1552 IEEE Transactions on Power Apparatus and Systems, Vol.PAS-98, No.
5 Sept/Oct 1979
AUTOMATIC CONTINGENCY SELECTION FOR ON-LINE
SECURITY ANALYSIS - REAL-TIME TESTS
G IRISARRI, MEMBER, IEEE
A. M. SASSON, SENIOR MEMBER, IEEE
AMERICAN ELECTRIC POWER SERVICE CORPORATION
2 BROADWAY
NEW YORK, N.Y. 10004
Abstract - Results of various real-time tests performed on the AEP-
EHV System with the automatic contingency selection algorithm proposed in [1]
are reported.
The first-order performance index sensitivities used in the algorithm
for line contingency ranking are rederived in a straight forward manner and, more-
over, the ranking algorithm is extended to accommodate second-order performance
index sensitivities. Many tests, however, indicate that - for the AEP-EHV Net-
work operating conditions - the automatic contingency ranking algorithms do not
perform reliably.
As an alternative, it is shown that a DC-Load Flow, although com-
putationally more demanding than the first-order automatic contingency selection
algorithm (but less computationally complex that the second-order extension), is
a reliable and efficient way of ranking line contingencies.
INTRODUCTION
The on-line steady-state contingency analysis packages presently
being used in energy control centers simulate transmission line contingencies
by either
- selecting the outages to be studied from a pre-prepared list of meaning-
ful contingencies, or,
- by operator selection depending on current system conditions as given
by a state estimator or any other power system monitoring facility.
The former approach is an off-line planning practice applied to the
on-line situation. This method fails to recognize that system conditions assumed
for the preparation of the contingency lists may be far from real time conditions.
In fact, what is understood to be a single contingency in a planning study may
very well be a double or even a triple contingency in a real time situation. The
second method - i.e. operator selection - may prove to be burdensome even to
experienced operators. As operating conditions in most systems get more strin-
gent, the number of contingency choices gets larger and noamountofexperience
guarantees proper selection of troublesome contingent situations.
As a consequence, it is necessary to develop methods for automatic
selection of meaningful contingency cases. A first approach in this direction is
the method proposed by Ejebe-Wollenberg [1] (see also reference [2]). In their
paper the authors propose a methodology for ranking transmission line contin-
gencies by evaluating the normalized sensitivities of a system wide performance
index with respect to line outages. The authors show that through the use of
Tellegen's theorem the first-order sensitivities of the performance index with re-
spect to changes in line admittance can be readily determined and efficiently
implemented. The normalized sensitivities (or first-order changes in the perform-
ance index), according to Ejebe-Wollenberg, can then be used to rank contin-
gencies from the most important (largest positive performance index change) to
the least important (largest negative performance index change).
A paper recommended and approved by the
F79 212-2
IEEE Power System Engineering Committee of the IrEE
Power Engineering Society for presentation at the IEEE
PES Winter Meeting, New York, NY, February 4-9, 1979.
Manuscript submitted September 1, 1978; made available
for printing November 6, 1978.
0018-9510/79/0900-1552$00.75 Q 1979 IEEE
Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
D. LEVNER, MEMBER, IEEE
Tests performed on the AEP-EHV system reveal, however, that the
automatic contingency ranking method of Ejebe-Wollenberg is unreliable. In fact,
for the case of real power flow performance index the method may produce ex-
tensive misclassification of transmission line contingencies resulting not only
in false alarms (i.e. unimportant outages classified high in the list) but, more
critically yet, meaningful contingencies may be classified as unimportant (low
in the list). The main reason for this unpredictable behaviour of the automatic
contingency selection method is that the performance index suggested in [1] is
not a monotonic function of the susceptance (DC-Load Flow formulation) of the
lines. This monotonicity condition seems to be at least a necessary condition
for the trustworthy performance of the method.
To improve the performance of the method, the calculation of the
normalized sensitivities of Ejebe-Wollenberg was extended to include second-
order effects. Even in this case, however, the automatic contingency selection
algorithm was not found to be sufficiently reliable.
However, the use of a DC-Load Flow calculation was determined to be
highly reliable and less computationally demanding than the second-order exten-
sion considered in this paper. It is concluded, then, that use of a DC-Load Flow
model may be a desirable alternative to the automatic method of Ejebe-Wollenberg
for contingency selection.
In the following sections of the paper the first-order and second-order
sensitivities of the real power flow performance index will be developed. The
approach presented herein for the development of these formulas is different from
that suggested by Ejebe-Wollenberg. Rather than using Tellegen's theorem, it is
shown that the same results can be obtained in a straight forward manner by the
use of elementary differential calculus. Also, results of the tests performed with
the first-order, second-order and DC-Load Flow methods are compared to the
"correct" ranking as given by full AC-Load Flow outage simulations.
CONTINGENCY SELECTION ALGORITHM
According to Ejebe-Wollenberg [1] (see also [4-5]), an intuitively
appealing index for quantifying the extent of line overloads may be defined in
terms of a real power performance index
NL
J Z ~~~~Wx
(W 2n
(1)
Jmw = L. 2n ~rii
..(= 1
where
Pt megawatt flow in line we
p mX megawatt capacity of line g
W, real non-negative weighting coefficient (assumed equal to unity in the
sequel)
NL number of lines in the system
n specified exponent (n = 1 used in the sequel)
The DC-f low in line a (between nodes m and n) is
P - 8B, 9 (2)
where Qf is the angle across the line
is the susceptance of line , eand
(6- m-8n).
 1553
Moreover, if it is assumed that the megawatt capacity of line,f(Pmx) applications, corresponds to the current angles at the nodes of the network given
is a linear function of the susceptance of the line,lhat is, by, for example, a state estimation monitoring facility. The first-order sensitivity
method only requires one forward and backward solution (once B is triangulated)
pmtaX = KX I BI (3) as evidenced by eq. (7).

then, substituting eqs. (2) and (3) in (1), results in the performance index Second-Order Sensitivities
The second-order sensitivity of the MW-flow performance index of eq.
NL (4) can be shown to be (see appendix 1)
,. 2 jMW 2 (M T B
jA
JMW )(4) -mMt) Ox Ox + -m;T . § B -I Yt 01 (9)
-
=

