Upload
others
View
9
Download
0
Embed Size (px)
Citation preview
Evolutionary Algorithm and
Fuzzy C-Means Clustering Based
Order Reduction of Discrete
Time Interval Systems Ch. Naga Sandhya1
, *M. Siva Kumar2
1 Department of Electrical and Electronics
Engineering, Gudlavalleru Engineering college,
Gudlavalleru, India. 2Department of Electrical and Electronics
Engineering, Gudlavalleru Engineering college,
Gudlavalleru, India.
E-mail:
[email protected],shivamangipudi
Abstract
This paper deals with a mixed method for model
order reduction of linear discrete time interval
systems. This method is based on Particle Swarm
Optimization; Fuzzy C means Clustering and
Kharitonov’s theorem. The reduced order interval
denominator coefficients are determined by using
Fuzzy C means Clustering technique while
Particle Swarm Optimization algorithm is used for
obtaining numerator coefficients by minimizing
ISE. This method is always generated to a stable
reduced order interval system if the original higher
order system is stable system. The proposed
method has to be applied to a typical numerical
example available in the literature. The results are
comparing with the results are determined by
using some other familiar methods.
Keywords: Interval system, Model order reduction,
Particle Swarm Optimization, Fuzzy C means Clustering,
Kharitonov’s theorem, Integral Square Error.
International Journal of Pure and Applied MathematicsVolume 114 No. 8 2017, 309-319ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu
309
1. Introduction Higher order model systems are complicate to handle due
to the computational complexities and they are also very
difficult to be used in real time problems. So it is
desirable that the higher model is to be replaced with
lower order model. Many number of methods for model
order reductions are available in literature [1]. Even
though several methods available, no approach gives the
best result for all systems. In literature, a very powerful
method which involves some simple algebraic calculations
i.e. moments matching continued fraction, and Pade
approximation [2], [3] are available. In some cases, Pade
approximant may turn out to be unstable even though the
original system is stable.
Routh approximation based methods are available
[2],[3],[6],[8] for continuous-time systems and for discrete-
time systems [4],[5],[7].There are more methods are
available for the order reduction of continuous time
systems, but very few methods extended to the discrete-
time systems. The methods for discrete-time systems are
classified into two types. The first one it consists of already
existed in continuous-time algorithm, G(z) into another
one G1(w), by using bilinear transformation z= (1+w)/(1-
w)[1]. Then one of known techniques for continuous-time
systems is applied to determine a reduced approximant
R1(w) for G1(w). Finally the inverse transformation w=
φ(z) converts R1(w) to the required approximant R(z).
Many systems the coefficients are constants but
uncertain within a finite range. Such systems are
classified as interval systems. Some methods like [9-15]
are proposed for order reduction of interval systems in
both continuous and discrete time systems. In [10],a
method for interval systems is γ − δ Routh Approximation
is proposed. A modified Routh approximation method is
presented for obtaining stable reduced models of
continuous-time systems. The reduced models give good
approximates of the initial transient response and the
steady-state response of the original system. In [11-16]
Kharitonov’s theorem is used for reduction technique of
linear interval systems Routh Approximation is used to be
generates stable reduced interval model.
In this paper, a proposed method for model order reduction
of Discrete time interval systems based up on Particle
Swarm Optimization (PSO) and Fuzzy C means Clustering
technique. The reduced interval model denominator
polynomials obtained by Fuzzy C means clustering method
International Journal of Pure and Applied Mathematics Special Issue
310
while the numerator coefficients are obtained by using
minimization of ISE by PSO. In this method produces a
stable reduced model if the original higher order system is
to be stable.
The proposed paper is to be organized as, section 2 consists
of a problem formulation, section 3 consists of proposed
method, section 4 consists of numerical example and
simulation results and section 5 consists of conclusion and
references.
3. Determination of Reduced Order Denominator polynomials:
For first Kharitonov's transfer function, Fuzzy C-Means
technique is applied to determine reduced order
polynomials.Fuzzy clustering method is a powerful
unsupervised method for analysis of data and construction
models. In many situations, fuzzy c means clustering is
natural compared to hard clustering. The boundaries
between many classes are not forced to fully belong to one
of the classes, but rather then the membership degrees
between 0 and 1 indicating the partial membership. Fuzzy
c-means algorithm is most widely used. The sum of
degrees of belongingness of the data point to all clusters is
always equal to unity: uij
ci=1 = 1 (1)
The cost function for FCM is generalization is
J= Jici=1 = um
ijnj=1
ci=1 d2
ij (2)
Where uij is between 0 and 1, ci is the cluster center of
fuzzy group i; dij = 𝑐𝑖 − 𝑥𝑗 is the Euclidean distance
between the 𝑖𝑡ℎ cluster center and the 𝑗𝑡ℎ data point; and
m Є [1,∞]. The necessary conditions for (9) to reach to its
minimum are
𝑐𝑖= 𝑢𝑖𝑗
𝑚𝑋𝑗𝑛𝑗=1
𝑢𝑖𝑗𝑚𝑛
𝑗=1
(3)
and 𝑢𝑖𝑗 =1
[ 𝑐𝑖−𝑋𝑗
𝑐𝑘−𝑋𝑗
2(𝑚−1)
𝑐𝑘=1
(4)
This algorithm works by assuming the membership to
each data point according to each cluster centers. The
distance between the cluster center and data point should
be more then the data nearer to the cluster center more is
its membership towards the particular cluster center.
