1, Which Is Mainly Com
1, Which Is Mainly Com
1, Which Is Mainly Com
5, MAY 1992
663
[16] -, Lecture Notes of Workshop on Robust Multivariable Control, Minneapolis, MN, June 1987. [17] G . W. Stewart, "On the perturbation of pseudo-inverse, projections and linear least squares problems," SIAM Rev., vol. 19, no. 4, pp. 634-662, Oct. 1977. [le] M. Vidyasagar, Control Systems Synthesis: A Factorization Approach. Cambridge, MA: M.I.T. Press, 1985.
for estimating the parameters of an excitation system in the Minghu pumped storage hydro power plant of the Taipower Company.
uk
can be
I. INTRODUCTION
Adequate information of excitation systems is very important in the study of power system stability and the controller design for excitation systems. However, data supplied by the vender were often in the form of characteristic curves of various elements, or the block diagrams based on the tests on individual elements. The reliable data were not generally available on the integrated operational system [ l ] . Hence, for obtaining the actual dynamic performance of excitation systems, a suitable technique for estimating the parameters of block transfer functions in an excitation system on-line is indispensable. Basically, there are two categories of approaches for estimating the continuous-time model [2], [3], namely the indirect method and the direct method. The former has the advantage of relatively less computation compared to the direct method [2]. However, it is always encountered that the order of the model may be overestimated, which makes the controller design much more complicated. Moreover, the frequency response characteristics of the continuoustime model may be distorted due to model transformation [4]. To solve the above-mentioned problems, an indirect method emphasizing order estimation and frequency response matching is proposed. First, an ARMA model is identified from the sampled data [5]. Then, the order of the model is determined by the dispersion analysis [6] and/or the accumulated dispersion analysis. Finally, the continuous-time model is obtained by matching its frequency response characteristics to that of the identified ARMA model in the sense of least squares. This proposed method is applied
Manuscript received December 4, 1989; revised February 16, 1990 and February 15, 1991. C. M. Liaw is with the Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan, R.O.C. T. S. Liu, A. H. Liu, Y. T. Chen, and C. J. Lin are with Taipower Company, Taiwan, R.O.C. IEEE Log Number 9107232.
(3)
and uk denotes the output noise or disturbance. For the measured input/output series { u k } and { y k } , the parameters p i and qi are estimated using an identification program which is written based on the maximum likelihood method [5].
where
n
Gj =
i= 1
giXi.
Gj is the sequence of the impulse response. The variance function of the model output xk defined as yk - uk can be found as
01
where a,' denotes the variance of the white noise input, and
(7)
0018-9286/92$03.00
0 1992 IEEE
664
where T is the sampling interval. If the numerical difference between G(e-jUT) and H ( j w ) is E( jco), then we have
G(e-jWT)= H ( j w ) Dj which accounts for the contribution of dynamic mode Xj to y(0) is used as a measurement of the importance of each dynamic mode. According to the dispersion analysis, if the dynamic modes with large dispersions at A,, X,, . ., X are retained, the order of the . , reduced model can be considered to be m. And the original system of (4) can be rewritten as
E(
jw).
(17)
The steady-state value of the continuous-time model is set equal to that of the discrete-time model, i.e.,
H ( 0 ) = bo = G ( B = 1 )
From (15) through ( 1 8 ) one can obtain the following equations: where
n
In the next section, G J B ) will be used to estimate the error bound of the reduced model. If the dispersion analysis fails to determine the order of the reduced model for the lack of dominant dynamic modes, the accumulated dispersion analysis is proposed as an alternative. According to (6), the variance of the system output xk can also be expressed as
y(o)
u;(G;
+ G: + G; + . . . ) .
m
(11)
ADj = ( G i
(12)
a ,
i = odd i = even
ADj can be regarded as the energy contribution by Go,G I ,* . Gj. Using ADj and the relationships of (I) and (4),the unknown orders of the reduced ARMA model G*(B) can be set as [ G ( B ) = G * ( B ) oforders ( m , m ) , if j = even m =j/2, G( B ) = G*( B ) of orders ( m , m
-
odd
i = even
(1
AD^) < E
i
1) ,
odd
i = even
= ( j
+ 1 ) / 2 , if j
odd (13)
i = odd
i = even
where E is a small positive number, e.g., a typical value of 5% can be chosen in practical applications.