= L T t -a B2't -,t
where
The assumption of eq. (3), although not mentioned explicitely, was
also made by Ejebe-Wollenberg in their paper [1]. In this paper results for both Q "adjoint" network susceptance matrix
performance indeces (eqs. (1) and (4)) will be given.
The adjoint susceptance matrix of eq. (9) is given by
The performance indeces of eqs. (1) and (4) are normalized functions
of the line flows (eq. (1)) and angles across the lines (eq. (4)) respectively. In NL
both cases, the performance indeces have a small value when all the line flows
-

32JMw
are less than the capability of the respective lines, and a high value whenever =j 1 lojl2 -j
(10)
there are line overloads.
The purpose of the automatic contingency selection method is to
efficiently identify the contingencies (line outages) that warrant further study For both cases - i.e. first-order and second-order sensitivities - the change in
by means of a full AC-Load Flow solution. This is achieved by calculating the the performance index AJt is calculated using the Taylor series expansion of
changes in the MW-flow performance index to changes in susceptance of the lines JMW as a function of the susceptance B,R of line,( (,=1,2,...,NL). That is,
in the network. These changes to the performance index can be obtained from
the sensitivities of JMW with respect to the susceptance of the lines in the net-
work. The first-order and second-order sensitivity expressions are given next
J = -aJMW(- B +1 M( )2 (11)
(see appendix I for details).

First-Order Sensitivities where the derivatives JMW/ZUB, and 32jMW/aB2 are given by eqs. (5) and (9)
respectively. Notice that the "change" in susceptance is precisely equal to the
The first-order sensitivity of the MW-performance index of eq. (4) is susceptance of the line so as to simulate a line outage. For the first-order
(similar formulas hold for the performance index of eq. (1)), effects only the first term in the right hand side of eq. (11) is needed. Notice
thatthe second-order method requires N L forward-backward solutions (see eq.(9)),
Bjmw - 4 01 J= It 2,
0't ... I NL (5)
-6 BX DC-Load Flow
where
The MW-flow in the lines of the network after a change in susceptance
Se angle across line,f (between nodes m and n) for the base case condition AB2 for lineA are given by (see appendix 1)
OX( angles across line,( in the "adjoint" network (Note: the same notation
SPk + b.( (MkT B- IM() B IkC.,<>
of [1] is used herein although the concept of adjoint network as used in PIt _ ,
(12)
that paper is not used in the present paper.)
L(p - AB<q( ) [I1 b, (Bla,3k,
The adjoint angle across line,t can be expressed as
A where Pk is the new flow in line k for a change in susceptance in lineX and
6A= MT 8 (6)
where
'A= (ABi + MT B1 E
-A
1 (13)

Mx column vector (NBxl) which is null accept for elements m and n which Therefore, the effect of the outage can be easily simulated by calculating the
are equal to 1 and -1 respectively (NB = number of nodes) approximate line flows from eq. (12) and substituting in the performance index
of eq. (1). Notice that the computational complexity of this approach is of the
8 vector of adjoint nodal angles. (NBx1) same order as that of the second-order sensitivity method as NL forward-back-
ward solutions are required.
In eq. (6) the vector of adjoint nodal angles can be obtained from
A A

BS = P (7) TEST CASES AND RESULTS

where The automatic contingency selection algorithm described above was
8 susceptance matrix (NBxNB) tested under real-time conditions on the AEP-EHV system. The base case con-
A
ditions corresponded to the system conditions given by the AEP State Estimator
P vector of adjoint injections (NBx1) at 6 AM on October 18, 1977. The reaction of the external system to changes
in the AEP-EHV internal system was considered by coupling to the internal
The adjoint injections vector can be calculated from system an updated REI equivalent of the external system [3]. A Ward equi-
valent (constant current) was also used to determine the effect of the equivalent
on the selection algorithm. The results obtained with this type of equivalent
A
were inferior to the results obtained with the REI equivalent.
= {% sign (Oj ) Mj } (8)
-L 1 j9 1 Results with both performance indeces (eqs. (1) and (4)) for the first-
order sensitivity, second-order sensitivity and DC-Load Flow methods are given
where 3JMW/-l;j can be calculated directly from eq. (4). in the sequel. The "correct" ranking, as given by a full AC-Load Flow, is used
as reference for comparison. Thus, the normalized sensitivities and approximate
Notice that the angles 0,( of eqs. (5) - (8) correspond to angles differ- DC-Load Flow performance index values are ordered according to the ranking
ences across the lines for the base case condition. This base case, for on-line given by the AC-Load Flow.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
 1554

TABLE 1: CONTINGENCY RANKING FOR PERFORMANCE INDEX OF EQ. (1) TABLE II: CONTINGENCY RANKING FOR PERFORMANCE INDEX OF EQ.(4)

Line AC-Load First Order Sec.-Order DC-Load Line AC-Load First-Order Second-Order
p q Flow Sensitivity Sensitivity Flow p q Flow Sensitivity Sensitivity

