Improved Modal Truncation Approximant: A Hybrid Approach
Improved Modal Truncation Approximant: A Hybrid Approach
Improved Modal Truncation Approximant: A Hybrid Approach
Abstract: A hybrid method using modal truncation and singular perturbation to derive the reduced
order model for a stable higher order system is presented. The approach is based on retaining the
dominant modes of the system and truncating comparatively the less significant once. As the
reduced order model has been derived from retaining the dominant modes of the original high
order stable system, the reduced order model preserves the stability. The strong demerit of the
modal truncation method is that, the steady state value of the reduced model does not match with
the original system. This demerit has been overcome by applying singular perturbation approach
of the balanced truncation method. Results obtained show the effectiveness of the proposed
technique. Numerical examples are carried out to illustrate the procedure. Further it has been also
applied to a system of order 1006 to check its applicability to large scale systems as well.
Keywords: Modal truncation, Singular perturbation, Steady state value, Eigen value, Large scale
systems.
8
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
neglected, the size of the original plant, and the reduced Hankel singular values (HSV) of the HOS is
plant. This criterion is judged by comparing the time computed and checked for comparatively
responses of various low order systems. An approach for significant HSV. The number of dominant HSV
selecting the order of the model for Davison technique was will be the order of reduction r .
proposed by Mahapatra [8]. Modal matrix P is calculated and partitioned as
In this present work the reduced order model has been below
obtained through modal truncation technique. The steady
state mismatch has been avoided by utilizing singular P P12
P 11 (5)
perturbation technique. The author has proposed to utilize P21 P22
the singular perturbation method [9] of balance truncation
to hybridize with modal truncation method. In singular where P11 is a matrix of order r r , P22 is a matrix
perturbation the states are grouped into two modes slow
of order n r n r , P12 is a matrix of order
and fast modes. The fast mode is made equal to zero thus
getting the slow modes of the system along with the r n r and P21 is a matrix of order n r r
.
contribution of the fast modes also. This is also known as Modal inverse matrix Q which is inverse of
residualization. With this the steady state of the reduced matrix P is obtained and partitioned as below
system matches with the higher order system. The
advantage of the proposed method is its applicability to Q Q12
Q 11 (6)
large scale systems which has been validated through an Q21 Q22
example of a system of order 1006 [10].
where Q11 is a matrix of order r r , Q22 is a
Statement of the problem
matrix of order n r n r , Q12 is a matrix of
Consider the linear time-invariant stable nth order state order r n r and Q21 is a matrix of order
space model described by following equations
x Ax Bu n r r .
(1)
y Cx Du Partition the matrices A, B, C and D as given
with the transfer function matrix below
1
G s C sI A B D (2) A A12
n m p
A 11
where x R , y R , u R A21 A22
the aim of the model order reduction is to obtain the low B
order representation described as B 1 (7)
B
2
xr Ar xr Br u
(3) C C1 C2
yr Cr xr Dr u
with transfer function matrix
1
Where A11 is a matrix of order r r , A22
Gr s Cr sI Ar Br Dr (4) is a matrix of order n r n r , A is a matrix
12
where xr R r , yr R m such that yr t is a close
of order r n r and A21 is a matrix of order
approximation to y t for all input u t . Moreover the
n r r
key features of the original system should be preserved in Now after the completion of section (i) the
the reduced model such as stability. reduced order model can be given as
2. PROPOSED MODEL REDUCTION ALGORITHM
Aˆ11 A11 A12 P21P111
The proposed algorithm is the result of hybridization of
Bˆ1 B1
modal truncation [1, 11-13] and singular perturbation [9] (8)
approach. It consists of two sections namely modal ˆ
C1 C1
truncation and singular perturbation. ˆ
D1 D
2.1 Modal truncation Due to mismatch in DC gain of the reduced order model
thus obtained with the original system, further singular
State space realization of the high order system perturbation is applied. This overcomes the DC gain
(HOS) as in (1) is obtained. mismatch problem. Singular perturbation steps are carried
out in continuation to modal truncation steps. It has been
All the Eigen values A of the HOS are explained in next sect
evaluated.