i = odd
i = even
V. PARAMETER ESTIMATION THE REDUCED-ORDER OF FROM THE IDENTIFIED CONTINUOUS-TIME MODEL DISCRETE MODEL
After determining the reduced-model order, the continuous-time model can be expressed as
B(s) H ( s ) = -A(s) 1
r s m
bo, Y r ( 0 ) = G , ( w ) . (20) For w = U ; , i = 1,2,. N , we can construct a system of 2 N equations, and parameters a,, i = 1,2, * . ., m and b;, i = 1; . ., r can be obtained by the least-squares fitting technique. The sum of the square of residual can be estimated as [6]
YR(w) =
GR(w)
a ,
1 E ( jw;) 1 <
N i= 1
I Gr(B=
i= 1
(21)
the parameters of (14) can be derived from the full-order discretetime ARMA model by the frequency response matching technique. The derivation procedure is basically deterministic, since the effect of the noise has already been considered in the modeling of the discrete-time ARMA model. The frequency response functions of the ARMA model and the continuous-time model can be expressed as
where Gr(B)defined in (10) is the discarded part of the identified full-order ARMA model. VI. PARAMETER ESTIMATION EXCITATION OF SYSTEM The proposed approach described in the previous sections is applied to the identification of an excitation system shown in Fig. 1. Firstly, we applied the white noise on-line at the input terminal denoted as 1 in Fig. 1 , and the signals at points 2, 3, 4 , and 5 are recorded, filtered, and sampled by a data acquisition system. These sampled data are then used to identify the transfer functions of the preamplifier block, the mixer block, and the damping block. Simi-
G ( B = e - j w T ) = G R ( w ) + jG,(w)
and
(15)
(16)
H(s=jw)
=HR(w)
+jH,(w)
665
POWER SYSTEM
EXCITATION SYSTEM
pre-amplifier
90R
VED
thyristor amplifier
synchronous generator
l-q-%m.yjv ,xp 1
~
0.2
1+0.5s
1.35X520
TABLE I THEESTIMATED CONTINUOUS-TIME FOR VARIOUS MODELS BLOCKS Estimated Continuous Model H*(s)
H*(s)
6.34
1
+ 0.01s
5.0
6.34
0.004s
3.56 * 0.2933 1 - 0.6763B 3.462 1 - 0.308B 0.2 - 0.2 B 1 - 0.9427B + 0.0025B2 0.2867 1 - 0.9853B
3.226
3.226
* *
O:O.O25a: 0.5~
0:o. loa:
0-0
0.2 x 0.5s 1 + 0.5s
5.0
0.004s
5.003
5.003
0.0
0.2
0.03s
0.0
0.206
0.0
0.206
0-0
20.0 20.0 1.35 x 520 X 5 10 1.35 *520
0.003s
19.55
19.55
0:O.Ol: 0.3
2.0
0.002s
2.043
2.043 1 + 0.0038s
2.043
140.4
0.002s
143.3
143.3
0:0.025s: 0.57
1.425 xo.01
0.002s
1.425
x 0.01
1.425
x0.01
0:0.025a : 0.5~
666
larly, the white noise is added at the input terminal denoted as 1* in Fig. 1, and the signals at 2*, 3*, and 4* are recorded and used to identify the transfer functions of blocks in the field voltage feedback loop. Determination of an adequate sampling interval is another difficult issue in performing system identification. A criterion is suggested to firstly select the sample interval
[3]
X , T
I 0.5
where X, is proposed to be the eigenvalue corresponding to the most dominant energy dynamic mode, rather than the magnitude of the largest eigenvalue suggested by Sinha and Puthenpura [2]. Then the sample interval T is trimmed until a minimum sum of squares of residual corresponding to the identified model is obtained. Each time, 1000 sampled data of the studied block are read and fitted by an identification program which is based on the maximum likelihood technique [5]. In which an ARMA model G( B ) of orders ( n , n - 1) is fitted, the order n is increased from an initial setting. The estimation process is iterated until a minimum sum of squares of residual is reached. Having found the full-order model, the reduced order of the continuous-time model is then obtained by the proposed techniques described above. Following the same procedures described above, all the transfer function models of various blocks in Fig. 1 are estimated from the recorded data. The average models of three sets corresponding to these models are listed in Table 1. The re?dlts indicate that except for the Preamplifier block ( l a - 2) and the field power amplifier block (1; - 2*), the estimated continuous-time transfer functions by