7 8 O .84406E+01 0.35623E+00 0.71958E+00 0.86371E+01 7 8 0.87901E+01 0.61241E+00 0. 11185E+01

11 10 0 .81172E+01 0.11882E+00 0.25049E+00 0.85318E+01 11 10 0.82872E+01 0.36905E+00 O0 77301E+00
5 4 0.76232E+01 -0.11607E+00 -0.13657E+00 0. 77251E+01 5 4 0. 79739E+01 0.10160E+01 0. 21348E+01
10 9 0. 75269E+01 -0.87468E-01 -0.63064E-02 0. 76032E201 10 9 0.77740E+01 0. 38803E+00 0.77703E+00
14 9 0.74670E+01 0.70404E-01 0.13016E+00 0. 75049E+01 14 9 0. 76912E+01 0O 20216E+00 0 .40064E+00
31 7 0. 74526E+01 0.33386E-01 0.14818E+00 0. 74665E+01 9 21 0.76599E+01 -0.71413E-02 0. 89425E-01
31 7 0. 74525E+01 0.31032E-01 0.14507E+00 0. 74666E+01 31 7 0.76517E+01 0.42261E+00 0.68580E+00
49 33 0.74243E+01 0.26069E+00 0.44455E+00 0. 77241E+01 31 7 0. 76516E+01 0.42545E+00 0 .69219E+00
30 31 0. 74206E.01 0.17424E+00 0.27030E+00 0. 74586E+01 30 31 0. 76368E+01 0.17635E+00 0. 27333E+00
9 21 0.74084E+01 -0.13946E+00 -0.16912E+00 0. 75072E201 14 15 0.76141E+01 0.20416E+00 0.30228E+00
14 15 0. 73728E+01 0.20059E+00 0.29771E+00 0. 74445E+01 49 33 0.75910E+01 0.23981E+00 0.41586E+00
11 12 0. 73393E+01 -0.55354E-01 -0.84896E-01 0. 73931E+01 11 12 0. 75802E+01 0.76755E+00 0.17061E+01
10 19 0. 73370E+01 -0.94446E-01 0.20769E-01 0. 78772E+01 10 19 0.75381E+01 0.79741E+00 0. 15221E+01
43 44 0. 72934E+01 0.10426E+00 0.14971E+00 0.72641E+01 43 44 0.74692E+01 0.95305E-01 0. 13814E+00
31 32 0.72519E+01 0.67503E-01 0.11781E+00 0.72380E+01 20 21 0.74643E+01 0.47570E+00 0. 10338E+01
21 22 0. 72477E+01 0.14570E-01 0.46449E-01 0. 72666E+01 31 32 0. 74576E+01 0.74886E-01 0. 12711E+00
20 21 o . 72275E+01 -0.92726E-01 -0.13446E+00 0. 72789E+01 21 22 0.74448E+01 0.23710E+00 0.42392E+00
29 30 0. 72269E+01 0.49955E-01 0.86273E-01 0. 72679E +01 55 25 0. 74268E+01 0.96784E-01 0. 12891E+00
55 25 0. 72262E+01 0.93268E-01 0.12432E+00 0. 72266E+01 29 30 0. 74258E+01 0 .48425E-01 0 . 83681E-01
55 25 0.72232E401 0.92792E-01 0.12353E+00 0.72255E+01 55 25 0.74231E+01 0.96290E-01 0. 12808E+00
25 5 0. 72119E+01 0.74389E-01 0.10640E+00 0.72163E+01 8 59 0. 74223E+01 0.53540E-01 0. 11225E.00
a 59 0.72100E.01 0.14557E-01 0.29611E-01 0. 72505E+01 25 5 0. 74079E+01 0.75099E-01 0. 10737E+00
25 1 0.71990E+01 0.89399E-01 0.11869E+00 0.72197E+01 14 13 0.73978E+01 0.61844E-01 0.86353E-01
14 13 0. 71924E+01 0.59447E-01 0.83199E-01 0.71845E+01 25 1 0.73964E+01 0. 87988E-01 0. 11686E+00
42 49 0. 71845E+01 0.62595E-01 0.91675E-01 0.72024E+01 42 49 0.73750E+01 0.59208E-01 0. 86945E-01
42 49 0. 71845E+01 0.62595E-01 0.91675E-01 0.72024E+01 42 49 0. 73750E+01 0.59208E-01 0. 86945E-01
32 33 0. 71701E+01 0.62379E-01 0.87572E-01 0.71894E+01 32 33 0. 73592E+01 0.57664E-01 0. 81429E-01
60 59 0.71601E+01 0.55968E-02 0.11187E-01 0.72797E+01 60 59 0. 73571E+01 0.19686E-01 0 .45459E-01
35 37 0. 71449E+01 0.16417E-01 0.28295E-01 0. 71460E+01 10 20 0. 73420E+01 0.37657E-01 0. 73582E-01
10 20 0. 71354E+01 0.30485E-02 0.88189E-02 0. 71336E+01 35 37 0. 73384E+01 0.14884E-01 0. 25691E-01
29 28 0.71321E+01 0.19436E-01 0.32363E-01 0. 71449E+01 29 28 0.73293E+01 0.18730E-01 O0 31199E-01
32 34 0. 71291E+01 0.21984E-01 0.34120E-01 0. 71344E+01 34 35 0.73218E+01 0.16568E-01 0.22688E-01
34 35 0.71279E+01 0.17699E-01 0.24188E-01 0.71148E+01 34 35 0.73210E+01 0.16344E-01 0.22299E-01
34 3S 0. 71271E+01 0.17459E-01 0.23773E-01 0.71142E+01 32 34 0.73210E+01 0. 19598E-01 0. 30655E-01
13 25 0. 71102E+01 0.11001E-01 0.17663E-01 0.71149E+01 13 25 0. 73076E+01 0.11569E-01 0. 18572E-01
37 38 0.71023E+01 0.42585E-02 0.69991E-02 0 70989E+01
O 13 58 0.72995E+01 0.78154E-02 0. 11270E-01
13 58 0.71014E+01 0.72042E-02 0.10392E-01 0.70999E+01 37 38 0. 72981E+01 0.38441E-02 0 .63205E-02
35 36 0.71003E+01 -0.39138E-02 -b.33840E-02 0. 70843E+01 35 36 0.72950E+01 -0.28353E-02 -0 -20590E-02
35 36 0.70996E+01 -0.38655E-02 -0.33487E-02 0. 70343E+01 35 36 0.72943E+01 -0.28003E-02 -0 . 20430E-02
39 38 0.70974E+01 0.22848E-02 0.38602E-02 0.70943E+01 39 38 0. 72927E+01 0.20820E-02 0. 35183E-02
42 43 0.70969E+01 0.42921E-02 0.77304E-02 0. 71036E+01 42 43 0.72899E+01 0.34418E-02 0. 63549E-02
37 40 0. 70907E+01 0.10769E-02 0.19343E-02 0. 70923E+01 45 46 0. 72884E+01 -0. 30920E-02 -0 . 32295E-02
45 46 0.70890E+01 -0.15026E-02 -0.84304E-03 0.70897E+01 37 40 0.72868E+01 0.91564E-03 0.16454E-02
49 34 0. 70874E+01 0.32664E-02 0.43041E-02 0.70918E+01 49 34 0.72834E+01 0.30653E-02 0 .40469E-02
48 50 0.70802E+01 -0.37575E-01 -0.41432E-01 0.70944E+01 6 59 0. 72807E+01 0. 87505E-02 0.19357E-01
7 6 0.70791E+01 -0.22947E+00 -0.19744E+00 0.75462E+01 48 50 0.72730E+01 0.11065E+00 0.20166E+00
6 59 0.70789E+01 -0.54482E-02 -0.86293E-02 0.70789E+01 49 48 0.72705E+01 0.13119E-01 0. 20805E-01
49 48 0.70753E+01 -0.24219E-01 -0.21754E-01 0.70734E+01 42 41 0. 72695E+01 0. 19645E-01 0.'46575E-01
42 41 0.70725E+01 -0.28515E-02 -0.51833E-02 0. 70821E+01 7 6 0.72692E+01 0.20136E+00 0.52972E+00
39 36 0.70724E+01 -0.13255E-01 -0.18022E-01 0. 70664E+01 39 36 0.72686E+01 -0.11687E-01 -0. 15817E-01
15 16 0.70711E+01 -0.53230E-02 -0.91143E-02 0.70685E+01 41 34 0. 72680E+01 0.21131E-02 0 .89855E-02
41 34 0.70708E+01 -0.18376E-01 -0.14235E-01 0. 70814E+01 15 16 0.72666E+01 -0.54422E-02 -0. 93191E-02
3 2 0. 70652E+01 -0.73650E-02 -0.10984E-01 0.70729E+01 3 2 0. 72649E+01 -0.68313E-02 -0. 10148E-01
57 13 0. 70639E+01 -0.82317E-02 -0.12427E-01 0. 70704E+01 57 13 0.72612E+01 -0.85849E-02 -0.12977E-01
32 50 0.70578E+01 -0.88433E-02 -0.12735E-01 0. 70727E+01 32 50 0.72547E+01 -0.92294E-02 -0.13343E-01
48 43 0. 70512E+01 -0.19804E-01 -0.25179E-01 0.70677E+01 48 43 0.72544E+01 0.12193E-01 0. 24545E-01
20 23 0. 70498E+01 -0.12012E+00 -0.21289E+00 0. 70012E+01 12 13 0.72480E201 -O .35650E-01 -0.40606E-01
12 13 0.70456E+01 -0.37555E-01 -0.43120E-01 0. 70457E+01 54 47 0.724032+01 0. 10222E-01 0. 23217E-01
54 47 0. 70368E+01 -0.27090E-01 -0.39538E-01 0. 70422E+01 45 44 0. 72046E+01 -0.85077E-01 -0. 10588E+00
45 44 0.69868E+01 -0.89442E-01 -0.11172E+00 0.69700E+01 46 44 0.71831E+01 -0.69885E-01 -0 . 10287E+00
46 44 0.69639E+01 -0.74783E-01 -0.11039E+00 0.69466E+01 20 23 0.71811E+01 0.74960E+00 0. 17572E+01
5 6 0.69528E+01 -0.76860E-01 -0.11893E+00 0.69825E+01 5 6 0 71792E+01
O -0.73462E-01 -0 .11309E+00
1 2 0.69303E+01 -0.87714E-01 -0.11790E+00 0.69587E+01 43 47 0.71340E+01 -0. 27767E-01 -0. 91518E-02
43 47 0.69060E+01 -0.17640E+00 -0.21566E+00 0 .68949E+01 1 2 0.71327E+01 -0.87234E-01 -0 . 11721E+00
24 23 0.68953E+01 -0.19120E+00 -0.23116E+00 0. 68834E+01 44 47 0.70783E+01 0.15984E+00 0. 36853E+00
27 26 0.68877E+01 -0.31343E200 -0.41344E+00 0. 68834E+01 27 26 0.70697E+01 0.35367E+00 0. 73553E+00
17 16 0.68775E+01 -0.10685E+00 -0.18662E+00 0.68031E+01 24 23 0.70592E+01 0.40874E-01 0.85872E-01
18 17 0.68770E+01 -0.61632E-01 -0.11321E+00 0. 67865E+01 27 28 0.70531E+01 -0.10987E+00 -0 .17270E+00
-27 28 0.68714E+01 -0.10666E+00 -0.16737E+00 0.68586E+01 52 51 0.70487E+01 -0. 32178E-01 -0 . 35494E-01
19 18 0. 68448E+01 -0.39565E-01 -0.71692E-01 0.71154E+01 17 16 0. 70467E+01 0.92637E-01 0.25026E+00
44 47 0.68432E+01 -0.11185E200 -0.18682E+00 0 .68226E+01 18 17 0.70464E+01 0. 46561E-01 0. 13451E.00
51 48 0.68377E+01 -0.18334E+00 -0.26929E+00 0.68402E+01 51 48 0. 70429E+01 0.14087E+00 0. 31636E+00
52 51 0.68351E+01 -0.81906E-01 -0.14398E+00 0.68500E201 19 18 0. 70145E+01 -0.44297E-01 -0 .80725E-01
61 4 0. 68106E+01 -0.38175E+00 -0.51053E+00 0.68181E+01 61 4 0.69649E+01 0.88879E+00 0.16946E+01
26 56 0.67847E+01 -0.17343E+00 -0.26639E+00 0 .67814E+01 26 56 0.69642E+01 0.19899E+00 0 .42312E+00
9 8 0.67372E+01 -0.15727E.00 -0.26412E+00 0. 70206E+01 48 53 0.68633E+01 0. 71581E+00 0. 15183E+01
48 53 0.66269E+01 -0.36883E200 -0.54612E+00 0 .67709E+01 9 8 0.68549E+01 0.66599E-01 0.22379E+00
1 4 0.64637E+01 -0.26587E+00 -0.42653E+00 0 .64698E+01 1 4 0.66604E+01 -0.10613E-02 0.72421E-01
25 26 0.62448E+01 -0.89742E+00 -0.12167E+01 0.62544E+01 25 26 0.63938E+01 0.52910E+00 0.12275E+01

