Model Order Reduction by Pade Approximation and

Improved Pole Clustering Technique

Kaushal Ramawat1, Manisha Bhandari2, Anuj Kumar3
1,2,3
Department of Electronics Engineering, University College of Engineering, Kota (Raj.)
1
kaushalramawat@yahoo.com

Abstract: The authors offered a mixed technique for Various mixed methods based on the clustering of poles and
reducing the order of the high order dynamic systems. A Padé approximations are also proposed [4]. Methods describe
pole-clustering based method to derive a reduced order the poles of the reduced system is use as the cluster centre of
approximation for a stable single-input single- output (SISO) the pole clusters of the original system which obtained by
continuous time system is presented. In this method, the Inverse distance measure (IDM) criteria. The choice of the
denominator polynomial of the reduced order model is clusters are either taken arbitrarily based on the order of
obtained by improved pole- clustering approach and the reduction or it is the investigator. In this proposed method, the
numerator polynomial is obtained through Padé reduced order denominator polynomial has been obtained
approximation technique. The reduced order model so using an a dominant pole based pole-clustering approach for
obtained by improved clustering algorithm guaranteed the reduced order model. The method uses the improved
stability in the reduced model and also preserves the clustering approach by deciding the value of the ratio of the
characteristics of the original system in the approximated residue to real parts of poles, taken in descending order and its
one. corresponding reduced order model was obtained through a
simple mathematical procedure. The model so obtained
Keywords: Order reduction, Padé approximation, pole preserves the stability. The clustering method proposed in this
clustering, Dominant pole, Mean Square Error, Transfer paper differs from the existing pole clustering technique by
function, IDM. considering the distance of system poles from the first pole in
the group clustering process and it gives better approximation
1. INTRODUCTION for order reduction.
Model order reduction are often take interest in system
modelling and design of high order systems. There are lots of 2. PROBLEM STATEMENT
methods proposed in the literature reflect the importance of Consider an linear SISO time invariant system of nth order .
producing a reduced order model for the system. It gives better Higher order transfer function be in the form
considerate of the physical systems, reduced computational
complexity, reduced hardware complexity, simplified      ⋯     
G(s)= (1)
controller design and cost effective solutions. The Padé      ⋯     
approximation technique has been successfully applied to find
the reduced system of higher order system, but sometimes this Where m ≤ n
method has the drawback of resulting in unstable reduced
order models of high order stable systems. () ∑     
G(s) = = ∑  (2)
()    
For the stable reduced order models, various methods based
on the preservation of dominant poles have been proposed [2]. Corresponding desired reduced order model of !  order
The reduced-order model retains the basic physical features should be given by
(such as time constants) of the original system and the stability
of the simplified model is guaranteed. In classical approach, $ $    ⋯%  % $%  %
the modes with the largest time constants i.e. slow modes or "# (s) = (3)
& & &   ⋯&' & ' &'  '
the poles nearest to imaginary axis are usually considered
dominant. This approach is good, still with some demerits. Where l ≤ r
Firstly slow modes may not be dominant, another in some
cases the system may have all the poles arise in a small region ' () ∑%  $(  (
in s-plane or modes with similar time constants, and lastly the "# (s) = = ∑'( ( (4)
' () (  &( 
for complex poles is not straightforward.

Energy Technology & Ecological Concerns: A Contemporary Approach

ISBN: 978-81-93024-71-3 103
Kaushal Ramawat, Manisha Bhandari, Anuj Kumar

New denominator )# (*) for reduced order model is obtain |,: | < |, | … … < |,< |. Cluster center for the reduced order
through pole clustering technique system is obtained through the following procedure. The
procedure described in first step is similar to the case of the
)# (*) = (* − ,- )(* − ,. ) … . . (* − ,# ) (5) method proposed in [5] but the pole cluster calculated in the
proposed method is based on the improved pole clustering
New numerator 1# (*) for reduced order model is obtain approach in that particular cluster center.
through Padé approximation technique as
Step 3.1 - Let the k number of poles available be arrange as
()
=
' ()
; 1# (*) = )# (*)*
()
(6) |,. | < |,: | < |, | … … < |,< |,
() ' () ()

Step 3.2 - Set M = 1,

Here important characteristics of original system retain by
obtained reduced order model through this mixed method of Step 3.3 - Calculate for the pole cluster as
model order reduction.
. 4.
3. DESCRIPTION OF PROPOSED METHOD => = ?@− |A | + ∑<7: −1/|, − ,. |D ÷ FG ,


Computation of reduced order model consists two steps are as

follows: Step 3.4 - Check if M = k, then the final cluster center is
=H = => and terminate the process else go for next step.
In first step find the denominator polynomial and in second
step find numerator polynomial. Step 3.5 - Again set M = M+1,