9
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
Partitioned forms as above can be used to construct sys
Magnitude (dB)
singular perturbation approximation [9]. The matrices of -50
the final reduced system can be given as -100
(9)
Phase (deg)
180
Cr Cˆ1 C2 A22 1 A21 90
Dr Dˆ C2 A22 1B2 0
-90
-2 -1 0 1 2 3
10 10 10 10 10 10
Frequency (rad/sec)
The proposed algorithm is the result of hybridization of
conventional modal truncation technique with singular Fig. 2. Bode magnitude and phase plot of the HOS
perturbation approximation method so as to utilize the The various steps to derive the reduced order model has
merits of the two methods. The technique has been applied been presented as follows.
to higher order systems which are discussed in the
preceding section. The eigen-values of the HOS are given as A .
A 1, 3, 5, 10
3. NUMERICAL EXPERIMENTS
As all the poles lie in the left half of s-plane so the HOS is
3.1 Example-1
stable.
This example is taken from [9]. It is a continuous, linear, Henkel singular value of the HOS has been calculated and
time-invariant and stable system of order four. The state- also plotted in Fig. 3, which gives the indication of the
space matrices are described as follows order of reduction. Comparatively the number of nonzero
dominant singular values is taken as order of the reduction.
0.43781 1.1685 0.41426 0.05098
Here, from Fig. 3, the first two singular values are
1.1685 3.1353 2.8352 0.32885
A significant and the last two singular values have become
0.41426 2.8352 12.4753 3.2492 insignificant.
0.05098 0.32885 3.2492 2.9516 0.016
Hankel Singular Values (State Contributions)
Stable modes
0.11814 0.014
0.1307
B 0.012
0.05634
0.006875
0.01
State Energy
0.004
The step response and bode diagram of the HOS is shown
in Fig. 1 and Fig. 2 respectively. 0.002
0
Step Response 0 1 2 3 4 5
0.03
State
sys
0.025 Fig. 3. Hankel singular values plot of the HOS
0.02
From the above plot order of the system to which it should
Amplitude
0.015 be reduced is decided as 2.
The algorithm for modal truncation has been applied to the
0.01
high order system (10).
0.005 The modal matrix P obtained is as:
1.0314 0 0 0
0
0 1 2 3 4 5 6 7 8 0 2.8885 0 0
Time (sec)
P (11)
0 0 5.1070 0
Fig. 1. Step response of the HOS
0 0 0 9.9731
Partitioned matrices P11 ,P12, P21 and P22 are given as
10
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
-1.1685 -3.1353 -2.8352 -0.3288
0.025
0.4143 2.8352 -12.4753 -3.2492
A21 ; A22 3.2492 -2.9516
-0.5098 -0.3288 (15) 0.02
sys
sysr1
Amplitude
C1 -0.1181 0.1307 ; C2 0.0563 -0.0069
0.005
Now applying the modal truncation only as given by (8)
the reduced system matrices obtained are given as 0
0 1 2 3 4 5 6 7 8
from the original high order system.
0.025
Step Response
0.03
0.02
0.025
0.015
sys
Amplitude
0.02 sysr
0.01
0.015
Amplitude
0.01 0.005
sys
0.005 sysrm
0
sysrmsp
0
-0.005
0 1 2 3 4 5 6 7 8
-0.005 Time (sec)
0 1 2 3 4
Time (sec)
5 6 7 8
Fig. 4. Step response of reduced system with HOS using Fig.6. Comparision of response of reduced system with
modal truncation only HOS by proposed algorithm and modal method
11
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
Magnitude (dB)
(18)
where e t y t yr t 20
0
y t is the step response of high order system and yr t is
-20
the step response of reduced order system. 90
sys
The ISE has been obtained for the above two cases. The 45
Phase ( deg)
table shows the values obtained. 0
-45
Table.1. Integral Square Error
-90
-1 0 1 2 3 4
10 10 10 10 10 10
S.No. Method used ISE Frequency (rad/sec)
1 Modal Truncation 1.27x10-6 Fig.8. Bode magnitude and phase plot of the HOS (19)
2 Proposed Approach 6.21x10-8
The eigen-values of the high order system are given as
ISE value by the proposed approach is much less as A and corresponding pole zero plot is given in Fig.9.