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
 1555
Table gives the results obtained for the performance index of eq. -mW
(1), while Table 11 gives corresponding results obtained for the performance -
7.4
index of eq. (5). The results obtained with the DC-Load Flow formulation are
shown in Table I and they are the same for both performance indeces. The base
case value of the performance index for the results of Table is 7.087 while for
the results of Table 11 is 7.2829.
From these results it is clear that although the ranking of contin-
gencies according to the first-order and second-order sensitivity methods is not
totally incorrect (as evidenced by the "clustering" of contingencies according 7.3
to importance), some of the possible line outages were misclassified. For in-
stance, for the performance index of eq. (1), the outage of lines (5-4), (10-9),
(9-21), etc., that appear to be important from the ranking given by the AC-Load
Flow, would have been classified as unimportant contingencies by both the
first-order and second-order sensitivity methods. Also, for the results of Table
11, line (9-21) would have been classified as unimportant based on first-order
sensitivities. Use of second-order sensitivities corrects this particular case
(it does not, however, provide extensive correction in other cases). Cases such 7.2
as these are particularly dangerous and any automatic contingency selection
method should avoid such situations. On the other hand, false alarms (unimport-
ant lines classified as important) are not as dangerous except if the misclassi-
fication is extensive. Indeed, if some unimportant cases are classified high in
the list, they may overshadow important cases which otherwise would have been
considered when running the full AC-Load Flows to further study the outages.
This situation is also to be avoided. Examples of false alarms can be seen
from Table 11. For instance, lines (61-4), (25-26), (48-53) would have been 7.1
classified high by both the first-order and second-order sensitivity methods
although the AC-Load Flow flagged them is unimportant. (OP)

