Alexandria Engineering Journal (2021) 60, 1137–1145

H O S T E D BY
Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

Fuzzy clustering to classify several time series

models with fractional Brownian motion errors
Mohammad Reza Mahmoudi b, Dumitru Baleanu c,d, Sultan Noman Qasem e,f,
Amirhosein Mosavi g,h,*, Shahab S. Band a,i,*

a
Institute of Research and Development, Duy Tan University, Da Nang 550000, Vietnam
b
Department of Statistics, Faculty of Science, Fasa University, Fasa, Fars, Iran
c
Department of Mathematics, Faculty of Art and Sciences, Cankaya University, Balgat 06530, Ankara, Turkey
d
Institute of Space Sciences, Magurele-Bucharest, Romania
e
Computer Science Department, College of Computer and Information Sciences, Al Imam Mohammad Ibn Saud Islamic
University (IMSIU), Riyadh, Saudi Arabia
f
Computer Science Department, Faculty of Applied Science, Taiz University, Taiz, Yemen
g
Environmental Quality, Atmospheric Science and Climate Change Research Group, Ton Duc Thang University, Ho Chi Minh
City, Vietnam
h
Faculty of Environment and Labour Safety, Ton Duc Thang University, Ho Chi Minh City, Vietnam
i
Future Technology Research Center, College of Future, National Yunlin University of Science and Technology, 123 University
Road, Section 3, Douliou, Yunlin 64002, Taiwan, ROC

Received 12 January 2020; revised 1 September 2020; accepted 7 October 2020

Available online 23 October 2020

KEYWORDS Abstract In real world problems, scientists aim to classify and cluster several time series processes
Classification; that can be used for a dataset. In this research, for the first time, based on fuzzy clustering method,
Fuzzy clustering; an approach is applied to classify and cluster several time series models with fractional Brownian
Fractional Brownian motion errors as candidates to fit on a dataset. The ability of the introduced technique is studied
motion; using simulation and real world example.
Non-stationary; Ó 2020 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria
Stationary; University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
Time series; licenses/by-nc-nd/4.0/).
RDI

* Corresponding authors at: Environmental Quality, Atmospheric Science and Climate Change Research Group, Ton Duc Thang University, Ho
Chi Minh City, Vietnam (A. Mosavi); Faculty of Environment and Labour Safety, Ton Duc Thang University, Ho Chi Minh City, Vietnam (A.
Mosavi); Future Technology Research Center, College of Future, National Yunlin University of Science and Technology, 123 University Road,
Section 3, Douliou, Yunlin 64002, Taiwan, ROC; Institute of Research and Development, Duy Tan University, Da Nang 550000, Vietnam (S. S.
Band).
E-mail addresses: [email protected], [email protected] (M.R. Mahmoudi), [email protected] (D.
Baleanu), [email protected] (S.N. Qasem), [email protected] (A. Mosavi), [email protected] (S. S.
Band).
Peer review under responsibility of Faculty of Engineering, Alexandria University.
https://doi.org/10.1016/j.aej.2020.10.037
1110-0168 Ó 2020 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1138 M.R. Mahmoudi et al.

