Applied Mathematical Modelling 40 (2016) 72257244

Contents lists available at ScienceDirect

Applied Mathematical Modelling

journal homepage: www.elsevier.com/locate/apm

Reduced-order modeling of linear time invariant systems

using big bang big crunch optimization and time moment
matching method
Shivanagouda Biradar, Yogesh V. Hote, Sahaj Saxena
Department of Electrical Engineering, Indian Institute of Technology Roorkee, Roorkee 247 667, Uttarakhand, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a new approach is proposed to approximate the high-order linear time in-
Received 29 September 2015 variant (LTI) system into its low-order model. The proposed approach is a mixed method
Revised 28 January 2016
of model order reduction scheme consisting of recently developed big bang big crunch op-
Accepted 9 March 2016
timization algorithm and the time-moment matching method. This proposed method is
Available online 24 March 2016
applicable to single-input single-output, multi-input multi-output system, and time de-
Keywords: layed LTI systems. The proposed approach is substantiated with various numerical ex-
Integral square error amples of low and high-order systems. The results show that the reduced-order models
Model-order reduction preserve both transient and steady state conditions of original systems. Further, the re-
Optimization sults are also compared with the existing approaches of reduced order modeling which
Routhian-array show exceptional improvement in integral square error (ISE) and other time domain
Step response specications.
2016 Elsevier Inc. All rights reserved.

1. Introduction

The concept of model order reduction (MOR) is basically practiced in the eld of systems and control engineering where
the properties of dynamical systems are analyzed to reduce their complexity and retain their inputoutput behavior as much
as possible. MOR simplies the understanding of the system, and minimizes the computational burden in the simulation
studies. Moreover unlike the design of complex H and synthesis based control schemes, it enables the control prac-
titioners to design simple control laws thereby making controller computationally and cost ecient [1]. Advanced robust
control concepts such as H control and synthesis are widely used in various engineering applications wherein the de-
sign schemes produce high-order controllers even for a simple second-order plant [2]. To cope with the aforementioned
challenges faced when dealing with large-scale dynamical systems, various MOR approaches have been proposed using va-
riety of concepts. Over thirty years of research in MOR for LTI system, the developed methods conceptualize features such
as dominant pole retention, singular value decomposition and Hankel norm based approximation, singular perturbation ap-
proximation, Krylov subspace method, H -optimization methods [36]. One of the important trends in these studies is the
optimization (minimization) of integral error criterion between the actual plant and its reduced model.
In recent years, there is a widespread interest and research in evolutionary optimization techniques mainly because of its
intuitiveness, ease of implementation, and the ability to effectively solve highly non-linear, mixed integer optimization prob-
lems of complex engineering systems. Now-a-days, these evolutionary optimization techniques unied with conventional

Corresponding author. Tel.: +918791638621.
E-mail addresses: shivmb@gmail.com (S. Biradar), yhotefee@iitr.ernet.in (Y.V. Hote), sahajsaxena11@gmail.com (S. Saxena).

http://dx.doi.org/10.1016/j.apm.2016.03.006
S0307-904X(16)30131-7/ 2016 Elsevier Inc. All rights reserved.
7226 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

MOR techniques have shown highly promising results. Among various evolutionary techniques, particle swarm optimization
and genetic algorithm are considered to be highly reliable algorithm for solving optimization problems [7]. On the same
lines, a novel universe inspired evolutionary technique so called Big Bang Big Crunch (BBBC) can be used for order reduction
of systems of various complexities [8]. Recently, a mixed method is proposed for order reduction of LTI systems for both
single-input single-output (SISO) and multi-input multi-output (MIMO) systems [9]. This method is based on unication of
BBBC and Routh approximation method. It is observed that this method is better than the existing method and has been
applied for low-order systems but a more accurate and improved results can also be achieved. Therefore, in this paper,
an alternative mixed method is proposed to obtain reduced model using BBBC and time moment matching method. In this
approach, numerator coecients of reduced-order transfer function model are optimized using BBBC technique and denomi-
nator is evaluated using the time moment matching method from full-order plants information. Here, this proposed method
is applied to SISO, MIMO and time delayed system. The comparison of the existing approaches and proposed approach has
been carried which show the remarkable improvement in integral square error (ISE) and other performance parameters.
The organization of this paper is as follows: Section 2 briey summarizes the preliminaries of the concepts involved in
MOR, formulation of the proposed MOR scheme is presented in Section 3, simulation studies of the proposed method has
been conducted on different types of systems in Section 4, and the conclusions are drawn in Section 5.

2. Background materials

In this section, we present the background material of the concepts that are involved in the proposed scheme.

2.1. Model reduction from control systems perspective

In control theory, MOR can be dened as follows.

Denition 1. Let G(s): u y be the transfer function of the full-order system with order n, then the model reduction
scheme seeks a reduced-order transfer function model G (s ) : u y with order n so that n < n and for the same input u(t),
we have y(t ) y(t ).

Remark 1. For the sake of deniteness, the type of systems considered in this paper is proper LTI systems. The proposed
method is limited to transfer function matrices such that
(1) G(s) are rational and stable, i.e., the poles lie on the left-half of the s-plane.
(2) In frequency domain, G(j) is nonzero for all including = .
In other words, MOR leads to optimization problem in the following manner.

