You are on page 1of 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/345446849

LTE Cell Traffic Grow and Congestion Forecasting

Conference Paper · April 2019


DOI: 10.1109/eStream.2019.8732145

CITATIONS READS

5 2,610

2 authors, including:

Darius Chmieliauskas
Vilnius Gediminas Technical University
4 PUBLICATIONS   5 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Darius Chmieliauskas on 16 November 2020.

The user has requested enhancement of the downloaded file.


LTE Cell Traffic Grow and Congestion Forecasting
Darius Chmieliauskas Darius Guršnys
Faculty of Electronics Faculty of Electronics
Vilnius Gediminas Technical University Vilnius Gediminas Technical University
Vilnius, Lithuania Vilnius, Lithuania
darius.chmieliauskas@vgtu.lt darius.gursnys@vgtu.lt

Abstract— Recent popularity grow of predictive analysis is • Total traffic per cell per day forecast using the
growing in many fields. Importance to forecast mobile traffic for fbProphet algorithm.
each LTE cell can bring mobile network planning advantages
for the operator. It can help operators to spend minimum • Daily traffic trend for 24 hours periods using
investing for new sites and cells but be able to guarantee fbProphet algorithm.
excellent service experience for mobile broadband users. In this
paper study of mobile traffic forecasting feasibility using • The relation between data traffic volume and
fbProphet algorithm developed by Facebook is presented. LTE cell load status.
Target is to have short term forecasting from which the operator
can proactively consider network expansion if the load is too
II. TOTAL DAILY TRAFFIC FORECASTING
high to satisfy user throughput demands. Five months of daily In LTE FDD networks main reason for the congestion is
traffic data used to train model and one month used for downlink traffic. For example, in 100 LTE sites, which were
forecasting and model testing. Also, 30 days of hourly data were used for the presented study, downlink traffic is 11 times
used for busy hour traffic forecasting. In the last part relation higher than uplink.
between data traffic carried in the LTE cell to cell load
explained. Due to downlink-uplink traffic asymmetry, only downlink
direction will be considered for forecasting and congestion
Keywords— LTE network load, time series forecasting, mobile evaluation.
data traffic, LTE cell capacity
A. LTE traffic forecasting method
I. INTRODUCTION fbProphet is a procedure for forecasting time series data
Based on Ericsson Mobility Report June 2018[1] “LTE based on an additive model where non-linear trends are fit
became the dominant mobile access technology at the end of with yearly, weekly, and daily seasonality, plus holiday
2017. The number of LTE subscriptions continues to grow effects. It works best with time series that have strong seasonal
strongly and is forecast to reach 5.5 billion by the end of 2023 effects and several seasons of historical data [2]. LTE daily
when it will account for more than 60 percent of all mobile traffic data perfectly fit fbProphet time series forecasting
subscriptions”. algorithm description.
As LTE continues to be the main technology for mobile LTE data traffic forecasting and evaluation were
data traffic, operators must consider the expansion of network performed on 100 randomly selected sites of the live LTE
capacity by adding new cells. Operators as units of business network. Each site may have a different amount of LTE cells
should target minimum investment, but keeping high Quality depending on current network plan. Selected sites have 141
of Experience (QoE) for mobile broadband users. Locations LTE Band 3 (LTE 1800MHz) cells and 237 LTE Band 20
with limited network capacity can lead to customer churn and (LTE 800MHz) cells. Selected sites make a significant
revenue decrease for operators with traffic volume-based proportion of live network and consist of different
tariff. Traffic suppression means a lower amount of data usage geographical areas and traffic patterns (e.g. city hot-spot,
which directly reflects mobile network operator income. The suburban with many fixed wireless clients, or sites for wide
most important factor is to guarantee data rate and latency rural coverage with low traffic). Such random selection gives
requirements for basic internet user services, e.g. watching insight about forecasting performance for different scenarios.
HD YouTube video or making Facebook messenger video The aim of the experiment is to forecast future traffic with
call. comparable or better accuracy to other forecasting methods.
For example, to be comparable to holiday traffic forecast
In this paper method to forecast LTE traffic grow for each results described in [3], where 84% of the data points had less
cell is verified. It could help to identify locations where LTE than 20% error relative to the actual value. This accuracy for
capacity expansion needed before clients will notice the 15-day ahead holiday forecast is generally accepted by the
degradation of the basic internet services usage experience. industry [3].
Machine learning method to evaluate and forecast network
bottlenecks would reduce manual workload for engineers. As Mobile network operators (MNO) store daily aggregated
well traffic increase forecasting would help to avoid performance indicator data for the long term and with higher
overinvestment for locations with low traffic demand. granularity for a shorter period (e.g. 1-hour granularity stored
for 30 days). Daily forecasting was performed from 180 days.
Many businesses are applying or starting to apply machine
Six months period were split to five months for training and
learning algorithms to improve operational efficiency. LTE
one month for forecast testing. In 180 days period, we have
network collects plenty of data from every node. Currently,
enough data to see weekly, monthly and seasonal trends.
mostly statistical evaluation is done. Machine learning
Forecast of 28 future days is enough time for the operator to
approach can be applied to improve evaluation and forecasting
act if cell congestion is forecasted. In the experiment
of various performance indicators.
forecasted 28 days were compared to real traffic and forecast
The presented study consists of 3 main parts: accuracy was evaluated using various forecast performance

