Copyright (c) 2005 IFAC.

All rights reserved

16th Triennial World Congress, Prague, Czech Republic

IDENTIFICATION OF A STATE SPACE MODEL FOR A SYNCHRONOUS MACHINE

Bin Wu, Peng Zhao, Om Malik

Department of Electrical and Computer Engineering, University of Calgary

Calgary, Alberta, Canada T2N 1N4

Abstract: Application of a recursive subspace identification method to derive a state space

model for a synchronous machine is described in this paper. Simulation studies show the
effectiveness of such an algorithm to identify on-line a synchronous machine model over
a wide range of operating conditions and disturbances. The model provides a foundation
for further study on a MIMO adaptive power system stabilizer. Copyright © 2005 IFAC

Keywords: Synchronous machines; State-space models; Recursive algorithms; MIMO

1. MOTIVATION connected to an infinite bus. In Sec. 5, simulation

results of MIMO identification are presented. Finally,
One aspect of power system stability is dynamic the conclusions are given in Sec. 6.
stability. The unstable situation in this case results in
long-term low frequency oscillations. Power system
stabilizer (PSS) is used to dampen such oscillations 2. INTRODUCTION OF THE ORDINARY MOESP
and it efficiently improves the stability of the power (OM) SCHEME
system.
Mathematically, a state space model can be
In recent years adaptive PSS (APSS) has been expressed by the following equations (Verhaegen and
developed to overcome inherent shortcomings of Dewilde, 1992):
conventional PSSs. Most of them utilize the x k +1 = Ax k + Bu k + w k
excitation control based on the single input single (1)
output model of a synchronous machine. As is well y k = Cx k + Du k + v k
known, governor can also improve the power system n´1
stability by maintaining the generator output and where x k Î R , u k Î R m´1 , y k Î R l´1 .The noise
frequency at predetermined values. processes are the process noise w k and the output
It is worth trying to coordinate excitation and measurement noise v k . They are assumed to be
governor controls together, that is, a multi-input discrete time, zero mean, white noise with
multi-output (MIMO) APSS. To obtain such a
appropriate dimensions.
controller, an accurate model of the plant is required.
State space model is a better choice than transfer For subspace-based algorithms, the following matrix
function model that is cumbersome with polynomial definitions are frequently used.
representations in the multivariable case. Thus,
· Hankel matrix
identification of a MIMO state space model is a key
step in designing a model based APSS. The Hankel matrices are constructed from the input
and output data. The following is an input Hankel
The paper is organized as follows. An ordinary matrix:
MOESP (MIMO Output Error State Space) é uk u k +1 ... u k + N -1 ù
algorithm is introduced in Sec. 2. Recursive subspace êu u k +2 ... u k + N úú
identification based on PAST (Projection U k ,i , N = ê k +1 (2)
Approximation Subspace Tracking) approach is ê ... ... ... ... ú
proposed in Sec. 3. Section 4 provides some ê ú
discussion concerning identification of a simple ëu k +i -1 u k +i ... u k +i + N - 2 û
power system consisting of a single machine where k is the starting time index, i is the number of

262
the rows and N is the number of columns. Estimation of B and D becomes a least squares
· Extended observability matrix Γi Ù Ù
regression problem with known A and C .
[
Γi = C CA ... CA i -1 ]T
(3) Mathematically
k -1
· Toeplitz matrix y (k ) = CA k x(0) + å CA k -1-τ Bu(τ ) + Du(k )
é D 0 ... 0ù τ =0

ê CB (15)
D ... 0 úú
Hi = ê (4) This equation can be rewritten in the following form
ê ... ... ... ...ú by using the Kronecker product Ä and stacking the
ê i-2 i -3 ú columns of B and D by vec( B ) and vec( C ).
ëCA B CA B ... Dû
é k -1 ù
· Matrix X k , N of state vectors y (k ) = CA k x(0) + êå u(τ ) T Ä CA k -1-τ ú vec(B)
ëτ = 0 û
X k , N = [x k x k +1 ... x k + N -1 ] (5)
[
+ u(k ) Ä I l vec(D)
T
]
For the special case of the absence of w k , v k , the
(16)
data equation can be written in a condensed form: Construct a vector with all output signals
Yk ,i , N = Γi X k , N + H i U k ,i , N (6) Y0, N ,1 = Γi [ Εy Ε u [x 0] vec(B) vec(D)]
T

It is obvious from (6) that the output from the system (17)
contains two parts. One is the zero input response
Γi X k , N , the other is the zero state response
where Y0, N ,1 = y (0) [ y (1) ... y ( N - 1)]
T

