You are on page 1of 13

97

IEEE Transactions on Power Apparatus and Systems, Vol.PAS-98,No.1 Jan/Feb 1979


AUTOMATIC CONTINGENCY SELECTION
G.C. Ejebe B.F. Wollenberg
Member IEEE Sr. Member IEEE
Power Technologies, Inc.
Schenectady, New York

ABSTRACT be studied by the on-line security analysis program


should be basedon the current operating conditions, and
A fast technique has been developed for the auto- should be adaptive and not a fixed list based upon off-
matic ranking and selection of contingency cases for a line studies.
power system contingency analysis study. A contingency
list is built containing line and generator outages This paper presents a methodology developed for
which are ranked according to their expected severity ranking transmission line outages and generator outages
as reflected in voltage level degradation and circuit according to the severity of their effects on bus voltage
overloads. An adaptive contingency processorcan be set or line flows. The method developed, utilizes Tellegen's
up by performing sequential contingency tests starting theorem [5] to generate the sensitivities of a system-
with the most severe contingencies at the top of the wide performance index with respect to the outages. The
list and proceeding down the list, stopping when the contingency ranking is accomplished by ordering these
severity goes below a threshold. Computational results normalized sensitivities fromthegreatest to the least.
of this technique applied to different test systems are The method uses the non-linear ac load flow equations
presented. for the evaluation of the effect of transmission and
generation outages on a voltage quality system-perfor-
INTRODUCTION mance index and uses the simple dc load flow model to
rank contingencies with respect to real power flows. As
In planning studies, the traditional approach for such, the method does not explicitly indicate whether a
steady-state contingency analysis is to test all con- contingency is going to give bus voltage or circuit
tingencies sequentially to evaluate system performance overload problems. Rather, it indicates the severity of
and reliability. This analysis consists of simulating each contingency relative to the others. Full ac load
outages of one ormore generating units and transmission flows must still be carried out--but onlyon the "pro-
facilities to investigate their effects on bus voltages blem" cases. Furthermore, what constitutes the set of
and line power flows. Fast computational techniques "problem" cases is easily determined by simply running
such as Stott's Fast Decoupled Load Flow [1], the lin- the load flow for each case starting at the top of the
earized ac load flow [2], etc. have been developed for list and stopping when the cases do not give problems.
contingency analysis and are well documented in a re- The cases below the stopping point need not be checked
cent survey [3]. However, exhaustive contingency test- since they are already ranked in severity below the ca-
ing including all conceivable contingencies, becomes ses at the stopping point. In this manner, the method
prohibitively long-running and costly when all primary constitutes an adaptive contingency processor since the
contingencies, each with additional levels of secondary number of cases solved will vary depending upon system
contingencies are involved. On the other hand, applying conditions. Such an adaptive contingency process is
contingency testing to a subset of contingency cases shown in Figure 1.
selected on the basis of the planner's experience and
intuition may be inadequate due to the possibility of
omitting some critical cases.
In the operational context, most control centers
are currently adapting load flow techniques for the
steady-state security analysis of contingencies to de-
termine whether the system is secure or insecure [4].
In the real time environment, it is hardly feasible to
perform exhaustive contingency testing to determine
how well the system, in its present state, can with-
stand these contingencies. At present, a contingency
list to be studied by the security analysis program is
created on the basis of operator experience and off-
line simulation studies. However, in real time, as
the system conditions change, the contingencies which
cause insecure operation may also change, and would be
different from those predicted by off-line simulation
studies. Therefore, the selection of contingencies to

F 78 228-9. A paper recannended and approved by


the Power System Engineering Ccwrittee of the
Power EngineerIng Society for presentation at the
IE.EE PES Winter Meting, New York, NY, January 29 - FIGURE 1
February 3, 1978. Manuscript submitted August 31,
1977; made available for printing November 11, 1977. ADAPTIVE CONTINGENCY PROCESSOR

0018-9510/79/0100-0097$00.75 © 1979 IEEE

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
98
SYSTEM PERFORMANCE INDICES synchronous condensers), the performance index PI v pro-
The standard approach for steady state contingency vides a good measure of the severity of abnormal volt-
checking is to run a load flow for the post-transient ages, as long as the generating units remain within
steady state condition following each outage. Some of their reactive power limits. However, it is possible to
the outages may result in system constraint violations encounter a contingency for which some generator reac-
such as load bus voltages outside their normal limits tive powers are driven to their limits. In this situa-
and transmission line (or transformer) overloads. tion, the standard ac 1 oad flow computes the bus voltage
using the limiting reactive powersatgenerator buses as
The voltage constraints at the load buses are usu- specified independent variables, and their voltages as
ally expressed in terms of a high and low limit. The dependent variables; as a consequence, there is a volt-
high limit is imposed by the maximum voltage value of age deviation from the scheduled voltage at these gen-
the system, and the low limit is a value below which it erator buses. Therefore, in order to reflect the reac-
is assumed that the load can no longer be supplied. tive power capability constraints of the generators in
Flows on transmission lines are usually constrained by the contingency selection for voltage analysis, we
thermal limits and sometimes by stability considerations define a generalized voltage-reactive power performance
on long lines. In light of these constraints, the sys- index by:
tem performance may be quantitatively evaluated in terms
of indices reflecting the severity ofout-of-limit volt- NB
PI i=l 2n1(
W4 V.1I
1 IV~pj1 )2n
-
age values or line overloads resulting froma particular
contingency. In defining the system performance indices, P'VQ AV;im
we treat the constraints on the load bus voltage and the
line flows as soft constraints, i.e., the violation of
these constraints, if not excessive may betolerated for
(2)
short periods of time. Therefore, the system performance
NG WQin- (Q 2n
index is defined as a penalty function to penalize se-
verely any violation of bus voltage constraints or line
+
(~RMax~
flow constraints.
where
System Performance Index for Voltage Analysis
Q = Reactive power produced at bus i
The voltage level performance index chosen toquan-
tify system deficiency due to out-of-limit bus voltages QMax
Q. = Reactive power production limit
is defined by
NG = Number of generating (reactive
NB WV |V. -IV- )2n production

i=l~~~~~~ (1) WQ.Q = Real non-negative weighting factor.

where The second summation, taken overall reactive pro-


duction units, penalizes any violations of the reactive
lvj = Voltage magnitude at bus i power constraints. The reactive power weighting factors
are set to zero if the effect of the reactive power de-
IVP1 I = Specified (rated) voltage magnitude
at bus i
ficit is not required.
It is perhaps important to emphasize here that the
AVLim
i = Vol tage devi ati on limi t, above whi ch contingency selection procedure developed in this paper
voltage deviations are unacceptable is not concerned with computing the system performance
index PIV (or PIVQ) defined above for each contingency--
n = Exponent of penalty function (n = 1 (this would necessitate performing a load flow for each
preferred) contingency). Rather, the procedure is primarily con-
NB Number of buses in the system cerned with computing the sensitivities of the index PIv
(or PIVQ) with respect to outages. As shown in the Ap-
WVVj = Real non-negative weighting factor. pendix, this is accomplished by only two load flow sol-
utions--one for the original network, and one for the
The voltage deviation AV1m represents the thres- adjoint network.
hold above which voltage level deviations are outside System Performance Index for Power Flow Analysis
their limit; any contingency load flow with voltage le-
vels outside this limit yields a high valueofthe index An index for quantifying the extent of line over-
Pi V On the other hand, when all the voltage level devi- loads may be defined in terms of a real power performance
ations from the rated (nominal) voltagearewithin AV im index:
the voltage performance index PIV is small. Thus, this WPI( )2n
PX
NL
index measures the severity of the out-of-l imit bus volt-
ages; and for a set of contingencies, this index pro-
t'W ITn
Pim = 2W P Lim~
1
(3)
vides a direct means of comparing the relative severity
of the different outages on the system voltage profile. where
It is pertinent to note, that sincethebus voltage P = The megawatt flow of line Q,
levels depend mainly on the reactive power flows, and (calculated by the dc load flow
therefore, on the reactive power production of the model)
generators (and reactive power production units, e.g.,

