You are on page 1of 7

Copyright © IFAC Control of Power Plants and

Power Systems. Munich. Germany. 1992

OBSERVER BASED DETECTION AND RECOGNITION


OF EXTERNAL DISTURBANCES FOR
LARGE TURBOGENERATORS
H.-J. l\ern, H.H . Soli man and H.A. ]';our Eldin
University of Wupperta l, Gro up of AuJomatic Control and Technical Cybernetics,
Fuhlrol/str . 10, D-5600 Wuppertal 1, Germany

Abstract: Turbogenerators are potentially subjected to a variety of disturbances (network voltage


disturbances, network impedance disturbances or network load disturbances). Therefore attention has
to be paid to monitor their operating conditions. In this contribution, a methodology of observer based
real time recognition of external disturbances for large turbogenerators is proposed. The internal states,
especially the rotor currents (damper), the polar angle, polar voltage and the rotor- and stator fluxes are
estimated by a real time nonlinear state observer. Furthermore, it has been shown that the observer is
robust against external disturbances like denoted above. The robustness against external disturbances
gives the potential to detect and recognize the type of present disturbance out of the performance of the
internal states. The data interpretation is achieved through advanced signal analysis methods (FFT) and
classification methods in the Pseudo Euclidean Space are used. The observer and the detectiOn/recogni-
tion system is realized on a transputer network consisting of 8 transputers.

1. Introduction

Usually, condition monitoring is involved in taking measure- achieved through advanced signal analysis methods (FFT)
ments to indicate the "condition" of the plant. At higher and classification methods in the Pseudo Euclidean space are
levels, monitoring includes status monitoring, limit monito- used. The observer and the detectionlrecognition system is
ring, trend monitoring etc., and usually diagnostics. In gene- realized on a transputer network consisting of 8 transputers.
ral, only measurable states are used for condition monitoring
of large turbogenerators, such as stator currents, field voltage 2. The Synchronous Generator Model and Its State Re-
and current, stator winding temperature, magnetic core status construction Observer
or rotor vibrations, cooling fluid- and metal temperature.
However, these states are measured without considering that The modelling of the dynamic behaviour of the synchronous
they may be subjected to existing external disturbances . The generator is well established [5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
crucial influence of these external disturbances on the inter- 15, 16, 17, 18, 19,20]. The first synchronous generator
nal states have not been taken into consideration. Even when observer of Arumugam and Ramamoorty, 1972, [21] was a
the internal state variables are estimated, linearized or redu- linearized observer using the Parks model. The subsequent
ced order observers have been proposed. The resulting dis- models of Ueda et. al. 1975 [22], Quintana and Moharram,
advantage should be considered serious as the observer 1978 [23], Urn and Shing, 1981 [24] and Yamashita and
operation depends on the turbogenerator operating point. In Taniguchi, 1985 [25] use either linearized or non linear
this paper a methodology of observer based real time reco- models of reduced order. The first non linear observer, of
gnition of external disturbances for large turbogenerators is Osman and Nour EldillI986 [1,2,3] applies the complete
proposed. The internal states, especially the rotor currents non linear generator i order model and shows robustness
(damper), the polar angle, polar voltage and the rotor- and against network voltage disturbances, network impedance
stator fluxes are estimated by a real time nonlinear state disturbances as well as driving moment changes.
observer. The performance of this state observer is inde-
pendent on the present operating point. Therefore, it is not
necessary to update the observer parameters through para- 2.1. The Synchronous Generator Model
meter estimation or to use adaptive feedback. This feature
allows to start up the observer during any arbitrary operating The synchronous generator model, which has been used in
point. Furthermore, it can be shown that the observer is this investigation, is the complete model of7th order (Park's
robust against external disturbances like network voltage model) . This has been chosen in order to develop the non
disturbance, network impedance or load change. The ob- linear state reconstruction observer for the synchronous gen-
server also behaves neutral towards the used control scheme erator, avoiding modelling errors under disturbances, for
(voltage control, load-frequency control). The robustness which reduced models are not exactly representative.
against external disturbances gives the potential to detect and The complete model for the synchronous generator connec-
recognize the type of a present disturbance out of the internal ted to a rigid network has been documented in [I, 2, 3, 4, 5,
states. This non linear real time state observer and the graphi- 6,10] . It consists of an electrical and a mechanical part The
cal animation of the internal states are realized on a transpu- electrical part represents stator, rotor and field currents and
ter network. It has been tested and applied on an 100 MW is basically linear (assumption: no saturation effects). The
turbogenerator in a combined gas and steam power station. mechanical part describes the generator slip and the power
In this paper, the data interpretation of the animated perfor- angle and is linear, too. Both parts are non linearly intercon-
mance of turbogenerators under external disturbances is nected through the electrical moment and the polar angle,

469
resp. The complete model of the synchronous generator used state reconstruction observer for the synchronous generator
when connected to the rigid network (UN. j) through an that will be used in this paper is a nonlinear observer with a
external impedance Ze is thus: constant feedback matrix (Fig.l). It has been shown in [3.
6]. that the electrical part and the mechanical part can be
1. The electrical subsystem regarded independently of each other.