T
é N -2
ù
H i U k ,i , N . Using the LQ factorization to get rid of E y = ê0 u(0) T Ä C ...
ë
å
τ =0
u(τ ) T Ä CA N - 2-τ ú
û
the zero state response from the output,
spancol (Γi ) can be estimated and hence the
matrices A and C .
[
E u = u(0) T Ä I l u(1) T Ä I l ... u( N - 1) T Ä I l ]
T

é U k ,i , N ù é L11 0 ù é Q1 ù Γi is the extended observability matrix as defined in

êY ú = LQ = ê úê ú (7)
ë k ,i , N û ëL 21 L 22 û ëQ 2 û (3). Solving (17)
where Q is an orthogonal matrix. [x 0 vec(B) vec(C)] = Γi
T
[ Εy Εu ]
ψ
Y0, N ,1
U k ,i , N = L11Q1 (18)
Yk ,i , N = L 21Q1 + L 22 Q 2 For the more general model with noise w k and v k ,
Yk ,i , N Q = L 21Q1Q + L 22 Q 2 Q = L 22 (8)
T
2
T
2
T
2 (6) becomes:
From (6) Yk ,i , N = Γi X k , N + H i U k ,i , N + G i Wk ,i , N + Vk , j , N
Yk ,i , N Q = Γi X k , N Q + H i U k ,i , N Q
T
2
T
2
T
2
(19)
where Wk ,i , N and Vk ,i , N are Hankel matrices
= Γi X k , N Q T2 + H i L11Q1Q T2 (9)
containing the process noise and the output
= Γi X k , N Q T2 measurement noise, respectively. Matrix G i shows
Comparing (8) and (9) the impact from w k on the output.
L 22 = Γi X k , N Q T2 (10) Introduction of IV (Instrumental Variable) can help
Doing a SVD factorization of L 22 , a consistent to get rid of the influence of noise without disturbing
the column space of Γi due to the following
estimate of span col (Γi ) can be obtained.
properties.
L 22 = USV2T (11)
[
1) Ε e(t )Ξ (t )
H
]= 0
[ ]} = n
Ù
Γi = U(:,1 : n ) (12) 2) Rank (M xξ ) = Rank{Ε x(t )Ξ (t )
H

It is easy to find the estimate of A and C from Γi .
Ù Ù
C = Γi (1 : l , :)
A can be estimated by solving the least squares
problem. ψ is the Moore-Penrose pseudo-inverse.
Ù Ù Ù Different source and construction of IV can be used
A = Γi (1 : (i - 1)l , :)ψ Γi (l + 1 : end , :) (14)
Ù
where e(t ) represents signal that needs to be
eliminated while x(t ) is the useful signal.
(13) Ξ (t ) Î R p´1 , p ³ n is an IV vector.
The rest of the algorithm to estimate system matrices
is the same as that used in the noise-free environment.
to handle different noise model (Lovera et al., 2000).

263
 3. RECURSIVE SUBSPACE IDENTIFICATION t 2
J ( W (t )) = å λt - k x(k )Ξ (k ) - W(t ) W(t ) H x(k )Ξ (k )
As mentioned before, it is not only necessary to k =1
2
derive a state space model, but also need to update t
» å λt - k x(k )Ξ (k ) - W (t )h(k )Ξ (k )
the model on line to adapt to the system variations. A k =1
bottleneck to the recursive implementation is the (23)
update of the SVD factorization. The idea of
applying subspace-tracking algorithms to the J ( W (t )) is minimized by
recursive subspace identification was introduced in Ù Ù
(Gustafson, 1998) to overcome the difficulty. A W (t ) = M xξ (t ) M hξ (t )ψ
(24)
successful subspace-tracking algorithm is the PAST More details about the algorithmic steps of a
approach (Yang, 1995). The basic idea is to treat the recursive formulation of (23) are given in (Gustafson,
signal subspace-tracking problem as the solution of 1998).
an unconstrained minimization task. Furthermore a
projection approximation reduces the minimization
to the well-known exponentially weighted least 3.3 Recursive updating span col (Γi ) in the OM
squares problem.
scheme

3.1 The PAST scheme Recursive estimation of A is equal to recursive

updating spancol (Γi ) . In the OM scheme a partial
m´1
Consider a random vector x Î R , and study the update of LQ factorization is required with the help
unconstrained criterion of the method of Givens Rotations (Lovera et al.,
2
J ( W(t )) = E x(t ) - W(t ) W(t ) H x(t ) (20) 2000).