For the purpose of demonstrating the reasons behind the unreliable

behaviour of the sensitivity methods, plots of JMW versus B/were made for a
few lines. From the plots it was discovered that the performance indeces (both
eqs. (1) and (4))are not always monotonic functions of B. Figure 1, for instance, _ 7.0
is a plot of JMW (eq. (4)) versus B/for line (19-18). Notice that for this parti- 320 240 160 80 0
cular line JMW is not a monotonic function of BN. Moreover, the operating point Si
(OP) is a point of negative slope and, within the region of interest (from OP to
the ordinate axis), the slope of the performance index function JMW changes sign. Figure 1: Plot of the Performance Index of Eq. (4) vs. Susceptance
for line (19 - 18)
(BI)
Therefore, if the first-order sensitivity (and, in this case, even the second-order
sensitivity because JMW is not a second-order function of B.t) is used to "pre-
dict" the value of JMW after outaging the line (B1,1 0), the prediction would be
far from the "true" value given by the AC-Load Flow. This behaviour of the
performance index function is, of course, system and operating point dependent studies. It may even mask important lines that otherwise would appear higher
and may or may not occur. This inconsistent behaviour is, precisely, what in the ranking than the misclassified cases. The latter case - misclassification
makes the method of automatic selection unreliable and, to some extent, un- of important outages - is, of course, highly unwanted in an on-line automatic
attractive. environment.

Another obvious reason for the unpredictable behaviour of the method
is that the approximation to the performance index function as given by eq. (11)
is, of course, valid only in the neighborhood of the "operating point", and may
not hold true (even with second-order effects) when the susceptance change is
large, as is precisely the case when simulating a line outage.
A way of overcoming these difficulties is provided by the use of the
DC-Load Flow model of eq. (12). The results given in Table for the DC-Load
Flow model (which also apply for the performance index of eq. (4)) show that
this approach is highly reliable and accurate. Minor misclassification resulting
in false alarms (as is the case of lines (10-19), (7-6), (19-18), and (9-8)) may
occur as shown in the results. It is expected, however, that these errors will
be more infrequent in cases where there is more separation between the impor-
tant and unimportant contingencies. The DC-Load Flow methods is, of course,
more demanding, computationally, than the first-order sensitivity method. It is,
however, less so than the second-order sensitivity method.
CONCLUSION S
The results of applying the automatic contingency selection method
of Ejebe-Wollenberg [11 to the real-time AEP-EHV system conditions were pre-
sented in this paper. It was found that this method, although conceptually high-
ly appealing, is unreliable and, under certain circumstances, may give incorrect
results. Two types of anomalous behavior were identified,
- classification of unimportant line outages as important (false alarms), and,
- classification of important outages as unimportant.
The former type of error, although not extremely dangerous, may in-
crease the computational load of running full AC-Load Flow cases for further
It was shown in the paper that the unreliable behaviour of the auto-
matic contingency selection method is due to the fact that the performance index
function is not a monotonic function of transmission line susceptances. This
is, in fact, at least a necessary condition for the acceptable performance of
the method. Even if the performance index is monotonic, a high degree of mis-
classification may result depending on the degree of convexity or concavity of
the performance index function. This, however, is highly system and operating
point dependent and cannot be ascertained a priori. All these factors make the
method, at best, unreliable.
For the purpose of enhancing the performance of the automatic con-
tingency selection, the method was extended to include second-order effects.
Although the results obtained with the second-order extension were better than
those obtained with the first-order method, the improvement was not significant.
It was shown that, not only the first-order sensitivity but the second-order sen-
sitivity formulas can be readily developed using elementary differential calculus
providing thus an alternative, and clearer way of developing the equations, than
the use of Tellegen's theorem.
It was also shown in the paper that the DC-Load Flow provides a
competitive alternative to the automatic contingency selection method. It is
apparent from the results obtained with the DC-Load Flow, that this method is
accurate and reliable. Therefore, for on-line contingency analysis, serious
consideration should be given to the possibility of using a DC-Load Flow model
for the selection of meaningful transmission line contingencies.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the various suggestions made by
Dr. B. Wollenberg of Power Technologies, Inc., and Dr. R. *Campo and Mr. P.
Ruiz of American Electric Power Service Corporation.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
 1556
REFEREN CES and using eq. (1.8) in (1.9) and the fact that MT B-1 M, = M B M, it is seen
[1] Ejebe G. C. and Wollenberg, B. F., "Automatic Contingency Selection", that
NL
paper F78 228-9, presented at the IEEE-PES Winter Meeting, New York,
NY, January 1979. ,ajmw = MT B- l , JMW sign (0j ) M ja (1.10)