1. Introduction functions with unknown parameters. In other words,

fi ; i ¼ 1;    ; k, are known time series models.
From (2.1), we have
Recently, statistical methods including regression and time ser-
ies models [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20, ProjectionðXt jXt1 ; Xt2 ;   Þ ¼ fi ðXt1 ; Xt2 ;   Þ;
21,22,23,24], mathematical methods including numerical anal-
which is a projection for Xt given Xt1 ; Xt2 ;   .
ysis and optimization [25–42],[43], and artificial intelligence
Some researches focused on testing H0 : f1 ¼    ¼ fk : In
methods including machine learning and deep learning [44–
other words, they considered the comparison of k processes.
50], frequently applied to model the natural phenomena. In
Some others studied the clustering of several independent pro-
data modeling problems such as time series modeling in agri-
cesses. But the clustering of several time series (specially, non-
cultural, biological, climatological, economic, financial, hydro-
stationary processes or time series with fractional Brownian
logical, management and signal processing researches,
motion noises) fitted on a specific dataset is lacked. In this
scientists aim to cluster and classify the stochastic mechanism
study, we are interested to know the rates of similarities of
of several time series that can be fitted on a specific dataset.
these k time series models and to cluster these models based
They usually use the values of different selection criterions
on their similarities.
likes R-Squared (R2), mean absolute or square error (MAE
Consider n observations X1 ;    ; Xn ; as a sample dataset of
or MSE), Akaike information criterion (AIC) and Bayesian
size n from Xt . We are interested in clustering f1 ;    ; fk . The
information criterion (BIC), to compare models and then the
following steps should be run:
model with the minimum MSE (or AIC or BIC) and maximum
R2 will be selected as the best. For example, in hydrological
Step (1): At first, all of time series models are fitted on data-
researches, Zarei and Mahmoudi [14], studied several models
set by
about the changes in a drought index. Also Zarei et al.
[51,52] studied the variability in drought’s trend and pattern
using different time series models.
The problem of comparison and classification of two or bðiÞ ¼ fbi ðXt1 ; Xt2 ;   Þ; i ¼ 1;    ; k:
X ð2:2Þ
t
several processes have been considered in many statistical
researches [6–8],[20,21],[53–85]. Recently, Mahmoudi et al. In other words, we estimate the unknown parameters of the
[6], developed a novel approach to compare two stationary known functions f1 ;    ; fk :
processes. Their proposed test was more powerful than other
methods. Step (2): The predicted values of X t is estimated based on all
Most of previous studies focused on two following cases: k models by
(1) comparison of two or several processes; (2) clustering of
several independent processes. Therefore the clustering of sev-
eral time series fitted on a specific dataset (specially, non- bðiÞ ;    ; X
bðiÞ ; i ¼ 1;    ; k:
X 1 n
stationary processes or time series with fractional Brownian
motion noises) was lacked. In this research, for the first time,
In other words, X bðiÞ ; j ¼ 1;    ; n; is the predicted value of
j
an approach is applied to classify and cluster several stationary
dataset Xj ; j ¼ 1;    ; n; based on im fitted time series model.
and non-stationary time series models fitted on a dataset, spe-  ð1Þ 
cially when the noises follow fractional Brownian motion pro- Now, we have k predicted values X b ;; Xbð1Þ , . . .,
1 n
cesses. At first, all of time series models are fitted on observed  ðkÞ 
Xb ;; X bðkÞ , for observed dataset X1 ;    ; Xn : Note that
dataset. Then, for observed dataset, the predicted values of all 1 n
 ð1Þ   ðkÞ 
models are estimated based on fitted models formulas. Finally, Xb ;; X bð1Þ , . . ., X b ;; X bðkÞ , can be considered as a
1 n 1 n
the fuzzy clustering technique is used to cluster the models
based on the similarities between predicted values. The ability sample of size n from f1 ;    ; fk ; respectively.
of the proposed technique is studied using simulation and real  ð1Þ 
world example. The proposed technique has many advantages. Step (3): Now, using the couples X b ð1Þ , . . .,
b ;; X
1 n
First, the power of this method is acceptable. Second, the  ðkÞ 
X b ðkÞ , the fuzzy clustering methods is applied
b ;; X
introduced approach is not too computational and time con- 1 n
suming. More ever, this technique can be applied to cluster dif- to know the rates of similarities of these k time series mod-
ferent dependent stationary and non-stationary time series els and also to cluster these models based on their com-
processes with dependent or independent noises. puted similarities.

2. Clustering several time series models As can be seen, this technique is not too computational and
time consuming and can be also applied to cluster different
Assume we have k time series models as candidates to fit on dependent stationary or non-stationary processes. In next sec-
Xt . Let write the equations of these k time series models by tion the power of the introduced technique is studied.

ð iÞ
3. Simulation study
Xt ¼ fi ðXt1 ; Xt2 ;   Þ þ BH ðtÞ; i ¼ 1;    ; k; ð2:1Þ
This section deals with studying the ability of the proposed
ðiÞ
where BH ðtÞ; i ¼ 1;    ; k; are independent fractional Brownian technique in simulated datasets. The simulation study is setup
motion errors, and fi ; i ¼ 1;    ; k, are known parametric as following:
Fuzzy clustering to classify several time series models 1139

Table 1 Empirical power of the introduced technique for Example 1.

H /1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9593 0.9624 0.9798 0.9957
0.25 0.1 0.1 0.5 2 0.9512 0.9617 0.9785 0.9927
0.25 0.1 0.1 0.9 2 0.9521 0.9680 0.9758 0.9954
0.25 0.1 0.5 0.1 2 0.9524 0.9660 0.9876 0.9915
0.25 0.1 0.5 0.5 2 0.9600 0.9691 0.9809 0.9949
0.25 0.1 0.5 0.9 3 0.9505 0.9602 0.9802 0.9955
0.75 0.1 0.1 0.1 1 0.9543 0.9627 0.9790 0.9978
0.75 0.1 0.1 0.5 2 0.9543 0.9671 0.9849 0.9983
0.75 0.1 0.1 0.9 2 0.9540 0.9673 0.9734 0.9969
0.75 0.1 0.5 0.1 2 0.9532 0.9670 0.9898 1.0000
0.75 0.1 0.5 0.5 2 0.9571 0.9639 0.9763 0.9950
0.75 0.1 0.5 0.9 3 0.9529 0.9623 0.9728 0.9944

Table 2 Empirical power of the introduced technique for Example 2.

H h1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9503 0.9632 0.9785 0.9939
0.25 0.1 0.1 0.5 2 0.9503 0.9613 0.9826 0.9966
0.25 0.1 0.1 0.9 2 0.9596 0.9670 0.9799 0.9956
0.25 0.1 0.5 0.1 2 0.9534 0.9658 0.9748 0.9913
0.25 0.1 0.5 0.5 2 0.9567 0.9685 0.9804 0.9965
0.25 0.1 0.5 0.9 3 0.9587 0.9672 0.9753 0.9948
0.75 0.1 0.1 0.1 1 0.9565 0.9621 0.9701 0.9940
0.75 0.1 0.1 0.5 2 0.9531 0.9634 0.9837 0.9922
0.75 0.1 0.1 0.9 2 0.9568 0.9616 0.9713 0.9906
0.75 0.1 0.5 0.1 2 0.9532 0.9696 0.9861 0.9925
0.75 0.1 0.5 0.5 2 0.9536 0.9664 0.9869 0.9918
0.75 0.1 0.5 0.9 3 0.9547 0.9681 0.9894 0.9994