L !i=Ri+~f+~Il(O) UN (La) r---------------------,I


I generotor
I
where
-i~
I
I i,
i = [if id iD iq iQ)T (I.b) I • I
I I
1 T Imm'
':!{= [ - 0 0 0 0] (Le) I
Tf I I
I
~,,(o) = [0 sin(o) 0 cos(o) 0 f (I.d) I I I

The matrices Land R are given in the appendix.


r-r
I ,
-------------~ 1>..

I I
I •
I I
2. The mechanical subsystem
d 1
II I
- s = - (mm-met) (Le)
dt Tm
d
-O=<os (1.1)
dt
where

3. Internal Variables

Flux
~(t) =Li(t) (2.a)
FIG 1.: The Synchronous GeneraJor Model and its
~ = [qJf qJd qJD qJq qJQ f (2.b) Non Linear Slate Reconstruction Observer

Polar voltage For the electrical part, the suitable output (which guarantees
global and local observability) is the field current, and the
well-established pole-placement method can be used to ob-
tain the feedback coefficients. The generator slip serves as
output variable for the mechanical subsystem. The corre-
sponding feedback parameters have to be found out by
stability investigations and simulation runs. The equations
for the observer state representation run as follows [1. 2. 3.
4.32. 33.34]:
+ [r+re r+re [~d]
xt!+xe -(XI{+xe)] Iq
+ [Sin(O)] UN
cos(o)
(2.d)
d !./!. = L-1R ..!../!. + L-I [ "f nIl(A)]
dt 0
h ( . /!.~
~ + _11 If- If! (3.8)
d/\ 1 /\ /\ /\
4. External variables dt S = Tm [mm - metl + 1162 (s - s) (3.b)
d ~
generator voltage dt 0 =<0 S + hn ( s - /\s )
/\
(3.c)

Ug = ...J Ugd
2 2
+ Ugq where
f!. [f!. f!. f!. f!. f!. )T
generated active power !. = If Id ID Iq IQ (3.d)

P get) = Ugd{t)iJ..t) + Ugq(t)iq(t) (2.g) '.Y= [~f 0 0 0 of (3.e)

generated reactive power ~,,(~) = [0 sin(~) 0 cos(~) 0 f (3 .1)

Qg(t) = Ugq(t)iJ..t) - Ugd{t)iq(t) (2.h)


(3.g)
/\
met = f!.If f!.Iq + ( Xd - Xq ) /!. /!.
Idlq + ( f!.IdlQ
/!. f!. f!.
- Iq ID
)
Equations (l.a...... l.g) and (2.a ...... 2.h) build the complete
dynamic model of the synchronous generator connected to a ~=[ Uf UNf (3.h)
rigid network. All state variables. internal and external vari-
ables of the generator can be computed and simulated [3. 4. !!1J = [hlJ h21 h31 141 h51 f (3.i)
5.6.32.33.34).
!!22 = [1162 hn f (3.j)
2.2. The Non Linear State Observer for the Synchro-
nous Generator 2.3. Simulation and Experimental Application Results

Generally. state reconstruction observers use a copy of the In [3. 5] it has been shown that the observer reconstructs the
system model equations to generate the reconstructed state. generator states independently of its initial conditions or the
When feeding back the output estimation error in a suitable operating point of the generator. Above all. this non linear
way. the resulting state error asymptotically vanishes. The observer can be started from zero initial conditions without

470
wo ~t*lW'f"""" (p.uJ
,i).t*M"" .... ,. eo .... ) ;Q.I;lbs.... v .... (p.u) 12~

0-181~-
a.o5O

~ " "" ,[J


D-OZ. ' .•.. . ' ' ' .' " ' .. . . . . . . . . . , . . .. . . ' _. ' . . . .••...•••..... . •.... • . .. •••...•... - .. .
, ., . .. . .
-o.oo~ . ~:;
Lt3. . . . .. . . . . . . . . . . . . . . . . . ' , . . . . . . . .. . . . : .. .
.: .. . :.. . '. . :" "", ' ; · . ..... .

::-----==- -
LO'. . . . :. . . . . .... .. ... -. .: . .' . .: . . . .: . . .

ILOt ..• ~ .. ., , . ' . . ...• . • . . . ' .


. , . . . "

L.,,+--~--+--4--~-+_--I--~--4-~~--
000 010 IJoD 2.4(1 3.20 .10 .eg 5.6g 6~ ' ,20 1.00
• to,
~ ~ ~ ~ ~ ~ ~ - - = -
::~
.- ... .. ... ... •.. .. •. ............ ~ :
:::~~
·.· ·.·· ······. ····.· ·· ···.······... ·· .'II
. . . .....

I~I' ...: . . . . . . . . . . . .
3800 . , . . .- . • '. , .. ' , . .. ..... ' . ... . •
· . . . . .
)520 . . . :, . .. . . . :., - . ' . . . '. ' . . ': IJ" .. . :. : : : : . .: . .: . .: . ; .
· . . .. . · . . . . . . .
JL60 ... - . . '. ' " . . . ', . . ... . . " " ..' ' .9'11 . .. . .... . .. . ................ - - - - •...
. . . . . . · . . . . . . . .