Denition 2. MOR denes the problem of nding the reduced-model G (s ) from the full-order plant G(s) using optimization
formulation such that

minimize ISE = [y(t ) y(t )]2 dt

subject to ISE <

where is an error tolerance.

Remark 2. We have selected ISE criterion because it quickly eliminates large errors in both transient and steady-state in
comparison to other integral error performance measures such as ISE, IAE, ITSE, etc.

2.2. Big bang-big crunch algorithm

The Big Bang-Big Crunch algorithm is basically an optimization method which was proposed by Erol and Eksin in 2006
[8]. This method is similar to GA with respect to generation of initial candidates. The random generation of initial candidates
is called Big Bang phase (BBP). In this phase, the candidate solutions are disseminated all over the search space in an uniform
manner but restricted under search space. The BBP is followed by contraction of solutions in Big Crunch phase (BCP) wherein
randomly distributed solutions are drawn into order, which can be named as center of mass. Here, the term mass refers to
the inverse of the tness function value. The center of mass C (x ) is function of position of each candidate (position vector)
in designed search space and is computed as
N xi
i=1 ISEi
C (x ) = N
 , (1)
1
i=1 ISEi

where xi is ith candidate in n-dimensional search space; ISEi is treated as an objective function value or tness function
value corresponding to ith candidate, which is dened as

M
[y(it ) y(it )] ,
2
ISEi = (2)
i=1
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7227

where M =  5T
t
. In (2), y(it) and y(it ) are sampled values of unit step response of original system and reduced-order LTI
system at time it and T is the time constant of the step response of the full-order system. In this paper, t is taken as
0.1 s. and N is population size in BBP. Population size must be optimally chosen which depends on the range of search space
and number of iterations. After BCP, N new candidates are generated in designed search space, to be used for BBP of next
iteration based on the knowledge of C (x ) as
xi new = C (x ) + i , (3)
where is deviation of a newly generated candidates from C (x ) and is calculated using following equation:
ri (xmax xmin )
i = . (4)
K
In the above equation, is the constant parameter generally taken to be [0.1, 0.3]. This parameter limits the size of
the solution space; ri is a random number from a standard normal distribution with decreasing standard deviation, which
changes for each candidate such that ri (0, 1]. Note that the greater the value of ri , the more scattered the candidate
solutions are around C (x ); Also, after extensive simulations and study, we came to conclusion that for problems which we
are taking into consideration work well when ri (0, 1]. If ri takes values more than 1, then convergence time increases
and accuracy of solution decreases. xmax and xmin are the upper and lower bounds on the values of the optimization vari-
ables, and K is a variable which increases with step size of one. This is done for better and faster convergence to solution.
After second explosion, C (x ) is recalculated. These successive explosions, i.e., BBP, and contraction steps, i.e., BCP are carried
repeatedly until a stopping criterion has been met. This procedure of BBBC algorithm is shown in Fig. 1.

2.3. Time moments matching method

Let the impulse response of high-order asymptotically stable system be g(t) then, we can write

G (s ) = g(t )est dt
0

Now using the power series expansion for est , G(s) can be written as
   
s2 t 2 s3 t 3 t2
G (s ) = g(t ){1 st + + ...}dt = g(t )dt s t g(t )dt + s2 g(t )dt ...
0 2! 3! 0 0 0 2!
or
G(s ) = c0 + c1 s + c2 s2 + ...,
where

(1 )i
ci = t i g(t )dt , i = 0, 1, 2, ...
i! 0
 
(1 )i
= [the ith time moment of g(t )].
i!

Thus the time moments of the reduced system G (s ) are proportional to power series expansion of G(s), about s = 0. In
time moment method, the idea is to match as many time moments of original and reduced system as much possible [11].

3. Main results

3.1. Statement of problem

In this subsection, we dene the mathematical expressions for asymptotically stable, LTI-SISO and MIMO plants.

3.1.1. SISO system

The original nth order system can be written in, usual notation, in transfer function form as
m
a sk a0 + a1 s + a2 s2 + a3 s3 + + am sm
G(s ) = nk=0 k k = , m n and k, m, n I, (5)
k=0 bk s b0 + b1 s + b2 s2 + b3 s3 + + bn sn
where G(s) is BIBO stable continuous-time original system of order n, ak and bk are constant coecients of s (complex
variable) in numerator and denominator polynomials, respectively. Our objective is to reduce the original full-order system
G(s) into its approximated reduced-order model G (s ) of form
m r
ak sk
G (s ) = nk=0 , mr nr and k, mr , nr I, (6)
r
k=0 k
b sk

where ak and b k are coecients of s in numerator and denominator, respectively.

7228 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Fig. 1. Schematic of BBBC algorithm.

