You are on page 1of 6

Available online at www.sciencedirect.


Procedia Computer Science 103 (2017) 483 – 488

XIIth International Symposium «Intelligent Systems», INTELS’16, 5-7 October 2016, Moscow,

Neural network models of time series of network traffic intensities

N. Gabdrakhmanovaa,*, A. Gabdrakhmanovb
RUDN University. 6, Mikluho-Maklaya str., Moscow, 117198, Russia
TransTelecom Private Corp.


The paper presents the results of building neural network predictive models of the occupancy of the channel packet data. The
problem is solved by the example of time series observations of the intensities at the commutator switch port. The above
algorithm for constructing a neural network model based on the determination of the fractal dimension of the time series. The
autoregressive model is constructed as a parallel model. The developed models are compared in accordance with the rated value
of entered criteria for assessing the accuracy of the forecast.
© 2017The
© 2017 TheAuthors.
Publishedby by Elsevier
Elsevier B.V.B.V.
This is an open access article under the CC BY-NC-ND license
under responsibility of the scientific committee of the XIIth International Symposium «Intelligent Systems».
Peer-review under responsibility of the scientific committee of the XIIth International Symposium “Intelligent Systems”
Keywords: neural network; network traffic; time series; correlation integral.

1. Introduction

The purpose of the research is to obtain valid conclusions on the approach to the construction of a mathematical
model in the considered applied area. The relevance of the work is defined as follows. The existence of a large
number of diverse services in one physical channel during the highest load hours can lead to an overload of
commutation and routing devices on the network backbone which can lead to a massive failure of the range of
services. To prevent situations leading to failure of the network backbone equipment and to optimize the use of
network resources, the proper use of the broadband of the data-transmission circuit becomes the most important
task. It is essential to use rational effective methods of traffic management and occupancy control, which would be

* Corresponding author.
E-mail address:

1877-0509 © 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
Peer-review under responsibility of the scientific committee of the XIIth International Symposium “Intelligent Systems”
484 N. Gabdrakhmanova and A. Gabdrakhmanov / Procedia Computer Science 103 (2017) 483 – 488

based on the data provided, for example, by an instrument of forecasting traffic intensities on the previous values
basis. There are different approaches to solve this problem 1,2. The authors propose an unconventional approach to
the solution by using the field data.
A commutator switch of the 2nd level organization focused on providing the main line services was taken as an
example for a research. Traffic arriving at each port of the device, is a summarized traffic from groups of clients
from a specified area. Figure 1 shows the schema of the study measurement. Statistics was prepared using Cacti
software, SNMP-Interface statistic protocol. The information about the occupancy level of the network is more
useful for practical purposes. The information about the packet count per unit time can be misleading. Therefore, as
an observable variable the aggregated value x (t) – traffic intensity (bit) at the moment t is chosen. Duration time of
data is 10080 points, or 7 days.

Fig.1. The schema of the study measurement

Figure 2 shows the intensities graph measured at the GE 0 port. Each point on the graph represents the number of
data bit transmitted on the network backbone channel in a time interval of 1 minute.

Fig.2. The traffic intensity graph at the GE 0 port

It is necessary according to the data from the observations ^ x t `t

, where N – amount of observation points, to
solve the problem of identification of the timing series and to build the forecast for m anticipation steps.
N. Gabdrakhmanova and A. Gabdrakhmanov / Procedia Computer Science 103 (2017) 483 – 488 485

2. Selection and Argumentation of the Neural Network Data Model

^ x t `t
Neural network model is developed using nonlinear dynamics methods. It is believed that 1
measurements of some observational scalar components d1-dimensional dynamical system y . Here, the
dimensionality and the behavior of the dynamic system all in all are unknown. For a given time series it is necessary
to construct a model that would drink in the considered dynamics responsible for generating experimental
observation x (t).
According to Takens theorem3, the geometric structure of the system dynamics, which depends on many
variables, can be retrieved based on experimental observation ^ x t `t
in D-dimensional space constructed on the
basis of a new vector

^ x t , x t  1 ,}, x t  D  1 `
z t ,

where D t 2 d1  1 . Evolution of the points z t o z t  1 in the reconstructed space corresponds to the evolution
of the points y t o y t  1 in the original space. The search procedure of a suitable D is called embedding. The
minimum number D, at which the dynamic recovery is achieved, is called measurement of embeddings. In this
research the Grassberger-Procaccia algorithm was used, which allows to estimate by the time series the fractal
dimension d fc of a strange attractor of the dynamical system. The search algorithm described by the following
steps. Suppose, the m1-dimensional reconstruction is constructed through the time series, and a set of vectors
z t ( x t , x t  1 , } , x t  m1  1 is at the disposal of researchers. A number G is set and an interval of a
phase space is divided into cubes with a G side. For selected m1 and G correlation integral is calculated3:

N § 1 N ·
¦ ¦
C G lim ¨ T (G  z n  z k ) ¸ ,
N of N ¨ N 1 ¸
n 1© k 1, k z n ¹

where T r ^1,0, rr ! 00
T ˜ - Heaviside step function,
˜ - no matter which norm (Euclidean norm in our case),
N - number of points.
C G - is a measure of the probability that two points z n and z k on the attractor are spaced on the
distance G . Number N should be sufficiently large.
With small G correlation integral is C G ~ G d k , that is why the correlation dimensionality d fc can be defined
by obliquity of the correspondence lnC G from lnG . In other words, the correlation dimensionality d fc can be
defined as

lnC G
d fc m lim .
G o0 lnG

With the growth of embedding m1 the correlation dimensionality d fc m increases. If in this time series there is
a demonstration of a determinate chaos, then, starting with any m1=m0, the value of the relevant d fc m1 stops
growing. The reached value is taken as the attractor’s dimensionality of an initial system d f d fc ( m0 ) . If the
486 N. Gabdrakhmanova and A. Gabdrakhmanov / Procedia Computer Science 103 (2017) 483 – 488

growth continues without saturation, this indicates that the considered signal is a noise signal. The final value d fc
means that the considered signal can be recreated with help of a determinate system. Figure 3 shows the dependency
graph of correlation dimensionality from embedding dimensionality for the row. In the graph X-axis single out
values of embedding m1 on Y-axis – value of correlation dimensionality d fC . The found contour shows that fractal
dimension of a time series of measurement values equal 4,53, it follows that embedding dimension for dynamic
system recoveries is D 2 ª¬ d f º¼  1 11 .

Fig. 3. Correspondence of correlation dimensionality from embedding dimension for the row

^ x t `t
The following problem can be formulated in the following manner. There is a timing series 1

^ x t , x t  1 ,}, x t  D  1 `
reconstruction parameters are set (in this case D=11). For N1 vectors z t the

values of required function are known: F t F z t (as it is known that next to z (t ) are elements of time series).
It is required to find the value of the required function in the new point z (t ) F z . Otherwise, it is required to
approximate the required function F on the given N1 points.
Artificial neural networks (NN) solve the problem of approximation of functions of more than one variable
according to the learning sample by dipping time series in the multidimensional space. To increase the accuracy of
forecasting of NN models it is proposed to use a dual level algorithm of model’s constructing. The problem of
constructing in the reconstructed space of the subdomains is solved on the first level, where it is possible to
construct a low-mode model. For this purpose, a multitude clustering method ^ z t `t
is used. The construction of
local NN models in each sub domain occurs on the second level. For this example, two clusters are designed. For
each cluster NN models are developed. The neural network type is a multilayer perceptron. When learning by back
propagation, the NN weight number is found by minimizing the functional 4:

R w E s w  P Ec w ,

where E s w - standard measure of neural network eff (mean square error), Ec w – penalty for complexity, P –
regularization parameter. By way of example,

1 wk
2 ³ wx
E c w, k Fnc x , w M x dx ,

Fnc x , w – executable model of function,

N. Gabdrakhmanova and A. Gabdrakhmanov / Procedia Computer Science 103 (2017) 483 – 488 487

M x – any function weighting, which defines the zone of input space where Fnc x , w should be differentiable.
For constructed models on a test variety are calculated using formulas: modulus average value of the relative
error of a forecast

x t  xˆ t

MSE , (1)
t 1
x t

Max and Min value of modulus of the relative error

x t  xˆ t
MaxE m ax , (2)
t x t

x t  xˆ t
MinE min , (3)
t x t

where x̂ t - value of the model x t .

The entered values criteria for constructed model: MSE= 0,011; MaxE=0,573; MinE=0,087.
Figure 4 shows the graph of collating of the forecasted data (in Output) with the field studies data (in). On the
graph singled out on the X-axis - the number of point, on the Y-axis – the standardized value of the traffic intensity.

Fig. 4. The graph of forecasted data (in Output) with the field studies data (in).

Table 1 shows the actual values x t , model of forecasted value x̂ t , model of a relative error of the forecast
x t  xˆ t
H t .
x t
As a parallel model, using the same data, autoregressive scheme was constructed. The comparation of calculated
values of the set criteria (1-3) for the estimation of the model’s forecasted properties indicates that the NN is the
most suitable for solving this problem.
488 N. Gabdrakhmanova and A. Gabdrakhmanov / Procedia Computer Science 103 (2017) 483 – 488

Table1. Fragments of the comparation of real and model indicator values on the test multiplicity

x t ˜ 10 xˆ t ˜ 10 H t ˜ 10
7 7 2

5,01 1,78 1,69 5,02

5,02 1,82 1,70 6,51
5,03 1,75 1,72 1,86
5,04 1,69 1,70 1,05
5,05 1,77 1,69 4,13
5,06 1,81 1,70 5,91
5,07 1,70 1,71 0,736
5,08 1,65 1,70 2,64

5,09 1,70 1,69 6,65

5,1 1,72 1,68 2,48

3. Conclusions

1. A statistical analysis of the actual data of the functioning network was carried out.
2. Based on analysis of time series as an observational nonlinear dynamical system the argumentation and
construction of neural network forecast model was carried out.
3. Positive results on network traffic forecasting by applying autoregressive scheme and NN models which can be
used for an efficient management and resource optimizing in the network functioning process were deduced.
4. Comparation of the forecasting results with models different approaches demonstrate the advantages of
dynamical approach when modeling with neural network technology. The conclusion is based on the basis of
calculations and comparation of the entered criteria values.


1. Heyman DP, Tabatabai A, Lakshman TV. Statistical analysis and simulation study of video teleconference traffic in ATM network s. IEEE
Transactions on Circuits and Systems for Video Technology 1992; 2, p.49-59.
2. Zhani MF, Elbiaze H. Analysis and Prediction of Real Network Traffic. Journal of Networks 2009; 4(9).
3. Potapov AB. Time-series analysis: when dynamical algorithms can be used. Proc. 5 Int. Specialist Workshop Nonlinear Dynamics of
Electronic Systems. June 26-27,1997; p. 388-393.
4. Haykin S. Neural Networks: A Comprehensive Foundation Second Edition 2006.