Clearly, combination of membership of each data point
should be equal to one.
International Journal of Pure and Applied Mathematics Special Issue
311
Fig 1 FCM algorithm
4 Determination of Reduced Order numerator polynomial:
PSO is an evolutionary algorithm that can be used
for solving the nonlinear equations. It is a kind of swarm
intelligence that is based on social-psychological principles
and provides insights into social behaviour, as well as
contributing to engineering applications. In PSO, the
velocity and position are randomly chosen for a set of
particles. During the start, the initial position is taken as
the best position and the velocity is updated.
The main purpose of this optimization method is
a) A global optimum for the nonlinear system may be
found,
b) It can produce a different number of substitute
solutions,
c) There are no mathematical limitations on the
formulation of the problem,
d) Comparatively very simple in execution and
e) Numerically strong.
In Table 1, the typical parameters for PSO optimization
routines, used in the present study are given.
TABLE1
Typical parameters used by PSO
Name Value(type)
Number of generations 100
Population size 60
Maximum Particle velocity 2
Epoch 100
Termination method Maximum Generation
International Journal of Pure and Applied Mathematics Special Issue
312
5. Numerical Example: Consider a higher order discrete time interval system
G(z) =
2.6 , 3 + 3.4 , 3.55 𝑧+ 3.15 , 3.45 𝑧2+ 2.05 , 2.15 𝑧3
+ 1.1 , 1.25 𝑧4+[0.85 , 0.95]𝑧5
2.2 , 2.25 + 1.3 , 1.6 𝑧+ 1.1 , 1.2 𝑧2+ 5.3 , 5.6 𝑧3+
6.2 , 6.65 𝑧4+ 8.1 , 8.45 𝑧5+[10.15 , 10.35]𝑧6
With the help of Kharitonov's theorem, the higher order interval
system can be represented as four Kharitonov's higher order transfer
functions are given as
𝐺61 𝑧 =
0.85𝑧5+1.1𝑧4+2.15𝑧3+3.45𝑧2+3.4𝑧+2.6
10.35𝑧6+8.1𝑧5+6.2𝑧4+5.6𝑧3+1.2𝑧2+1.3𝑧+2.2
𝐺62 𝑧 =
0.95𝑧5+1.1𝑧4+2.05𝑧3+3.45𝑧2+3.55𝑧+2.6
10.35𝑧6+8.45𝑧5+6.2𝑧4+5.3𝑧3+1.2𝑧2+1.6𝑧+2.2
𝐺63 𝑧 =
0.85𝑧5+1.25𝑧4+2.15𝑧3+3.15𝑧2+3.4𝑧+3
10.15𝑧6+8.1𝑧5+6.65𝑧4+5.6𝑧3+1.1𝑧2+1.3𝑧+2.25
𝐺64 𝑧 =
0.95𝑧5+1.25𝑧4+2.05𝑧3+3.15𝑧2+3.55𝑧+3
10.15𝑧6+8.45𝑧5+6.65𝑧4+5.3𝑧3+1.1𝑧2+1.6𝑧+2.25
Step 1:
Consider the first Kharitonov's transfer function
𝐺61 𝑧 =
0.85𝑧5+1.1𝑧4+2.15𝑧3+3.45𝑧2+3.4𝑧+2.6
10.35𝑧6+8.1𝑧5+6.2𝑧4+5.6𝑧3+1.2𝑧2+1.3𝑧+2.2
(4.6)
The poles of 𝐺61 𝑧 are
𝜆1 = -0.0773+0.8618i 𝜆4 = -0.7546-0.2622i
𝜆2 = -0.0773-0.8618i 𝜆5 = 0.4406+0.5008i
𝜆3 = -0.7546+0.2622i 𝜆6 = 0.4406-0.5008i
The second order reduced model is desired to be approximated for
the system as
𝐺𝑟1 𝑧 =
𝑏2𝑧2+𝑏1𝑧+𝑏0
𝑎2𝑧2+𝑎1𝑧+𝑎0
By using FCM algorithm the poles of the reduced order system are
found to be
-0.241567± 0.5475i and the denominator of the ROM is obtained
as
D2(s) = 𝑧2 + 0.4831𝑧 + 0.3581
The reduced order model is:𝐺21 𝑧 =
𝑏12𝑧2+𝑏11𝑧+𝑏10
𝑧2+0.4831𝑧+0.3581
Repeat the FCM algorithm for remaining Kharitonov’s transfer
function the reduced order models are:
𝐺22 𝑧 =
𝑏22𝑧2+𝑏21𝑧+𝑏20
𝑧2+0.5412𝑧+0.3692 ;𝐺2
3 𝑧 =𝑏32𝑧
2+𝑏31𝑧+𝑏30
𝑧2+0.4936𝑧+0.3689;𝐺2
4 𝑧 =
𝑏42𝑧2+𝑏41𝑧+𝑏40
𝑧2+0.5484𝑧+0.3793
International Journal of Pure and Applied Mathematics Special Issue
313
Step2 :
The reduced order numerator coefficients are determined by the
minimization of ISE using PSO algorithm for 4 Kharitonov’s
transfer functions and there given as