3.1.2. MIMO system

The original nth order MIMO system can be written in transfer function matrix form as

a11 ... a1m
1
[G(s )] = n .. ..
.
.. , (7)
k . .
k=0 bk s
a p1 a pm
where [G(s)] is original system and dimension of [G(s)] is p m. Eq. (7) can also be represented as

g11 ... g1m
[G(s )] = [gi j ] pm = ... ..
.
.. , i, j = 0,
. (8)
g p1 g pm
where
ai j A0 + A1 s + A2 s2 + + Am sm
g i j = n = , m n. (9)
k=0 bk sk b0 + b1 s + b2 s2 + + bn sn
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7229

The model reduction of (7) gives

a11 ... a1m
  1
G (s ) = n .. ..
.
.. , (10)
r
b s k . .
k=0 k
a p1 a pm

where [G (s )] is reduced system and dimension of [G (s )] is p m. In other words, (10) can be written as

g11 ... gm
[G (s )] = [gi j ] pm = ... ..
.
.. ,
. (11)
g p1 g pm

where

ai j A 0 + A 1 s + A 2 s2 + + A w sw
gi j = z = , w z, (12)
k b 0 + b 1 s + b 2 s2 + + b z sz
k=0 bk s

where w < m, z < n.

3.2. Proposed algorithm

In this subsection, we discuss the procedure of our proposed MOR algorithm. Let us consider a full-order plant described
in (5). The required form of reduced-order model having order p, p < n can be written as

x0 + x1 s + + x p1 p1
G (s ) = , (13)
1 + a13 s + a12 s2 + + a1 p s p

such thatp < n. The proposed algorithm constitutes two parts:

3.2.1. Determination of denominator polynomial using time moment matching method [10]
We apply the time moment matching method to obtain the denominator polynomial of reduced model.
Step 1: In (5), divide the numerator and denominator coecient of G(s) with b0 to get
m
a sk a0 /bo + a1 /bos + a2 /bo s2 + a3 /bo s3 + + am /bo sm
G(s ) = nk=0 k k = , mn (14)
k=0 bk s 1 + b1 /bo s + b2 /bo s2 + b3 /bo s3 + + bn /bo sn

which can be further written as

a21 + a22 s + a23 s2 + a24 s3 + + a2m sm

G (s ) = . (15)
1 + a12 s + a13 s2 + a14 s3 + + a1n sn
Step 2: Now, arranging the coecient of above rational polynomial function of (15) in Routhian array form [11] as

where ak,l = ak1,1 a1,l+1 ak1,l+1 a1,1 k = 3, 4, ... and l = 1, 2, ... . Next, the values obtained from Routhian array can
be used as,

G(s ) = a21 a31 s + a41 s2 a51 s3 + (16)

Step 3: Using ci = (1 )k+2 ak1 k = 2, 3, ... and i = 0, 1, 2, ..., the matrix can be structured into the form
7230 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

The above formed matrix so called partitioned matrix can be written as

(17)

where C11 , C21 , C22 are of order (m + 1 ) n, n n, n (m + 1 ), respectively.

Step 4: Lastly calculate A 1 = C 1 21 C2 = [a12 a13 .... a1n ]T . If the second order model is required, then denomi-
nator can be written as D(s ) = a12 s + a13 s + 1.
2

3.2.2. Determination of numerator polynomial using BBBC

Once the denominator polynomial of form: D (s ) = a11 s p + a12 s p1 + ... + a1 p1 s2 + a1 p s + 1 where p < n is calculated.
(s ) as can be inferred from (13) using BBBC such
Now, our objective is to calculate the coecients [x0 , x1 , . . . , x p1 ] of N
that the ISE get minimized. As discussed earlier in Section 2.2, the steps to obtain the numerator coecients can be further
illustrated from the following algorithm.

Algorithm

Let G(s ) be given.

Step 1: Initialize r, , N, t and it er (it eration ) = 1;
Step 2: Generate population (containing N particles)
Xkiter = {xiter
1 , x2 , . . . , xN } [xmax , xmin ],
iter iter
k = 1 . . . N;
Step 3: Evaluate objective function for each particle
 2
ISEiiter (xi ) = [yiter (it ) yiter (it )] , i = 1 . . . N;
i

Step 4: leastISE = min(ISEiiter ) and i f (leastISE )

Then jump to step 9
Otherwise Go to step 5;
Step 5: Calculate the center of mass as dened in (1)
Step 6: it er it er + 1;
Step 7: Generate new solution around center of mass
new
Xkiter = {new xiter
1 , N } = C ( x k ) + i
x2 , . . . , new xiter
new iter  iter

Step 8: Go to step 3
Step 9: Return Xr = {x1, x2 , . . . , xr }

4. Numerical studies

In this section, the proposed MOR method is applied to six different problems and the results are compared with other
recently developed schemes.

4.1. Example 1

Consider a fourth-order system expressed in transfer function form as

s3 + 7s2 + 24s + 24
G1 ( s ) = (18)
s4 + 10s3 + 35s2 + 50s + 24
On dividing the numerator and denominator by 24 in (18), we have
1 + s + 0.292s2 + 0.0416s3
G1 ( s ) = (19)
1 + 2.083s + 1.458s2 + 0.4167s3 + 0.0416s4
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7231

Table 1
Comparison of various MOR techniques with respect to ISE.