compared to modal truncation method. So it confirms the
effectiveness of the proposed approach. A 1, 2,..., 1000, 1 100 j , 1 200 j , 1 400 j
Pole-Zero Map
400 sys
300
100
Imaginary Ax is
This example is taken from [10]. It is a dynamical system 0
of order 1006. The state-space matrices are given by -100
sys A, B, C , D -200
-300
A1 0 0 0 -400
0
A2 0 0 -1000 -900 -800 -700 -600 -500 -400 -300 -200 -100 0
A Real Axis
0 0 A3 0
Fig.9. Pole-zero plot of the high order system (19)
0 0 0 A4
1 100 1 200 1 400 (19) Henkel singular value of the high order system has been
A1 ; A ; A
2 200 1 3 400 1 ; calculated and also plotted in Fig.10, which is indication
100 1
of the order of reduction. Comparatively the number of
A4 diag 1,2,...,1000
nonzero dominant singular values is taken as order of the
system to be reduced. Here the first eight singular values
; D 0
T
B C 10,10,...,101,1,...,1 are significant and after that ninth singular values onwards
6 1000
have decayed sharply [14], becoming insignificant. So the
The step response and bode diagram of the original high order of reduction is decided as eight.
order system is shown in Fig.7. and Fig.8. respectively.
Hankel Singular Values (State Contributions)
60
7 40
State Energy
6
30
5
Amplitude
4
20
3
2 10
sys
1
0
0 0 2 4 6 8 10 12
0 1 2 3 4 5 6 State
Time (sec)
Fig.7. Step response of the high order system (19)
Fig.10. Henkel singular value plot of the HOS (19)
From the plot of Fig.11., order of the system to which it
should be reduced is decided as 8.
12
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
The Modal truncation method alone has been applied to Further, proposed approach has been applied to the
the high order system (19). The reduced system matrices system. The reduced system matrices obtained are given as
obtained are given as sysr Ar , Br , Cr , Dr sysr1 Ar1 , Br1 , Cr1 , Dr1
1 100 0 0 0 0 0 0
100 1
0 0 0 0 0 0 6.112 1 .276 80.26 0.3737 69.88 180.3 9.427 9.229
0 0 1 200 0 0 0 0 1 .276 0.2662 361.8 0.07799 44.53 81.91 1.968 1.927
0 0 200 1 0 0 0 0 80.26 361.8 0.0003607 1 2.95 0.0004454 0.00143 0.0791 0.07371
Ar
0 0 0 0 1 400 0 0
0.3737 0.07799 12.95 0.02285 1 54.6 50.48 0.5769 0.5648
0 0 0 0 400 1 0 0 (20) Ar1
69.88 44.53 0.0004454 154.6 0.00055 0.001766 0.09769 0.09104
0 0 0 0 0 0 1 0
180.3 81.91 0.00143 50.48 0.001766 0.005669 0.3137 0.2924
0 0 0 0 0 0 0 2 9.427 1.968 0.0791 0.5769 0.09769 0.3137 90.56 120.5
Br 10 10 10 10 10 10 1 1
T
9.229 1.927 0.07371 0.5648 0.09104 0.2924 120.5 189.3
Cr 10 10 10 10 10 10 1 1
24.73
Dr 5.9855
5.159
0.1899
The step response of the reduced system (20) with high
order system (19) is compared in the Fig.11, which is seen 1.511
Br1
to be approximant of the system with dc gain mismatch. 0.2344
0.7526
Step Response
19.91
9
sys
8 sysr1
19.03
7
Cr1 [24.73 5.159 0.1899 1.511 0.2344 0.7526 19.91 19.03]
6
Dr1 5.9855
Amplitude
4
(21)
3
The step response of the reduced system (21) with high
2
order system (19) is compared in the Fig.13, which is seen
1
to have close approximant with steady state matching from
the original high order system.