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
99
pLim
2. = The megawatt capacity of line Q lines to overload. However, computational experience
with these two methods shows that they do not always
NL = The number of lines in the system provide a proper selection of contingencies.
n = Specified exponent(n = 1 preferred) CREATING ORDERED CONTINGENCY LISTS
W = Real non-negative weighting coeffi- An ordered contingency list can be created for any
cient; may be used to reflect the of the three options below.
importance of some lines.
PERFORMANCE
The performance index PIMW contains all line flows OPTION INDEX OUTAGE TYPE
normalized by their limits. These normalized flows are
raised to an even power (by setting n = 1, 2...); thus, 1 | Piv or PIVQ Line and/or Generator
Outages
the use of absolute magnitude of flows is avoided. This
index PIMW has a small value, when all line flows are 2 Pi Line Outages
within their limits; and a high value when there are
line overloads. Thus, it provides a good measure of the 3______ PiMW Generator Outages
severity of line overloads for a given state of the
power system. CONTINGENCY LIST OPTIONS
Since the purpose of a contingency selection process One may wish to concentrate on only one option for
is to identify quickly (without performing a complete a particular application, or all three may be desired.
load flow or predictive analysis for each outage) those When more than one option is desired, the user simply
critical outages that need to be studied more rigorous- repeats the procedure shown in Figure 1, with the appro-
ly, the absolute value of the index PIMwforeach outage priate contingency ranking chosen to match the option
is not significant, and in fact, is not available. What desired.
is needed is the relative change of the MW performance The ordered list of contingency cases is based upon
index with respect to these outages. In a manner analo- the relative ranking of the sensitivities of the perfor-
gous to the ac case, these sensitivities are generated mance index to the outages. The sensitivities are the
from the results of two dc load flow solutions--one for first derivative of the performance index with respect
the original (base case) power system and the other for to a change in admittance for a line outage or with re-
the "adjoint" power system. spect to a change in net power output for a generator
Other Contingency Ranking Methods outage. As such, the sensitivities only give the incre-
mental change in performance index to an incremental
The use of fast and approximate algorithms has been change in line admittance or generator output. The full
suggested [3] as a way to prescreen from all probable effect of a contingency is then approximated by multi-
contingencies, those contingencies which are prone to plying the derivative by the full line admittance for a
result in unacceptable voltage and line overloads. The line outage and by the full net power being delivered
distribution factor method which is significantly faster for the generator to be outaged. For Option 1, line and
than any other technique, but less accurate, can be used generator outages may be ranked together or separately.
to prescreen contingencies for the more accurateac load Option 2 ranks line outages only while Option 3 ranks
flow analysis. However, the distribution factor method generator outages allowing for redispatch of the lost
does not provide voltage prediction and, hence, cannot generation.
be used to screen out contingencies to be studied for The computation of the derivatives is shown in the
voltage analysis. Appendix and involves the use of Tellegen's Theorem.
During the development of the method proposed in This theorem allows rapid computation of agradient vec-
this paper, we investigated the use of other contingen- tor containing the derivatives needed. Each derivative
cy ranking methods, particularly with respect to line is multiplied by the appropriate line admittance or gen-
overloads. One such method is to utilize the base case erator output. The resulting numbers, called normalized
loadflowsand rank outages of lines on the basis of sensitivities, represent the approximate API for each
their normalized MW flows. i.e., rank lines from the contingency, and are ordered from greatestto least (al-
most heavily loaded to the least heavily loaded; accord- gebraically). The performance index itself is always a
ing to: positive number, but the API's may be positive or nega-
tive indicating an increase or decrease in the perfor-
PO mance index for any particular outage. Many contingency
( Lim) (4) cases do not have any appreciable effect on the perfor-
mance index (small API), some may actually decrease it
(negative API). For this reason, we always order alge-
where braically--putting the outages with greatest positive
API at the top of the list and those with little or even
P09. = Base case megawatt flow in line Q negative effect at the bottom.
pLim = The megawatt capacity of line 2. The normalized sensitivities (API's) are only a
PI first order approximation of the change in performance
Another ranking scheme investigated is the use of index with a contingency and the performance index is a
the base case absolute magnitudes of line flows to gen- highly non-linear function of the flows and voltages in
erate a priority list of lines to be outaged; i.e., the network. Therefore, the ranking will be incorrect
rank lines according to actual MW carried on base case. whenever the linear approximation does not correctly
The assumption underlying these two methods is that the predict the true relative valueof the performance index
loss of a heavily loaded line (or a line with heavy MW for each contingency. Our results using the ranking al-
flow) would invariably result in a large increase in gorithm, shown in the next section, indicate that there
flow on other lines in this system, thus, causing some are often misorderings, especially among two or three

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
100
contingency cases which are close in the list. This is a) An 11-bus EHV backbone of the ITAIPU transmis-
shown graphically in the "effectiveness profiles." A sion system [6] shown in Figure 2
perfect ranking would produce a monotonically decreas-
ing profile which is rarely ever obtained. However, to b) A 29-bus system (which is a modified version
be used in the adaptive contingency processor, we only of the IEEE 30-bus system) shown in Figure 3.
need a profile which is generally downward sloping,
even if some "bumps" occur. This is possible because c) A 10-bus CIGRE test system [6].
the non-perfect ranking is accounted for in the stopp-
ing criteria. The 11-bus network is one of the five alternative
schemes designed to transfer power from generating fa-
NUMERICAL EXAMPLES cilities on the Parana River to the EHV receiving net-
work in South-Central Brazil over long 800 kV lines.
The ability of the contingency ranking technique Previous steady-state reliability studies [6] performed
to adequately order contingencies according to their on this system indicate that the system would tend to
expected severity was tested on a number of systems. exhibit voltage problems following the loss of critical
The results of these tests on three systems with dif- lines; the voltage problem severity was assessed in
ferent characteristics are presented to illustrate the terms of the power transfer reduction (curtailment) re-
effectiveness of the contingency ranking algorithm. The quired to avoid overloading of reactive production
systems used are: units. This system, endowed with synchronous condensers

11

FIGURE 2 - 11 -BUS SYSTEM

Line Numbers shown inside the Brackets.