Model reduction technique Reduced order model ISE

0.28389s+1.0 0 043
Proposed method 0.3986s2 +1.3744s+1
1.1478 103
0.6997s+0.6997
Sikander& Prasad [12] s2 +1.45771s+0.6997
27.7989 103
0.8058s+0.7944
Desai and Prasad [9] s2 +1.65s+0.7944
2.8358 103
7s2 +24s+24
Truncation method [13] 35s2 +50s+24
70.138 103
20.57s+24
Routh Hurwitz [14] 30s2 +42s+24
97.41283 10-3
0.7311s+2.5088
Pad approximation [15] s2+3.4481s+2.5088
1.234 103
0.02957s2 +0.6925s+2.501
Singular perturbation [16] s2 +3.395s+2.501
9.449
0.8216s+0.4542
Balanced realization [17] s2 +1.268s+0.4663
14.19 103
0.7442575s+0.6991576
Parmar method [18] s2 +1.45771+0.6997
17.4381 103
0.6997s+0.6997
Chen et al. [19] s2 +1.4577s+0.6997
27.7989 103
s+34.2465
Pal [20] s2 +239.882s+34.2465
19.4603994
s+0.6997
Parthasarathy and Jayasimha [21] s2 +1.45771s+0.6997
342.0218 103
s2 +2
Davision [22] s2 +3s+2
2735.7076 103
s+2.3014
Sheih and Wei [23] s2 +5.7946s+2.3014
1474.5453 103
s+5.403
Saifonov and Chiang [24] s2 +8.431s+4.513
1786.6177 103

Now, on arranging the above transfer function (19) in Routhian array form [11], we get

The above partitioned matrix gives c0 = 1, c1 = 1.083, c2 = 1.0899, c3 = 1.0663. Thus we get

(20)

1.083 1 1.3744
On comparing (20) with (17), we get C21 = [ ] and therefore A 1 = (C21 )1 C2 = [ ]. After this
1.0899 1.083 0.3986
step, the BBBC algorithm is employed to nd the coecients of numerator such that objective functions (ISE) get minimised
and the results obtained are x1 = 0.28389 and x2 = 1.0 0 043. Hence, the reduced- order model of a original system is ob-
tained as
1.0 0 043 + 0.28389s
G 1 (s ) = (21)
1 + 1.3744s + 0.3986s2
The comparison of original model with proposed reduced models and other recently developed reduced order models
in time domain and frequency domain is performed, which is shown in Figs. 2, and 3, respectively. It is observed that in
both domains, proposed model is a better approximation with respect to original model in comparison with other existing
models. Further in order to validate this, ISE are calculated for all the models. They are shown in Table 1. It is observed that
the ISE of the proposed model is the least in comparison to other existing models. Furthermore, the qualitative comparison
is also carried out in Table 2. For qualitative comparison, rise time, settling time, and peak overshoot are considered. Again,
it is found that the proposed method brings closer approximation to the original system.
7232 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Fig. 2. Step response comparison for Example 1.

S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7233

Fig. 3. Frequency response comparison for Example 1.

7234 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Table 2
Qualitative comparison with respect to transient response.

Model order reduction technique Rise time (Tr ) Settling time (Ts) Peak overshoot (Mp )

Original system 2.260 3.931 0

Proposed method 2.268 3.958 0
Sikander & Prasad [12] 2.301 3.410 1.0722
Desai and Prasad [9] 2.278 3.619 0.274
Truncation method [13] 2.737 4.080 0.564
Routh Hurwitz [14] 1.926 5.587 3.595
Pad approximation [15] 2.260 3.947 0
Singular perturbation [16] 2.259 3.917 0
Balanced realization [17] 2.089 5.355 2.059
Parmar method [18] 2.188 3.219 1.301
Chen et al. [19] 2.301 3.410 1.072
Pal [20] .381 27.361 0
Parthasarathy and Jayasimha [21] 1.591 5.563 4.099
Davision [22] 1.696 3.257 0
Sheih and Wei [23] 4.961 8.834 0
Safonov and Chiang [24] 3.801 6.747 0

4.2. Example 2

Let us now take an eighth order system:

18s7 + 514s6 + 5982s5 + 36380s4 + 122664s3 + 222088s2 + 185760s + 40320
G2 ( s ) = (22)
s8 + 36s7 + 546s6 + 4536s5 + 22449s4 + 67284s3 + 118124s2 + 109584s + 40320
On dividing the coecients of both numerator and denominator by 40320 in (22), we get
1 + 4.607s + 5.508s2 + 3.042s3 + 0.9023s4 + 0.1484s5 + 0.01276s6 + 0.0 0 04464s7
G2 ( s ) = (23)
1 + 2.718s + 2.93s2 + 1.669s3 + 0.5568s4 + 0.1125s5 + 0.01354s6 + 0.0 0 08929s7 + 0.0 0 0 0248s8
Arranging the coecients of (23) into Routhian array form [11], we get

Now, for the second-order reduced model, the partitioned matrix is

(24)
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7235

Fig. 4. Step response comparison for Example 2.

Fig. 5. Frequency response comparison for Example 2.

