells within a specified radius (2.5 km)
- Downlink and uplink transport capacity
- Use of higher-order modulations
The frequency band division is crucial, as radio propagation differs across various frequency bands. Inter-cell interference, a significant factor affecting system performance, is more pronounced in cells operating at lower frequencies due to increased radio signal propagation and interference with neighboring cell signals.
Another key factor is the number of neighboring cells within a specific radius, indicating network density. A denser network tends to be more sensitive to interference. The study explores the average PRB consumption, both determined by neighboring cells and aggregated across neighboring cells, encapsulating the neighborhood count in an overall score. Different radii (2.5 km, 5 km, and 10 km) are compared, with the best performance observed using a 5 km radius, although similar performance is seen with a 2.5 km radius.
Furthermore, the load in neighboring cells, represented by the average PRB consumption, is defined for profiling different network scenarios with higher traffic. The use of higher-order modulation is selected as an indicator that distributes users in space and cells based on average radio conditions, represented as the percentage of 64QAM and 256QAM usage.
The study also considers traffic types, such as video streaming or heavy FTP downloads, which heavily utilize LTE/5G-NR network resources. Data-heavy factors are challenging to capture with available PM system parameters. Therefore, the authors attempt to correlate data-heavy parameters with traffic load and PRB consumption. Despite the difficulty in capturing traffic volume accurately, a significant correlation exists between heavy data parameters and higher-order modulation features with spectral efficiency.
B.Data and Network Model Validation
In this article, the required data for the Huawei vendor and 4G technology system during the summer period from June 1 to September 1 has been collected. The frequency bands for all records are 15 and 20, and since similar frequency bands act as noise during data normalization, they have been excluded from the model training process. The columns of the used data are as follows:
- Physical resources for download and upload
- Heavy data factor in the payload of the system per hour for download and upload
The heavy data for LTE cells in the Tehran city area for modulation averages are used as input data for the proposed deep learning network model, and the spectrum efficiency performance is illustrated in Fig. 2.
Fig. 2. Proposed Deep Learning Network Model
As mentioned earlier, the required data for each cell has been collected and prepared using PM systems. The data for this study is obtained from the operator “Hamrah Aval” in Iran, including LTE network data from 6 stations operating in different frequency bands: 1800 MHz and 2100 MHz. Although it is possible to increase the number of stations operating at lower frequency bands, such as 900 MHz, due to hardware limitations, currently, this number of stations and data are sufficient. The input data is prepared in two forms: one-hour time clustering and one-day time clustering (24 hours) and is provided as separate input for two distinct models.
The data, collected over a three-month period with hourly intervals, contains over 2000 records for each cell. In the initial preprocessing step, the consumption of Physical Resource Blocks (PRB) in neighboring cells, the number of neighboring cells, the average PRB consumption in neighboring cells, download and upload capacity, and higher-order modulation are considered from raw data. The proposed deep learning model is trained using these input features to predict the average Channel Quality Indicator (CQI).
During the experiment, the model’s performance is evaluated on unseen data. 80% of the data is used as training data, and the remaining 20% is selected as test data for evaluation. The neural network is implemented using the Keras library, with an architecture comprising 9 dense layers for the first model with hourly input and 4 dense layers for the second model.
Given that the problem is a regression type, the Mean Absolute Error (MAE) with a value of 23 is used as the performance metric, calculated according to Equation 1.
| (1) |
Fig. 3. Comparison of Training Data and Validation Data Graphs, and Dispersion Plots in Different K-folds Over One-hour Time Intervals
Fig. 4. Graph of Training and Evaluation alongside Dispersion Plot in Different K-folds for the 24-hour Data Model
C.The results of the proposed model and the validation credibility
Considering the limited number of base station cells and to prevent overfitting, a 24-fold cross-validation with five layers and subsets has been employed for further learning. The cross-validation, as a model evaluation method, assesses the extent to which the results of statistical analysis on a dataset generalize and are independent of the training data. This method is particularly useful for predictive models in order to determine how useful the intended model will be in practice. In general, one round of K-fold cross-validation involves partitioning the data into two complementary subsets, performing an analysis on one subset (training data), and validation analysis using the other subset (validation or test data). To reduce variability, the cross-validation process is repeated several times with different partitions, and the results of these validations are averaged. In this study, K=5 was utilized.
Subsequently, in Figures 3, scatter plots of prediction data and evaluation data based on Mean Absolute Error (MAE) in various K-folds are provided. As evident from the scatter plots in different K-folds, the proposed model exhibits convergence across different input data types for training and evaluation.
In Figures 4, scatter plots of dispersion in different K-folds are presented. As indicated by the scatter plots in various K-folds, regardless of the input data type for training or evaluation, the proposed model has the ability to predict data dispersion in different cells.
VI.Conclusion
The main focus of the research presented in this article is the utilization of deep learning for addressing network design issues, capacity planning, and particularly, modeling the spectral efficiency evaluation in growing network load conditions. The results of two models, presented in one-hour and 24-hour time intervals in this article, include information from six cells and have the potential for expansion to a greater number of cells in the future. On one hand, several performance indicators for the network were introduced with this idea to enhance the accuracy of the deep neural network model analysis. A strong correlation between CQI and PRB consumption, payload, and modulation is observed. Additionally, a stronger correlation with PRB consumption is noticed when neighboring cells are added. Nevertheless, the authors decided to introduce the average consumption and the number of neighboring cells within a 2.5-kilometer radius as separate features.
References
[1] Ericsson Mobility Report, November 2021.
Available online: https://www.ericsson.com/en/
reports-and-papers/mobility-report (accessed
on 13 January 2022).