[2] Clements K. A. and Ringlee R. J., Final Report - Research on Adequacy Let
l 3JMW
Assessment of Interconnected Electric Power Systems, Worcester Poly- NL
technic Institute, Electrical Engineering Department, December 1977. P =
-a II sign (Ij) Mj (1.11)

[31 Dopazo J. F., Irisarri G., and Sasson A. M. "Real-Time External System j =I
Equivalent for On-Line Contingency Analysis", paper F78 674-4, pre- and
sented at the IEEE-PES Summer Meeting, Los Angeles, CA, July 1978. A A
8= B-1 P (1.12)
[4] Fischl R. and Puntel W. R., "Efficient Method for Computing Electric
Power Transmission Network Sensitivities", paper C72 167-0, IEEE Using eq. (1.12) in (1.10) and letting Ol = M'_ it is finally obtained
Winter Power Meeting, New York, NY, January 1972.
,ajmw
[5] Fischl R. and Puntel W. R., "Computer-Aided Design of Electric Power
Transmission Networks," paper C72 168-8, ibid.
-T19 (1.13)

which is thefth (.( = 1,2, ..., NL) element of the gradient vector of eq. (1.1).
APPENDIX I
Second-Order Sensitivities
First-Order Sensitivities
The gradient of the performance index of eq. (4) with respect to From eq. (1.9) it can be shown that the second order derivative
changes in susceptance of the lines in the network is (3JMW/3 Bk X B,() is given by

(VB JMW)T = [JmW j] (1.1) a Jmw JMw 3 + 'JMW

F
leiI alej 'leji
Bk Bk Z.29Ol Bk3BA -aO1j2 Bk -3 J
where thefth element is §= (1.14)

iJ3-MW = [ -MW ]T[ (1.2) In this equation the only term that needs further elaboration is the derivative
(32 1Ojl//BkB B..). The other terms can be obtained from previous results.
with 101 a vector of angle differences across the lines in the network. Notice It can be easily seen that
that the elements of vector [3JMW/l.2I] in eq. (1.2), can be readily obtained
from eq. (4). On the other hand, element jth of [ lOI/BB ] is equal to
-aBk-aB
e :- sign (Oj) j
Bk-aBL
, Oj t (1.15)
li = sign(0j) -, 9j P O (1.3)
-a B -a Bx
From eq. (1.8), and using similar manipulation as that used for the
To calculate the derivative 3aj/aB, in eq. (1.3),consider the DC- first-order case, it can be shown that eq. (1.15) can be written
Load Flow formulation
I2 I 1 ) (O
-B 8
T
P
-

= (1.4) s

iBn)( k+MM (1.16)

where 8 is the vector of nodal angles and P is the vector of nodal injec
From eq. (1.4) it is clear that Notice that this is a general formula that can be used to simulate the effects
-S = _ -B-1 of double contingencies. However, for the case of single contingencies, eq.
lB B (1.16) reduces to
a2l - 2 sign (Oj) (MT B" M )(MTB U) (1.17)
but the derivative of the inverse susceptance matrix is equal to

3 Bt D B,t
(1.6) Using eq. (1.17) in (1.14) (with k =t) it is obtained

using eq. (1.6) and noticing that BD B,B - M, MJc it can be easily sho wn that a JMW - 2 ( B e) (M1B ) P df
'aBBa-B -B1 M,.
M R
X
(.18)
+ (M, B' ) V,f-
; - ~I1 Oj22
m
j j(B 9)0
and from eqs. (1.3) and (1.7) it is finally seen that
a sign (Oj
s
)
(MT BI M0) , (1.8)
Let
A 'Jmw
B = ,...~]{ ' 3l0j 2 Mj MT } (1.19)
j i2J
To calculate the elements of the gradient vector, eq. (1.2) can be
written as Substituting eq. (1.19) into (1.18) and simplifying it is finally obtained
NL that
2 A

BDB, _ /az (1.9) Jm

v jMW = Ilj l0jBB
A
-.0 .0

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.

