Conference PaperPDF Available

Abstract

With the constant development of networking technologies and the increase of internet userbase, traffic prediction is becoming a vital part of today’s network optimization. In this paper, we propose a method for network traffic prediction based on the PROPHET model. We examine its different parameters find their best configuration for diverse traffic types. Our research has shown that PROPHET is an accurate solution for backbone optical network traffic forecasting for a 14-day horizon.
Forecasting the network traffic with PROPHET
Pawe l Cembaluk, Jakub Aniszewski, Aleksandra
Knapi´nska[0000000326544893], and Krzysztof Walkowiak[0000000316863110]
Department of Systems and Computer Networks, Wroc law University of Science and
Technology, Wroc law, Poland
aleksandra.knapinska@pwr.edu.pl
Abstract. With the constant development of networking technologies
and the increase of internet userbase, traffic prediction is becoming a
vital part of today’s network optimization. In this paper, we propose a
method for network traffic prediction based on the PROPHET model. We
examine its different parameters find their best configuration for diverse
traffic types. Our research has shown that PROPHET is an accurate
solution for backbone optical network traffic forecasting for a 14-day
horizon.
Keywords: Traffic prediction ·Application-aware network ·Timeseries.
1 Introduction
The overall traffic in today’s backbone optical networks is seeing tremendous
growth in the last few years, especially during the COVID-19 pandemic. Vari-
ous network-based services are widely used in many areas, including education,
business, and entertainment. In turn, the summary traffic consists of multiple
low-bitrate flows and thus is characterized by strong daily and weekly season-
ality with periodically recurring patterns [3]. In light of the inevitable capacity
crunch, various solutions are being developed for more efficient use of the ex-
isting network resources. Multilayer application-aware network optimization [6]
is seen as a promising approach, in which different types of traffic related to
various services and applications are treated according to their specific qual-
ity of service (QoS) requirements. The knowledge regarding future volumes of
traffic used in network optimization algorithms can improve their efficiency and
decrease bandwidth blocking [7].
In this article, we present a traffic prediction method based on PROPHET – a
solution proposed by Facebook [5]. The PROPHET is a forecasting procedure for
time data series based on an additive model where non-linear trends are fit with
yearly, weekly, and daily seasonality, plus holiday effects. This method proved
its effectiveness in the prediction of cellular network traffic [1,2,8] but, to the
best of our knowledge, has not been studied in the context of backbone optical
This work was supported by National Science Centre, Poland under Grant
2019/35/B/ST7/04272.
P. Cembaluk et al.
networks. To fill this research gap, we propose a throughout study of various
parameters of PROPHET and evaluate it on datasets corresponding to diverse
types of backbone optical network traffic. The rest of the paper is organized as
follows. In Sec. 2, we define the problem and describe the chosen method. Sec.
3describes conducted numerical experiments. Sec. 4concludes this work.
2 Problem definition and method description
The problem considered in this work concerns the prediction of traffic in back-
bone optical networks. The created regression model attempts to predict future
volumes of a specific traffic type with a given sampling rate based on historical
data. The chosen forecast horizon is 14 days.
The PROPHET algorithm [5] provides automated timeseries forecasts that
can be tuned. In this paper, we explore the impact of three of the PROPHET
parameters on the prediction quality of various network traffic types. The first
parameter is the changepoint prior scale, and it determines how much the trend
changes at changepoints. Its value is a tradeoff between trend under- and over-
fitting. The second parameter is the seasonality prior scale. Its large value allows
the seasonality to fit large variations, and a lower one decreases the magnitude
of the seasonality. Because of the density of input data, as the last parameter,
we chose the number of automatically placed changepoints.
As a reference approach, we propose a Linear Regression model, which proved
to be an accurate method for network traffic forecasting [4]. Our implementation,
adjusted for long-term traffic forecasting, takes four input features: the hour of
the day, minute of the day, second of the day, and weekday.