Step 1: For each time series model, three samples of size n For the first, the second and the third models, we assumed
are separately generated. /1 ¼ 0:1; /1 2 f0:1; 0:5g; and /1 2 f0:1; 0:5; 0:9g respectively.
Step 2: The proposed approach was applied to cluster the Also, we consider the Hurst parameter H 2 f0:25; 0:75g:
three samples.
Step 3: Step1 and 2 were repeated 10,000 times. Example 2:. Consider the MA(1) model,
Step 4: The empirical power ð b
p Þ was computed by
Xt ¼ BH ðtÞ þ h1 BH ðt  1Þ:
For the first, the second and the third models, we assumed
T h1 ¼ 0:1; h1 2 f0:1; 0:5g; and h1 2 f0:1; 0:5; 0:9g respectively.
p¼
b ;
10000 Also, we consider the Hurst parameter H 2 f0:25; 0:75g:
such that T is equal to the number of runs that the value of the
estimated number of clusters after using the proposed Example 3:. Assume the ARMA(1,1) model,
approach is similar to actual number of clusters.
Xt ¼ /1 Xt1 þ BH ðtÞ þ h1 BH ðt  1Þ:
Remark 1:. If the parameters of three populations are the We assume /1 ¼ 0:2, for all models and for the first to the
same, the true value of number of clusters is equal to 1. If only third models, we let h1 ¼ 0:1; h1 2 f0:1; 0:5g; and
the parameters of two populations are the same and the h1 2 f0:1; 0:5; 0:9g respectively. Also, we consider the Hurst
parameters of other population differs, the true value is 2, and parameter H 2 f0:25; 0:75g:
if the parameter settings of three populations are different, the
true value is 3. Example 4:. Consider the PAR(1) process

Example 1:. Assume the AR(1) process, Xt ¼ /t Xt1 þ BH ðtÞ;

Xt ¼ /1 Xt1 þ BH ðtÞ: such that

1140 M.R. Mahmoudi et al.

Table 3 Empirical power of the introduced technique for Example 3.

H h1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9599 0.9656 0.9770 0.9985
0.25 0.1 0.1 0.5 2 0.9521 0.9628 0.9701 0.9985
0.25 0.1 0.1 0.9 2 0.9514 0.9656 0.9774 0.9979
0.25 0.1 0.5 0.1 2 0.9589 0.9685 0.9758 0.9940
0.25 0.1 0.5 0.5 2 0.9566 0.9606 0.9813 0.9912
0.25 0.1 0.5 0.9 3 0.9521 0.9649 0.9846 0.9906
0.75 0.1 0.1 0.1 1 0.9585 0.9636 0.9842 0.9985
0.75 0.1 0.1 0.5 2 0.9544 0.9667 0.9743 0.9935
0.75 0.1 0.1 0.9 2 0.9552 0.9600 0.9877 0.9919
0.75 0.1 0.5 0.1 2 0.9590 0.9606 0.9735 0.9951
0.75 0.1 0.5 0.5 2 0.9533 0.9622 0.9708 0.9919
0.75 0.1 0.5 0.9 3 0.9548 0.9671 0.9726 0.9973

Table 4 Empirical power of the introduced technique for Example 4.

H /1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9569 0.9657 0.9800 0.9985
0.25 0.1 0.1 0.5 2 0.9536 0.9652 0.9809 0.9903
0.25 0.1 0.1 0.9 2 0.9509 0.9603 0.9801 0.9902
0.25 0.1 0.5 0.1 2 0.9502 0.9648 0.9871 0.9938
0.25 0.1 0.5 0.5 2 0.9589 0.9645 0.9735 0.9939
0.25 0.1 0.5 0.9 3 0.9588 0.9615 0.9762 0.9987
0.75 0.1 0.1 0.1 1 0.9553 0.9649 0.9765 0.9985
0.75 0.1 0.1 0.5 2 0.9515 0.9632 0.9801 0.9950
0.75 0.1 0.1 0.9 2 0.9554 0.9679 0.9720 0.9927
0.75 0.1 0.5 0.1 2 0.9586 0.9657 0.9743 0.9952
0.75 0.1 0.5 0.5 2 0.9526 0.9669 0.9821 0.9993
0.75 0.1 0.5 0.9 3 0.9594 0.9644 0.9888 0.9940

Table 5 Empirical power of the introduced technique for Example 5.

