- 1 -
中国科技论文在线
A Novel Satellite Traffic Flow Prediction Scheme Based on
Wavelet Neural Network and Genetic Algorithm
Li Ning, Wang Pengfei, Han Ke, Deng Zhongliang
*
5
(School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing
100867)
Brief author introduction:Ning Li, female, ., associate professor at the School of Electronic Engineering,
Beijing University of Posts and Telecommunications. Research Interests: Information Security, Satellite
Communications. E
Abstract: Satellite communications network is a kind of delay tolerant network with characters of
intermittent connection, long packet queuing time, complex and uncertain delay time, etc. The call
congestion rate of new business is higher in delay tolerant network which leads to the reduction of 10
communication quality and the high error code rate. The existing traffic flow prediction scheme is hard
to meet new business and individual requirements. It becomes more and more important to study the
most effective traffic prediction scheme so as to increase the satellite bandwidth utilization. In this
paper, BP neural network (BPNN) has been used as regression model, and the Genetic Algorithm (GA)
has been used to search the weights matrix and thresholds of BP neural network. Then the satellite 15
traffic flow prediction scheme based on wavelet neural network and genetic algorithm (WNNGA) was
put forward. Simulation results with the real traffic traces show that the proposed method can more
accurately predict the future of the satellite network traffic and decrease the processing time compared
with previous model method.
Key words: Traffic Prediction; Satellite Communications; Wavelet Neural Network; Genetic 20
Algorithm
0 Introduction
With the development of the wireless communication technology, the demand for broadband
communication services has increased rapidly. Satellite communications network plays an 25
important role not only by providing broadband access directly, but also being part of the
worldwide core network
[1]
. In order to guarantee satellite network resources reasonable use, it is
necessary to realize efficient bandwidth allocation among users for broadband satellite
communications system. To accurately predict network traffic on a satellite network design and
management, conflict control and dynamic bandwidth allocation plays a vital role. Therefore, it 30
becomes more and more important to study the most effective traffic prediction scheme so as to
increase the satellite bandwidth utilization.
For traffic prediction, many researchers have proposed a variety of effective schemes. Secchi
and Barsocchi propose the satellite network service rate allocation method based on linear
quadratic control which can more accurately predict the future of the satellite network traffic
[2]
. 35
But the network traffic is modeled as linear dynamic system driven by independent random
process and it has a lot of limits. Yueqiu J proposes the traffic prediction method based on BP
(Back Propagation) neural network
[3]
. BP neutral network can learn from examples and has a
strong ability of studying. However, to make neutral network's results satisfied, it needs a lot of
training data which led to a low efficiency of the traffic prediction. Delli designs the bandwidth 40
allocation mechanism on the basis of MMPP (Markov Modulated Poisson Process) model
[4]
.
Nevertheless, Markov model is a random process with no aftereffect so it is very difficult to be
used widely in practice due to its high complexity.
As the rapid development of satellite communications, satellite network traffic has become a
nonlinear and time scale system. Since self-similarity, long-range correlation, chaos and breaking 45
characteristic are the common properties of traffic in the satellite networks
[5]
, it is crucial to
- 2 -
中国科技论文在线
consider these impacts on network management and traffic control.
1 System Model
The system configuration is shown in . As shown in , in the satellite uplink, the
data of UTs are collected by STs. Then via onboard transponder, the data are transmitted to the GS 50
(Gateway Station) in the ground. At the end, data reach to the ground IP backbone network. The
transmission in satellite downlink is reverse. During the whole transmission, many operations are
centralized processed by NCC (Network Control Center) in the ground, such as traffic control.
Compared with ground network, the most significant difference lies in longer transmission delay.
From ST to NCC, transmission delay of single hop is about 250ms in GEO (Geostationary Earth 55
Orbit) broadband satellite communications system. To a large extent, one of the consequences
caused by delay is that the bandwidth allocation orders provided by NCC are based on the traffic
prediction of next period. This architecture also results in the demand for accurate network traffic
prediction
[6]
.
60
The system model
With the rapid increase in satellite network applications, satellite network traffic will also be
a very substantial increase. Satellite network traffic has become a nonlinear and time scale system
with the properties of self-similarity, long-range correlation, chaos and breaking characteristic
[7]
.
As it is difficult to evaluate the relationship between each factor and the network traffic function 65
by precise mathematical model, the network traffic prediction has become a system problem of
complex time series regression.