the proposed method are rather close to those supplied by the vender. From Fig. 1, and Table I, one can observe that the dc gain of the block l a - 2 obtained by the proposed method is about one half of that supported by the vender. And the model of block 1; - 2* obtained by the proposed method is a first-order transfer function with time constant equal to 0.2002s, but the model proposed by the vender is a pure gain constant. However, the correctness of the models obtained by the proposed method can be verified from the circuit configurations supplied by the vender. Moreover, the recorded data are also inputted to the SD380 signal analyzer, the gain constant of block l a - 2 estimated from the frequency response is 3.056, which is rather close to that obtained from the proposed method. Other transfer function models estimated from the frequency responses of SD380 are also rather close to the results listed in Table I. VII. CONCLUSIONS An indirect stochastic parameter estimation method and its application for estimating excitation system parameters have been introduced. In this method, the ARMA model of the system is first identified from the sampled data. Then the order of the reduced model is determined by the dispersion analysis. Finally, the continuous model is obtained by matching the frequency responses of the discete-time model and the continuous-time model. The estimated transfer function models of the studied excitation system are rather close to that supported by the vender and obtained from the SD380 signal analyzer.
REFERENCES [ l ] M. J. Gibbard and Q. H. Kaan, Identification of excitation system parameters, IEEE Trans. Power Apparat. S W . , vol. PAS-94, pp. 1201-1207, July/Aug. 1975. [2] N. K. Sinha and S. Puthenpura, Choice of the sampling interval for the identification of continuous-time systems from samples of 0018-9286/92$03.00
input/output data, ZEE Proc. Pt. D , vol. 132, no. 6, pp. 263-267, 1985. S. Puthenpura and N. K. Sinha, Identification of continuous time systems using instrumental variables with application to an industrial robot, ZEEE Trans. Industrial Electron., vol. E-33, no. 3, pp. 224-229, 1986. Katz Paul, Digital Control Using Microprocessor. Englewood Cliffs, NJ: Prentice-Hall, 1981. ID. G. F. Franklin and .. Powell, Digital Control of Dynamic Systems. Reading, MA: Addison Wesley, 1980. M. Ouyang, C. M. Liaw, and C. T. Pan, Model reduction by power decomposition and frequency response matching, ZEEE Trans. Automat. Contr., vol. AC-32, no. 1, pp. 59-62, 1987. Z. S. Elrazaz, The choice of the reduced-model order, in Proc. Amer. Contr. Conf., Arlington, VA, 1982, pp. 1063-1067. J.-S. Lai and J. C. Hung, Practical model reduction methods, ZEEE Trans. Industrial Electron., vol. E - 3 4 , no. 1, pp. 70-77, 1987.
I. INTRODUCTION The approximation of complex models of physical systems has attracted the attention of many researchers in various engineering areas. In particular, there is a wide literature in the control engineering field about model reduction of linear high-order dynamical systems: for example, Pade [l] and Routh approximations [2] and their different extensions, the aggregation method [3], singular perturbation [4],moments matching techniques [5], balanced realization [6], and recently L-optimization techniques [7]. In this note, the model reduction problem is handled in the state-space domain. We propose a new procedure which can be applied to a general linear stable or unstable minimal system. Our approach is based upon a straightforward system partition using a numerically reliable algorithm, namely the Schur decomposition. Very recently, this algorithm has been used by Safonov [8], however, his method is completely different from the one proposed herein. Also, we present a general formula for the reduced model to recapture the original system static gain. ILPROPOSEDMETHOD Let us consider a linear system described under the state-space form
A is ( n
n), B(n
(1)
0 1992 IEEE