H h1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9593 0.9672 0.9841 0.9938
0.25 0.1 0.1 0.5 2 0.9546 0.9640 0.9804 0.9974
0.25 0.1 0.1 0.9 2 0.9550 0.9676 0.9717 0.9916
0.25 0.1 0.5 0.1 2 0.9556 0.9642 0.9758 0.9916
0.25 0.1 0.5 0.5 2 0.9520 0.9640 0.9771 0.9933
0.25 0.1 0.5 0.9 3 0.9503 0.9600 0.9854 0.9900
0.75 0.1 0.1 0.1 1 0.9597 0.9683 0.9827 0.9969
0.75 0.1 0.1 0.5 2 0.9594 0.9620 0.9884 0.9924
0.75 0.1 0.1 0.9 2 0.9588 0.9624 0.9704 0.9977
0.75 0.1 0.5 0.1 2 0.9548 0.9608 0.9720 0.9924
0.75 0.1 0.5 0.5 2 0.9576 0.9620 0.9704 0.9961
0.75 0.1 0.5 0.9 3 0.9595 0.9668 0.9837 0.9979

/t ¼ 0:6 þ /1 cosð2pt=TÞ Xt ¼ BH ðtÞ þ ht BH ðt  1Þ;

For the first, the second and the third models, we assumed such that
/1 ¼ 0:1; /1 2 f0:1; 0:5g; and /1 2 f0:1; 0:5; 0:9g respectively. ht ¼ 1 þ h1 cosð2pt=TÞ
Also, we consider the Hurst parameter H 2 f0:25; 0:75g:
For the first, the second and the third models, we assumed
Example 5:. Consider the PMA(1) process h1 ¼ 0:1; h1 2 f0:1; 0:5g; and h1 2 f0:1; 0:5; 0:9g respectively.
Also, we consider the Hurst parameter H 2 f0:25; 0:75g:
Fuzzy clustering to classify several time series models 1141

Table 6 Empirical power of the introduced technique for Example 6.

H h1 Actual Number of Clusters n
First Second Third 50 75 100 500
0.25 0.1 0.1 0.1 1 0.9550 0.9633 0.9772 0.9966
0.25 0.1 0.1 0.5 2 0.9535 0.9670 0.9750 0.9965
0.25 0.1 0.1 0.9 2 0.9592 0.9613 0.9768 0.9960
0.25 0.1 0.5 0.1 2 0.9543 0.9672 0.9777 0.9957
0.25 0.1 0.5 0.5 2 0.9534 0.9685 0.9874 0.9988
0.25 0.1 0.5 0.9 3 0.9583 0.9670 0.9754 0.9991
0.75 0.1 0.1 0.1 1 0.9515 0.9612 0.9770 0.9928
0.75 0.1 0.1 0.5 2 0.9504 0.9682 0.9739 0.9984
0.75 0.1 0.1 0.9 2 0.9517 0.9663 0.9873 0.9997
0.75 0.1 0.5 0.1 2 0.9587 0.9685 0.9800 0.9945
0.75 0.1 0.5 0.5 2 0.9592 0.9696 0.9752 0.9970
0.75 0.1 0.5 0.9 3 0.9524 0.9693 0.9870 0.9981

Series

2.

1.

0.

-1.

-2.

5 10 15 20 25 30

Fig. 1 Annual RDI time series of Zahedan synoptic station (1980–2010).

Sample ACF Sample PACF

1.00 1.00

.80 .80

.60 .60

.40 .40

.20 .20

.00 .00

-.20 -.20

-.40 -.40

-.60 -.60

-.80 -.80

-1.00 -1.00
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40

Fig. 2 Sample ACF/PACF for annual RDI time series of Zahedan synoptic station (1980–2010).
1142 M.R. Mahmoudi et al.

Example 6:. Consider the PARMA(1,1) process

Xt  /t Xt1 ¼ BH ðtÞ þ ht BH ðt  1Þ;

such that
/t ¼ 0:6 þ /1 cosð2pt=TÞ and ht ¼ 1 þ h1 cosð2pt=TÞ.
We assume /1 ¼ 0:2, for all models and for the first to the
third models, we let h1 ¼ 0:1; h1 2 f0:1; 0:5g; and
h1 2 f0:1; 0:5; 0:9g respectively. Also, we consider the Hurst
parameter H 2 f0:25; 0:75g:
The values of the estimated power ð b p Þ for Examples 1–6,
are reported in Tables 1–6, respectively. As it can be seen,
the values of b p are very close to the one, specially when the
value of n grows. In other words, the introduced approach
can nicely discriminate and cluster different fitted models. It
can be concluded that the introduced approach has many
advantages. First, the power of this method is acceptable. Sec-
ond, the introduced approach is not too computational and
time consuming. More ever, this technique can be applied to
cluster different stationary and non-stationary time series
processes.

Fig. 3 Fuzzy clustering method to classify the time series models

4. Real data
fitted on the RDI.

Now, the behavior of the introduced procedure is studied in

real world applications. The values of the annual reconnais-
sance drought index (RDI) from an Iranian synoptic station
(Zahedan) from 1980 to 2010 have been considered. Figs. 1
and 2 indicate the dataset, sample auto-correlation (ACF)
and partial auto-correlation functions (PACF) for RDI time
series of Zahedan synoptic station. As can be seen in ACF/
PACF plot, AR(1), MA(1) and ARMA(1,1) models can be
candidate models to fit on this dataset.
Table 7 shows the results for the fitted models. As this table
indicates, based on AICC and variances, the MA(1) model is
better than other models.
We are interested in clustering these three models. To do
this aim, the proposed technique is applied. By using Kaiser
Index (the number of eigen-values of correlation matrix that
are more than 1), the number of clusters was considered to
be 2. Table 8 and Figs. 3 and 4 provide the results of the fuzzy
clustering technique in. As it can be seen, these three models