FIGURE 3 - 29-BUS SYSTEM

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
101
and reactors, was therefore chosen as a good system on 10.0
which the automatic selection of contingencies with re-
spect to the voltage performance index can beeffective- or
ly tested. (Option 1 with line outages only.)
c

The 29-bus system is a modified version of the IEEE ply 1 1V11 - yP2n
30-bus system and was used to check the ranking of line Exponent n = 1
outage contingencies for line overloads and out-of-
limit voltages. (Options 1 and 2 with line outages.) 11-Bus System
The 10-bus CIGRE system [7] with seven generating co
plants was used to test the ranking of generator outages
for line overloads and out-of-limit voltages. (Options
1 and 3 with generator outages.) I.1

The performance of the contingency ranking tech-


nique is checked against the results of ac and dc load
flows. First, the contingency ranking algorithm is used
to generate a list of ordered line or generator contin-
gencies based on the computation of the normalixed sen- 10 12
sitivities of the appropriate performance index. For 2
LINE
4
OUTAGE
6 8
CONTINGENCIES ORDERED BY RANKING ALGORITHM
voltage analysis, an ac load flow is performed for each
line or generator outage to compute the true value of
the voltage performance index PIv (or PIVQ). For the FIGURE 4 - EFFECTIVENESS PROFILE FOR VOLTAGE
PERFORMANCE INDEX
real power flow performance index, the simpler dc load
flow is solved for each outage to compute PIMW. The va-
lues of the appropriate index for all the outages are TABLE 2: A COMPARISON OF VOLTAGES AND VOLTAGE INDICES FOR WORST
then ordered according to their magnitudes toobtain the THREE OUTAGES ON THE 11-BUS SYSTEM
true priority list which is compared against the list OUTAGE OF
predicted by the contingency ranking algorithm.
BUS BASE CASE LINE 7 LINE 8 LINE 9
Table 1 shows a comparison of the ranking of line VOLTAGES
contingencies on the 11-bus system, according to the 1 .9950 0.9950 0.9950 0.9950
voltage problem severity. The effectiveness of the con- 2 1.0000 1.0000 1.0000 1.0000
3 0.9807 0.9693 0.9875 0.9810
TABLE 1: A COMPARISON OF AC LOAD FLOW AND CONTINGENCY RANKING ALGORITHM 4 0.9900 0.9346 0.9547 0.9398
FOR THE VOLTAGE INDEX ON THE 11-BUS SYSTEM
5 0.9517 *0.9021 *0.9338 0.9669
LINE OUTAGE RANKING LINE OUTAGE RANKING BY 0.9380
6 0.9469 *0.9354 *0.9295
BY AC LOAD FLOW CONTINGENCY SELECTOR 7 0.9443 *0.9199 *0.9392 *0.9291
ORDERED VOLTAGE WORST PERCENT ORDERED NORMALIZED 0.9700 0.9655 0.9700
LINE PERFORMANCE OF*OUT-OF- LINE SENSITIVITY 8 0.9700
NUMBERS INDEX LIMIT VOLTAGE NUMBERS (API) 9 0.9657 0.9665 0.9700 0.9700
7 1.9697 1.24 7 0.2676 0.9782 .9782 0.9792
10 0.9778
8 1.4341 0.97 8 0.2475 0.9900 0.9900
11 0.9900 0.9900
9 1.127 0.93 9 0.1784
VOLTAGE
5 0.9878 0.78 5 0.1445 INDEX 1.9697 1.4341 1.127

0.8073 0.72 6 0.0659 out of Limit


4 -

6 0.6182 0.64 12 0.0364


12 0.4861 0.67 11 0.0322
In ranking contingencies from the standpoint of
11 0.4797 0.64 10 0.0314
0.4654 0.67 4 0.0236
voltage quality, the maximum voltage deviation AVm was
10
3 0.4374 0.60 15 0.0022 set at 0.075 p.u. for all buses; thus, any voltage drop
(or rise) which is more than 0.075 p.u. from normal
2 0.4310 0.60 13 0.0002
0.4273 0.61 1 -O.2504E-5
(rated) voltage VSP
1 is unacceptable and is consequently
13
15 0.4271 0.60 2 -0.1295E-4 penalized severelyin the voltage performance index. The
choice of + 7.5 % voltage threshold is cons i dered adequate
14 0.4252 0.60 3 -0.2171E-4 and consonant with normal operating guidelines of + 5 %
1 0.4198 0.59 14 -0.2101E-4 - + 10 %; although the threshold value is not critical
since the voltage constraints are treated as soft con-
strai nts.
tingency ranking algorithm in the example is shown gra- The ranking of line contingencies according to their
phically in Figure 4 which shows the values of the volt-- projected severity on line overloads is showninTable 3
age performance index (computed using the ac load flow) for the 29-bus system. Figure 5 shows the effectiveness
plotted for the corresponding contingencies as ordered profile of the contingency ranking algorithm with respect
by the contingency ranking algorithm. to the real power flow performance index PIMw. Values
Table 2 shows a summary of the bus voltage magni- of this index computed using the results of the dc load
tudes for the worst three line contingencies of the 11- flow are plotted for the corresponding contingencies as
bus system. ordered by the contingency ranking scheme.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
102
TABLE 3: A COMPARISON OF RANKING BY DC LOAD FLOW AND CONTINGENCY TABLE 4: A SUMMARY OF GENERATOR OUTAGES ON THE 10-BUS
RANKING ALGORITHM FOR'LINE,CONTINGENCIES ON THE 29-BUS SYSTEM CIGJE SYSTEM