2 The proposed method
In the proposed method, BPNN had been used as core algorithm, and GA had been used as
search parameters tool to search the best weights matrix and thresholds of BPNN. The following 70
parts would introduce the theories of both BPNN and GA.
The basic theory of BPNN
BPNN is a more layers hierarchical neural network with upper neurons full associated with
lower neurons. When a couple of learning samples is supplied to the network, the transferred value
is propagated from the input layer through middle layer to the output layer, and we can get neural 75
network input response from neurons in output layer. Along the direction of reducing the error
between expected output and actual output, connection weights are adjusted from the output layer
to every middle layer, and ultimately to the input layer. With the ongoing amendment by this
back-propagation, the correct rate for the network response to input also increases continuously.
As BP algorithm implements middle hidden layer and has a corresponding learning rules to follow, 80
- 3 -
中国科技论文在线
it has the ability to identify the non-linear pattern
[8]
.
Based on characteristics of neurons, we can get typical neuron models, shown in Figure 2. In
this figure, f() is a non-linear function, x1 , x2 ,…,xn are n inputs related to neurons; y1 is the
output; wj1 , wj2 ,…, wjn are weight factor of x1 , x2 ,…,xn ,respectively; j is the threshold of
the output. 85
Neural network model
Mathematical expression of neuron model:
( )j jy f S (1)
1
n
j ji i j
i
S w x
(2) 90
Where, column vector X is input vector, row vector Wj is the connection weight vector for
neuron j, Sj represents the input of neurons.
Typical BP network has three layers, input layer, hidden layer (middle layer) and output layer.
Full connection is applied between layers.
Consider that there are P pairs of patterns, when the Pth pattern is in act, the output layer 95
error Ep is defined as:
2
1
1
( )
2
m
p jp jp
j
E t y
(3)
The input of jth unit in output layer is:
0
q
j jk k
k
S w z
(4)
p
yj
j
E
S
(5) 100
Expression (6) is defined to be error signal term of output layer. According to expression (3),
As long as f() is differentiable, then we have :
( ) ( )
p
yj j j yj j
j
E
t y f S
S
(6)
( )yj jf S is the derivative of transfer function of jth neuron in output layer to its clear input
Sj .And: 105
- 4 -
中国科技论文在线
1 1 1
1
1
1
1
( ) ( )
( ) ( )
( ) ( )
m
yj j z
j
z
m
T
k y z yj jk zk k zk z
j
zqm
yj jq zq q
j
w f S
W f w f S
w f S
(7)
The Theory of GA
Intelligent algorithms such as genetic algorithms and simulated annealing algorithms have
widely been applied to the field of large scale data analysis and data processing. It is potential for
the high-performance distributed computing technologies or platforms to further increase the 110
execution efficiency of these traditional intelligent algorithms
[9]
. The evolution of GA usually
starts from a population of randomly generated individuals, and is an iterative process, with the
population in each iteration called a generation. In each generation, the fitness of every individual
in the population is evaluated; the fitness is usually the value of the objective function in the
optimization problem being solved. The more fit individuals are stochastically selected from the 115
current population, and each individual's genome is modified (recombined and possibly randomly
mutated) to form a new generation. The new generation of candidate solutions is then used in the
next iteration of the algorithm
[10]
.
Fitness function
The weights matrix and thresholds is obtained by individuals and the output of system could 120
be predicted by the training data. Then the fitness value F is calculated by expression (8):
1
( ( ))
n
i j
i
F k abs y S
(8)
Where n is the number of network output unit, i
y
is the expected output of ith unit, j
S
is
the prediction value of ith unit’s output and k is a coefficient used to adjust the fitness value range.
Selecting operation 125
In this paper, we choose the roulette wheel selection for selecting potentially useful solutions
for recombination. The probability of selection with each individual ip is calculated by:
1
i n
i
j i
k
p
k
F
F
(9)
Where iF is the fitness value of ith individual, k is coefficient and n is the number of
individuals in the population. 130
Crossover operation
The crossover operation of kth individual chromosome ka and lth individual chromosome
ja is shown in expression (10) and (11):
- 5 -
中国科技论文在线
(1 )kj kj lja a b a b (10)
(1 )lj lj kja a b a b (11) 135
Where b is a random number in [0,1].