Table 7 Best models for annual RDI time series of Zahedan

synoptic station (1980–2010).
Model Model’s Formula AICC
AR(1) Xt ¼ 0:3869Xt1 +BH ðtÞ 87.5374
MA(1) Xt ¼ BH ðtÞ þ 0:5033BH ðt  1Þ 86.7242 Fig. 4 Fuzzy clustering plot to classify the time series models
ARMA(1) Xt ¼ 0:6538Xt1 þ BH ðtÞ  0:2635BH ðt  1Þ 91.4619 fitted on the RDI.

can be divided in two distinct clusters; cluster 1: MA(1), and

cluster 2: AR(1) and ARMA(1,1). In other words, MA(1) is
best model to describe the annual RDI time series of Zahedan
Table 8 The probabilities of membership in two clusters using synoptic station (1980–2010). Also, there is no significant dif-
fuzzy clustering. ference between AR(1) and ARMA(1,1) models.
Model Cluster 1 Cluster 2
5. Conclusion
AR(1) 0.38232564 0.617674356
MA(1) 0.99402695 0.005973045
ARMA(1,1) 0.02219271 0.977807293 Classifying and clustering several time series models fitted on a
dataset is an important problem in data analysis. The cluster-
Fuzzy clustering to classify several time series models 1143

ing of several dependent time series fitted on a specific dataset [7] M.R. Mahmoudi, M.H. Heydari, R. Roohi, A new method to
(specially, non-stationary processes or time series with frac- compare the spectral densities of two independent periodically
tional Brownian motion noises) was lacked. In this research, correlated time series, Math Comput. Simulat. 160 (2019) 103–
we tried to solve this important problem. At first, all of time 110.
[8] M.R. Mahmoudi, M.H. Heydari, Z. Avazzadeh, Testing the
series models were fitted on dataset. Then, for observed data-
difference between spectral densities of two independent
set, the predicted values of all models were estimated based on periodically correlated (cyclostationary) time series models,
fitted models formulas. Finally, the fuzzy clustering technique Commun. Stat. Theory Methods 48 (9) (2019) 2320–2328.
was applied to cluster the models based on the similarities [9] M.R. Mahmoudi, M. Mahmoudi, E. Nahavandi, Testing the
between predicted values. The ability of the proposed tech- difference between two independent regression models,
nique was studied using simulation. The results indicated that Commun. Stat. Theory Methods 45 (21) (2016) 6284–6289.
for small sample size (n ¼ 50), the power of the proposed [10] M.R. Mahmoudi, M. Maleki, A. Pak, Testing the equality of
method to classify different models was more than 95%. The two independent regression models, Commun. Stat. Theory
power was increased to more than 99% as the sample size grew Methods 47 (12) (2018) 2919–2926.
to (n ¼ 50). Consequently the introduced approach could [11] M.R. Mahmoudi, On comparing two dependent linear and
nonlinear regression models, J. Test. Eval. 47 (1) (2018) 449–
nicely discriminate and cluster different fitted models. The
458.
introduced approach had many advantages. The power of this [12] P. Ji-jun, M.R. Mahmoudi, D. Baleanu, M. Maleki, On
method was acceptable. Moreover, the introduced approach comparing and classifying several independent linear and non-
was not too computational and time consuming. Furthermore, linear regression models with symmetric errors, Symmetry 11 (6)
this technique could be applied to cluster different dependent (2019) 820.
stationary and non-stationary time series processes. The appli- [13] M. Bahrami, M.J. Amiri, M.R. Mahmoudi, S. Koochaki,
cability of the introduced procedure in real world applications Modeling caffeine adsorption by multi-walled carbon
was also investigated by studying RDI values from an Iranian nanotubes using multiple polynomial regression with
synoptic station. Three candidate models AR(1), MA(1) and interaction effects, J. Water Health 15 (4) (2017) 526–535.
ARMA(1,1) have been fitted to dataset. Then the introduced [14] A.R. Zarei, M.R. Mahmoudi, Evaluation of changes in RDIst
index effected by different Potential Evapotranspiration
approach was used two cluster the fitted models. Based on
calculation methods, Water Resour. Manage. 31 (15) (2017)
the results, these models were classified into 2 clusters; cluster
4981–4999.
1: MA(1), and cluster 2: AR(1), and ARMA(1,1). [15] M.H. Heydari, Z. Avazzadeh, M.R. Mahmoudi, Chebyshev
cardinal wavelets for nonlinear stochastic differential equations
CRediT authorship contribution statement driven with variable-order fractional Brownian motion, Chaos,
Solitons Fractals 124 (2019) 105–124.
[16] M.R. Mahmoudi, M. Maleki, A new method to detect
Mohammad Reza Mahmoudi: Data curation, Validation.
periodically correlated structure, Comput. Statist. 32 (4) (2017)
Dumitru Baleanu: Conceptualization, Methodology, Software, 1569–1581.
Supervision. Sultan Noman Qasem: Validation, Visualization. [17] A.R. Nematollahi, A.R. Soltani, M.R. Mahmoudi, Periodically
Amirhosein Mosavi: Validation, Visualization. S Band: Valida- correlated modeling by means of the periodograms asymptotic
tion, Visualization. distributions, Stat. Pap. 58 (4) (2017) 1267–1278.
[18] M.R. Mahmoudi, M.H. Heydari, Z. Avazzadeh, On the
asymptotic distribution for the periodograms of almost
Declaration of Competing Interest
periodically correlated (cyclostationary) processes, Digital
Signal Process. 81 (2018) 186–197.
The authors declare that they have no known competing [19] M.R. Mahmoudi, M.H. Heydari, Z. Avazzadeh, K.H. Pho,
financial interests or personal relationships that could have Goodness of fit test for almost cyclostationary processes, Digital
appeared to influence the work reported in this paper. Signal Process. 96 (2020) 102597.
[20] M.R. Mahmoudi, M. Maleki, K. Borodin, K.H. Pho, D.
Baleanu, On comparing and clustering the spectral densities of
References several almost cyclostationary processes, Alexandria Eng. J. 59
(4) (2020) 2555–2565.
[1] H. Haghbin, M.R. Mahmoudi, Z. Shishebor, Large sample [21] R. Zhou, M.R. Mahmoudi, S.N.Q. Mohammed, K.H. Pho,
inference on the ratio of two independent binomial proportions, Testing the equality of the spectral densities of several
J. Math. Ext. 5 (1) (2011) 87–95. uncorrelated almost cyclostationary processes, Alexandria
[2] M.R. Mahmoudi, J. Behboodian, M. Maleki, Inference on the Eng. J., Article (in Press) (2020).
ratio of means in two independent populations, J. Stat. Theory [22] M.R. Mahmoudi, D. Baleanu, B.A. Tuan, K.H. Pho, A novel
Appl. 16 (3) (2017) 366–374. method to detect almost cyclostationary structure, Alexandria
[3] M.R. Mahmoudi, M. Mahmoodi, Inferrence on the ratio of Eng. J. 59 (4) (2020) 2339–2346.
variances of two independent populations, J. Math. Ext. 7 (2) [23] R. Roohi, M.H. Heydari, M. Aslami, M.R. Mahmoudi, A
(2014) 83–91. comprehensive numerical study of space-time fractional bioheat
[4] M.R. Mahmoudi, M. Mahmoodi, Inferrence on the ratio of equation using fractional-order Legendre functions, Eur. Phys.
correlations of two independent populations, J. Math. Ext. 7 (4) J. Plus 133 (2018) 412.
(2014) 71–82. [24] M.R. Mahmoudi, A.R. Nematollahi, A.R. Soltani, On the
[5] M.R. Mahmoudi, R. Nasirzadeh, M. Mohammadi, On the ratio detection and estimation of the simple harmonizable processes,
of two independent skewnesses, Commun. Stat-Theor. Methods Iranian J. Sci. Technol. (Sciences) 39 (2) (2015) 239–242.
48 (7) (2019) 1721–1727. [25] A. Babaei, H. Jafari, S. Banihashemi, Numerical solution of
[6] M.R. Mahmoudi, M. Maleki, A. Pak, Testing the difference variable order fractional nonlinear quadratic integro-differential
between two independent time series models, Iran J. Sci. equations based on the sixth-kind Chebyshev collocation
Technol. A (Sciences) 41 (2017) 665–669. method, J. Comput. Appl. Math. 112908 (2020).
1144 M.R. Mahmoudi et al.