with a full When a new data point arrives, the LQ must be

rank ( = n) matrix W (t ) Î R
m´ n , m > n . updated as
é Q 1 (k ) 0ù
éU k ,i , N +1 ù é L 11 (k ) 0 Φ uf (k + 1) ù ê
Applying the theorem êY ú=ê ú êQ 2 (k ) 0úú
“The global minimum of J ( W(t )) is attained if and ë k ,i , N +1 û ëL 21 (k ) L 22 (k ) Φ yf (k + 1)û ê 0 1úû
ë
m´ n
only if W = U s T , where U s Î R contains the (25)

[
n dominant eigenvectors of E x(t )x(t ) T and T is ] where the new set of I-O data vectors is

an arbitrary unitary matrix.” [

Φ uf (k + 1) = u Tk + N ... u Tk +i + N -1 ] T
(26)

(k + 1) = [y ] T
From (Yang, 1995), the columns of W (t ) which Φ yf T
k+N ... y Tk +i + N -1 (27)
minimize J ( W (t )) form an orthonormal basis of Using Givens Rotations matrix P(k + 1) to
the signal subspace. annihilate Φ uf ( k + 1) .

For a practical algorithm, considering forgetting é P11 0 P12 ù

factor λ , replace (19) with P(k + 1) = êê 0 I 0 úú (28)
t 2
J ( W(t )) = å λ t -k
x(k ) - W(t ) W(t ) x(k )
H
êëP21 0 P22 úû
k =1
(21) é L11 (k ) 0 Φ uf (k + 1) ù
ê ú P(k + 1)
ëL 21 (k ) L 22 (k ) Φ yf (k + 1)û
The key issue of the PAST is to replace
W (t ) H x(k ) with h(k ) = W(k - 1) H x(k ) . (29)
é L11 (k + 1) 0 0 ù
Hence the original fourth-order function of W(t ) is =ê - ú
reduced to a quadratic problem. J ( W (t )) is ëêL 21 (k + 1) L 22 (k ) Φ yf (k + 1)ûú
( )
-
minimized by
-1
Φ yf = Γi X k , N Q1T P12 + x k + N P22
æ t öæ t ö (30)
W (t ) = ç å λt - k x(k )h(k ) H ÷ç å λt - k h(k )h(k ) H ÷ + E k ,i , N Q1T P12 + Φ ef P22
è k =1 øè k =1 ø
Ù Ù E k ,i , N denotes a Hankel matrix that consists of noise
= M hx (t ) M hh (t ) -1
(22) in the past. Φ ef represents current noise item at time
instant (k + 1) .
3.2 The IV- PAST scheme
The new information contained in
Using IV, similar results can be obtained as from the
standard PAST.
[Φ uf (k + 1) T Φ yf (k + 1) ]
T T
is condensed into a