Denominator polynomial computation Step 3.6 - The improved cluster center from
The reduced denominator polynomial computed by improved
pole clustering approach based on dominant pole and also by => = −I,. ∗ =>4.
classical approach as shown below:
Step 3.7 - Check M = k . If no then go to step 3.5 otherwise for
In dominant pole based pole-clustering approach Consider the next step,
given system in (1) can be shown as
Step 3.8 - The final cluster center is =H = => .
3
"(*) = ∑67. ( (7)
45 ( Here if the system to be reduced have pole lie on imaginary
axis that should be retain as the cluster center as a single pole
By the given equation corresponding to every pole λi is the and other remaining are clustered in another clusters based on
ratio of residue to pole intended as algorithm. The pole cluster centres computed by the proposed
method are more dominant than the pole cluster centres as
|3( |
Ratio of residue to poles = obtained from the method proposed in[5]. While calculating
|39(5( )|
the rth order system to be obtained, cluster center are obtained.
And finding the cluster center values, we have the following
The poles λi arrange in descending value of ratio of residue to three cases as
poles and then n-poles λ1 λ2 λ3.....λn
arranged in group as most dominant pole for r-clusters such Case(1) - If all denominator poles or cluster center obtained
that the poles are equal or in maximum number of clusters. are real. The denominator polynomial of reduced order in the
Number of cluster and cluster center one from each cluster is form as:
depend upon the order of the reduced system. The maximum
number of poles in per cluster are not limited but for the
)# (*) = (* + =K. )(* + =K: ) . … (* + =K# ) (8)
computation of cluster center have none of repeated pole in
same cluster center. In cluster formation first cluster have
The improved cluster values given as =K. =K: … . . =K# and r is
most dominant poles next cluster have next most dominant
the order of reduced system.Or we can represent it as given in
poles and so on.
equation(5).
Proposed algorithm for cluster center:
Case(2) - If all poles or cluster center obtained are complex.
Let k number of poles available in one cluster group as: P1, P2, The denominator polynomial of reduced order in the form as:
P3, ........ Pk cluster center obtained by inverse distance
measure. The arrangement of poles are as follows |,. | <

Energy Technology & Ecological Concerns: A Contemporary Approach

ISBN: 978-81-93024-71-3 104
 Model Order Reduction by Pade Approximation and Improved Pole Clustering Technique

Let k complex conjugate poles in single cluster group be 4. MEAN SQUARE ERROR(MSE)
(,. ± M . , ,: ± M : , … … ,< ± M < ). Where |,. | < |,: | <
For further effectiveness of the proposed approach MSE is
|, | … … < |,< | are the obtain through same algorithm is
determine. The MSE is as
proposed above, it follows separately for real and complex
poles. Then the improved cluster center is as _
βi = , ± M 1
MSE = Z(y- (t ] ) − y^ (t ] )):
N
]7.
Corresponding reduced order polynomial is
)# (*) = (* + |O. |)(* + |O: |) . … (* + |O |) (9)
Here y- (t) is the step response of original higher order system
and y^ (t ] ) is step response of reduced order model.
Case(3) - If some cluster center are real and remaining are in
the complex form.Then applying the algorithm separately for
Numerical example
real part and then for complex terms. To get cluster center for
reduced system combine both of them to find denominator
Let us consider the 8th order system given in [8] described by
polynomial.
the transfer function as:
Classic dominance based pole clustering approach: ()
In classic dominance based approach pole cluster formation is G(s) = =
()
made by most dominant pole first, and most dominant pole is
decided through the modes with the largest time constants or `-:-.abc- :::-aa  .::` d 
….
nearest pole to the origin. And further cluster center is `-:- .-eba`..a.:` c:a` d 
obtained through the same algorithm proposed.
+ 36380* ` + 5982* b + 514*  + 18* c
..
Denominator polynomial computation + 22449* ` + 4536* b + 546*  + 36* c + * a

Denominator polynomial is find by Pade approximation in it It can be represent in the form of (7) as
equate the original higher order system transfer function with
4.---` `.-``` 4.cb- ...bbb 4.a-c
generated reduced system transfer function. The denominator G(s) = + + + + +
. :  ` b
polynomial of reduced system obtained from first step is use 4..: 4c.- e.:b`
here to get the unknown coefficient of reduced system. As + +
 c a
given
Poles of the system are
1(*) 1# (*) -1, -2, -3, -4, -5, -6, -7, -8
=
)(*) )# (*)
Dominant pole based pole clustering Approach
     ⋯     
=
     ⋯      By the dominance method of ratio to residue the poles are in
$ $    ⋯%  % $%  % order as given below
10)
& & &   ⋯&' & ' &'  ' -1, -4, -3, -2, -8, -7, -5, -6

By cross multiplying the equation and comparing the To reduce the model into third order by DPPCA there are
coefficient of same power of ‘s’and we get the new numerator three cluster are formed as
coefficient as Cluster (1) – Poles (-1, -4, -3)
P- Q- = R- S- Cluster (2) – Poles (-2, -8, -7)
P- Q. + P. Q- = R- S. + R. S- Cluster (3) – Poles (-5, -6)
P- Q: + P. Q. + P: Q- = R- S: + R. S. + R: S-
PT Q# = R6 SU Cluster center calculated by improved pole clustering method
as in algorithm
By solving it the unknown coefficient S- , S. , S , … . . SU can get
easily. The numerator polynomial of reduced order system is Reduced order model of system is as
shown in form of
1# (*) = Q- + Q. * + Q: * : + …. "# (s) =
' ()
=
.-.cb    :.   c.`e
' ()  d  ..:    .:.`a   c.`e