[26] A. Babaei, B.P. Moghaddam, S. Banihashemi, J.A.T. Machado, based model development for detecting artery disease, in: ICIT,
Numerical solution of variable-order fractional integro-partial 2019, pp. 800–805.
differential equations via Sinc collocation method based on [46] S.M.J. Jalali, S. Ahmadian, A. Khosravi, S. Mirjalili, M.R.
single and double exponential transformations, Commun. Mahmoudi, S. Nahavandi, Neuroevolution-based Autonomous
Nonlinear Sci. Numer. Simul. 82 (2020) 104985. Robot Navigation: A Comparative Study, Cognit. Syst. Res.
[27] H. Jafari, A. Babaei, S. Banihashemi, A novel approach for (2020).
solving an inverse reaction–diffusion–convection problem, J. [47] S.J. Mousavirad, G. Schaefer, S.M.J. Jalali, I. Korovin, A
Optim. Theory Appl. 183 (2) (2019) 688–704. benchmark of recent population-based metaheuristic algorithms
[28] A. Babaei, H. Jafari, M. Ahmadi, A fractional order HIV/AIDS for multi-layer neural network training, in: Proceedings of the
model based on the effect of screening of unaware infectives, 2020 Genetic and Evolutionary Computation Conference
Math. Methods Appl. Sci. 42 (7) (2019) 2334–2343. Companion, 2020, July, pp. 1402–1408.
[29] R.M. Ganji, H. Jafari, D. Baleanu, A new approach for solving [48] S.M.J. Jalali, S. Ahmadian, P.M. Kebria, A. Khosravi, C.P.
multi variable orders differential equations with Mittag-Leffler Lim, S. Nahavandi, Evolving artificial neural networks using
kernel, Chaos, Solitons Fractals 130 (2020) 109405. butterfly optimization algorithm for data classification, in:
[30] H. Jafari, M. Saeidy, D. Baleanu, The variational iteration International Conference on Neural Information Processing,
method for solving n-th order fuzzy differential equations, Open Springer, Cham, 2019, December, pp. 596–607.
Physics 10 (1) (2012) 76–85. [49] M. Maleki, M.R. Mahmoudi, D. Wraith, K.H. Pho, Time series
[31] H. Jafari, M. Saeidy, J. Vahidi, The homotopy analysis method modelling to forecast the confirmed and recovered cases of
for solving fuzzy system of linear equations, Int. J. Fuzzy Syst. COVID-19, Travel Med. Infect. Dis. 101742 (2020).
11 (4) (2009) 208–313. [50] A.R. Abbasi, M.R. Mahmoudi, Z. Avazzadeh, Diagnosis and
[32] W. Gao, P. Veeresha, D.G. Prakasha, H.M. Baskonus, Novel clustering of power transformer winding fault types by cross-
dynamic structures of 2019-nCoV with nonlocal operator via correlation and clustering analysis of FRA results, IET Gener.
powerful computational technique, Biology 9 (5) (2020) 107. Transm. Distrib. 12 (19) (2018) 4301–4309.
[33] K.S. Al-Ghafri, H. Rezazadeh, Solitons and other solutions of [51] A.R. Zarei, M.M. Moghimi, M.R. Mahmoudi, Analysis of
(3+ 1)-dimensional space–time fractional modified KdV– changes in spatial pattern of drought using RDI index in south
Zakharov–Kuznetsov equation, Appl. Math. Nonlinear Sci. 4 of Iran, Water Resour. Manage. 30 (11) (2016) 3723–3743.
(2) (2019) 289–304. [52] A.R. Zarei, M.M. Moghimi, M.R. Mahmoudi, Parametric and
[34] J. Singh, D. Kumar, Z. Hammouch, A. Atangana, A fractional non-parametric trend of drought in arid and semi-arid regions
epidemiological model for computer viruses pertaining to a new using RDI index, Water Resour. Manage. 30 (14) (2016) 5479–
fractional derivative, Appl. Math. Comput. 316 (2018) 504–515. 5500.
[35] Y.D. Zhang, S.H. Wang, X.J. Yang, Z.C. Dong, G. Liu, P. [53] A.M. Alonso, D. Pena, J. Romo, Forecasting time series with
Phillips, T.F. Yuan, Pathological brain detection in MRI sieve bootstrap, J. Statist. Plann. Inference 100 (2002) 1–11.
scanning by wavelet packet Tsallis entropy and fuzzy support [54] A.M. Alonso, D. Pena, J. Romo, On sieve bootstrap prediction
vector machine, SpringerPlus 4 (1) (2015) 716. intervals, Statist. Probability Lett. 65 (2003) 13–20.