1.2434
For calculating denominator using A 1 = C 1 21 C2 , we get A 1 = [ ]. Now, we use BBBC algorithm to search optimal
0.2075
coecients of numerator which gives x1 = 3.1084 and x2 = 1.0 0 05. Hence, the reduced-order model obtained is
3.1084s + 1.0 0 05
G 2 (s ) = (25)
0.2075s2 + 1.2434s + 1
The ISE of above reduced order model in (25) is 38.1244 103 . The comparison of above proposed model is carried out
with Desai and Prasad mixed method [9], which gives
2.06774s + 0.43184
G DP (s ) = (26)
s2 + 1.17368s + 0.43184
with ISE =19.2386. Similarly, the reduced model determined by Sikander and Prasads mixed method [12] is
16.92s + 5.263
G SP (s ) = (27)
s2 + 6.893s + 5.262
with ISE = 69.569 103 . From the ISE analysis, it is found that that the proposed method gives least ISE, i.e. modelling
error is minimized. Fig. 4 shows the comparison of the step response of original 8th order system G(s) with reduced second
order system obtained using the proposed method G 2 (s ), Desai and Prasad mixed method G DP (s ) [9] and Sikander and
Prasad mixed method G SP (s ) [12]. Further, from the frequency response characteristic shown in Fig. 5, it can be concluded
that the proposed method is nearly in tune with frequency response characteristics of the original model.

4.3. Example 3

Consider an example of two-input, two-output (TITO) system from [9] as

 2(s+5 ) (s+4 )

G11 G12 (s+1 )(s+10 ) (s+2 )(s+5 )
[ G ( s )] = = (s+10 ) (s+6 ) (28)
G21 G22
(s+1 )(s+20 ) (s+2 )(s+3 )
7236 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Table 3
Comparison between various MOR techniques with respect to ISE.

Model Method Transfer function ISE

G 11 (s ) Proposed method 0.2023s+1

0.10224s2 +1.10237s+1
7.91154 107
0.7938s+0.6181
Sikander and Prasad s2 +1.34952s+0.6181
0.1672
0.9475s+0.7091
Desai and Prasad s2 +1.548267s+0.7091
0.0701
G 12 (s ) Proposed method 0.10084s+0.4
0.10097s2 +0.7021s+1
1.27668 108
0.42728s+0.24717
Sikander and Prasad s2 +1.34952s+0.6181
0.0958
0.4892s+0.2837
Desai and Prasad s2 +1.548267s+1
0.0602
G 21 (s ) Proposed method 0.05025s+0.5
0.05052s2 +1.0505s+1
5.6205 109
0.37952s+0.309
Sikander and Prasad s2 +1.34952s+0.6181
0.03121
0.455s+0.3546
Desai and Prasad s2 +1.548267s+0.7091
0.0101
G 22 (s ) Proposed method 0.1676s+1
0.1673s2 +0.8343s+1
8.4009 108
0.93382s+0.6181
Sikander and Prasad s2 +1.34952s+0.6181
0.2004
1.126s+0.7091
Desai and Prasad s2 +1.548267s+1
0.0636

a (s ) a12 (s )
[G(s)] can be further written as [G(s )] = 1
( 11
Dn ( s ) a ( s )
) where Dn (s ) = s6 + 41s5 + 571s4 + 3491s3 + 10060s2 + 13100s +
21 a22 (s )
60 0 0 and
a11 (s ) = 2s5 + 70s4 + 762s3 + 3610s2 + 770 0s + 60 0 0
a12 (s ) = s5 + 38s4 + 459s3 + 2182s2 + 4160s + 2400
(29)
a21 (s ) = s5 + 30s4 + 331s3 + 1650s2 + 3700s + 3000
a22 (s ) = s5 + 42s4 + 601s3 + 3660s2 + 9100s + 6000

The reduced model of [G(s)] using our proposed scheme is presented in Table 3. For comparison, we consider model
obtained using Desai and Prasads mixed method [9] and Sikander and Prasad mixed method [12]. The obtained results
show that the proposed method gives closer approximation in comparison to these two methods. This is also observed
through step response and frequency response characteristics as shown in Figs. 6, and 7, respectively.

4.4. Example 4

Let us take another example of TITO power system plant of order 10 to strengthen the reliability of proposed method
as:

0.5517 0 0.3091 0 0 0 0 0 0 0.1695
0.0410 0 0.3050 0 0 0 0 0 0 0
0 314.1593
0 0 0 0 0 0 0 0
9.5540 0 0.8660 20 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0.0421 0.0328
A=
0.1962 10.8696 0.1672 0 0 10.8696 0 0 0 0
0.9386 51.9849 0.7999 0 0 41.1153 10.8696 0 0 0

0.9386 51.9849 0.7999 0 0 41.1153 10.8696 0.1 0 0
0 0 0 10 0 0 10 0 0 0 0 10 0 0 20 0