RANKING BY DC LOAD FLOW RANKING BY


___ICONTINGENCY SELECTOR
RANKING FOR VOLTAGE ANALYSIS
ORDERED OVER- ORDERED NORMALIZED
LINE INDEX PI WORST LOAD LINE SENSITIVITY
NUMBERES MW LINE % NUMBERS (API) AC LOAD FLOW CONTINGENCY SELECTOR
26 0.48999E+00 25 1.17 25 0.19516E+00 ORDERED VOLTAGE ORDERED NORMALIZED WORST
11 0.35401E+00 25 0.94 24 0.11700E+00
24 0. 33118tE+oo 25 0.91 26 0.89359,-01 GENERJATOR INDEX ENERATOR SENSITIVITY BUS
25 0.32813E+00 26 1.08 23 0.65127E-01 NUMBERS PiV NUMBERS APIv VOLTAGE
23 0.31317E+00 25 O.92 11 0.45630E-01
22 00.25857L+00 25 0.90 29 0.41855E-01
34 0.24408E+00 25 0.94 13 0.34547E-01 3 O.9543E01 3 0.4832 0.818
13 0. 22660E+00 24- 0.92 22 0.28970E-01
19 0.16821E+00 25 O.86 34 0.2740tE-01 5 0.9215EO1 5 0.1849 0.834
29 0.16142t+00 24 0.87 39 0.36951E-02
7 0,1608HE+00 13 0.81 9 0.35319E-02 6 O.6912EO1 6 0.1383 0.886
5 0.14988E+00 V 0.80 2 0.33801E-02
9 0.14478E+00 5 0.80 16 0.32214E-02
21 0.13142E+00 25 0.84 3 0.29135E-02 7 O.3136EO1 4 0.1165 0.965
17 0.13139E+00 25 0.84 5 0.22260E-02
39 0.11183E+00 25 0.83 8 0.16378E-02 4 O.3010EO1 2 0.0065 0.970
2 0.11002E+00 25 0.80 21 0.15351E-02
1 0.1095SE+00 25 0.79 10 0.14508E-02 2 O.1373EO1 7 -0.3059 0.983
8 0.10849E+00 25 0.80 15 0.12976E-02
4 0.10770E+00 25 0.80 20, 0.12011E-02 1 *
20 0.10749E+00 25 0.82 4 0.72247E-03
10 O.10545E+00 25 0.82 1 0.55163E-03
16 0i10511E+00 25 0.82 19 0.49377E-03 * Swing bus generator excZuded frzom voZtage
3 0.99940E-01 25 0.80 36 0.19867E-03
15 0.97954E-01 25 0.80 35 0.13463E-03 problem ranking
36 0.97119E-01 25 0.80 37 0.23515E-04
35 0.96752E-01 25 0.80 32 0.38378E-05
37 0.95840E-O1 2S 0.80 14 O.OOOOOE+00
28 0.95722E-01 25 0.80 12 -0.12383E-12
12 0,95401E-01 25 0.80 18 -0.36667E-03 RANKING FOR LINE OVERLOADS
18 0.95139E-01 25 0.80 28 -0.40012E-03
38 094882E-01 25 0.80 38 -0.49860E-03
6 0.948'_5E-01 25 0.79 17 -0.20753E-02 DC LOAD FLOW CONTINGENCY SELECTOR
14 0.93421E-O1 25 0.80 6 -0.35195E-02
30 0. 89522't-01 25 0.75 33 -0.43336E-02 ORDERED INDEX ORDERED NORMALIZED
32 0.84051E-01 25 0.77 30 -0.62557E-02 GENERATOR PIMW GENERATOR SENSITIVITY
33 0.78866E-01 25 0.76 7 --0.787848--02 NUMBERS NUMBERS APIMW
31 0.69493E-01 24 0.72 27 -(0.13098E-02
27 0.66651E-01 24 0.71 31 -0.14140E-01
3 1.6932 3 0.4699
7 0.7985 4 0.1683
a

U. 1. 4 0.6589 6 0.1442
C~ NL 1 (IpI )2n
PiW = X -( 6 0.6157 7 0.1418
w
mw
=91 2n PI im

>1 5 0.4818 5 0.0386


co
Exponent n 3
C 1 0.3188 1 -0.3328
L> 29-Bus System .2 0.1935 2 -0.9597
:i
ui
C,

_'
w,§O .

oE
list until no limit violations were found on several ca-
ses in succession. In the case of the study shown in Ta-
5 10 15 20 ble 3, a stopping criterion of three "no violation" ca-
LINE OUTAGE CONTINGENCIES ORDERED BY RANKING ALGORITHM ses in succession would have been adequate. The user
would be required to set the number of no violation ca-
FIGURE 5 - EFFECTIVENESS PROFILE FOR REAL POWER ses to be used in the stopping criterion, and this will
PERFORMANCE INDEX require experience gained with using the technique on
the user's system. A margin of error can also be added
Table 4 shows the ranking of generator outages on to the number of no violation cases if desired.
the 10-bus CIGRE system, with respect to both the volt-
age problem severity and line overloads. Only complete Finally, a different stopping criterion altogether
outages of generating plants were simulated. could be used by simply asking the question: "How many
cases should be run to be sure to cover the worst N ca-
STOPPING CRITERIA FOR THE ADAPTIVE CONTINGENCY ses?", where N would typically be between 1 and 20. The
PROCESSOR stopping criterion, which must also be built by experi-
ence, would simply give a corresponding N' for each
The simplest stopping criteria foradaptive contin- value of N. For instance, by always running ten cases
gency processor would exit as soon as the load flow for (N' = 10) off the top of the list, one might be reason-
one case showed no out-of-limit conditions. This would ably assured of getting the worst five (N = 5).
work in the case of the line outage contingencies shown
in Table 1 where the first case results in a voltage be- Both of the above criteria are used in the bulk
low low limit and subsequent cases show no out-of-limit power adequacy assessment program reported in reference
voltages. However, the simple single case stopping rule [8]. In this program, the contingency ranking algorithm
will not work forthe line outage contingencies shown in is used to rank both primary and secondary outages. The
Table 3, where the first and fourth cases have line over- number of primary outages is fixed by user input, and
loads, and the second, third, and remaining cases do the number of secondary outages is determined by using
not. the adaptive process with a default number of no viola-
tion cases set to five.
A better stopping criterion would require the adap-
tive processor to continue to execute load flows for the

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
103
CONCLUSIONS tion and Transmission of Electric Energy, Rio de
Janeiro, Brazil, September 1977.
The automatic contingency selection algorithm pre-
sented in this paper represents a refinement of contin- [7] F. Ariatti, et.al., "Optimization Studies for the
gency analysis methodology which will increase the ef- Operation of Power Systems," 1972 CIGRE Conference
fectiveness of existing techniques. The fact that the Proceedings, Paper 32-12, 1972.
algorithm can provide an ordered list of contingencies
extends the capability of a planning study tofind those [8] P.L. Dandeno, G.E. Jorgensen, W.R. Puntel, R.J.
combinations of equipment outages which will cause fu- Ringlee, "A Program for Composite Bulk Power Elec-
ture trouble on a system. This is accomplished by cre- tric System Adequacy Assessment," IEE International
ating a list for all primary contingencies and then Conference on Reliability of Power Supply Systems,
subsequent lists for secondary contingencies. In real London, February 1977.
time systems, use of the contingency ranking algorithm
will enable system operators to locate the areas of
weakness under the present operating conditions. It is APPENDIX
even possible to allow a real time system to study pri-
mary and secondary outages so as to be better prepared CALCULATION OF NORMALIZED SENSITIVITIES
in the event of a primary contingency actually happen-
ing. The ranking algorithm is based upon ordering of
the normalized sensitivities which give the first or-
The ranking algorithm does not give a perfect order- der change in performance index to a change in line ad-
ing and requires the user to tune the adaptive stopping mittance or generator output. The normalized sensitiv-
criterion to balance the overall execution time with the ities can then be expressed as:
desired margin of error to assure that all possible
troublesome combinations are covered. API or S A I (A.1)
St = Ay;- Ay9. oSg AG G g
The contingency ranking algorithm has thus far been
applied to single outage contingency cases, especially and the total change in performance index becomes:
with respect to voltage problem analysis. Further work
is continuing to investigate the use of composite sen- (A. 2)
sitivities for ranking common-mode multiple outage con- APITotal = Z
9. SQ + g
g 9
tingencies; suchasthat involving multi-terminal lines,
common-mode line and generator outages. Tellegen's Theorem states a relationship between
the voltages and currents of two electrical networks,
ACKNOWLEDGEMENTS a base network and an adjoint network. The base and
adjoint networks may differ in admittance, voltages
The authors would like to thank R.J. Ringlee of and currents but must have the same topological struc-
Power Technologies, Inc. for his valuable discussions ture. Tellegen's Theorem can also be used to state
and suggestions towards the development ofthe presented relationships between perturbations in either of the
algorithm. two networks, and reduces to the following form:
The work was supported by the National Science (A.3)
9. (AVkIk
i - AI9V9) = 0
Foundation under grant AER74-19588.
REFERENCES Q over all node to node and node to ground
branches.
[1] B. Stott and 0. Alsac, "Fast Decoupled Load Flow,"
IEEE Transactions on Power Apparatus and Systems, This equation gives a basic relationship between
Vol. PAS-93, pp. 859-869, May 1974. the voltage and current shifts AV9 and AI in the
[2] N.M. Peterson, W.F. Tinney and D.W. Bree, "Itera- original network and the voltages and currents V and
tive Linear AC Power Flow Solution for Fast Ap- I in the adjoint network. Equation (A.3) is basic to
proximate Outage Studies," IEEE Transactions on
Power Apparatus and Systems, Vol. PAS-91, p. 2048, all of contingency ranking options described in the
Sept./Oct. 1972. paper. The voltage and current shifts are related to
contingency effects while the adjoint voltages and
[3] A.S. Debs and A.R. Benson, "Security Assessment of currents are related to derivatives of the performance
Power Systems," Proceedings Engineering Foundation index function.
Conference onSystems Engineering for Power: Status
and Prospects, Henniker, N.H., August 1975, pp. 144- Sensitivity of PIv and PIVQ
176.
We can define the performance index by equation
[4] T.E. Dy Liacco, "System Control Center Design," Pro- (1) or (2) which is of the following form:
ceedings Engineering Foundation Conference on Sys-
tems Engineering for Power: Status and Prospects, NB
Henniker, N.H., August 1975, pp. 196-229. (A.4)
PI = X
i=l
f (V ,Qi)
[5] S.W. Director and R.A. Rohrer, "The Generalized Ad-
joint Network Sensitivities," IEEE Transactions on and the total change in the Performance Index becomes:
Circuit Theory, Vol. 16, No. 3, pp. 318-323, August
1968.
af. af.
[6] B.P. Lam, N.D. Reppen and R.J. Ringlee, "Reliability API = i1 Re [aV;
v 1 AVji
+ 1AQ ]
yAQi
(A.5)
Criteria and Predictions for the Transmission Sys-
tem of Itaipu," 4th National Conference on Produc-