978-1-7281-2499-5/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: Vilnius Gediminas Technical University. Downloaded on November 16,2020 at 08:23:16 UTC from IEEE Xplore. Restrictions apply.
evaluation metrics. In the second step, using the available 30 TABLE I. LTE TRAFFIC FORECASTING ACCURACY:
OVERALL RESULT
days of hourly data, traffic trend for 24 hours is defined. From
the long term forecast, we forecast total daily traffic and LTE Traffic forecasting accuracy for all cells
knowing the hourly trend we can forecast traffic amount for MAPE Explained variance score R² score
every hour during the day. Forecasted hourly traffic compared
44.22% 0.90007 0.8980
to the spectral efficiency of LTE cell and evaluated if the load
threshold is reached. Combining all steps explained here we
can detect future LTE cell congestion during the busy hour. Explained variance and R² score are close to 1 (best
B. LTE traffic forecasting results possible score). Mean absolute percentage error is 44%.
Forecasting made up to 28 days, 44% MAPE is comparable to
In Fig. 1 example of real and forecasted traffic is shown. results presented in [3], where the forecast was done for 15
Forecasted traffic has similar decreasing trend and spikes days.
similar to real traffic.
Fig. 2 shows the MAPE and Fig. 3 shows the explained
variance and R² score for all cells.

Fig. 1. Example of daily cell traffic forecast. First 152 days show trend
obtained using the fbProphet algorithm, last 28 days show forecasted traffic.
Trend and forecasts compared to real traffic.
Fig. 2. Mean absolute percentage error vs forecast day. The chart shows the
calculated forecast results for 100 LTE sites (378 LTE cells) value for every
Following methods were used to evaluate the accuracy of forecast day from 1 to 28.
forecasting:
• Explained variance score (1)
• R² score (2)
• Mean Average Percentage Error (MAPE) (3)

_ , 1 , (1)

samples

, 1 samples
, (2)

samples
where ∑
samples

% samples | |
MAPE , ∑ (3)
samples
Fig. 3. R² score and explained variance score vs forecast day. The chart
shows the calculated forecast results for 100 LTE sites (378 LTE cells) value
where y is real data volume in bits, and ŷ is forecasted data for every forecast day from 1 to 28.
volume in bits. All of our accuracy evaluation was
implemented using scikit-learn [4], an open-source package As expected forecasting up to 1 week is giving better
for Python. results with larger accuracy which decreases for later days.
Table 1 is showing MAPE, explained variance score and
In Fig. 4 scatter of forecasting accuracy (MAPE) vs traffic
R² of the overall forecast (28 days of all cells comparing to
volume for every cell.
real traffic per day per cell).