..... •
e!51.oo 0-*..::----:',.=-0--,-1.601-:----:......::--,:1:.,,-,
0 - - •I.,::-- I::----:,lc60:---.:I:."':---:!,.=-0-~8JlO
. ... D.R10}~::----:1,~I-:----:.~~:---.:I:,~.-~,.~,.-.,..t.~,-~~::-.-.,..~~:---,:I:»:-~,~-~."
't tu
p ["""] p.0I0"~W'r'" Dol ... } "IF Ir.olri ........... [p. .... )
8500 ".37" T----:::--,---.--:....----,-----.--.--..---.--...,

~
alOO

1J1.C1J
. ..•..

.
·
• :" '
.. .
.
: '
. .. .
.
.•
.

.•. : '
. , . ..
'

.,
. ...
'

, ' •.
. . ..
,

.'
, . .'
.
. . : ~ .: • .
.. .. "
.
J~.
. - -

:
.
.
-

.
.,ZZ!I

...o~
. ..•...
. .
. . . :. . . :
' .'

:
.
.
.: ..: .
.
.
."
.
. : .. .. : . . . .: . . . .: . . .
. ,
.
. , _ -
.
. - • '
.
. .

7'.00 . . . . . . :. , ":" - , , " :. . , .: . . ~.~ . : . . : .: . . J.~ . . . :. . :, . . :. . . . : . . . . : . . . . : . . . .: . . . .: . . . .: . . •

:::~~ ~ ~ . . . . :::~
~ ~ ~ ~ - ~ ~ ~ ~ ~ ~.w ~ ~ ~ ~ - ~ ~ ~ ~ ~ ~
test

F [G. 2: Transformator Step F [G. 4: Sinudated Network VoI/age Change of -20%

any stability or starting problems and shows robustness electrical torque as well as the comparison between the
against external disturbances. The results of the experimental measured and the reconstructed active power. These curves
application of the observer system in a 260 MW combined are included here to illustrate that real time state reconstruc-
gas and steam power station are shown inFig.2 - FigJ. This tion for turbogenerators is now technological feasable. In
power station consists of two gas turbosets and a steam Fig.4 and Fig. 5, the robustness against external disturban-
turboset, one gas and the steam turboset were used for the ces is demonstrated for a simulated network voltage change
observer application at 45% - 95% nominal load. The com- of -20% resp. a simulated network impedance change of
parison of the measured and observed values shows accurate -20%. The simulation results are valid for a turbogenerator
conformity. In Fig.2 the results are shown for a transformer connected to an infinite voltage bus with voltage magnitude
step-down with the values of the reconstructed damper cur- and external impedance Xc. It can be observed that these
rents (p.u.) iQ, iD and the measured and reconstructed re- disturbances imply 50 Hz-Eigenfrequency of the generator's
active and activepower. electrical part. However, the reconstruction of the states is
In Fig. 3 the results are shown for a driving completed quite fast. For the network impedance change, it
moment change of 10% (valve step), the diagram demonstra- takes 40 ms to reconstruct the states (le has to be reconstruc-
tes the performance of the reconstructed polar angle and the ted, too). The network voltage is a measurement variable.

"Wt.oiJ_,. ..... [M .... ) ug ui-dH ... ...... [p .... 1


85.00 L211 ,..:.--,---,--=--,-----.--.---,---.---,---,---
.""~
.. .
- . . ........
71.00
.. . . ... . ... ... ..
. . . . '-
.... . . -. ...
. , ' , - , . ' - . ' .'. . . . . . . . . . . -" , -" - - ,
02.,~
... ~. . . .. ....................... ..........
U89 . , . '. .

. . ........... . . " ... : . . .
· ' . . .. . . . . . . . . . . .
73.00 " ' :. . . . :. . . . :., . . •• . , . :. . . . . . ' '. " '. " , " 1178 ..... . . . . . ..... . ... : .. .
· . . . , , . . " .
(,q .oo · -,. . .. . . , ... . . ... . .. . . . -'.. U67 .. .• . .. ' . . . . . ....... . .. , . . . . . . . .. . . . ... .
. ' . . ' . . . . . . . . .
6',00 U~ +---+--_+--+-__+--1__-+_-+---+---+---
000 tOO 200 31)(1 . 00 SOC 1'.00 100 a.oo 9 ,00 IODO 1,0:1 O,fl.4 O.DB '-le! Ol6 020 Q.t4 ~~ '.32 0.3' 0.•0
t(s]

.. ,a. ooc.......... [p.u.l


\.i.'5l

""'~
~
.. . .... .
1116 .. . ~. . . . ' . . .' . .. : '. .:' - , -: ' . ' : ' , , ~ - - ,

,....,
<'f.hO
. '
"
~. '"
",
'..
. .,
. .
. "
.:. . .
..
.
... ' .
." .... . . . .:.
. . . . . . . , ..
:. .
. . .:.
,. -" ,
..
.
.
,
-....
:
..
...
.. .
LIlt . . . . . . . . '. ., . . . . . . . . . . . . . . ,. . . . : .,.
. . .
"- .... . . • . . ...
.
...
,