Mutation operation
The mutation operation of ith individual’s jth chromosome ija is obtained by (12):
max
min
( ) ( ),
( ) ( ),
ij ij
ij
ij ij
a a a f g r
a
a a a f g r
(12)
Where maxa is the upper bound of chromosome ija , mina is the lower 140
bound. 2 max( ) (1 / )f g r g G , 2r is a random number, g is the number of iterations, maxG is
the max number of evolution, r is a random number in [0,1].
The process of WNNGA
The flow chart of WNNGA 145
The process of WNNGA can be divided into three parts, which are determining the structure
of BPNN, optimizing by GA and the prediction by BPNN. First of all, the structure of BPNN is
determined by the input and output of fitting function which is used to determine the length of the
individuals in GA. Then it is need to optimize the weights and threshold in BPNN by GA. The
information of weights and threshold is contained in each individual and the individual fitness 150
value is calculated by the fitness function. Through selection, crossover and mutation operation,
the GA is able to find the best individual which is corresponding to the fitness value. At last, the
BPNN need to initial weights and threshold with the best individual selected by GA and then the
- 6 -
中国科技论文在线
BPNN could be trained and used to predict the output of function.
The WNNGA prediction model is a combination of WNN model and GA. In this model, 155
network traffic is resolved into low and high frequency data at first. Low and high frequency
traffic flow data contains the train data and test data respectively and then the BPNN is trained by
train data. Finally, after putting the test data into trained BPNN and the prediction output of
original test data could be obtained by the superposition of high and low frequency data. The flow
chart of WNNGA is shown in . 160
3 Simulations and analysis
In order to measure the performance of the proposed mechanisms compared with simple
ARIMA model and BP neutral network model, the experiments are carried out based on one
segment of CERNET (China Education and Research Network) traffic data which is taken from
the router node in Beijing network on July 1st, 2015. The trace statistics summary of CERNET is 165
shown in Table 1.
Table 1 Trace statistics summary of CERNET
Bandwidth Total packets Total flows Utilization Flows in progress MTU
100Gbps 1384M 5461K 34% 84600 12000
The program is constructed in OPNET and simulates the full operation of satellite network
shown in , as well as the DVB-RCS for the return link of a satellite network. Three error
indexes are employed to estimate the accuracy performance of all the involved forecasting models, 170
including the Mean Absolute Error (MAE), the Mean Absolute Percentage Error (MAPE) and the
Standard Deviation (SD).
Mean Absolute Error:
1
1 ˆ
M
i
MAE X i X i
M
(13)
Mean Absolute Percentage Error: 175
1
ˆ1
ˆ
M
i
X i X i
MAPE
M X i
(14)
Standard Deviation:
2
1
1 ˆ
1
M
i
SD X i X i
M
(15)
Where X t is the measured time series, Xˆ i is the forecasted time series and ‘M’
is the number of the terms in the X t series. 180
- 7 -
中国科技论文在线
Bandwidth utilization performances versus number of UTs per ST
presents the bandwidth utilization percentage of the three different mechanisms for an
increasing number of UTs per ST. The BP neutral network model has much better performance
than the ARIMA due to its high self-adaptive ability. WNNGA has slightly better performance 185
especially if the number of USs is between four and eight. Compared to BP neutral network model,
WNNGA has an improvement of 5% for the most part. The three algorithms have identical
performance after 18 UTs, as the number of requested slots is larger than the one that the NCC can
give to the STs. So, all proposed mechanisms get the same number of available slots and have the
same performance. 190
- 8 -
中国科技论文在线
Requests for different number of SS. A Request for 4 SS, B Request for 9 SS, C Request for 18 SS
195
- 9 -
中国科技论文在线
Results of the traffic prediction error by different mechanisms for eight UTs
shows the requests for different number of SS. Where A represents request for 4 SS, B 200
represents request for 9 SS, and C represents request for 18 SS. shows the difference
between real needed number of bytes and predicted ones by the three different mechanisms for
eight UTs per ST. These figures show that the proposed mechanism attains much better
performance than ARIMA model and BP neutral network, as the number of predicted bytes is
closer to the needed one (real queue). From the presented results, we conclude that WNNGA is 205
preferred for satellite network traffic and it has also the lowest computational requirements and the
lowest requirements in amount of memory, as raised from the computational analysis made. The
results of the accuracy and the time performance estimation for these predictions are given in
Tables 2 and 3, respectively.