which can be used, together with eq. (1.13), to calculate the changes in the per-
formance index to changes in the susceptance of linefusing eq. (11).
DC-Load Flow
From eq. (1.4) the nodal angles, after a change in susceptance for
line,( of AB/, are given by
Ut=- (B + AB M T)a lP
ivro i Substituting
Using the well known matrix inversion lemma it can be shown that
84 = 8 -
b B- aO_
M, (1.22)
where b,Ais given by eq. (13) and 8 is the vector of nodal angles corresponding
to the base case conditions.
D. Levner, photograph and biography not available at the time of
publication.
G. Irisarri, photograph and biography not available at the time of
publication.
Alberto M. Sasson (S'65, M'69, SM'70) was
born in Barranquilla, Colombia in 1938. He
enrolled in Engineering, Universidad de los
Andes, Bogota in 1957. He received the B.Sc.,
degree and the M.Sc., degree from the Universi-
ty of Illinois in 1961 and 1963 respectively. In
1968 he received the Ph.D. degree from Imperial
College, London.
He joined the Instituto Technologico y de
Estudios Superiores de Monterrey as Instructor,
in the electrical engineering department,
Monterrey, Mexico. In 1969 he was appointed Professor of Electrical
Engineering, at the Instituto Technologico de Monterrey. In 1970 he
was a Lecturer,short courses, at the Universidad de los Andes, Bogota,
Purdue University, Commonwealth Edison, Chicago. In 1971 he joined
the American Electric Power Service Corporation (AEP), as Senior
Engineer, in New York, to work on real-time computer applications to
power systems. He became Chairman, of Power Systems, Conference
on Systems Networks and Computers, Oaxtepec, Mexico. He later was
promoted to Staff Engineer at AEP with responsibilities in real-time
power system security problems. In 1974 he worked with artist Ramirez
Villamizar in the use of computers to create and execute drawings. He
was a Lecturer on short courses in Spain, Romania, Mexico,
Guatemala, Panama and Colombia. He currently works with artist
Felguerez on the development and use of mathematical models of the
aritst's style, using system identification techniques.
In 1965, publication on power system design with computers, he
was awarded first prize for research paper, Panamerican Congress,
Mexico.
Dr. Sasson is or has been a consultant to a number of utilities in
Mexico, Colombia and Venezuela and has published nearly fifty papers
inthe fields of power system analysis, network theory, non-linear pro-
gramming, diakoptics, real-time monitoring and application of iden-
tification theory.
Discussion
T. A. Mikolinnas, W. R. Puntel, R. J. Ringlee, and B. F. Wollenberg
(Power Technologies, Inc., Schenectady, NY): The authors should be
commended for presenting some challenging ideas about the use of con-
tingency selection techniques. In their paper, they have reported on
problems they encountered with the use of gradient methods for rank-
ing circuit contingencies. They concluded from their tests that linear
power angle, "DC," flow models offered more dependable results than
gradient methods.
1557
The new line flows are
-Bk Ok , k J,
(1.23)
S t +A ) from,k=q1t
(Bin Be
S'k4in eq. (1.23) from eq. (1.21) it is finally obtained
Pk +be (MTB-' MJBk8A
pA=
[ X (k B- , k (1.24)
(% -AB't, ) [I - k(MAT- I Mt)] , k =,4
The new flows given by eq. (1.24) can be substituted in eq. (1) to calculate the
value of the performance index.
We do not disagree with the results reported in the paper but we do
think it appropriate to raise some questions about the tests on circuit
outages and the conclusions drawn by the authors. We also wish to
report on some additional findings.
First, it will be noted from Table I and from page 3 of the text that
the change in performance index from the 6 AM October 18, 1977 AEP
system base case for the highest ranked contingency (loss of branch 7-8)
is quite small. Apparently this represents a light load case. Would the
authors please indicate how many overloads were present, and for each
overloaded circuit, the flow expressed as a percentage of that circuit's
rating, for the base case and several of the higher ranked contingencies?
We are particularly curious about the cases where branches 5-4 and 10-9
were dropped.
Second, we would appreciate clarification of the statement in the
third paragraph in the section Test Cases and Results which reads "The
results obtained with the D.C. Load Flow formulation are shown in
Table I and they are the same for both performance indices." While we
agree that the D.C. load flow solutions are of course identical, the
numerical values of the two performance indices given by equations (1)
and (4) are not the same when a circuit has been dropped. In particular,
as BK goes to zero, so does the associated term in eq. (1), whereas in eq.
(4), the corresponding term goes to a nonzero constant. The final values
of the two performance indices are then related by:
lim J + lim J _ (wfb.ol_fMt.l)
(911 8,11
B,-+ (eq. 4) B,-+ (eq. 1)
where circuit K was spanned by buses i and j. As a result, the rankings
based on the respective indices, both calculated from the same D.C.
load flow solution, are different.
Alternatively, in the base case, where no circuits have yet been
dropped, eqs. (1) and (4) should give the same numerical value for both
performance indices. Would the authors please indicate why the two in-
dices have different values?
Third, according to the authors, the D.C. load flow provides a
competitive alternative to contingency selection methods. Indeed, for
systems of modest size, the cost in terms of computational effort and
elapsed time may be tolerable. However, neglecting the time required to
build and factor the system susceptance matrix, a minimum of LT
seconds is required, where L is the number of branches in the system
and T is the time required for a single forward backward solution. For
large systems this time can become prohibitive. For example, Table 1
gives the times required to perform a DC load flow and check for
overloads, exploiting sparsity, for each single circuit outage, on each of
the three systems on a Prime 400 computer. Clearly, with the additional
time required to perform a number of AC load flows, the total time re-
quired for computations can become excessive. For this reason, we feel
that it is more practical to use the DC load flow in conjunction with
rather than as an alternative to some means of contingency ranking.
A frequently cited arguement is that it is unnecessary to run L DC
load flows because the system operator can provide a shorter list of con-
tingencies to analyze. However, as the authors correctly point out "As
 1558
System # of Buses I of Branches LT(cpu-sec) LT(elapsed min)
A 569 953 471
B 857 1369 866
C 628 1099 663
Table 1. Timing Data for D.C. Load Flows
operating conditions in most systems get more stringent, the number of
contingency choices gets larger and no amount of experience guarantees
proper selection of troublesome contingent situations." Indeed, this is
one of the major reasons for interest in contingency ranking methods.
A rhetorical question which can then be asked is: If the operator cannot
guarantee the proper selection of troublesome contingent situations,
then how can he guarantee that his shortened list will contain all such
troublesome situations? We are not disputing the operator's ability to
select and eliminate from consideration many of the circuit outages
which have little or no adverse effect on system performance. We are
merely pointing out that if the process is carried too far, he may er-
roneously eliminate a "troublesome contingency" from the list of con-
tingencies to be analyzed.
Fourth, it is not clear whether the authors interpret the perfor-
mance indices as summations over individual circuits or equivalent
branches (where parallel susceptances have been combined into a single
equivalent susceptance). Perhaps they could clarify this point, since
each interpretation results in a different performance index.
Note that interpreting J as a sum over equivalent branches does not