0 0 0 0 0 0 0 0 1.0526 0.8211
 T
0 0 0 0 0 0 0 0 10 0 0 0
B=
0 0.0926 0 0 0 0.4428 2.1179 2.1179 0 0
   
0 0 1 0 0 0 0 0 0 0 0 0
C= ,D= (30)
0.4777 0 0.0433 0 0 0 0 0 0 0 0 0

The above given state space form is converted into transfer function form as

  29.08s8 + 1868s7 + 46100s6 + 545900s5 + 31850 0 0s4 + 87020 0 0s3
G11 (s ) 1 + 12060 0 0 0s2 + 76060 0 0s + 648300
G21 (s )
=
D (s )
1.26s 85.17s7 2089s6 25680s5 109800s4 711800s3
8 ,
10830 0 0s2 29720 0s 19430
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7237

Fig. 6. Step response comparisons of (a) G11 (s), (b) G12 (s), (c) G21 (s), and (d) G22 (s) for Example 3.
7238 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Fig. 7. Frequency response comparison of (a)G11 (s), (b)G12 (s), (c)G21 (s), and(d)G22 (s) for Example 3.
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7239

where
D(s ) = s10 + 64.21s9 + 1596s8 + 19470s7 + 126800s6 + 503600s5 + 15690 0 0s4
+ 3240 0 0 0s3 + 40610 0 0s2 + 29050 0 0s + 25310 0
and
29.08s8 + 1868s7 + 46100s6 + 545900s5 + 31850 0 0s4 + 87020 0 0s3 + 12060 0 0 0s2 + 76060 0 0s + 64830 0
G11 (s ) =
D (s )
(31)
1.26s 85.17s 2089s 25680s 109800s 711800s 10830 0 0s 29720 0s 19430
8 7 6 5 4 3 2
G21 (s ) = . (32)
D (s )
For G11 (s) in (31), after arranging it in Routhian [11] form, we get as c0 = 2.561, c1 = 0.652, c2 = 0.9327, c3 = 1.836, c4 =
17.7491, c5 = 179.2176 and c6 = 1792.3375. Using these values, we calculate denominator coecients A 2 (as mentioned
in step 4, Section 3.2.1) and thus, reduced order denominator is obtained as
D 11 (s ) = 2.3108s3 + 2.7911s2 + 10.2646s + 1
The numerator is obtained using BBBC algorithm which is found out to be
11 (s ) = 17.42257s2 + 24.6483s + 2.8875
N
Hence, the third-order reduced model is
N 11 (s ) 17.42257s2 + 24.6483s + 2.8875
G 11 (s ) = = (33)
D 11 (s ) 2.3108s3 + 2.7911s2 + 10.2646s + 1
The ISE for (33) is 68.26931. The only problem with above system is that DC gain obtained does not match with the
original system model. The DC gain shown by original system is G(0 ) = 2.561 whereas the reduced model shows a DC gain
of G (0 ) = 2.8875. Since, DC gain is not matching, we made slight modication in algorithm. We x the G (0 ) same as G(0)
and then calculate the rest unknown numerator coecients using BBBC. Thus, we get
11dc (s ) = 16.8003s2 + 25.6068s + 2.561
N
and G 11 (s ) becomes
11dc (s )
N 16.8003s2 + 25.6068s + 2.561
G 11dc (s ) = = , (34)
D11dc (s )
2 . 3108 s3 + 2.7911s2 + 10.2646s + 1
wherein DC gain of both original system and reduced model matches accurately but with marginal increase of performance
function (ISE) which come out to be 69.1327. For the same system, we applied Desai and Prasad mixed method [9] and
found the reduced order model as
11DP (s )
N 3.2023099s2 + 2.081667s + 0.2056
G 11DP (s ) = = 3 , (35)
D11DP (s )
s + 0.85591s2 + 0.92131s + 0.08027
which shows an ISE of 147.3 and it is more than the proposed model. The step and frequency response characteristics shown
in Fig. 8, have considerable differences in both models but the proposed method is far better than Desai and Prasads mixed
method. It can be observed in frequency response characteristic (Fig. 8(b)) that in working range of frequency (5060 Hz),
the proposed model imitates original model very well and hence results into more robust controller design keeping in view
minor deviation but with heavily reduced numerical task.
Now, let us take G21 (s) of (32), the obtained ci  s are c0 = 0.07676, c1 = 0.2931, c2 = 0.31714, c3 = 0.76656, c4 =
7.1841, c5 = 72.3491 and c6 = 723.5899. Using these values, we reduce system to third- order model. The obtained de-
nominator is D 21 (s ) = 4.8812s3 + 7.6535s2 + 10.672s + 1. Now, using BBBC algorithm, we nd out optimal numerator which
21 (s ) = 4.0 0 055s2 0.95028s 0.0909 and hence we get
is N
N 21 (s ) 4.0 0 055s2 0.95028s 0.0909
G 21 (s ) = = . (36)
D 21 (s ) 4.8812s3 + 7.6535s2 + 10.672s + 1
This reduced transfer function gives ISE of 0.61606. In this case also, DC gain of original system and reduced system does
not match, therefore we follow the same procedure as mentioned earlier and hence the realized reduced transfer function
is
N 21dc (s ) 4.1s2 0.9594s 0.07676
G 21dc (s ) = = . (37)
D 21dc (s ) 4.8812s3 + 7.6535s2 + 10.672s + 1
The ISE of this transfer function is 0.6278, which is slightly more than G 21 (s ). For the same problem, Desai and Prasad
mixed method [9] is applied, the reduced order transfer function is obtained as
21DP (s )
N 0.3287s2 0.0716s 0.00615
G 21DP (s ) = = 3 , (38)
D21DP (s )
s + 0.8559s2 + 0.92131s + 0.08027
7240 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Fig. 8. (a) Step response and (b) frequency response comparison for Example 4.