Table 2 Analysis of the accuracies of the predictions 210
Indexes WNNGA BP neutral network ARIMA model
MAE
MAPE
SD
Table 3 Analysis of the time performance of the predictions
Indexes WNNGA BP neutral network ARIMA model
Times(s)
From Tables 2, it can be analyzed that the proposed WNNGA model has the best forecasting
accuracy in all step predictions among the three built forecasting models and table 3 indicates that
WNNGA has a shorter processing time than ARIMA model and BP neutral network. 215
4 Conclusion
Following a bibliographic search for traffic prediction in satellite networks, in this paper we
select BPNN model and GA as the most suitable algorithm in order to improve the prediction of
traffic in satellite networks. BPNN has unique advantages in describing the characteristics of 220
satellite network traffic with the ability of abstract conclusion and the self-adaptive ability while
GA is a statistical method used to describe variability among observed which can resolve the
problem of BPNN model in the internal parallel computation. Simulation results with the real
traffic traces show that the proposed method can more accurately predict the future of the satellite
network traffic and decrease the processing time compared with previous model method. 225
- 10 -
中国科技论文在线
References
[1] Markos P. Anastasopoulos, Dionysia K. Petraki, et al., Call Admission Control Scheme for Multiclass Services
under Rain Fading for Satellite Networks, IEEE Transaction on Wireless communications, 8 (2009), No. 5, 230
2473-2483
[2] Secchi R, Barsocchi P, Davoli F, Linear quadratic control of service rate allocation in a satellite network,
Communications, IET, 4 (2010), No. 13, 1580-1593
[3] Yueqiu J, Hang L, Study on Bandwidth Allocation of Broadband Multimedia Communication Satellite System,
3rd International Conference on Intelligent Networks and Intelligent Systems, (2010) ,506-508 235
[4] Delli Priscoli F, Pompili D, A demand-assignment algorithm based on a Markov modulated chain prediction
model for satellite bandwidth allocation, Wirel. Netw., 15 (2009), 999-1012
[5] Ullah F, Khan G M, Mahmud S A. Intelligent Bandwidth Management Using Fast Learning Neural
Networks[J]. IEEE International Conference on High Performance Computing and Communication and IEEE
International Conference on Embedded Software and Systems, 2012, 248(4):867 - 872. 240
[6] Gao B, Zhang Q, Liang Y, et al., Predicting self-similar networking traffic based on EMD and ARMA, Journal
on Communications, 32 (2011), No. 04, 47-56 (in Chinese)
[7] Li Song, Liu Lijun, Liu Yingpeng. Prediction for chaotic time series of optimized BP neural network based on
modified PSO. Computer Engineering and Application, 2013, 49(6):245-248..
[8] Zhu Li; Qin Lei; Xue Kouying; Zhang Xinyan, "A Novel BP Neural Network Model for Traffic Prediction of 245
Next Generation Network," Natural Computation, 2009. ICNC '09. Fifth International Conference on , , no.,
,38, 14-16 Aug. 2009
[9] Kewen Liu; Haiming Zhou; Zhanyong Yang; Fumin Qu, "Application of BP Neural Network for Line Losses
Calculation Based on Quantum Genetic Algorithm," Computational Intelligence and Design (ISCID), 2011 Fourth
International Symposium on , , no., ,7, 28-30 Oct. 2011 250
[10] Satsri S, Ardhan S, Chutchavong V, et al., ANN based NGN IP traffic prediction in Thailand, International
Conference on Control, Automation and Systems, (2007), 2154-2157
基于小波神经网络和遗传算法的卫星255
网络视频流量预测方法
李宁，王鹏飞，韩可，邓中亮
（北京邮电大学电子工程学院，北京 100876）
摘要：卫星通信网络是一种具有间歇性连接、数据包排队时间长、复杂和不确定等特性的网
络。由于新业务的呼叫阻塞率较高，导致卫星系统通信质量下降以及误码率的提高。现有的260
流量预测方案很难满足新的业务和用户需求。因此，本文利用 BP 神经网络（BPNN）作为
回归模型，同时将遗传算法（GA）用于搜索 BP 神经网络中的权重矩阵和阈值，提出了基
于小波神经网络和遗传算法的卫星网络视频流量预测方法。通过仿真结果表明，与以前的模
型方法相比，该方法可以更准确地预测卫星网络流量，同时减少系统的处理时间。
关键词：流量预测；卫星通信；神经网络；遗传算法 265
中图分类号：