imply that all circuits whose susceptances were combined must be
outaged simultaneously.- The gradient still gives aJ/aBK. However, J is a
different function than when a sum over circuits is-considered and BK is
interpreted as a susceptance between buses i and j which- span
equivalent circuit K. When a J/ a BK is multiplied by AB corresponding to
the outage of any one of the circuits between buses i and j, it gives a
prediction of the corresponding change in J. While it may be more in-
tuitively satisfying to consider the summation over circuits, our ex-
periences indicate that a summation over equivalent branches results in
higher capture rates. (We define the capture rate, r,I, as the fraction of
the worst N contingencies appearing in the first I*N entries in the rank-
ing.)
It has been our experience with the gradient method (requiring one
forward backward solution T sec.) that capture rates in the range of
0.60<rNO0.80 are attainable. Table 2, which summarizes the capture
rates for the example from the paper, shows that even for this
presumably light load case, such capture rates can be reached.
N r N r2N
1.00 1.00
5 0.40 0.40
10 0.40 0.70
20 0.65 0.70
30 0.77 X the respective performance index equations. The outaged line,of
40 0.85 * *
* I*N > L/2
Table 2. Summary of Capture Rates for AEP System
It is nevertheless dissatisfying to see the third worst contingency in
the example misranked. We concur with the authors that when using
the gradient method, the non-monotonic nature of the performance in-
dex may occassionally result in such misrankings.
Fifth, we are of the opinion that using another computationally ef-
ficient method of ranking contingencies in conjunction with the gra-
dient method could prove superior to running an exhaustive list of DC
load flows. An alterntive ranking algorithm currently being investigated
shows great promise. Requiring approximately 8T seconds for execu-
tion, it is based on weighted equivalent injections, and recognizes the
discrete change which occurs when a circuit is dropped. Capture rates
for the three systems described earlier are given in Table 3. The range of
values given is for N varying from N = 10 to N 0.2*L.
Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
9.11
System rIN r2N
_ r3N
A 0.47 - 0.67 0.50 - 0.96 0.70 - 0.99
16.7
B 0.25 - 0.50 0.40 - 0.74 0.50 - 0.86
13.0
c 0.50 - 0.65 0.65 - 0.92 0.85 - 0.98
Table 3. Capture Rates of Weighted Equivalent Injection Method
Thus far, only single circuit outages have been considered. The
problem of unit outages still remains. In our experiences the gradient
method handles these quite well. Have the authors investigated this
aspect of the gradient method?
In conclusion, we agree with the authors that under certain cir-
cumstances the gradient method has its shortcomings. However, we feel
it can be used to advantage, alone or with other rankings techniques, to
significantly reduce the time required to determine and analyze the
worst contingencies given an initial system state.
Manuscript received February 23, 1979.
G. Irisarri, D. Levner, and A. M. Sasson: We would like to thank the
discussers for their interest in our paper and for their valuable and
perceptive comments. We would like to address each of the questions
raised separatedly and in the order they were posed.
The results given in the paper do indeed correspond to a light load
case as the discussers correctly surmised. However, to compensate for
the light load conditions, the MW capability of the transmission lines
was reduced to half of their design value forcing thus a heavy loaded
condition. The results obtained with the gradient-type selector,
however, did not improve by this limit reduction. This again cor-
roborates our main contention in the paper, that is, the major limitation
of the gradient-type methods is a theoretical limitation springing from
the very choice of the performance index wherein the gradient method
is based. Incidentally, we searched for other meaningful types of per-
formance indeces and particularly those that would overcome the
limitations of the performance indeces of eqs. (1) and (4); however, our
search proved to be unsuccessful.
For the particular data of October 18, 1977 used in the paper there
were no overloads for the base case data. Reducing the MW capability
of the lines-which may be considered a valid operating deci-
sion-produced overloads in some of the lines. For the base case, for
example, line 7-8 was carrying 903.5 MW (capacity 2000 MW), line 5-4
was carrying 430.0 MW (capacity 800 MW) and line 10-9 was carrying
3N 659.1 MW (capacity 2000 MW). If the reduced MW capabilities are
used, line 5-4 will show an overload of 8% over the 400 MW limit.
1.00
However, again, citing particular cases does not enlighten, we feel, the
basic problems of the gradient-type method.
0.60 The results given by the DC (linear) load flow we believe must be
0.70
the same for both performance indeces as mentioned in the paper. The
main reason for this is given by eqs. (2) and (3) used to obtain the per-
formance index of eq. (4) from the performance index of eq. (1). These
equations are precisely the DC load flow equations used in calculating
the angles 0, and the flows P, which were subsequently substituted in
course, must not be included in the performance index evaluation once
the line is "dropped". On the other hand, the values of the perfor-
mance indeces for the base case are different because the angles J, used
in eq. (4) are the true (SE given) angles for the base case while the P,
used in eq. (1) were calculated using eq. (2) which is, of course, only an
approximation. If the flows P, had been calculated using the full AC
equations, the results would have been identical as the discussers
rightfully point out.
We still feel that the DC (linear) load flow is indeed, at present and
for the lack of a better technique, a good alternative to the gradient-
type methods. The discussers oppose the use of the DC load flow on the
grounds of excessive computational complexity. However, systems of
the sizes cited in their discussion (Table 1) are average system sizes for
planning applications not for real-time applications wherein the typical
system sizes are in the order of 100-150 nodes and 200-250 lines. The
AEP EHV system, for instance, has 71 nodes and approximately 112
lines. Therefore, only 112 linear load flow solutions (forward-backward
on a triangulated coefficient matrix) are needed. CPU times obtained
for systems of this size are well within the requirements of present day
control center applications. In fact, for the 118 node, 180 lines IEEE

test system we obtained the following CPU times for two different com-
puters: 4.2 secs. on an IBM 370/165 and 24.7 secs. on a TANDEM
minicomputer. Even if the AEP 138KV system (300 nodes 500 lines)
were included as part of the monitored system (plans to do so are
presently underway) the size will not be a great problem because of the
138KV system will be "torn" in three subsystems of roughly the same
size and each subsystem will be studied separatedly.
We do not foster the idea of relying solely on pre-prepared contingen-
cy lists as the intent of our paper clearly shows. We feel that the DC
load flow model is able to produce the desired answers in acceptable
Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 11:11 from IEEE Xplore. Restrictions apply.
1559
time with present day technology. We are definitely open minded to
new approaches and are looking forward to study and possibly test the
discussers new method of automatic contingency selection. Selecting
meaningful contingencies automatically we regard as a very important
and useful tool in an energy control center environment and new more
efficient methods are, of course, welcome. At the same time, due to the
delicate nature of the real-time operating decisions, a foremost condi-
tion to be met by these methods is reliability and trustworthyness.
Manuscript received April 19, 1979.