,
.. .
.
. . ....
.
.
.
. .. .
.
I.~ ... :
: .. .
: ..'
.
'.'
, :. .
.
.
.
.
. ...
. . .
''''~~
lZZO " , . . ,'. , . . . ... ' .. . . ... ., . ' ." . . -. . . •...
. . , , . . . .
.,"'.+----+----+---+~--I--~I----+---~---+----+--~
~ ill ~ _ ~ _ ~ ~ ~ _ ~ 087. ~~-__I_-f__-+--+---+--_+-__+-__I--I__~
UIl 0-0<1 C.De 012 Ol6 020 024 O.ze 0.32 o.l' 0.40

,... If" i~ ........... [P .... J

::~~
. .. .. ..... ......... .. . .... . . .. ...... . . .. .
16~ .. .. . .. ", . . . - . . - . . . . .. ....
-
.
J.'
3.&40
'''~
. .. .- ....• . . . ... . ....... ...•....•..
. ....
·
. ..
. ...•..
.
, . ..
.
.
.
.
.
.
,
... . •.....•
.
.

7t.60 .... " . - , . , .. .. . . . , . . . .. . . . .. ..... .


.. .
3 .7" . . . : . . . . . .' ... ... .' " . . .. .
. . . · . . . . , . . .
68.9:1 .. • . . . " ,.' . " " • . . . ' . . . ' . .. .. . . . •... J .7Zl ... - . '. . . • '. . . . ' .... - - " ' , .. ' . . . .. . . . •
· . . . . . . .
.'''' .+----+----+---+----I--~I----+---~---+----+--~ lA'1 +-_ _+_-_+--+-__+--I__-+_-+---I---+--~
- - - ~ ~ ~ - - ~ = ~
C'(c I 00 ~1Xl 300 400 5,00 6.00 1,00 &.DO '.00 10.00
t(s]

FIG. 3: Driying Momenl Change 10% FIG. 5: Sinudated Network Imptdance Change of -20%

471
3. Generation of the Disturbance Feature Vector and This training set is used for the following classification in the
the Training Set Pseudo Euclidean spaces [28].

The scope of the investigation in this paper is the detection 4. Geometrical Classification In the Pseudo-EucUdean
and recognition of four main classes of disturbance types. 1. Spaces
positive and 2. negative network voltage change. 3. positive
and 4. negative network impedance change based on datasets Geometrical classification methods have been widely used
of the simulated synchronous generator and its non linear in many disciplines [26]. They are related to the Principal
observer. For disturbance detection alone the reconstructed Coordinate Analysis. In contrary to cluster analysis. geome-
damper currents iQ and iD are used. because they are only trical classification does not impose a structure on the feature
different from zero in transient state. With this trigger point set tlSed [26.27.28] . Further. each pattern is represented in a
the f~llowing 6 reconstructed s!atesl.\aret.Pse<VOl;,geRerating low-dimertsion space.
the dISturbance feature vector: ID . Iq. IQ. s. O. PO . The
values are subjected to a FFT-transformation with a sample The principal analysis methods. which are sometimes refer-
of 2048 points. In Fig. 6 and Fig. 7 the FFTresults are shown red to as either Q- or R-mode [26]. have been successfully
for network voltage and impedance disturbances. It used in numerous technical applications. especially in geo-
is obvious. that two frequencies are dominant, the eigenfre- physics. Goldfarb [27.29] has widened such applications by
quency of the mechanical part of the synchronous generator introducing the classification in Pseudo-Euclidean spaces.
fm '" 1 Hz and the eigenfrequency of the electrical part The main advantage of the Pseudo-Euclidean space is that it
fe '" 50 Hz. Therfore these two frequency components are is domain independent. It helps to develop an analytic basis
used for generating the disturbance feature vector. The for taking decisions based on the set of pattern vectors only.
feature is computed out of the ratio of the 3 "slow" frequency and independent of the form of pattern representation. Anot-
lines (2.3.4) and 3 "fast" frequency lines (102.103.104): her important advantage is that the classofPseudo-Euclidean
(P.E.) spaces is sufficient to accommodate data of any gene-
12+[3+/4 rality [29]. This is in contrast to the classical geometrical
x = -::-,""-....,;.c.--",-;:-- (3.1) classification from which apriori conditions should be satis-
/HY2 + /103 + /104 fied [30].
Having p features for classification. the data for n sample
With this feature and the 6 chosen reconstructed states the vectors can be represented in the (n,p) feature matrix X of
disturbance feature vector dimertsion is of 6 th order. In order such samples as
to generate the training set for the used classification method T
X=[~1~2 ... ~,I: • •• ~I\] (4.1)
in the Pseudo Euclidean Space for each of the 4 classes 4
different disturbance levels are used (0.5%. 1.5%. 3%. 5%). where ~,I: : feature vector of the k-th sample. Dim. 2U = p.
This training set is shown in table 3.1. The n sample vectors can be represented as n points in the
p-dimensional feature space. In this space. the p-dimensional
feature difference vector ~ij.
Table 3-1 Feature vectors for the training samples
~ij = ( ~ - ~) (4.2)
iQ s /) po
i iD io can be represented as lines connecting points in the feature
Un+ (Oass 1) space. The feature distance dij of these feature difference
1 0.8233 2.2952 2.7801 52.7992 58.1195 2.6436 vectors ~ij. are dermed as
2 0.8270 2.2461 2.7299 52.1554 56.0154 2.6904
3 0.8294 2.1790 2.6630 50.9781 53.6693 2.7509 dij= ~'!j . ~ij
4