3 Numerical experiments
The datasets in our experiments contain data based on the information from
the Seattle Internet Exchange Point (SIX), collected between 22 X 2019 and 23
XII 2019, with a sampling rate of 5 minutes. To simulate diverse traffic types
in a network, we added some fluctuations into the original data. To measure
how the traffic in created datasets differs from the collected aggregated Seattle
measurements, we use the mean absolute percentage error (MAPE). In this pa-
per, we consider three datasets of diverse profiles: traffic a (MAPE = 3.39%),
traffic b (MAPE = 8.21%) and traffic c (MAPE = 13.35%). Low MAPE val-
ues imply fewer fluctuations since the traffic is the most similar to the original
aggregated SIX traffic. Intuitively, high MAPE values mean more fluctuations.
The provided MAPE values are averaged across all the samples in each dataset.
For more details regarding the datasets creation, we refer to [4].
We explore the impact of PROPHET parameters on each traffic type, to
obtain the most accurate prediction. For the changepoint prior scale, we inves-
tigate the values of 0.0005, 0.001, 0.01, 0.1, and 0.5. For the seasonality prior
scale, the default value in the PROPHET is 10.0, which means almost no tuning.
In our experiments, the tested values are 0.1, 1.0, and 10.0. For the number of
216
Forecasting the network traffic with PROPHET
automatically placed changepoints, the default value in PROPHET is only 25,
and in our experiments, the tested values are 25, 50, and 100. Overall, for each
traffic type, we examine 45 parameter configurations.
Table 1: MAPE values for three best PROPHET configurations and the reference
approach for each traffic type
changepoint prior scale seasonality prior scale number of changepoints MAPE
0.001 1.0 25 0.07777
0.001 0.1 100 0.07780
0.001 10.0 25 0.07797
traffic a
linear regression 0.22504
0.001 1.0 100 0.11837
0.001 0.1 100 0.11845
0.001 10.0 100 0.11874
traffic b
linear regression 0.23897
0.001 1.0 25 0.15992
0.001 10.0 25 0.16004
0.0005 10.0 25 0.16007
traffic c
linear regression 0.25212
In Table 1, we present the results of the three best PROPHET configurations
for each traffic type together with their parameters, alongside the reference Lin-
ear Regression model. Intuitively, the most accurate forecasts were obtained for
the least fluctuating traffic a, and the highest prediction errors were noted for the
most difficult traffic c. Performing the experiments, we noticed that the most sig-
nificant PROPHET parameter is the changepoint prior scale. Its smallest value
resulted in the lowest MAPE across traffic types. In particular, for traffic b, the
difference between the worst result for the changepoint prior scale of 0.001 and
the best for 0.5 was as high as four percentage points. After setting its value
too low, i.e., 0.0001, the optimization algorithm failed – the model had to fall
back to the Newton algorithm, yielding worse results than its default one. The
seasonality prior scale did not seem to affect the results as much. In general,
lower values, i.e., 0.1 and 1.0, yielded lower errors than 10.0. The differences
between them were, however, only fractions of a percent. Furthermore, the num-
ber of changepoints did not appear to have any significant impact on the results
either. Its influence was the most noticeable for cases with a changepoint prior
scale of 0.01, where a higher number of changepoints resulted in more accurate
forecasts. Once again, the results differ by only a fraction of percent, whereas
more changepoints require more computational power for calculating the traffic
prediction. Thus, it may not be worth it to increase this parameter. Overall, the
predictions made by the PROPHET were of significantly higher quality than the
reference Linear Regression. Though this approach appears in the literature as
a prominent solution for short-term traffic forecasting (e.g., [4]), the PROPHET
is versatile for long-term network traffic prediction.
217
P. Cembaluk et al.
(a) (b)
Fig. 1: PROPHET performance for traffic a. (a) Rolling window MAPE for best
prediction model; (b) Prediction of the best model made for a 14 days horizon
An illustration of the results can be found in Fig. 1a and 1b where we present
plots for the best model that we received during the experiments for traffic a.
Fig. 1a shows MAPE with a 10% rolling window over the results. The grey points
are the errors for each predicted point in the horizon. Figure 1b shows the input
traffic data (in black) along with its predictions (in blue), and a 14-day forecast.