264
rank one modification of L 22 ( k ) . Feeding
-
Φ yf (k + 1) directly to the PAST algorithm to
update spancol (R 22 ( k ) ) would result in a biased
estimation due to the noise it contains. IV-PAST is
used to rescue this problem. Previous I-O data are
usually chosen as IV Ξ (t ) Î R
p´1
, p ³ n.
4. SYNCHRONOUS MACHINE
IDENTIFICATION
The power system model consists of a synchronous
machine connected to an infinite bus through a
double circuit transmission line. The cylindrical rotor
machine is simulated by seven first order differential
equations in the d-q frame of reference (Anderson et
al., 1977). The AVR used is IEEE standard type
ST1A. Electro hydraulic governor and steam turbine
are also included in the model. System parameters
are given in the Appendix.
The system configuration for identification is shown
in Fig.1. For identification, the input signals are
V_pss and V_gov that are connected to the AVR and
governor summing junctions. Deviation of rotor
speed and active power, dw and dPe, are chosen as
the plant output signals.
Variance Accounted for (VAF) is used as a standard
to measure the accuracy of identification.
æ var iance( y measured - y estimated ) ö
VAF = çç1 - ÷÷ ´ 100%
è var iance( y measured ) ø e(k + 1) = ( A - KC)e(k )
(31)
GOVERNOR
SYNCHRONOUS
MACHINE
AVR
V_pss dw
V_gov IDENTIFIER dPe
Fig. 1. System configuration.
5.1 Selection of the model order n
Actually power system is a high order non-linear
system. However, for control purposes, it is not
necessary to identify a detailed model representing
the whole system. For on-line identification the
model is required to just stress the desirable features
of the system, damping the low frequency
oscillations of the mechanical mode of the
interconnected system in the present case. A model
that can represent this kind of oscillation accurately
is required. A third order model can usually provide
a pair of complex poles that represent the dominant
265
oscillation frequency of the system and a single real
pole that represents the free damping part of the
system response. Therefore the model order n is
selected as 3.
5.2 Scaling output data
Better identification results are obtained from a well-
conditioned problem than an ill conditioned one. The
results will be more sensitive to noise and
perturbations for an ill-conditioned estimation
problem.
dPe is nearly 100 times larger than dw. After scaling,
the two output signals are brought to the same order
of magnitude. Simulation studies show using scaled
data leads to more accurate result than just using
original data especially in the case of disturbance.
5.3 Observer for state estimation
Although state estimate is not a part of identification,
it plays an important role for output prediction as
well as the model parameters. The state space model
is identified directly from the I-O data. Therefore, the
states in the model have no physical meanings. An
observer can be established to estimate all the states
based on the following equations (Franklin et al.,
1998).
x(k + 1) = Ax(k ) + Bu (k ) (32)
Ù Ù Ù
x(k + 1) = A x(k ) + Bu (k ) + KC[x(k ) - x(k )]
(33)
Subtracting (32) from (33) yields the error vector
(34)
The solution to this homogeneous first-order vector
difference equation is given by
e(k ) = ( A - KC) k e(0) (35)
If the matrix ( A - KC ) has eigenvalues inside the
unit circle, then ( A - KC) approaches the zero
k
matrix as the time index k gets large.
Based on the pole placement algorithm, the
eigenvalues of ( A - KC ) are placed close to the
origin. Therefore, ( A - KC) approaches zero in
k
about n steps, where n is the system order.
5. SIMULATION RESULTS
In the simulation studies, the synchronous machine is
operating in steady state at Pe = 0.7 p.u. , 0.85 power
factor lag. The infinite bus voltage is 1.0 p.u. Using a
sampling frequency of 20Hz and a variable forgetting
factor (Park,1991) to improve fast tracking ability of
the algorithm, identification results were obtained
under steady state as well as under various
disturbances.
A model from off line identification was used as the
 initial parameters for recursive subspace Case 4: Disconnection of one transmission line, Fig.5.
identification. Consider the following situations, Case 5: A three phase to ground fault is applied at the
where all disturbances occur at 70s: middle of one transmission line and cleared 50ms
Case 1: Steady state, Fig.2. later by opening the breakers at both ends, Fig.6.
Case 2: Reference terminal voltage increased by 5%, Case 6: The same three phase to ground fault, using
Fig.3. original output data without scaling, Fig.7.
Case 3: Reference torque increased by 10%, Fig.4. Case 7: The same three phase to ground fault, using
fixed forgetting factor instead, Fig.8.
-3 -3
x 10 x 10
3 4
dwestimate

Deviation of rotor speed (p.u.)

dwestimate
Deviation of rotor speed (p.u.)

2 dwtrue dwtrue
2
1

0 0

-1
-2
-2

-3 -4
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

0.2 0.2

Deviation of active power (p.u.)

dPeestimate
Deviation of active power (p.u.)

dPeestimate
dPetrue dPetrue
0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

Fig. 2. Identifier response to a steady state. Fig. 5. Identifier response to disconnecting one
transmission line.
-3
x 10
6 0.02
dwestimate
Deviation of rotor speed (p.u.)

dwestimate
Deviation of rotor speed (p.u.)

4 dwtrue dwtrue
0.01
2

0 0

-2
-0.01
-4

-6 -0.02
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

0.4 2
Deviation of active power (p.u.)

dPeestimate
Deviation of active power (p.u.)

dPeestimate
1.5 dPetrue
dPetrue
0.2
1

0 0.5

0
-0.2
-0.5

-0.4 -1
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

Fig. 3. Identifier response to a 5% increase in Fig. 6. Identifier response to a three phase to ground
reference terminal voltage. fault.
-3
x 10
3 0.02
dwestimate
Deviation of rotor speed (p.u.)

dwestimate
Deviation of rotor speed (p.u.)