5
0.8315

0.8220
2.0899

2.3393
2.5753
Un-
2.8223
49.2267
(Oass 2)
53.6889
50.8158

59.3284
2.8284

2.6064
= ±
.1:=1
(Xi.\:-Xj.l:)2 .iJ = 1.2... . •n (4.3)
6 0.8199 2.3886 2.8740 54.4736 61.2674 2.5616
7 0.8156 2.4578 2.9433 55.6944 63.9412 2.4999 Following Goldfarb [29] to achieve a representation in the
8 0.8088 2.5477 3.0345 57.2721 67.2248 2.4261
n-dimensional sample space. the scalar product dij in the
Ze+ (Oass 3)
feature space is used to build the (n.n) dimensional feature
9 0.6737 5.2809 5.4891 242.5377 88.0218 2.0657
1 )
10
11
12
0.6813
0.6938
0.7095
5.2791
5.2865
5.2892
5.4835
5.4867
5.4858
Ze-
243.4492
244.4496
244.9615
(Oass 4)
90.2067
92.8662
96.2497
2.0479
2.0290
2.0043
distance matrix D [30].
D == 1dij I= 1~'!j . ~ij I = ±
.1:=1
(Xi.\: - Xj,l:)2 (4.4)

13 0.6654 5.2589 5.4669 241.7234 87.3115 2.0708 The matrix D is symmetric. having its centre at
14 0.6568 5.2623 5.4741 240.2950 85.3302 2.0858
11\1\
15 0.6437 5.2386 5.4555 237. 1675 82.7629 2.1055
16 0.6269 5.2159 5.4397 232.7228 79.6818 2.1301 do == 2' L L dij (4.5)
n i= 1 j=1

3 .9

3 .J.2

!! 23. SI 1.8

1.50 1_Z

0.78
00

0 o L-'""!I=
!lill!;;,=".
,,,"". ""..""
.... ""..""
. .......
. ' '''.''''.'U11':W;":J!:."''''''',,",'""".,,,,,,,
,,,,,,,,
"I,,,,
I!lI:J!:,
! !!k""
!!II,,,,
:!i!:!!!J
:Il';o:
;: ' " ilIC
' a:: ' ' ""- -.J
.... .......
0 25 12.5 25 37 .5 50 625
frequen cy [Hz) frec;uet"lcy (H=)

FIG.6: FFT·Resul/s , V.·Disturbance FIG. 7: FFT-Resul/s, le-Disturbance

472
In geometrical classification [30]. it is recommended to use Table 5-1. The first three eigenvectors of the matrix A.
a classification matrix whose centroid coincides with the i Yil Y..u Y..n ~
origin for the principal coordinates. The matrix A defined as
1 +0.2513 -0.0234 -0.0406 +1.4449E+5
A= l aij! = l -~(dij+ (do- di. - d.j» ! (4.6) 2
3
+0.2539
+0.2579
-0.1334
-0.2522
-0.1274
-0.2465
+323.63952
+0.18758
4 +0.2636 -0.3939 -0.4014 +0.00046
where di. and d.j are the row and column mean respectively. 5 +0.2485 +0.0355 +0.0399 +0.00006
satisfies tl,!ese requireme',!ts 11 11
6 +0.2457 +0.1352 +0.1270 +0. ()()()() 1
7 +0.2414 +0.2716 +0.2407 +0.0()()()()
L aij = 0 = L aij ; L L aij = 0 (4.7)
8
9
+0.2360
-0.2541
+0.4385
-0.0118
+0.3918
-0.0243
+0.0()()()()
+0.0()()()()
j=1 i=1 i=lj = 1 10 -0.2574 +0.1004 -0.1398 +0.0()()()()
11 -0.2610 +0.2378 -0.2507 +0.0()()()()
This centred feature distance matrix A remains real symme- 12 -0.2637 +0.4192 -0.3766 -0.0()()()()
tric and can be represented by its real eigenvalues- and 13 -0.2517 -0.0438 -0.0313 -0.0()()()()
eigenvectors matrices Z and Y through [26.30.31] 14 -0.2472 -0.1403 +0.1059 -0.0()()()()
15 -0.2380 -0.2544 +0.2543 -0.0()()()()
A = y . Z . yT (4.8) 16 -0.2252 -0.3853 +0.4790 -0.0()()()()