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
104
where it is necessary to take the real part since the Then (A.9) reduces to:
voltages are complex. The adjoint network has the
same admittance matrix as the base network. The ref-
erence bus voltage is set to zero. Relationships are
API = I
21
Pz
9 ABP (A. 1 1)
defined between the base and adjoint voltages and cur-
rents to model the generator MW output and voltage and this equation also corresponds to equation (A.2)
magnitude and load MW and MVAR in the same way as a for the line outage case.
load flow. On the basis of these relationships equa-
tion (A.5) becomes (for line outages): Ranking Generator Outages
API = - X Re{V
e
V Ay } (A.6) The ranking of generator contingencies according
to the severity of line overloads or out-of-limit volt-
ages is accomplished by the same basic approach out-
Equation (A.6) corresponds directly to equation lined. However, the computation for generator outages
(A.2) with the exception of generator outages which are requires the application of a generation redistribution
dealt with at the end of the Appendix. function to allocate lost generation to the remaining
generators. Different redistribution rules have been
Sensitivity of PIMW suggested [3]; such as (a) allowing the slack generator
to absorb all lost generation, (b) redistribution of
The sensitivity of the line powerflowperformance lost generation among the remaining generators in pro-
index is based upon the dc load flow model giving line portion to their normalized inertia, (c) redistribution
on the basis of spare capacity, and (d) redistribution
power flow,,P, as: based on economic dispatch criterion.
P= BP(9i-G ) (A.7) In this work, the generation redistribution rule
(b) is used, and it is assumed that the rotating iner-
where tia of a generating unit is proportional to its maxi-
mum power output.
B = the line susceptance of line Q
Two types of generator outages may usually be con-
Gi2,j = the phase angles at buses i and j. sidered, partial loss of generating units or a com-
plete shutdown of a generating plant. In order to re-
The performance index, Equation (3), can then be flect more severe effects, especially from the view-
stated as: point of bus voltages, only complete loss of generat-
ing plants has been considered.
NL
PI = I hz(Gz) (A.8) In contingency analysis, the solution of an ac
jz= 1 load flow for a complete shutdown of a generating
plant treats the voltage magnitude at the generating
and the total change in performance index becomes: bus as load bus voltage which is free to vary. Here
the ranking of generator outages with respect to volt-
NL ahz age problems is achieved by computing and ranking the
API = Xa AO normalized composite sensitivities of the voltage per-
formance index with respect to the incremental changes
of generator transformer - connection admittances, and
The following substitutions are made to equation incremental changes in the power generation.
(A.3):
V e-9- Phase angle across line Q
I ÷P Power flow through line Q 2
2
resulting in: The drawing above shows a generator transformer
(A. 9) connection where bus 2 is the generator bus specified
I (AGzPz - APz@z) = 0 in the load flow, by a PV model. The complete loss of
generation at bus 1 is modeled by the outage of the con-
nor+;inn rAncfnrmar
rlClc Liriy fLr-curib ur-111r .
The adjoint network used in the real power case
has the same line susceptances and the phase angle at The ranking of generator outages with respect to
the reference bus is set to zero. The adjoint injec- line overloads requires the evaluation of the normal-
tions are: ized sensitivities of MW performance index with respect
ah to these outages. This is a real power flow analysis
for which the dc load flow model is used. The genera-
Pi I O (A.10) tion redistribution function proportional to the nor-
Q over malized inertia of the remaining plants is applied to
all lines obtain the changes in the remaining generators. The
connected incremental change in the MW performance index for
to bus i each generator outage is computed using the adjoint
network to account for the composite effects of the
incremental changes in generation.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
105
Discussion tion, at no additional computational cost. Practical implementation of
this algorithm on an interactive graphical minicomputer system proved
H. B. Puttgen and R. L. Sullivan (University of Florida, Gainesville, that the proposed software requires only little extra memory when com-
FL) H. B. Puttgen now with California State University, Fresno): An pared with the AC load-flow system and when programmed using the
interesting application of Tellegen's theorem to the problem of power overlay technique.
systems analysis is offered in this paper. As very closely related work (1) In conclusion we believe that the "Automatic Contingency Sec-
has been carried out at the University of Florida, there are some major tion" method presents an interesting but rather restricted application of
aspects of this method that we would like to discuss. Tellegen's theorem to power system's analysis. A more general ap-
The algorithm presented is aimed at obtaining a hierarchical proach was taken using the complex sensitivity coefficients approach as
classification of successive first order contingencies in a power system. outlined in the forthcoming IEEE paper and in reference (1).
It is important to note that only a classification of the relative impor-
tance of the contingencies considered is obtained. No evaluation of the REFERENCES
system's behavior under the particular contingencies is offered. Thus,
for each contingency considered an AC load-flow calculation is still re- (1) H. B. Puttgen "A User Oriented Method for Transmission System
quired to compute the system's actual state. Indeed, it is perfectly possi- Planning Using Interactive Graphics", Ph.D. Dissertation,
ble that the contingency identified as being the most important, is far as University of Florida, June 1976.
the impact on the system's voltages and line flows is concerned, is (2) S. W. Director and R. A. Rohrer "The Generalized Adjoint Net-
perfectly acceptable when the results of the corresponding AC load- work and Network Sensitivities", IEEE Trans. Circuit Theory,
flow calculations are compiled. The method proposed only provides an CT 16, pp. 318-323, 1969.
educated guess as to which first order contingencies should be analysed (3) R. Rischland and W. R. Puntel "Computer Aided Design of Elec-
first when screening the most damaging states of the system. tric Power Transmission Networks", IEEE paper C72 168-8, 1972.
A performance index is used to quantify the relative importance of (4) H. B. Puttgen "Le Calcul des Parametres de Sensitivitie d'un
each individual contingency considered. This performance index is a Reseau de Transport d' Energie Electrique" Bull. ASE/UCS 68,
quadratic penalty function taking mainly bus voltage magnitudes and Feb. 1977, p. 196-199.
line real power flows into consideration. The bus voltage phase angles
are ignored as are the magnitudes of the line current flows which are the Manuscript received March 22, 1978.
true line flow constraints. The paper states that for the line real power
flow constraints, the linearized power flow equations, P = B 6, are
used. The paper does not say whether the voltage magnitude constraints M. S. Sachdev and S. A. Ibrahim (University of Saskatchewan, Saska-
are handled using a linearized model of the system's equations or if a
toon, Sask., Canada): The authors are to be complimented for present-
non-linear model was retained. We would like to have some further in- ing an interesting approach for automatic selection of serious con-
formation concerning the handling of the voltage magnitude constraints tingencies. Tellegen's Theorem has been used in this approach to
from a computational point of view. generate the sensitivities of a performance index with respect to the con-
The proposed algorithm could not be used in the synthesis phase of tingencies. The contingencies are ranked by arranging the sensitivities in
power systems planning as the possible impacts of nonexisting branches
a descending order.
on the system state cannot be evaluated by the method as outlined in the
Will the authors comment on the following remarks and questions?
paper.
1. In equation (1), the authors have used the weighting factors W,,
Closely related work has been performed at the University of divided by 2n. Is it possible to choose the weighting factors in such a
Florida since 1974. The basic idea was to attempt to use the method of manner that the division by 2n is avoided? What will be the index if
"Computer Aided Design" primarily developed for electronic circuits IVi axi - IVisPi does not equal IViVPI IVimini?
-