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (sec)
9
Step Response
Fig.11. Step response of reduced system with high order sys
8 sysr
system using Modal truncation only
Frequency response of the reduced system with the higher 7
order system is also presented in Fig.12, which reveals the 6
behaviour as shown in Fig.12. 2
1
Bode Diagram
60 0
0 1 2 3 4 5 6
40 Time (sec)
Magnitude (dB)
20
0
-20
Fig.13. Step response of reduced system with high order
90
system by proposed technique
45
Phase (deg)
-45 Frequency response of the reduced system with the higher
-90
10
-1
10
0
10
1 2
10
3
10
4
10
order system is also shown in Fig.14, which reveals the
Frequency (rad/sec) reduced system to be close approximation of the high
Fig.12. Frequency response of reduced system with high
order system using modal truncation only
order system.
13
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
Bode Diagram
60
40
REFERENCES
Magnitude (dB)
0
simplifying linear dynamic systems. In: IEEE
-45
Transactions on Automatic Control
-90
(Correspondence) AC-12: 119-121. 1967
-1 0 1 2 3 4
10 10 10 10 10 10
Frequency (rad/sec)
3. Chidambara M. R. and Davison E. J.: Further
Fig.14. Frequency response of reduced system with high remarks on simplifying linear dynamic systems
order system by proposed technique IEEE Transactions on Automatic Control
(Correspondence) AC-12: 213-214. 1967.
Further another step response of the reduced system with
the high order system for one second has also been shown 4. Chidambara M. R. and Davison E. J.: Further
in Fig.15 to clearly view the transient behaviour remarks on 'A method for simplifying linear
comparison. From Fig.15, the response show considerable dynamic systems. In: IEEE Transactions on
deviation from the high order response but it lasts for 0.03 Automatic Control (Correspondence) AC-12: 799-
seconds only. This short high dynamic behaviour, if 800,1967a.
ignored the rest of the response is quite close. 5. Marshall, S. A.: An approximate method for
Step Response
9
sys
reducing the order of a linear system. In:
8
7
sysr
International journal of Control: 642-643, 1966.
6
6. Mitra D. : The reduction of complexity of linear
5
time-invariant systems. In: Proceedings of 4th IFAC
Amplitude
3
technical series 67, (warsaw): 19-33 1969.
2
7. Kumar A. and Chandra D.: Improved Padé-Pole
1
0
Clustering Approximant. In: International
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec) Conference on Computer Science and Electronics
Engineering, Dubai, UAE,17-18 November, 2013.
Fig.15. Step response of reduced system with high order
8. Mahapatra G. B.: A Note on Selecting a Low-Order
system for short duration (one second)
System Davison's Model Simplifcation Technique.
IEEE Transactions on Automatic Control AC-22(4):
Integral Square Error (ISE) has been calculated to show
677-678., 1977.
the effectiveness of the approach. The ISE obtained
ignoring the short transient is 0.1966, which ensures the 9. Liu Y. and Anderson B.D.O. : Singular perturbation
effectiveness of the method. approximation of balanced systems. In:
International journal of control 50: 1379-1405,
4. CONCLUSION
1989.
A hybrid method utilizing modal truncation and singular
10. Chahlaoui Y. and Dooren P. Van. : A collection of
perturbation approximation approach has been applied
benchmark examples for model reduction of linear
successfully to determine the reduced replica of a high
time invariant dynamical systems. SLICOT Working
order system. One of the important demerits of the modal
Note, 2002.
truncation method is dc-gain mismatch i.e., steady state
value of the reduced order system do not agree with high 11. Antoulas, A.C.: SIAM Book series, Advances in
order system. This demerit has been removed in the Design and Control, DC-06. 2008.
proposed approach. Further the method has been tested on
high order system of order 1006. The results obtained 12. Antoulas, A.C.: A new result on passivity preserving
confirm the validity of the approach. Applicability to large model reduction. Systems and Control Letters 54(4):
scale systems makes the method more advantageous. The pp.361-374.2005.
high frequency transients show approximation error 13. Antoulas, A. C., Sorensen, D.C. & Zhou, Y.: On
though it is of very short duration which is open for further the decay rate of Hankel singular values and
investigation. related issues. In: Systems and Control Letters,
46(5), pp.323–342. 2002.