where Y = [VI. y2. .. .. VII ] is the normalized orthogonal


matrix of eigenvectors Or A. and Z = diag.14! is the Table 5-2. The first column of the coordinates matrix Y.
diagonal matrix of eigenvalues of A with Un+ Un- Ze+ Ze-
+95.5356 +94.4685 -96.5878 -95.6730
for k= 1.2. .... n (4.9) +96.4984 +93.3921 -97.8258- -93.9560
Yil +98.0261- +91.7707- -99.2253 -90.4693-
+100.1999 +89.7019 -100.2539 -85.6022
Let the positive eigenvalues are ordered according to their For our purpose. we aim to select a representative for each
values of the four classes out of the 16 samples of the training set.
(4.10) The centroid (mean) coordinate of each class will be deter-
and the negative eigenvalues ordered according to their mined first. Then, the nearest element in each class to the
magnitude corresponding class centroid is selected to represent this
IAo+11 ~ IAo+21 ~ ... ~ I~I (4.11) class. For the given training set, the third sample is the nearest
then. the diagonal matrix Z should be. to the centroid of class 1. while samples 7.10 and 15 are the
nearest to the corresponding classes 2. 3. and 4 respectively.
=
Z diag . ( AI. 1..2..... Ao; 1Ao+ 11 . ... 1Ap I; O.... ) (4.12) Accordingly, they are selected as representatives for these
This arrangement of the eigenvalues in Eq.(4.15) renders the classes.
resulting vector representation a distance-preserving map-
ping in the Pseudo-Euclidean space [29.30.31]. This em-
bedding algorithm constructs a domain-independent vector 6. Direct On-line Classification in the P.E. Space for Ze
and Un Disturbances
representation in P.E. space. It includes the presentation in
Euclidean space as a special case.
With the methods introduced in the previous sections. it is
The new (n.n) coordinate matrix Y can be evaluated from the possible to represent disturbance classes by class repre-
matrices Y and Z through sentatives. The class representatives are represented by their
Y =
Y Z 112 [~·~I • ..fi:2·~2• .... ~,~,,] (4.13)
= coordinates in the low dimension space. Therefore. for on-
Depending on the magnitude of the eigenvalues. a reduced line applications. a complete on-line classification algorithm.
representation in the P.E. space. can be selected by applying applied in the lower dimensiona1ity space. will be introdu-

trace Z = L" 11.01 (4.14)


cedhere.
In order to avoid solving an eigenvalue problem during the
m i=1 on-line classification, a special algorithm is used, so that new
K =~Z
trace
. L 1Ai 1 ~ M ;M ~ 1 ; and m < n. (4.15) feature vectors can directly represented on-line in the P.E.
space. This algorithm will not be discussed here further. For
i=1
This means that only the first m eigenvalues are considered more details. refer to [31].
whose sum contributes significantly to the (trace Z). Thus The knowledge-base required for on-line classification of
the n samples can be represented in the P.E. space by the first new patterns. will contain some information extracted during
m columns of Y alone instead of all n columns. the off-line training phase. The organization and contents of
such knowledge-base depend on the classifier used. For
5. Selection of Class representatives for the disturbance classification. the knowledge-based classi-
Training Set fier used in our approach will need the following data to be
extracted during the training phase:
The feature extraction and geometrical classification techni- i) Number of classes k and the size of feature vector p
ques described in the previous sections will be applied here
ii) Feature vectors of all class representatives
on the training set of 16 feature vectors which are selected to
represent the four classes of disturbances. Table 3.1 shows iii) The mean coordinate YII; of each class i • expressed as
the elements of the (16.6) feature matrix X. In the training YmaXi + Ymin;
set, the four classes of disturbances are represented equally YII; = (6. 1)
2
by four samples each.
where Ymax and)'min are the maximum and minimum coor-
The ordered eigenvalues as well as the corresponding first
dinates within the i-th class in the P.E space.
three eigenvectors of the centred distance matrix A are
listed in Table 5-1. iv) The range of coordinates R; within each class i
Choosing K = 0.98 for representation of samples in a lower Ri =I
Ymax; - Ymini I (6.2)
dimension space means that only the contribution by the first v) The Direct Projecting Vectors (DPV) for the training set
eigenvalue would be sufficient and the corresponding coor- [31]. which are used to direct representation of new
dinates of the training set in the reduced one-dimensional samples in the P.E space.
space are given by the first column of the corresponding Such a knowledge-base as extracted for the given training
matrix Y as shown in Table 5-2. set is listed in Table 6-1.

473
6.1. Direct On-line Classification Algorithm 6.3 Results of Classification or New Samples

For on-line classification of new samples. the following steps The classification algorithm. discussed above. was applied
are carried out: for classifying 12 new disturbance samples. The samples
a) The feature extraction have to be performed as in Sec- have been characterized by their feature vectors (Table 6-2.)
tion (3). to deliver the feature vector 31. extracted according to Section (3). In Table 6-3 •• the direct
projected coordinates of all samples together with the clas-
b) The extracted feature vector will then be projected di- sification results are shown. It is clear. from Table 6-3. that
rectly to the P.E. space [31] using the DPY's in the all the samples are successfully classified.
knowledge-base as.
bI) Compute the squared distance d1i between the feature Table 6-2: Feature vectors of the test samples.
vector 31 and each of class representatives Xk in the fea- 0
i io iq iQ s pO
ture space
p 2 1 0.8212 2.3630 2.8479 54.1548 60.4413 2.5824
dli= L(XIj - Xij) .i= 1• ... •k (6.3) 2 0.8185 2.4104 2.8952 54.9203 62.0795 2.5410
3 0.8122 2.5033 2.9895 56.4706 65.6245 2.4615
j=1 4 0.8261 2.2708 2.7547 52.6349 56.8556 2.6690
b2) using the DPY' s in the knowledge base. the required 5 0.8279 2.2242 2.7080 51.7808 55.1942 2.7107
6 0.8306 2.1340 2.6186 50.1355 52.1936 2.7912
coordinate YI of the unknown pattern in the reduced 7 0.6602 5.2651 5.4768 240.8928 85.7717 2.0824
P.E. space is calculated through 8 0.6525 5.2550 5.4683 239.3525 84.3968 2.0929
YI = Cl dll + C2 da + ... + Cl; dlk (6.4) 9 0.6353 5.2304 5.4505 235.1928 81.2471 2.1178
where Ci is the i-th component of the corresponding DPY. 10 0.6772 5.2807 5.4860 243.3331 89.1401 2.0570
11 0.6849 5.2795 5.4828 243.5637 90.9603 2.0434
12 0.7014 5.2863 5.4842 244.6875 94.6166 2.0163