2 dwtrue dwtrue
0.01
1

0 0

-1
-0.01
-2

-3 -0.02
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

0.2 2
Deviation of active power (p.u.)

dPeestimate
Deviation of active power (p.u.)

dPeestimate
1.5 dPetrue
0.1 dPetrue
1
0
0.5
-0.1
0

-0.2 -0.5

-0.3 -1
69 69.5 70 70.5 71 71.5 72 69 69.5 70 70.5 71 71.5 72
Time (s) Time (s)

Fig. 4. Identifier response to a 10% increase in Fig. 7. Identifier response to a three phase to ground
reference torque. fault without scaling output data.

266
 7. APPENDIX
0.02
dwestimate

Deviation of rotor speed (p.u.)

0.01
dwtrue
Generator (Cylindrical rotor machine) Parameters In
0
p.u.
-0.01
ra = 1.1e - 3 x d = 1.70 x q = 1.64
-0.02
69 69.5 70 70.5 71 71.5 72
rD = 1.3e - 2 x D = 1.61
rQ = 5.4e - 2 xQ = 1.59
Time (s)

r f = 7.4e - 4 x f = 1.65
Deviation of active power (p.u.)

dPeestimate
1.5 dPetrue

0.5 x md = 1.55 x mq = 1.49 H = 2.37( s)

0
Transmission Line Parameters In p.u.
-0.5

-1
69 69.5 70 70.5 71 71.5 72
re = 0.01 xe = 0.2
Time (s)
AVR Parameters
Fig. 8. Identifier response to a three phase to ground K R = 1.0 TR = 0.001
fault with a fixed forgetting factor
TC = 1 TB = 10 K C = 0.08
Table 1 Summary of identification results under K A = 190 T A = 0.001
different cases
Vmax = 7.8 Vmin = -6.7
Case Vaf1 (%) Vaf2 (%)
Scaled VFF Electro Hydraulic Governor parameters
dw dPe dw dPe output K g = 25.0 Tg = 0.1
1 99.8 99.1 99.8 99.5 Ö Ö C max = 0.5 C min = -0.5
2 99.8 99.1 95.8 91.9 Ö Ö Pmax = 1.0 Pmin = 0.0
3 99.8 99.1 99.6 97.2 Ö Ö
Steam Turbine parameters
4 99.8 99.1 93.2 84.4 Ö Ö
TCH = 0.2 TRH = 4.0 TCO = 0.4
5 99.8 99.1 97.4 89.5 Ö Ö
6 99.3 98.2 85.4 51.3 ´ Ö FHP = 0.3 FIP = 0.4 FLP = 0.3
7 99.8 99.1 96.8 88.3 Ö ´
REFERENCES
*: Vaf1 is the value of VAF before disturbance Anderson, P. M. and Fouad, A.A. (1977). Power
Vaf2 is the value of VAF after 0.5s of disturbance System Control and Stability, (1st Ed.), The
VFF means Variable Forgetting factor IOWA State University Press, AMES, IOWA,
U.S.A.
It is obvious from Table 1 that a lower order linear Franklin, G. F., Powell, J. D. and Workman, M.
state space model can be used to accurately represent (1998). Digital control of dynamic systems, (3rd
the synchronous machine behaviors under steady Ed.), Addison-Wesley, Don Mills, Ontario.
state. The recursive subspace algorithm also Gustafsson, T, (1998). Instrumental variable
performed well under small and large disturbances. subspace tracking using projection
Using variable forgetting factor can improve the approximation. IEEE Trans. on Signal
identification results to some extent. Proper scaling Processing, Vol. 46(3), 669-681.
of the output data makes the results more attractive. Lovera, M., Gustafsson, T., and Verhaegen, M.
(2000). Recursive subspace identification of
linear and non-linear Wiener state-space.
6. CONCLUSIONS Automatica, 36(11), 1639-1650.
Park, D. J., Jun, B. E. and Kim, J. H. (1991) Fast
Use of recursive subspace identification for a state tracking RLS algorithm using novel variable
space model of a synchronous machine is presented forgetting factor with unity zone, Electronics
in this paper. The approach is based on IV ideas and Letters, Vol. 27(23), 2150-2151.
on the use of subspace tracking for the update of the Verhaegen, M., and Dewilde, P. (1992). Subspace
SVD. It has been tested with synchronous machine model identification part 1. The output error
operating in steady state and under various state space model identification class of
disturbances. Accuracy of identification can be algorithm. International Journal of Control, Vol.
increased by incorporating properly scaled I-O data 56(5), 1187-1210.
and by using a variable forgetting factor. The Yang, B. (1995). Projection approximation subspace
proposed approach provides a reasonable model for tracking. IEEE Trans. on Signal Processing, Vol.
designing a MIMO APSS. 43(1), 95-107.

267