and as outlined for example in (2). A first attempt to use Tellegen's 2. We agree that ranking of line outages according to the base case
theorem and CAD techniques for power systems planning was carried MW or normalized MW flows, will not be the best approach. But we do
out by Fischl and Puntel (3), but using the linearized power flow model, not agree that real power flow is a good measure for a line being
P = B 6. Our work was centered around avoiding the linear model and overloaded, only MVA or current flows provide proper information on
using the full non-linear AC load-flow equations including the non- overloading.
linear power sources, S = VI*. A first set of results of this research are 3. The authors have stated that the exponent n of penalty function
contained in (1) and are briefly summarized in (4). A paper concerning should perferrably be one; but the value of n used in the study of Figure
this work was submitted to IEEE/PES for publication at the 1976 sum- 5 is 3. How was this value chosen?
mer meeting, and is still being processed by IEEE.
4. To make the contingency selector fully adaptive, the stopping
criteria needs to be investigated to eliminate the operator experience
A major result of this research is the proof of the fact that one can factor. Have the author's investigated this aspect?
indeed numerically determine a set of complex sensitivity coefficients 5. The contingencies are ranked according to the sensitivity of the
relating the change of any complex voltage (in magnitude and phase) performance index to the changes introduced by these contingencies. It
with the change in admittance (in real and imaginary parts) of any ex- NB
isting or fictitious branch. The technique used in determining these sen- NB
seems that the sensitivity of the function E W, V,2, or E: W,jVj, is also
sitivity coefficients is a major extension to the concept of adjoint net- i=l i=l
work. Indeed, in order to accomodate the non-linearity of the AC load- likely to give the same ranking but with considerably less computing ef-
flow equations (which are due to the non-linear power sources) we used fort.
two different adjoint networks, having the same topology but different We, once again, congratulate the authors for a valuable contribu-
branch contents. It was further shown that these two adjoint networks tion to the adaptive selection of power system contingencies.
can be simultaneously solved using only one forward-backward
substitution on a "super sparse" set of linear equations for one set of Manuscript received March 9, 1978.
sensitivity coefficients (sensitivity of one voltage with respect to any ex-
isting or fictitious branch admittance).
Using these sensitivity coefficients one can greatly simplify the en-
tire process of transmission planning. The sensitivity coefficients are R. Fischl (Drexel University, Philadelphia, PA): This paper presents a
not only used to quantify the relative importance of contingencies but new algorithm for quickly identifying the set of critical contingencies
also, and more important, to actually evaluate the state of the system which must be considered in steady-state security assessment, in par-
under these contingencies. We arrive at results that closely correlate ticular the problem of identifying those contingencies which are likely
those obtained using the classical AC load-flow method, with dif- to create voltage problem. The effectiveness of the algorithm depends
ferences of only 1-5% on voltage magnitudes and line current on the choice of the performance index (PI), and the efficiency of the
magnitudes. Yet only one AC load-flow is performed on the unperturb- algorithm is based on the use of the adjoint method for computing the
ed system, all the contingencies are then computed using simple sensitivities of PI with respect to the lines admittances or generator
forward-backward substitutions. The sensitivity coefficients are furtherpowers. My comments and questions are concerned with the choice of
the PI's and the adjoint method for computing sensitivities described in
used in the synthesis stage to guide the planning as one can evaluate the
sensitivity of any complex bus voltage (and thus of any complex line the Appendix.
flow) with respect to any fictitious admittance and thereby truely pro- In order for the proposed contingency selection algorithm to be ef-
vide valuable synthesis information, in addition to analysis informa- fective, it is not sufficient to select the same performance indices as