Table 6-1. Knowledge-Base extracted from the training set. Table 6-3. Classification results of the test samples.
No. of classes k 4 y, al a2 a3 ~ To True
Class Class
Size of feature vector p 6 +93.836 1.729 0.735 104.884 36.635 2 Un-
+92.826 2.162 0.311 104.333 36.434 2 Un-
1 0.8294 2.1790 ... 2.7509 +90.743 3.055 0.563 103.197 36.021 2 Un-
+95.895 0.846 1.599 106.007 37.044 1 Un+
Featu re vector of 2 0.8156 2.4578 ... 2.4999 +96.997 0.373 2.061 06.608 37.263 1 Un+
rep. of class i : +99.089 0.524 2.939 107.749 37.678 1 Un+
3 0.6813 5.2791 ... 2.0479 -94.615 82.534 78.337 2.077 0.790 4 Ze-
-92.880 81.791 77.609 3.023 0.445 4 Ze-
4 0.6437 5.2386 ... 2.1055 -88.284 79.820 75.680 5.530 0.468 4 Ze-
-97.546 83.791 79.567 0.477 1.372 3 Ze+
Ranges of co-ordinates -98.055 84.010 79.780 0.200 1.473 3 Ze+
Ri 4.6643 4.7666 3.6661 10.0710
-99.731 84.728 80.483 0.714 1.806 3 Ze+
Mean co-ordinates YI4 Classification results 100'10 1= Un+ 2= Un-
+97.868 +92.085 -98.421 -90.638 3=Ze+ 4=Ze-
Direct Projecting Vector -0.00146 -0.00116 +0.00173 +0.00090

6.2 Knowledge-Based Classification of Samples in the


P.E.spaces 7. Acknowledgement

To assign a new disturbance pattern to one of the known The authors would like to acknowledge the technical and
classes. the following procedure is used fmancial support from the "Deutsche Forschungsgemein-
a
1) The normalized distances 1i (i=I •... ,k). between the coor-
schaft" DFG and also the cooperative support of KWU
Erlangen concerning the execution of the measuring tests in
dinate of the new pattern YI and the mean coordinate
the gas and steam power station (MUnchen Sild).
YIJ; of each class will be determined by
a ==
1i 2· I YI- YIJ; I
Ri
(6.5) 8. Appendix, Abbreviations

where Ri is the range of the i-th class.


1
-xd(I-<14') -(l-jJ.f) 0 0

a
2) If one of the normalized distances 1i has value less than
or equal ·One·. The tested pattern will be assigned to
1
-(xct+-xe)
L= -(l-llD) xd(1-<1dD)
-1
1
0
0
0
0
0 0 0 -(xq+~) -1
the class corresponding to that distance.
0 0 0 Xq(I-<1q) 1
3) When all of the distances are greater than ·One·. the te-
sted pattern lies between the classes. It will be defmed -1
as 'between classes'. then 0 0 0 0
Tf
3a) The position of the pattern relative to the classes in the
P.E. space will be examined. 0 ro(r+re) 0 -<O(xq+~) -(0

3b) Once the relative position of the tested pattern is found. -1


it is possible to affiliate the pattern as mixed type pat- R= 0 0 0 0
tern which belongs to two classes i and j . The ratio to TD
the i-th class can computed by
alj (6.6)
-(0 ro(xct+- Xe) (J) ro(r+re) 0