[2] Tomi´c, I.; Davidovi´c, M.; Bjekovi´c, S. On
the downlink capacity of LTE cell. In Proceedings
of the 23rd Telecommunications Forum
TELFOR, Belgrade, Serbia, 24–25 November
2015; pp. 181–185. [CrossRef]
[3] Dahlman, E.; Parkvall, S.; Skold, J. Overall
Transmission Structure. In 5G NR: The Next
Generation Wireless Access Technology;
Academic Press: London, UK; San Diego, CA,
USA, 2020; pp. 103–131.
[4] Santos, G.L.; Endo, P.T.; Sadok, D.; Kelner, J.
When 5G Meets Deep Learning: A Systematic
Review. Algorithms 2020, 13, 208. [CrossRef]
[5] Zhang, C.; Patras, P.; Haddadi, H. Deep
learning in mobile and wireless networking: A
survey. IEEE Commun. Surv. Tutor 2019, 21,
2224–2287. [CrossRef]
[6] Gijón, C.; Toril, M.; Luna-Ramírez, S.; Marí-Altozano,
M.L.; Ruiz-Avilés, J.M. Long-Term Data
Traffic Forecasting for Network Dimensioning in
LTE with Short Time Series. Electronics 2021,
10, 1151. [CrossRef]
[7] Bastos, J. Forecasting the capacity of mobile
networks. Telecommun. Syst. 2019, 72,
231–242. [CrossRef]
[8] Li, R.; Zhao, Z.; Zheng, J.; Mei, C.; Cai, Y.;
Zhang, H. The learning and prediction of
application-level traffic data in cellular networks.
IEEE Trans. Wirel. Commun. 2017, 16,
3899–3912.
[9] Hua, Y.; Zhao, Z.; Liu, Z.; Chen, X.; Li, R.;
Zhang, H. Traffic prediction based on random
connectivity in deep learning with long shortterm
memory. In Proceedings of the 2018 IEEE
88th Vehicular Technology Conference (VTCFall),
Chicago, IL, USA, 27–30 August 2018;
- 1–6.
[10] Evolved Universal Terrestrial Radio Access
(E-UTRA); Physical Layer Procedures, 3GPP
TS 36.213 15.7.0. Available online: https://
www.etsi.org/deliver/etsi_ts/136200_136299/
136213/15.07.00_60/ts_136213v150700p.pdf
(accessed on 31 January 2022).
[11] Tomi´c, I.; Luki´c, Ð.; Davidovi´c, M.; Draji´c,
D.; Ivaniš, P. Statistical analysis of CQI reporting
and MIMO utilization for downlink scheduling
in live LTE mobile network. Telfor J. 2020,
12, 8–12. [CrossRef]
[12] Kumar, V.; Mehta, N.B. Modeling and Analysis
of Differential CQI Feedback in 4G/5G OFDM
Cellular Systems. IEEE Trans. Wirel. Commun.
2019, 18, 2361–2373. [CrossRef]
[13] Yin, H.; Guo, X.; Liu, P.; Hei, X.; Gao, Y.
Predicting Channel Quality Indicators for 5G
Downlink Scheduling in a Deep Learning
Approach. Available online: https://arxiv.org/
pdf/2008.01000.pdf (accessed on 31 January
2022).
[14] Rassa, E.H.R.; Ramli, H.A.M.; Azman,
A.W. Analysis on the impact of outdated
channel quality information (CQI) correction
techniques on real-time quality of service
(QoS). In Proceedings of the IEEE Student
Conference on Research and Development
(SCOReD), Bangi, Malaysia, 26–28 November
2018.
[15] Torres J., G.; Bustamante, R. Analysis of the
effects of CQI Feedback for LTE Networks on
ns-3. IEEE Latin Am. Trans. 2015, 13,3538–
- [CrossRef]
[16] Tomi´c, I.; Davidovi´c, M.; Draji´c, D.; Ivaniš,
- On the impact of network load on CQI reporting
and Link Adaptation in LTE systems. In
Proceedings of the IcEtran, Staniši´ci, Bosnia
and Herzegovina, 8–10 September 2021; pp.
612–624.
[17] Djuri´c, K.; Tomi´c, I.; Neskovi´c, A. On the
impact of Network density on correlation between
Network load and Link adaptation in MIMO-
OFDM based Mobile Broadband Networks.
In Proceedings of the 29th Telecommunications
Forum TELFOR 2021, Belgrade, Serbia, 23–24
November 2021; 2021. [CrossRef]
[18] Weisberg, S. Applied Linear Regression;
John Wiley & Sons: Hoboken, NJ, USA, 2013.
[19] Pedregosa, F.; Varoquaux, G.; Gramfort, A.;
Michel, V.; Thirion, B.; Grisel, O.; Duchesnay,
- Scikit-learn: Machine learning in Python. J.
Mach. Learn. Res. 2011, 12, 2825–2830.
[20] Chen, T.; Guestrin, C. XGBoost: A Scalable
Tree Boosting System. In Proceedings of the
22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, San
Francisco, CA, USA, 13–17 August 2016; pp.
785–794. [CrossRef]
[21] XGBoost Python Package. Available online:
https://xgboost.readthedocs.io/en/stable/python/
index.html, (accessed on 13 January 2022).
[22] Fine, T. Feedforward Neural Network Methodology;
Springer: New York, NY, USA, 1999.
[23] Chollet, F. Keras. Available online: https://github.
com/fchollet/keras. (accessed on 13 January
2022).
[24] Live:https://machinelearningmastery.com/kfold-
cross-validation
بدون دیدگاه