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
106
those used in transmission expansion or security-constrained dispatch REFERENCES
problems. The performance indices must satisfy additional criteria to
those mentioned in the paper, namely: [9] R. Fischl and W. R. Puntel "Efficient Method for Computing
(1) The PI must reflect the severity of the outage, i.e. it must be a linear Electric Power Transmission Network Sensitivities", Paper No.
(or monotonically increasing) function of the change in parameters, so C72-167-0, IEEE Winter Power Meeting, New York, N.Y., Jan
that the post-outage performance index can be approximated by the 1972.
first order Taylor-series approximation, namely [10] R. Fischl and W. R. Puntel "Computer-Aided Design of Electric
PI' -PIo+ 7- SI + Sg(1) Power Transmission Network", Paper No. C72-168-8, IEEE
where PI" is the preoutage value of the performance index and SI and S, Winter Power Meeting, New York, N.Y., 1972.
are defined in Eq. (A. 1). How good such an approximation is, depends [11] H. K. Kesavan and B. H. Bhat, "Sensitivity Analysis of Load
on the choice of PI. If higher order terms in the Taylor-series expansion Flow Solutions", Paper No. C73-465-2, IEEE Summer Power
are significant, then the effectiveness of using SI (or S5) to predict the Meeting, Vancouver, Can., July 1973.
severity of an outage is poor. Our experience with the authors' index [12] H. P. Puttgen and R. L. Sullivan, "A Novel Comprehensive Ap-
PIMW has been that it is a highly nonlinear function of the branch, ad- proach to Power Systems Sensitivity Analysis", submitted to the
mittances, yi, so that I was not surprised to find the poor correlation in 1978 IEEE Summer Power Meeting, July 1978.
the results shown in Fig. 5. [131 F. F. Wu "A Twin Nonlinear Resistive Circuit Model for Power
(2) There should be a minimum value of the normalized sensitivities, System Steady-State Analysis, 1977 Allerton Conference on Cir-
API"i", for which one can guarantee that for all APIS < API"in, "no- cuits and System Theory, Monticello, Ill., 1977.
violation" cases exist. [14] B. Stott and 0. Alsac, "Fast Decoupled Load Flow", IEEE
With reference to the above comments, I would like the authors to Trans. of Power Apparatus and Systems, Vol. PAS-93, May/June
comment on their motivation for selecting the PI's. Specifically, the 1974, pp. 859-69.
choice of the weighting factors Wi = 1 and exponent n = 1 for PI,, and [15] R. Fischl and R. G. Wasley, "Efficient Method for Power and
exponent n = 3 for PIMW in the examples given. It seems to me that Voltage Sensitivities to Power Network Parameter Changes" sub-
both the weighting factors W1 and exponent n should be chosen so as to mitted to the 1979 IEEE Winter Power Meeting.
achieve the linear approximation given in Eq. (1) above. The choice of [161 R. Fischl, G. C. Ejebe and J. A. DeMaio "Identification of Power
these constants must take into account the system structure and so their System Steady-State Security Regions Under Load Uncertainty",
values must be tuned to the system under study. Paper No. A76-495-2, IEEE Summer Power Meeting, Portland,
Although the authors mention that they use an adjoint method for Ore., July 1976.
computing the sensitivities of PI, and PIQ, the description given in the
Appendix leaves much to question. This discussor has no questions con- Manuscript received February 27, 1978.
cerning the computation of the PIMW sensitivities since we have
developed it, and its derivation is well documented [9, 101. The area of
interest is, how to set up the adjoint network for the AC load flow Kavuru A. Ramarao (Cleveland Electric Iluminating Co., Cleveland,
model used in the computation of API, and APIQ. The application of OH): We congratulate the authors for their timely paper.
the Tellegen's theorem to this problem is not a straight forward applica- The objective of Security Analysis is to study every possible "next
tion of the results of Ref. [5] as the authors state in Eq. (A.6). The contingency" and determine any abnormal conditions. Such a predeter-
reason for this is that although Tellegen's theorem can be applied mined contingency set should be exhaustive and the adaptive contingen-
directly to nonlinear electronic networks with independent voltage and cy processor should adapt to the prevailing operating conditions and
current sources, the nonlinearities encountered in power networks are select the contingencies (without any exception) which cause the system
not due to the nonlinearities in the network elements (as the case is in insecure.
electronic circuit analysis in which the elements are nonlinear functions This paper is a step in the right direction.
of voltages or currents), but are due to the nonlinearities of inputs The contingency ranking algorithm is based upon the relative rank-
which are independent nonlinear sources involving real and reactive ing of the sensitivities of the performance index (PIMW) to the line
powers. Therefore, in order to develop an adjoint network for com- outages.
puting the AC load flow model sensitivities in accordance to the method We have the following comments in the use of the simplified D.C.
described in Ref. [5], it is necessary to approximate the independent real model and PIMW.
and reactive sources by either impedances, or independent current or * The contingency ranking algorithm appears satisfactory if used in
voltages sources as suggested in Ref. [11, 12]. There are, however, conjunction with the Security Analysis by the method of distribution
several problems associated with this approach. First, the adjoint net- factors.
work, so defined, represents only an approximation to the actual ad- * This technique may not be satisfactory if used with A.C. Security
joint since the PV and PQ bus constraints may not be satisfied under Analysis. This algorithm may not detect contingencies which may cause
network pertubations, and this gives the wrong value of Vj used in Eq. overloads due to predominant MVAR flow component in a line.
(A.6). Second, the solution of the resulting adjoint network requires the If the performance index is treated as a function of either MVA or
solution of a new system of equations (representing the adjoint net- current magnitude or both real and reactive power and the problem
work). Another approach for obtaining the adjoint network is to use a reformulated, we may obtain more accurate contingency ranking.
twin nonlinear resistive circuit model [13]. The problem with this ap- We are looking forward to authors' further work on the contingen-
proach is that its accuracy depends on the amount of coupling that ex- cy ranking involving multi-terminal lines which form a major portion in
ists between the real and reactive power flows in the network, so that any realistic contingency set.
when the coupling is low (i.e. when the assumptions used in defining the We appreciate the authors' comments on the above.
decoupled AC load flow [141 hold), then the results are accurate. In
summary, we would like the authors to expand on their adjoint method Manuscript received February 27, 1978.
for computing sensitivities, since we have developed such a method
[15], which not only corrects for some of the problems discussed above,
but also uses the Jacobian to compute VI and 01. L. S. VanSlyck, G. Irisarri and A. M. Sasson (American Electric Power
In conclusion, althouZh the algorithm presented in this paper is ex- Service Corporation, New York, NY): The automatic contingency
tremely useful, further work is needed to determine a PI which selection method presented by the authors can be applied to both off-
guarantees that the stopping criterion used will identify all the con- line planning studies and on-line security assessment as well. Our ques-
tingency cases which may give out-of-limit conditions while guarantee- tions and remarks, however, are addressed to the latter type of applica-
ing that those case eliminated from the set will give "no-violations". As tion.
a last comment, I would like to know if the authors considered using a Present on-line contingency evaluation programs used in energy
steady-state security region identification algorithm (sch as that describ- control centers run the various contingencies by either: selecting the
ed in Ref. [16]) under Option 2 and 3 of their contingency selection outages from a pre-prepared list of contingencies, or, by operator selec-
algorithm, rather than using APIMW. The advantage of such an iden- tion depending on current system conditions as given by a state
tification algorithm is that it not only identifies all the contingencies estimator or any other power system monitoring facility. The former
which give out-of-limits conditions but also gives a redispatch strategy approach is an off-line planning practice applied to the on-line situa-
in order to relieve this condition if such a strategy exists. Have the tion. This method fails to recognize that the system conditions assurned
authors implement an automatic redispatch strategy when taking line for the preparation of the contingency lists may be far from the real
outage contingencies, and if so what kind? time conditions. In fact, a single contingency on a planning study may