~ 0 0 0 0
-1
and the ratio of the j-th cl~s
TQ
(6.7)
ali+a1j

474
it = field current cal Machine", lEE (British), Monograph. Cambridge
id,q = stator current d-q-axis Univ. Press, London 1970
iD.Q = rotor, damper current d-q-axis (idd.qq) [16] Anderson P.M .
Tf,D,Q.m = time constants "Analysis of faulted Power Systems", Iowa Stae Un-
o = polar angle iv. Press. 1973
m.1 = electrical torque [17] Nelles D.
mm = driving torque "Die Beschreibung der Synchronmaschine fUr Aus-
hll = feedback vektor electrical part gleichsvorgange in Drehstrornnetzen", Wiss. Ber.
1162 ,hn = feedback coefficients mechanical part AEG-Telefunken,46,1973
p v = thermal losses
[18] Anderson P.M., Fouad A.A.
xd,q •• = reactances
"Power System Control and Stability", Vol. 1, 1977
r,r. = resistances
Z"U = network impedances [19] Leonhard W .
IlDf.Od,q = leakage factors "Regelung in der elektrischen Energieversorgung",
Teubner Verlag, Stuttgart 1980
s = slip
[20] Stepina I.
"Raurnzeiger als Grundlage der Theorie der elektri-
9. References schen Maschinen", ETZ-Arch. Bd. 88,1967
[21] Arurnugam M., Ramamoorty M.
[1] Nour Eldin H.A., Osman N. "A Dynamic Observer for a Synchron Machine", Int.
"Robust Nonlinear State Observer for the Synchro- I. Control, 1972, vol. 15, No 6,1129-1136
nous Generator", Congress "Real Time Control of In-
dustrial Processes", Warsaw 1986 [22] Ueda R., Takata H. et. al.
I.E.E.E. Trans. Power Appar. Syst., 94, 2135,1975
[2] Nour Eldin H.A., Osman N.
"Robust Nonlinear Observer for the Synchronous [23] Quintana V.H., Moharrem O.E.
Generator", Foundations of Control Engeneering ed. "Minimal-Order Observers for Synchronous Machine
by T.Kaczorek, Vo112, NO. I, 11-29, 1987 Control", IEEE PES Winter meeting, New York, Ja-
nuary 29-February 3, 1978
[3] OsmanN.
"Nichtlinearer Zustandsbeobachter rur die Spannungs- [24] Choi San Shing, Lim Choo Min
regelung des Synchrongenerators", Dissertation, Un- "On the Design of Luenberger-Type Observers for
iv . ofWuppertal, 1990 Single-Machine Infmite-Bus", Electric Power Sy-
stems Research, 4 271-276,1981
[4] Nern H.-I., Nour Eldin H.A.
"Survey in Methodology and Real Time Realization [25] Yamashita K., Taniguchi T.
of State Observers of ElectriSal Generators", IFAC, "On the Estimation of the Transient State of a syn-
IMACS-SymposiumMIM S ,Bruxelles, 1990 chronous machine by an optimal observer", Int. I.
Control, vol41, No 2, 417-428, 1985
[5] Nour Eldin H.A., Lerch E. et. al.
"Digitale Simulation der Synchronmaschine mit Zu- [26] Gordon, A.D."Classification", Chapman and Hall
1981. '
standsraumdarstellung", ETZ-Archiv, Bd. 2,1980
[6] Lopez L. [27] Goldfarb, L. "A unified approach to pattern recogni-
"Modellierung und systematische Modellreduktion tion", Pattern recognition, Vol. 17, No. 5,1984.
von nichtlinearen Netzmodellen mittels Singular Per- [28] Nour Eldin H.A., Alam Eldin A.T.
turbation", Diss. BUGH Wuppertal, 1987 "Robot vision: Geometric classification for automated
[7] Crary S.B. inspection", in Intelligent Robotic Systems, Tzafestas
"Power System Stability", vol2, Wiley, New York S.G. (Ed.), Marcel Dekker, 509-526,1991.
1947 ' [29] Goldfarb, L., "A new approach to pattern recogni-
[8] Westinhouse Electr. Corp. tion", Progress in pattern recognition IT, ed. L.N. Ka-
Electrical Transmission and Distribution Reference nal and A. Rosenfeld, Elsevier Science Publishers
Book, Pittsburgh, Pa., 1950 B.V,1985.
[9] Concordia, C. [30] Alam Eldin, A.T. and Nour Eldin, H.A.," Automated
"Synchronous Machines", Wiley New York, 1951 inspection of web type products in Pseudo-Euclidean
spaces", 10. DAGM-symposium 1988, ZUrich.
[10] Laible, Th.
"Theorie der Synchronmaschine im nichtstationiiren [31] Soliman, H.H., "Systematic Approach for Automated
Betrieb", Springer Verlag, Berlin 1952 Visual Inspection", Internal report, atk Group, Univ .
ofWuppertal, Germany, 1991.
[11] Kimbark E.W.
"Power System Stability", vol1, Wiley, New York [321 Nern, H.-I., Fischer F., Nour Eldin H.A.
"Distributed Realtime State Reconstruction and Vi-
1956
sualization for Control and Monitoring of Turbo gene-
[12] Lynn T.W. rators", IFAC International Symposium on
"Tensor Analysis of a Synchronous Two-Machine Sy- Distributed Intelligence Systems, DIS '91, 12-15 Au-
stem", lEE (British), Monograph. Cambridge Univ. gust, Arlington, Virginia, USA
Press, London, 1961
[33] Nern H.-I., Nour Eldin H.A.
[13] Taylor G.D. "Observer Based Control Scheme of Large Turboge-
"Analysis of Synchronous Machines Connected to nerators on a Distributed Processor System", IMACS-
Power Network", lEE (British), Monograph. Cambrid- IF AC International Symposium on Parallel and
ge Univ . Press, London, 1962 Distributed Computing in Engineering Systems, Iune
[14] Adkins B. 23-28,91, Corfu, Greece
"The General Theorie of Electrical Machines", Chap- [34] Nern H.-I., Nour Eldin H.A.
man and Hall, London, 1964 "Animation of Synchronous Generator Dynamics and
[15] Harris M.R. its State Reconstruction Observers on a Transputer
"Per Unit Systems: with special reference to Electri- Network", IFAC - IMACS Symposium McrS, May
7-10,1991, Lille, France

475

You might also like