4 Conclusion
In this paper, we investigated the topic of optical backbone network traffic pre-
diction. The developed model based on Facebook’s PROPHET proved its high
performance making 14-day forecasts for diverse types of network traffic. For
each traffic type, various parameter configurations were tested to find the most
accurate model. We found that the most important parameter in the PROPHET
model for network traffic prediction is changepoint prior scale. In future work, we
plan to use traffic forecasts for the optimization of application-aware networks.
References
1. Chakraborty, P., et al.: A comparative study for time series forecasting within soft-
ware 5g networks. In: 14th Int. Conf. on Signal Proc. and Comm. Systems (2020)
2. Li, Y., et al.: Prophet model and gaussian process regression based user traffic
prediction in wireless networks. Sc. China Information Sciences 63(4) (2020)
3. NOKIA: The Nokia Deepfield Network Intelligence Report 2020. Tech. rep. (2020)
4. Szostak, D., et al.: Machine learning classification and regression approaches for
optical network traffic prediction. Electronics 10(13), 1578 (2021)
5. Taylor, S.J., et al.: Forecasting at scale. The American Statistician 72(1), 37–45
(2018)
6. Tomkos, I., et al.: Application aware multilayer control and optimization of elastic
wdm switched optical networks. In: 2018 OFC Conference (2018)
7. Troia, S., et al.: Deep learning-based traffic prediction for network optimization. In:
2018 ICTON Conference. IEEE (2018)
8. Xu, M., et al.: Hybrid holiday traffic predictions in cellular networks. In: NOMS
2018 IEEE/IFIP Network Operations and Management Symposium. IEEE (2018)
218
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Rapid growth of network traffic causes the need for the development of new network technologies. Artificial intelligence provides suitable tools to improve currently used network optimization methods. In this paper, we propose a procedure for network traffic prediction. Based on optical networks’ (and other network technologies) characteristics, we focus on the prediction of fixed bitrate levels called traffic levels. We develop and evaluate two approaches based on different supervised machine learning (ML) methods—classification and regression. We examine four different ML models with various selected features. The tested datasets are based on real traffic patterns provided by the Seattle Internet Exchange Point (SIX). Obtained results are analyzed using a new quality metric, which allows researchers to find the best forecasting algorithm in terms of network resources usage and operational costs. Our research shows that regression provides better results than classification in case of all analyzed datasets. Additionally, the final choice of the most appropriate ML algorithm and model should depend on the network operator expectations.
Conference Paper
Full-text available
In recent years, researchers realized that the analysis of traffic datasets can reveal valuable information for the management of mobile and metro-core networks. That is getting more and more true with the increase in the use of social media and Internet applications on mobile devices. In this work, we focus on deep learning methods to make prediction of traffic matrices that allow us to proactively optimize the resource allocations of optical backbone networks. Recurrent Neural Networks (RNNs) are designed for sequence prediction problems and they achieved great results in the past years in tasks like speech recognition, handwriting recognition and prediction of time series data. We investigated a particular type of RNN, the Gated Recurrent Units (GRU), able to achieve great accuracy (<7.4 of mean absolute error). Then, we used the predictions to dynamically and proactively allocate the resources of an optical network. Comparing numerical results of static vs dynamic allocation based on predictions, we can estimate a saving of 66.3% of the available capacity in the network, managing unexpected traffic peaks.
Article
User traffic prediction is an important topic for wireless network operators. A user traffic prediction method based on Prophet and Gaussian process regression is proposed in this paper. The proposed method first employs discrete wavelet transform to decompose the user traffic time series to high-frequency component and low-frequency component. The low-frequency component bears the long-range dependence of user network traffic, while the high-frequency component reveals the gusty and irregular fluctuations of user network traffic. Then Prophet model and Gaussian process regression are applied to predict the two components respectively based on the characteristics of the two components. Experimental results demonstrate that the proposed model outperforms the existing time series prediction method.
Forecasting at scale
  • S J Taylor
Taylor, S.J., et al.: Forecasting at scale. The American Statistician 72(1), 37-45 (2018)
Application aware multilayer control and optimization of elastic wdm switched optical networks
  • I Tomkos
Tomkos, I., et al.: Application aware multilayer control and optimization of elastic wdm switched optical networks. In: 2018 OFC Conference (2018)