𝑁21 𝑧 = 0.05534 𝑧2 +0.3731z+0.2846; 𝑁2
2 𝑧 =
0.05348 𝑧2 +0.4874z+0.2012
𝑁23 𝑧 = 0.0212 𝑍2 +0.4095z+0.3021 ;𝑁2
4 𝑧 =
0.0211𝑍2+0.4157z+0.3191
The four reduced order Kharitonov’s transfer functions are:
𝐺21 𝑧 =
0.05534𝑧2+0.3731𝑧+0.2846
𝑧2+0.4831𝑧+0.3581; 𝐺2
2 𝑧 =
0.05348𝑧2+0.4874𝑧+0.2012
𝑧2+0.5412𝑧+0.3692
𝐺23 𝑧 =
0.0212𝑧2+0.4095𝑧+0.302
𝑧2+0.4936𝑧+0.3689 ;𝐺2
4 𝑧 =
0.0211𝑧2+0.4157𝑧+0.3191
𝑧2+0.5484𝑧+0.3793
Therefore, the reduced order interval model is obtained as
G2(z) =
0.0211 , 0.05534 𝑧2+ 0.3731 , 0.4913 𝑧
+[0.2017 0.3191]
1,1 𝑧2+ 0.4831,0.5484 𝑧+[0.3581, 0.3793]
Comparing the proposed model with reduced order model obtained
by the method given in [9]
G2(z) =
0.02285 , 0.0791 𝑧2+ 0.3665, 0.4956 𝑧
+[0.2067 ,0.2999]
1,1 𝑧2+ 0.4286,0.712 𝑧+[0.2478, 0.399]
Fig. 3: Step Response comparison for 1st Kharitonov’s TF
International Journal of Pure and Applied Mathematics Special Issue
314
Fig. 4: Step Response comparison for 2nd
Kharitonov’s TF
Fig. 5: Step Response comparison for 3rd
Kharitonov’s TF
Fig. 6: Step Response comparison for 3rd
Kharitonov’s TF
The step responses of the four reduced order Kharitonov's models
are compared with the method given in [9]. It has been observed
International Journal of Pure and Applied Mathematics Special Issue
315
from figure 3 to figure 6 that the step response of four Kharitonov's
transfer functions of higher order time interval system and the
respective reduced models obtained by the present method are
closely matching than the method given in [9].This method also
retains the stability of the reduced order interval model since the
original system is stable. It is observed that the reduced order model
obtained by the present method has better matching of transient and
steady state response than the method in [9]. From Table 1 it is
observed that the ISEs of lower and upper bounds of reduced
interval model determined by the present method are less than the
method given in [9].
Table 1
Comparison of ISEs for reduced interval system model
MOR ISE for lower bound ISE for upper bound
Proposed Method 0.292 0.3061
Pade & Pole clustering
Method[11] 0.4052 0.3545
5. Conclusion In this proposed paper, the model order reduction
for a linear discrete time interval system is done by
combining the pole clustering technique and an
evolutionary algorithm based soft computing technique i.e.
PSO algorithm. The reduced model is determined by using
the Kharitonov’s theorem and Fuzzy C-means Clustering
algorithm for denominator coefficients, while numerator
coefficients are determined by using PSO algorithm. The
main objective to find the numerator is minimization of
ISE. The use of interval arithmetic has the possibility of
producing an unstable reduced model. To avoid this
problem, Kharitonov’s theorem is used to make reduced
interval models robustly stable. It is observed that, since
the original discrete time interval system is stable the
reduced order polynomials are also stable, has better
matching response. Therefore, the original higher order
interval system is reduced to second order interval model
by minimizing the ISE.
References:
[1] L.Fortuna, G. Nunnari, and A. Gallo, “Model Order
Reduction Techniques with applications in Electrical
Engineering", (1992)SpringerVerlag, Lomndon.