14
Journal of Electrical Engineering ISSN 1582-4594 - Vol 22 No 1 (2022)
14. Soni M.G., Chitra D. R. and Pooja Soni :Model D=[0]
Order Reduction- A Time Domain Approach In:
Special Issue of International Journal of Computer % Algorithm for Modal truncation in
Applications (0975-8887) on Electronics, time domain
Information and Communication Engineering- A=input('A=')
ICEICE (2): 6-9,2011. B=input('B=')
C=input('C=')
15. AvadhPati, Kumar Awadhesh and Chandra Dinesh
D=input('D=')
Suboptimal Control Using Model Order Reduction.
E=eigs(A)
In: Chinese Journal of Engineering, Hindawi
M=modreal(A)
Publishing Corporation Article ID 797581,
n=size(M)
2014(2014),pp.1-5, 2013.
r=input('r=')
16. Awadhesh Kumar, Bhoomika Maurya and Dinesh M1=M(1:r,1:r)
Chandra: Dimension reduction and controller M2=M(1:r,r+1:n)
design for a waste water treatment plant In: IEEE M3=M(r+1:n,1:r)
International Conference on Power and Advanced M4=M(r+1:n,r+1:n)
Control Engineering (ICPACE), Pages:413-417, N=inv(M)
13-14 August, 2015, IEEE Conference A11=A(1:r,1:r)
Publications DOI: 10.1109/ICPACE.2015.7274983 A12=A(1:r,r+1:n)
A21=A(r+1:n,1:r)
17. Deepak Gupta and Awadhesh Kumar: A22=A(r+1:n,r+1:n)
Approximation of Large Scale Systems by Balanced B1=B(1:r,:)
Truncation And Singular Perturbation Method In: B2=B(r+1:n,:)
i-manager’s Journal on Instrumentation and Control C1=C(:,1:r)
Engineering, Volume: 4, No. 2, Issue : Feb- C2=C(:,r+1:n)
Apr 2016, Pages : 1-6, 2016.
% Applying Only Modal Truncation
APPENDIX Ar=A11+A12*M3*inv(M1)
Br=B1
% Program for the Modal Truncation and
Cr=C1
Singular Perturbation
Dr=D
Dr=D-C*inv(A)*B+Cr*inv(Ar)*Br
% Defining FOM real time model from
benchmark examples
% Applying Singular Purturbation to
A1=[-1 100;-100 -1]
match steady state
A2=[-1 200;-200 -1]
A11n=A11+A12*M3*inv(M1)
A3=[-1 400;-400 -1]
Ar=A11n-A12*inv(A22)*A21;
A4=diag(-1:-1:-1000)
Br=B1-A12*inv(A22)*B2;
A=blkdiag(A1,A2,A3,A4)
Cr=C1-C2*inv(A22)*A21;
C1a=[10 10 10 10 10 10]
Dr=D-C2*inv(A22)*B2;
C2a=ones(1,1000)
C=[C1a C2a]
B=C' % Ploting rsponse and comparision
D=[0] sys=ss(A,B,C,D)
sysr=ss(Ar,Br,Cr,Dr)
fig1=figure
% Eigen analysis and Hankel singular
step(sys,'r',sysr,'bo')
values
fig2=figure
eig(A)
bode(sys,'r',sysr,'bo')
sys=ss(A,B,C,D)
hankelsv(sys)
% ISE CALCULATION..
t=0:0.1:6;
% Defining Example-1 4th order model
[y,t]=step(sys,t)
A=[-0.43781 1.1685 0.41426 0.05098;
[y1,t]=step(sysr,t)
-1.1685 -3.1353 -2.8352 -0.32885;
e=y-y1
0.41426 2.8352 -12.4753 -3.2492;
j=e.^2
-0.5098 -0.32885 3.2492 -2.9516]
ISE=sum(j)
B=[-0.11814;-0.1307;0.05634;-0.006875]
% END of program%
C=[-0.11814 0.1307 0.05634 -0.006875]
15