Authorized licensed use limited to: Vilnius Gediminas Technical University. Downloaded on November 16,2020 at 08:23:16 UTC from IEEE Xplore. Restrictions apply.
Using fbProphet algorithm for hourly data traffic trend for
each cell can be created. Traffic trend reveals a busy hour of
each cell and the proportion of busy hour traffic compared to
total daily traffic. Combining total daily and hourly traffic
forecasting we can forecast future traffic per LTE cell per
hour.
IV. LTE CELL CAPACITY EVALUATION
In this part previously forecasted hourly data traffic will
be used to evaluate if the LTE cell is capable to handle such
traffic amount without QoE degradation due to congestion and
competition for spectrum resource.
A. LTE single-user throughput
In general LTE single-user throughput depends on 3 major
Fig. 4. Mean absolute percentage error of forecasted traffic vs transferred components:
traffic per cell. Cells with more traffic tend to have better forecasting results.
• The number of scheduled physical resource block
Evaluating forecasting results for each cell we can (PRB) from LTE time-frequency radio resources grid.
compare forecast MAPE versus total carried traffic volume in PRB allocation and downlink scheduling
each cell. From fig. 4 it is clear that cells which carry more described in [5].
traffic usually are more predictable. If traffic in a cell is low, • Signal to Interference & Noise Ratio (SINR), which
any individual peak or heavy data user can make quick is reported by UE as Channel Quality Indicator (CQI).
variations of total traffic per day and such cell traffic The UE determines CQI such that it corresponds to
distribution has more randomness which is difficult to predict. the highest Modulation and Coding Scheme (MCS)
allowing the UE to decode the transport block with
III. HOURLY TRAFFIC TREND error rate probability not exceeding 10% [6].
Usually, the MNOs store daily performance indicator
values for the long term. However, to evaluate congestion, • Multiple-input multiple-output (MIMO) usage
shorter period of peak usage (e.g. 1 hour) should be performance. The use of (MIMO) antennas that can
considered. further improve the spectral efficiency gain by using
spatial division multiplexing [7].
A. LTE traffic forecasting hourly method
MIMO usage proportion and CQI depends on user
Using fbProphet peak hour and the proportion of daily location, overall interference levels, and user equipment
traffic amount can be obtained for every cell. Using proportion receiver quality. LTE eNodeB based on UE reports about
of daily traffic, we can calculate how much traffic will be downlink channel quality defines MCS based on Table
carried each hour and later check if this traffic reach capacity 7.1.7.1-1: Modulation and TBS index table for PDSCH [7].
threshold for that cell. To obtain the daily traffic trend 30 days Knowing the MCS index and number of scheduled PRBs for
of hourly traffic data were used. UE we can get transport block size (TBS) for each
B. LTE traffic forecasting hourly results Transmission Time Interval (TTI) using Table 7.1.7.2.1-1:
Transport block size [7]. LTE cell with 20MHz bandwidth can
Fig. 5 represents real hourly data traffic over LTE cell and
schedule a maximum 100 PRBs for single UE. If 100 PRB
trend from the fbProphet algorithm. Traffic volume trend
scheduled for UE, using MCS 28, 75376 kbit/s per MIMO
(proportion of total daily traffic), accurately follows real
stream or approximate 150 Mbit/s for MIMO 2x2 QAM64 UE
traffic pattern but removes random peaks or decreases and
can be achieved. If same 100 PRB is scheduled with MCS 0
allows to forecast traffic volume for each hour.
and transmit diversity only 2792 kbit/s can be achieved.
Depending on channel quality for the same amount of time-
frequency resources LTE downlink throughput can vary 54
times. It also shows that to transfer the same data amount time-
frequency radio resources can be very different. In a live LTE
network, users are in different radio conditions and we cannot
define direct relation between individual user experienced
throughput to total traffic carried per cell.
B. LTE network level throughput to traffic relation
The number of resources which can be scheduled to a
single user depends on cell load and scheduler activity.
The threshold can be defined to evaluate every cell for
every hour if it is considered as limited QoE based on criteria
from available performance indicators. This criterion can be
the average downlink experienced user throughput. It is
Fig. 5. One week of LTE cell traffic with one-hour granularity compared to calculated by the 3GPP recommended formula.
traffic trend modeled using the fbProphet algorithm. Modeled traffic trend
represents the proportion of traffic for each hour of the day. Average user throughput is measured as

Authorized licensed use limited to: Vilnius Gediminas Technical University. Downloaded on November 16,2020 at 08:23:16 UTC from IEEE Xplore. Restrictions apply.
explained before every cell has unique user and traffic
(4)
distribution, so for the same amount of data different spectrum
resources will be utilized. Here it is suggested to evaluate
where TotPDCPVolDataExclLastTTIs is the total Packet Data every cell individually, as it would show the more accurate
Convergence Protocol Service Data Unit volume transferred amount of traffic until the high load threshold will be reached.
per cell during Reporting Output Period and Using regression we can calculate spectral efficiency for every
TotEffectiveTimeExclLastTTIs is the time used to send the cell and find how much traffic can be carried until PRB
information [8]. Measurements are aggregated for all users in threshold will be reached.
a cell. Fig. 6 shows the relation between Average user
throughput and cell traffic volume. If average user throughput
is considered as criteria, live network cells show what
correlation between cell traffic volume and average user
throughput is very low.

Fig. 7. The test set LTE cells PRB utilization vs average user throughput of
1 busy hour. Linear correlation is visible, as PRB load increases average user
experienced throughput decreasing.

Fig. 8 shows an example of different LTE cells with


Fig. 6. The test set LTE cells traffic volume versus Average user throughput different spectral efficiency.
of 1 busy hour. Every dot represents 1 cell during the busy hour. Correlation
between traffic amount and average user throughput in the cell is low.