[36] S. Wang, X. Yang, Y. Zhang, P. Phillips, J. Yang, T.F. Yuan, [55] A.M. Alonso, J.R. Berrendero, A. Hernandez, A. Justel, Time
Identification of green, oolong and black teas in China via series clustering based on forecast densities, Comput. Stat. Data
wavelet packet entropy and fuzzy support vector machine, Anal. 51 (2006) 762–776.
Entropy 17 (10) (2015) 6663–6682. [56] J. Caiado, N. Crato, D. Pena, A periodogram-based metric for
[37] M. S ß enel, B. S
ß enel, C.A. Havle, Risk analysis of ports in time series classification, Comput. Stat. Data Anal. 50 (2006)
Maritime Industry in Turkey using FMEA based intuitionistic 2668–2684.
Fuzzy TOPSIS Approach, in: ITM Web of Conferences (Vol. [57] J. Caiado, N. Crato, D. Pena, Comparison of times series with
22, p. 01018). EDP Sciences, 2018. unequal length in the frequency domain, Comm. Statist.
[38] K. Yeh, C.W. Chen, Stability analysis of interconnected fuzzy Simulation Comput. 38 (2009) 527–540.
systems using the fuzzy Lyapunov method, Mathematical [58] D.S. Coates, P.J. Diggle, Tests for comparing two estimated
Problems in Engineering, 2010. spectral densities, J. Time Ser. Anal. 7 (1986) 7–20.
[39] S. Wang, Y. Li, Y. Shao, C. Cattani, Y. Zhang, S. Du, Detection [59] P.S.P. Cowpertwait, T.F. Cox, Clustering population means
of dendritic spines using wavelet packet entropy and fuzzy under heterogeneity of variance with an application to a rainfall
support vector machine, CNS & Neurol. Disorders-Drug time series problem, The Statistician 41 (1992) 113–121.
Targets (Formerly Current Drug Targets-CNS & Neurological [60] G.R. Dargahi-Noubary, Discrimination between Gaussian Time
Disorders) 16 (2) (2017) 116–121. Series based on Their Spectral Differences, Commun. Statist.
[40] A. Yokusß, S. Gülbahar, Numerical solutions with linearization (Theory and Methods) 21 (1992) 2439–2458.
techniques of the fractional Harry Dym equation, Appl. Math. [61] P. De Souza, P. Thomson, Lpc distance measures and statistical
Nonlinear Sci. 4 (1) (2019) 35–42. tests with particular reference to likelihood ratio, IEEE Trans.
[41] D. Kumar, J. Singh, D. Baleanu, A new numerical algorithm for Accoust. Speech Signal Process 30 (1982) 304–315.
fractional Fitzhugh-Nagumo equation arising in transmission of [62] H. Dette, T. Hildebrandt, A note on testing hypothesis for
nerve impulses, Nonlinear Dyn. 91 (1) (2018) 307–317. stationary processes in the frequency domain, J. Multivariate
[42] J. Danane, K. Allali, Z. Hammouch, Mathematical analysis of a Anal. 104 (2011) 101–114.
fractional differential model of HBV infection with antibody [63] H. Dette, T. Kinsvater, M. Vetter, Testing nonparametric
immune response, Chaos, Solitons Fractals 136 (2020) 109787. hypotheses for stationary processes by estimating minimal
[43] W. Gao, P. Veeresha, H.M. Baskonus, D.G. Prakasha, P. distances, J. Time Ser. Anal. 32 (2010) 447–461.
Kumar, A new study of unreported cases of 2019-nCOV [64] H. Dette, E. Paparoditis, Bootstrapping frequency domain
epidemic outbreaks, Chaos, Solitons & Fractals 109929, 2020. tests in multivariate time series with an application to
[44] S.M.J. Jalali, M. Karimi, A. Khosravi, S. Nahavandi, 2019, comparing spectral densities, J. Roy. Statist. Soc. B 71
October, An efficient neuroevolution approach for heart disease (2009) 831–8571.
detection, in: 2019 IEEE International Conference on Systems, [65] P.J. Diggle, N.I. Fisher, Nonparametric comparison of
Man and Cybernetics (SMC), IEEE, pp. 3771–3776. cumulative periodograms, Appl. Stat. 40 (1991) 423–434.
[45] S.M.J. Jalali, A. Khosravi, R. Alizadehsani, S.M. Salaken, P.M. [66] P.J. Diggle, I. Al Wasel, Spectral analysis of replicated
Kebria, R. Puri, S. Nahavandi, Parsimonious evolutionary- biomedical time series, Appl. Stat. 46 (1997) 31–71.
Fuzzy clustering to classify several time series models 1145