+Q#4. * #4. + Q# * # (11)

Graphical representation is shown in fig. 1

Energy Technology & Ecological Concerns: A Contemporary Approach

ISBN: 978-81-93024-71-3 105
Kaushal Ramawat, Manisha Bhandari, Anuj Kumar

Now by classical approach TABLE 1: MSE VALUE OF STEP RESPONSES

In this case poles of same system are sorted by classical
approach as in form S. No. Approach MSE
-1, -2, -3, -4, -5, -6, -7, -8 1. DPPCA 0.0071
2. Classical dominance 0.0015
Cluster formation based on dominance criteria a
Cluster (1) – Poles (-1, -2, -3) TABLE 2: COMPARISION OF PARAMETERS OF
Cluster (2) – Poles (-4, -5, -6) RESPONSES
Cluster (3) – Poles (-8, -7)
Approach Rise Settling Peak Peak
Cluster center calculated as same algorithm as proposed. Time Time Time
Reduced order model of system is Original 0.064 4.820 0.449 2.203
' () ..`    `-..   ...`b system
"# (s) = =
' ()  d  a.-.e    .a.`c   ...`b
DPPCA 0.103 4.463 0.636 2.237
Graphical representation of this system is shown in fig. 2 Classical 0.082 4.711 0.531 2.242
dominance
5. CONCLUSION
In the proposed mixed method of reducing the model of higher
order using improved pole clustering approach applying with
the DPPCA and classical dominance approaches. Here
denominator polynomial is derived by pole clustering and
numerator polynomial through Pade approximation to get the
reduced order model.

The comparison of MSE shows that proposed improved pole

clustering approach with classical dominance perform better
over DPPCA. The stability and characteristics of original
system are retain in reduced order model. This method can be
more improve easily by further modification as it
i is let go to
research.

REFERENCES
Fig. 1. Step response comparison of original and reduced system [1] R. Komarasamy, N. Albhonso & G. Gurusamy “Order
with improved dominance reduction of linear systems with an improved pole
clustering” Journal of Vibration and Control18(12)
Control
1876–1885, 2011.
[2] Y. Shamash, “Linear system reduction using Padé
approximations
imations to allow retention of dominant poles,
poles ”
Int. J. control, vol. 21, no. 2,
2 pp 257-272, 1975.
[3] Y. Shamash, “Viability of methods for generating
stable reduced order models,
models ” IEEE trans. Automatic
control, vol. 26, no. 6, pp 1285-1286,
1285 1981.
[4] Vishwakarma C.B., Prasad R., R. “System reduction using
modified pole clustering and Padé approximation,
approximation ”
conference NSC-2008 .
xxxii national systems conference,
[5] Vishwakarma CB & Prasad R (2009). “MIMO system
reduction using modified pole clustering and genetic
algorithm”, Modelling
odelling and Simulation in Engineering.
DOI:10.1155/ 2009/540895.
[6] A.kumar D. Chandra “Improved
Improved Padé-pole
Padé clustering
approximant” International Conference on Computer
Fig. 2. Step response comparison of original and reduced system Science and Electronics Engineering.
with classical dominance

Energy Technology & Ecological Concerns: A Contemporary Approach

ISBN: 978-81-93024-71-3 106
 Model Order Reduction by Pade Approximation and Improved Pole Clustering Technique

[7] W. T. Beyene, “Pole-clustering and rational- [10] V. Singh, D Chandra, & H. Kar, "Improved Routh-
interpolation techniques for simplifying distributed Padé approximants: a computer-aided approach, "
systems, ” IEEE Trans. on Circuits and Systems-I: IEEE, Trans. on automat. Contr., Vol. 49, pp. 292-296,
Fundamental Theory and Applications, vol. 46, no. 12, 2004.
pp.1468-1472, 1999. [11] Parmar G, Prasad R and Mukherjee S (2007). “Order
[8] M. F. Hutton, & B. Friedland, “Routh approximation reduction of linear dynamic systems using stability
for reducing order of linear time invariant system”, equation method and GA”. International Journal of
IEEE Trans. Auto. Cont., AC-20, pp. 329-337, 1975. Computer, Information, and Systems Science, and
[9] B. Vishwakarma, & R. Prasad, “Clustering method for Engineering 1(1): 26–32.
reducing order of linear system using Padé [12] Gupta DK, Bhagat SK & Tewari JP (2002). “A mixed
approximation, ” IETE Journal of Research, vol. 54, method for the simplification of linear dynamic
no. 5, pp. 326-330, 2008. systems”. In Proceedings of International Conference
on Computer Applications in Electrical Engineering,
Roorkee, India, Feb. 21–23, pp. 455–459.

Energy Technology & Ecological Concerns: A Contemporary Approach

ISBN: 978-81-93024-71-3 107