Cell congestion or low average user throughput may be


calculated from the PRB load. The relation between PRB
utilization and user throughput explained in [9]. Fig. 7 shows
the live network test cell set PRB load versus average user
throughput. Clear linear correlation is visible. Moving closer
to 100% PRB load average user throughput gets low and
deviation also decreases. The PRB load threshold, above
which cell is considered as congested, is defined depending on
commercial requirements. LTE 800MHz cell with 10MHz
bandwidth can be considered as loaded if utilization is 60% or
20 PRB unused, and LTE 1800MHz with 20MHz bandwidth
PRB utilization is above 80% (20 PRB left).
From PRB load to traffic relation we can calculate spectral
efficiency. Using spectral efficiency we can calculate how
much traffic cell is capable to carry until reaching the PRB
load threshold.
Fig. 8. Two example cells from the same LTE eNodeB and same LTE Band
(5) with different spectral efficiency. Every dot represents carried traffic amount

and average PRB used to carry it for the same cell per hour. Total of 720
points per cell (30 days 24 hours) used to create a regression.
Here SE is the spectral efficiency bits/Hz/second, Data
Volume is the traffic carried in the cell during the
V. CONCLUSIONS
measurement period in bits, PRB Util is average PRB
Utilization during the measurement period, Cell Bandwidth is In this article the possibility to use time series forecast
LTE cell bandwidth in Hz, measurement period - data algorithm - fbProphet for LTE networks evaluated. Such an
collection period in s (e.g. 1 hour). Knowing spectral approach highly used for internet services, but not common
efficiency we can calculate traffic amount until the PRB for mobile networks evaluation. Using fbProphet algorithm, it
threshold will be reached. In industry average network-level is possible to forecast daily cell traffic with R² and explained
approach is used to evaluate and classify cells, but as variance score 0.9 accuracies;

Authorized licensed use limited to: Vilnius Gediminas Technical University. Downloaded on November 16,2020 at 08:23:16 UTC from IEEE Xplore. Restrictions apply.
The Facebook fbProphet algorithm can forecast daily LTE [4] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O.
traffic and support calculation of future traffic for each hour; Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J.
Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E.
Forecasted traffic amount can be compared to individual Duchesnay, "Scikit-learn: Machine learning in Python," Journal of
Machine Learning Research, vol. 12, pp. 2825–2830, 2011.
spectral efficiency (bits/s/Hz) of every cell and future load in
physical resource block utilization can be forecasted; [5] F. Capozzi, G. Piro, L. A. Grieco, G. Boggia and P. Camarda,
"Downlink Packet Scheduling in LTE Cellular Networks: Key Design
Knowing future LTE cell load can help mobile network Issues and a Survey," in IEEE Communications Surveys & Tutorials,
vol. 15, no. 2, pp. 678–700, Second Quarter 2013.
operator to guarantee acceptable user experience and avoid
overinvestment. [6] M. T. Kawser, N. I. B. Hamid, N. Hasan, M. S. Alam, M. M. Rahman,
"Downlink SNR to CQI mapping for different multiple antenna
techniques in LTE", International Journal of Information and
REFERENCES Electronics Engineering, vol. 2, no. 5, 2012.
[7] Evolved UTRA; Physical layer procedures (Release 8). 3 GPP TR
[1] Ericsson Mobility Report, June 2018. [Online]. Available: 36.213 V8.3.0, May 2008.
www.ericsson.com/mobility-report. [8] “Key Performance Indicators (KPI) for Evolved Universal Terrestrial
[2] S. J. Taylor and B. Letham. (2017). Forecasting at scale. Facebook Radio Access Network (E UTRAN): Definitions,” TS 32.450, Jun.
Research Tech. Rep. [Online]. Available: 2010, Version 9.1.0 Release 9
https://facebookincubator.github.io/prophet/static/prophet_paper_201 [9] Salo, Jari, and Eduardo Zacarías. "Analysis of LTE Radio Load and
70113.pdf. User Throughput.", International Journal of Computer Networks &
[3] M. Xu, Q. Wang and Q. Lin, "Hybrid holiday traffic predictions in Communications (IJCNC) Vol.9, No.6, November 2017
cellular networks," NOMS 2018 - 2018 IEEE/IFIP Network Operations
and Management Symposium, Taipei, 2018, pp. 1–6.

Authorized licensed
View publication stats use limited to: Vilnius Gediminas Technical University. Downloaded on November 16,2020 at 08:23:16 UTC from IEEE Xplore. Restrictions apply.

You might also like