which shows an ISE of 1.12626. The comparison between step response of original system (G21 (s)), reduced third order model
by proposed method (G 21 (s )) and the reduced order model by Desai and Prasad mixed method (G 21DP (s )) are shown in Fig.
9. The step and frequency responses depicted have deviations but our proposed method is better then Desai and Prasads
mixed method [9]. And again it can be observed that our model is emulating frequency response characteristics in working
range of frequency very well, hence this model can be used successfully keeping in mind a difference in step response.

4.5. Example 5

Time delays often appear in many real-world engineering systems either in state, the control input, or the measurements
(output). Now let us consider in this example a time delayed system of order 7 as
(40 0 0s + 50, 0 0 0 )e0.3s
G (s ) = . (39)
s7 + 69s6 + 1764s5 + 20, 280s4 + 102, 500s3 + 221, 375s2 + 187, 500s + 50, 0 0 0
We approximate the delay time in (39) upto third order using Pad approximation as

120 60sT + 12(sT ) (sT )

2 3
R3 ( s ) = (40)
120 + 60sT + 12(sT ) + (sT )
2 3

and using (40), the resultant transfer function is

40 0 0s4 + 110, 0 0 0s3 666, 700s2 15560, 0 0 0s + 22220 0, 0 0 0

GPade (s ) = .
s + 109s + 5191s8 + 141, 300s7 + 2396, 0 0 0s6 + 25680, 0 0 0s5 + 16750 0, 0 0 0s4 + 61050 0, 0 0 0s3
10 9

+11110 0 0, 0 0 0s2 + 86670 0, 0 0 0s + 22220 0, 0 0 0

(41)
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7241

Fig. 9. (a) Step and (b) frequency response comparison for Example 4.

The proposed method is applied to transfer function in (41), the obtained ci  s are c0 = 1, c1 = 3.9705, c2 = 10.484 and
c3 = 23.7888. Using these values, we reduce the system to a second-order model. The denominator of this second order
model is D (s ) = 2.928s2 + 3.378s + 1. Now, using BBBC algorithm, we nd out an optimal numerator as N (s ) = 0.238432s2
0.54329s + 1. Using this numerator and denominator, the reduced order model can be written as

N (s ) 0.238432s2 0.54329s + 1
G (s ) = = (42)
D (s ) 2.928s2 + 3.378s + 1
If Desai and Prasads mixed method [9] is applied on (41), the obtained transfer function is
0.0989s + 0.2328
G DP (s ) = . (43)
s2 + 0.90802s + 0.2328
The transfer function in (42) and (43) shows an ISE of 0.0329 and 0.5928, respectively which clearly shows that proposed
method is highly reliable. Fig. 10 shows the comparison between original system (G(s)), Pad approximated model (GPade (s)),
proposed reduced model (G (s )), and Desai and Prasads reduced model (G DP (s )). Note that here GPade (s) and G(s) in
Fig. 10 exactly imitates each other, hence distinction is not possible. Step response in Fig. 10(a) clearly shows that proposed
method is highly reliable in time domain and performs better than Desai and Prasads method, whereas it is observed that
in frequency response (Fig. 10(b)) the reduced order models are unreliable at higher frequency but performs well in low
frequency region.

4.6. Example 6

Example considered in (39) has a delay of 0.3 seconds which is negligible, therefore let us consider another example
having a delay of 2 seconds as

(40 0 0s + 50, 0 0 0 )e2s

G (s ) = . (44)
s7 + 69s6 + 1764s5 + 20, 280s4 + 102, 500s3 + 221, 375s2 + 187, 500s + 50, 0 0 0
7242 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

Fig. 10. (a) Step and (b) frequency response comparison for Example 5.

Using the Pad approximation in (44), the resultant transfer function is

40 0 0s4 26, 0 0 0s3 + 240, 0 0 0s2 690, 0 0 0s + 750, 0 0 0

GPade (s ) = (45)
s + 75s + 2193s8 + 31914s7 + 251, 620s6 + 1167, 0 0 0s5 + 3357, 0 0 0s4 + 6032, 0 0 0s3
10 9

+6433, 0 0 0s2 + 3563, 0 0 0s + 750, 0 0 0

The reduced model is obtained using proposed mixed method which is given as

1.27s2 0.898s + 1
G (s ) = . (46)
5.622s2 + 4.2859s + 1

In order to show comparative results, the reduced order model is obtained using existing Desai and Prasads mixed
method which is given as

0.01336s + 0.14526
G DP (s ) = . (47)
s2 + 0.6901s + 0.14526

The reduced order transfer function (46) and (47) show ISE of 0.8063 and 1.7405, respectively. Fig. 11(a) and (b) show step
and frequency response comparison of original system (G(s)), Pad approximated model (GPade (s)), proposed reduced order
model (G (s )), and Desai and Prasads reduced order model (G DP (s )) respectively. It can be observed from step response in
Fig. 11(a) that performance of proposed method is better than Desai and Prasads method, whereas in frequency response
in Fig. 11(b), the reduced order models show divergence in higher frequency region while show strong resemblance in low
frequency region.
 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244 7243

Fig. 11. (a) Step response and (b) frequency response comparison for Example 6.