[67] M. Eichler, Testing nonparametric and semiparametric [78] E.A. Maharaj, Comparison of non-stationary time series in
hypotheses in vector stationary processes, J. Multivariate the frequency domain, Comput. Stat. Data Anal. 40 (2002)
Anal. 99 (2008) 968–1009. 131–141.
[68] K. Fokianos, A. Savvides, On comparing several spectral [79] F. Pattarin, S. Paterlini, T. Minerva, Clustering financial time
densities, Technometrics 50 (3) (2008) 317–331. series: an application to mutual funds style analysis, Comput.
[69] S. Fruhwirth-Schnatter, S. Kaufmann, Model-based clustering Stat. Data Anal. 47 (2004) 353–372.
of multiple time series, CEPR Discussion Paper No. 4650, 2004. [80] D. Piccolo, A distance measure for classifying ARIMA models,
[70] C. Jentsch, A new frequency domain approach of testing for J. Time Ser. Anal. 11 (1990) 153–164.
covariance stationarity and for periodic stationarity in [81] B. Potscher, E. Reschenhofer, Discriminating between two
multivariate linear processes, J. Time Series Anal. 33 (2012) spectral densities in case of replicated observations, J. Time
177–192. Ser. Anal. 9 (1988) 221–224.
[71] C. Jentsch, M. Pauly, A note on using periodogram-based [82] G.E. Salcedo, R.F. Porto, P.A. Morettin, Comparing non-
distances for comparing spectral densities, Statist. Probab. Lett. stationary and irregularly spaced time series, Comput. Stat.
82 (2012) 158–164. Data Anal. 56 (12) (2012) 3921–3934.
[72] C. Jentsch, M. Pauly, Testing equality of spectral densities using [83] J. Timmer, M. Lauk, W. Vach, C.H. Lucking, A test for the
randomization techniques, Bernoulli 21 (2) (2015) 697–739. difference between spectral peak frequencies, Comput. Stat.
[73] Y. Kakizawa, R.H. Shumway, M. Taniguchi, Discrimination Data Anal. 30 (1999) 45–55.
and clustering for multivariate time series, J. Am. Stat. Assoc. 93 [84] U. Triacca, Measuring the distance between sets of ARMA
(1998) 328–340. models, Econometrics 4 (3) (2016) 32.
[74] M.F. Macchiato, L. La Rotonda, V. Lapenna, M. Ragosta, [85] M.R. Mahmoudi, M. Mahmoudi, A. Pak, On comparing,
Time modeling and spatial clustering of daily ambient classifying and clustering several dependent regression models,
temperature: an application in Southern Italy, Environmetrics J. Stat. Comput. Simul. 89 (12) (2019) 2280–2292.
6 (1995) 31–53.
[75] E.A. Maharaj, A significance test for classifying ARMA models, Further reading
J. Stat. Comput. Simul. 54 (1996) 305–331.
[76] E.A. Maharaj, Comparison and classification of stationary
multivariate time series, Pattern Recogn. 32 (1999) 1129–1138. [86] M.R. Mahmoudi, M.H. Heydari, K.H. Pho, Fuzzy clustering
[77] E.A. Maharaj, Clusters of time series, J. Classif. 17 (2000) 297– to classify several regression models with fractional Brownian
314. motion errors, Alexandria Eng. J. 59 (4) (2020) 2811–2818.