[2] Y Shamash, “Stable reduced-order models using Padé type
approximations”, IEEE Trans. Auto. Control, Vol. AC-19,
International Journal of Pure and Applied Mathematics Special Issue
316
No.5, (1974)pp. 615-616.
[3] Maurice F.Hutton and Bernard Friedland, “Routh-
approximation for reducing order of linear, time-invariant
systems” IEEE Trans. Auto. Contr. Vol. AC-20, No.3, (1975)
pp. 329-337.
[4] G.Vasu,M.SivaKumar and M.RamalingaRaju, A novel
method for optimal model simplification of large scale
linear discrete time systems, International
journals.Automation and control, vol.10,no.2(2016).
[5] G.Vasu,M.SivaKumar and M.RamalingaRaju,Optimal Least
Squares Model Approximation For Large Scale Discrete-
Time Systems, Transactions of the
InstitueofMeasurementandcontrol,SAGEpublications,(201
6),DOI:1177/0142331216649023.
[6]Vasu.G,SivaKumar.M and RamalingaRaju.M,A novel Model
Order Reduction Technique for Linear Continuous-Time
Systems Using PSO-DV Algorithm, Journal of Control,
Automation and Electrical Systems,SpringerVol.28,issue:1,
(2017)pp.68-77,.
[7] C.B.Vishwakarma and R. Prasad “Clustering method for
Reduction Order of Linear System using Pade
approximation”, IETE Journal of Research, vol 54,
issue5(2008).
[8]N.Sai Dinesh, Dr.M.Siva Kumar, D.SrinivasaRao“Model
Order Reduction of Higher Order Discrete Time Systems
using Modified Pole Clustering Technique “International
Journal of Engineering Research & Technology Vol. 2 Issue
11(2013).
[9] Vinay Pratap Singh, and Dinesh Chandra, “Reduction of
discrete interval systems based on pole clustering and
improved Pade approximation: a computer aided
approach”,AMO-Advanced Modeling and Optimization,
volume 14, Number 1(2012).
[10] Bandyopadhyay B AvinashUpadhye and Osman Ismailγ-δ
R Routh-Pade approximation for interval systems, IEEE
Trans. on Automatic Control,Vol. 42,no.8, (1997) pp.1126-
1130.
[11] Sastry G V K R and SivaKumar MDirect Routh
Approximation method for linear SISO uncertain systems
Reduction ,International Journal of Applied Engineering
Reasearch ,vol.5,no.1. (2010) pp.99-98.
[12] SivaKumarMVijayaAnand N and Srinivas Rao R,Model
Reduction of Linear Interval Systems Using Kharitonov’s
Polynomial ,IEEE International Conference on Energy
Automation and Signal(ICEAS),2011.
International Journal of Pure and Applied Mathematics Special Issue
317
[13] G. V. Nagesh Kumar , J. Amarnath and B.P.Singh„
"Behavior of Metallic Particles in a Single Phase Gas
Insulated Electrode Systems with Dielectric Coated
Electrodes", IEEE International Conference on “Condition Monitoring and Diagnosis” CMD-2008 Beijing, P. R. China. , April 21-24, 2008, Page(s):377-380.
[14] Kumar, G.V. Nagesh ,Amarnath, J. ,Singh, B.P.,
Srivastava, K.D., " Particle initiated discharges in gas
insulated substations by random movement of particles in
electromagnetic fields“, International Journal of Applied
Electromagnetics and Mechanics, vol. 29, no. 2, pp. 117-
129, 2009
[13] SivaKumarMVijayaAnand N and Srinivas Rao R, Impluse
Energy Approximation of Higher Order Interval systems
using Kharitonov’s polynomials Transactions of the
Institute of Measurement and Control ,vol 39,no
6(2016),pp 1225–1235.
[14] SivaKumar M and Gulushad BEGUM A New Biased Model
Order Reduction for Higher Order Interval
Systems.Advances in Electrical and Electronics
Engineering ,volume 14,no 2, (2016) pp.145-152.
[15] Siva kumar M and and Gulushad BEGUM model order
reduction of time Interval Systems using Stability
Equation Method and a Soft computing technique
Advances in Electrical And Electronics Engineering,
volume 14,no 2, (2016)pp.153-161
[16] B.Venkateswara Rao , G.V.Nagesh Kumar , M.Ramya
Priya , and P.V.S.Sobhan, "“Implementation of Static VAR
Compensator for Improvement of Power System
Stability”", International Conference on Advances in
Computing, Control, and Telecommunication Technologies,
ACT 2009 organized by ACEEE and CPS, Trivandrum,
Kerala, India, , 28-29 December, 2009, Pages: 453-457.
International Journal of Pure and Applied Mathematics Special Issue
318
319
320