5. Conclusions

In this paper, we have proposed a new scheme of model-order reduction for SISO, MIMO and time delayed LTI systems.
The proposed scheme unies the concept of time moment matching and BBBC optimization to obtain the reduced model. In
this scheme, time moment matching scheme is applied to the denominator of the original system which helps in retaining
the dominant properties of the original systems into the reduced model. Further, BBBC algorithm is utilized to evaluate the
numerator of the reduced model. Finally, simulation studies conrm the eciency of the proposed scheme in both time and
frequency domain analysis.

References

[1] W.H.A Schilders, H.A. van der Vorst, J. Roomes, Model Order Reduction: Theory, Research, Aspects and Applications, Springer-Verlag, Berlin, Heidelberg,
2008.
[2] G. Obinata, B.D.O. Anderson, Model Order Reduction for Control System Design, Springer-Verlag, London, Berlin, Heidelberg, 2001.
[3] R. Eid, Time Domain Model Reduction by Moment Matching PhD thesis, Technische Universitat Munchen, 2009.
[4] L. Fortuna, G. Nunnari, A. Gallo, Model Order Reduction Techniques with Applications in Electrical Engineering, Springer-Verlag, New York, 1992.
[5] P. Benner, V. Mehrmann, D.C. Sorensen, Dimension Reduction of Large-Scale Systems, vol. 45, Springer, 2005.
[6] Muscato, G., G. Nunnari, and L. Fortuna, Singular perturbation approximation of bounded real balanced and stochastically balanced transfer matrices,
Int. J. Control, 66(2), pp.253270
[7] R. Hassan, B. Cohanim, O. de Weck, A comparison of particle swarm optimization and the genetic algorithm, in: Proceedings of the
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dyanamics and material conference Austin, Texas, April 2005, pp. 18971910.
[8] O.K. Erol, I. Eksin, New optimization method: big BangBig Crunch, Adv. Eng. Softw. 37 (2006) 106111.
[9] S. Desai, R. Prasad, A novel order diminution of LTI systems using big bang big crunch optimization and Routh approximation, Appl. Math. Model. 37
(2013) 80168028.
[10] V. Zakian, Simplication of linear time-invariant systems by moment approximants, Int. J. Control 18 (1973) 455460.
[11] S Janardhanan, Model Order Reduction and Controller Design Techniques, http://web.iitd.ac.in/janas/courses/material/eel879/sp_topics_01.pdf.
[12] A. Sikander, R. Prasad, Linear time-invariant system reduction using a mixed methods approach, Appl. Math. Model. 39 (2015) 48484858.
[13] Y. Shamash, Truncation method of reduction, a viable alternative, Electron. Lett. 17 (2) (1981) 9799.
[14] V. Krishnamurthy, V. Sheshadri, Model reduction using Routh stability criterion, IEEE Trans. Automat. Control 23 (1978) 729731.
7244 S. Biradar et al. / Applied Mathematical Modelling 40 (2016) 72257244

[15] T.C. Chen, C.Y. Chang, K.W. Han, Stable reduced order Pad approximants using stability equation method, Electron. Lett. 16 (9) (1980) 345346.
[16] P. Kokotovic, H. Khalil, J.O. Reilly, Singular Perturbation Methods in System and Control: Analysis and Design, IEEE Press, New York, 1986.
[17] B.C. Moore, Principal component analysis in linear systems: controllability, observability and model reduction, IEEE Trans. Automat. Control AC 26
(1981) 1731.
[18] G. Parmar, R. Prasad, S. Mukherjee, Order reduction of linear dynamic systems using stability equation method and GA, Int. J. Comput. Inf. Eng. 1 (1)
(2007) 2632.
[19] T.C. Chen, C.Y. Chang, K.W. Han, Stable reduced order Pad approximants using stability equation method, Electron. Lett. 16 (9) (1980) 345346.
[20] J. Pal, Improved Pad approximants using stability equation method, Electron. Lett. 19 (11) (1983) 426427.
[21] R. Parthasarathy, K.N. Jayasimha, System reduction using stability equation method and modied Cauer continued fraction, Proc. IEEE 70 (10) (1982)
12341236.
[22] E.J. Davison, A method for simplifying linear dynamic systems, IEEE Trans. Automat. Control AC 11 (1966) 93101.
[23] L.S. Shieh, Y.J. Wei, A mixed method for multivariable system reduction, IEEE Trans. Automat. Control AC 20 (1975) 429432.
[24] M.G. Safonov, R.Y. Chiang, Model reduction for robust control: a Schur relative error method, Proc. Int. J. Adapt. Control Signal 2 (1988) 259272.