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
107
very well be a double or triple contingency in a real time situation. The reliability-critical cases should never be mis-ranked badly enough to
latter method-i.e. operator selection-may prove to be burdensome escape attention. This depends on two factors:
even to experienced operators. In fact, as operating conditions get more (i) the analytic performance-index function must reliably
and more stringent, the number of contingency choices gets larger and discriminate critical from non-critical operating conditions, and
larger and no amount of experience is enough for selecting troublesome (ii) the estimated changes in the index should be accurate, or at
contingent situations. The method of the authors combines common least they should all have consistent errors, in order to rank the cases
sense (i.e. adequate performance indices) with sound mathematical correctly.
tools (i.e. Tellegen's theorem) and real-time system conditions (i.e. a In relation to point (i), it appears that it will be necessary to use
real-time data base). It will provide the operator with a small set of im- higher-order penalty function exponents, to weight violations heavily
portant and realistic contingencies to be tested with a load-flow pro- enough. To illustrate, with exponent n = 1, a non-critical case with ten
gram. lines each loaded to 70% of rating is ranked the same as a very critical
We would like to ask the authors how are they proposing to use the case where nine of the lines are lightly loaded and one is overloaded by
method in a real-time environment. Should the contingency selection 120%/. However, with n = 2, the index for the latter case hecomes ten
program be executed every time a new state estimation solution is ob- times that for the former, which seems more reasonable.
tained in the control center? Or is it better perhaps to design an "on-line The frustrating aspect of the method is that, as the noniinearity is
reliability function" that will trigger the execution of the contingency increased, the accuracy and reliability of the index-change estimates are
selection program whenever the system wide reliability drops below a decreased. Therefore the method as it stands represents a compromise
minimum accepted level? This reliability function will, of course, de- on nonlinearity, hoping that neither of the properties (i) and (ii) are
pend on the voltage levels, line flows, etc. Could the performance in- weakened sufficiently to give dangerously erratic results. The approach
dices defined by the authors be used for that purpose? will become a real breakthrough if some way of resolving the above
The performance indices are produced by summing over busses, conflict can be found, without sacrificing the present very high com-
generators and/or lines of the network. The most significant terms in putational efficiency too much.
the summation will be those for which the voltage magnitude (or power Security analysis is the most demanding in real-time application,
flow) is outside the limit. Two contingencies with the same value of the since here high computing speed is a technical and not merely economic
performance index might have it made up a widely different number of requirement. At present, for example, 100 contingency cases for a
significant terms in the summation. If the number of significant terms is 155-bus Brazilian system are taking about 5 CPU seconds in off-line
small or large respectively we might say that the contingency is cor- simulation on the IBM 370/158, using the fast decoupled load flow. As
respondingly local or global. fast as this may seem, automatic contingency selection is a practical
If the ranking of the contingencies is near the stopping threshold it necessity when we consider larger systems, a slower real-time machine
is likely that the local contingency is more severe than the global. The with programs competing for CPU, high security standards where the
argument would be as follows. Since there are many significant terms in contingency list is extensive, and frequent security-analysis runs (in-
the performance index of the global contingency, no bound can be cluding during preventive control calculations). At the same time, it
greatly exceeded. On the other hand, with few large terms contributing should be anticipated that in the not-too-distant future, the computing
to the ranking of the local contingency, the bounds must be violated by requirements for real-time static security analysis are unlikely to expand
a more significant amount. at anything like the rate at which fast hardware is becoming cheaper.
The above argument might be used to enhance the stopping Also, contingency evaluation lends itself naturally to the application of
criteria. If the ranking algorithm counts the number of terms out-of- parallel processing.
bounds in the performance index, this count can be used as another The authors are complimented on an excellent piece of original
parameter to be considered by the stopping criteria. research, which with further refinement may provide a practical answer
The ranking of the contingencies is based on the sensitivities of the to the contingency-selection problem.
performance indices rather than the values of the indices themselves.
This seems to work well whenever the performance indices are either Manuscript received February 21, 1978.
monotonically increasing or decreasing but not both. However, the per-
formance index of eq. (1) of the paper is not a strictly increasing or
decreasing function. In fact, the sensitivity with respect to changes in Walter 0. Stadlin (Leeds & Northrup Company, North Wales, PA):
voltage is positive for voltages above the upper limit and negative for The authors have succeeded in developing a powerful technique for the
voltages below the lower limit. Should eqs. (A.5) and (A.6) be modified selection of contingency cases for study in a real-time computer system
to reflect this fact? environment. Their approach minimizes the need for exhaustive com-
Finally, would the authors case to compare their method with the puter analysis and a priori case selection. It is interesting to note that
method suggested by J. L. Carpentier (A). Tellegen's Theorem was also applied by Fischl and Puntel (1, 2) to solve
Notice that his method can be used for single as well as multiple the problem of selecting transmission lines for planning. Apparently,
contingencies. the same sensitivity coefficients can be used to predict the effects of
either removing or adding a transmission line. Performance index sen-
REFERENCE sitivity coefficients can be derived by direct differentiation of the net-
work equations and observing the relationship among the resulting vec-
(A) J. L. Carpentier, "Differential Injections Method-A General tors and matrices. Evaluation of the sensitivity coefficients leads to the
Method for Secure and Optimal Load Flows", 1973 PICA Con- analogy of solving a second load flow on an "adjoint" network which I
ference, Minneapolis, Minnesota, June 1973. interpret to be the transpose of the original network. This is probably a
minor point and may only show up in a network with phase shifters.
Manuscript received February 24, 1978.
REFERENCES
B. Stott, 0. Alsac and M. Thome (CEPEL, Rio de Janeiro, Brazil): The (1) R. Fischl and W. R. Puntel, "Efficient Method for Computing
emphasis in algorithmic development for security analysis has been on Electric Power Transmission Network Sensitivities," (Abstract:
efficient contingency-evaluation techniques, to enable large numbers of (C72 167-0); T-PAS72 Jul/Aug, 1728 (7EO9).
outage cases to be solved as fast as possible. This timely paper focusses (2) R. Fischl and W. R. Puntel, "Computer-Aided Design of Electric
attention on the complementary issue-how, in the first place, to avoid Power Transmission Networks," (Abstract: C72 168-8); T-PAS72
having to solve large numbers of cases. The problem is a difficult one, Jul/Aug, 1728.
as we ourselves have been finding, and there seem to have been very few
ideas on the subject in the literature. Yet a reliable and efficient method Manuscript received February 15, 1978.
for contingency selection is likely to give much more computational sav-
ing than any further improvements in the speed of contingency-
evaluation algorithms.
The method in this paper is very innovative, both in the idea of G. C. Ejebe and B. F. Wollenberg: The authors appreciate the oppor-
ranking, via the sensitivities between a performance-index function and tunity to respond to the discusser's comments. Because many of the
the branch admittances, and in the elegant application of Tellegen's discussers' comments overlap, general comments will be made followed
theorem. The most important practical requirement is by specific responses to the discussers' remaining points.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
108
GENERAL COMMENTS sided" in that only one limit would be modeled. For example, for the
1) The authors wish to correct an oversight in the reference section of voltage performance function we would have:
our paper. Proper credit should have been given to W. R. Puntel and R.
Fischl whose work formed the foundation of the work reported. The
use of Tellegen's theorem and a performance index were first introduc-
ed by Puntel and Fischl (see References 9 and 10 of discussion by
Fischl). PIV(l) 2nNB
1 ( 7W Li)2n
2) Several of the discussers indicate a desire for more details of the
derivation of the a.c. version of Tellegen's theorem. In the a.c. version
of Tellegen's theorem all bus injections are defined in terms of normal
load flow conventions. This results in a pair of complex equations at i=l \vSP _+ Lim
each bus. These equations are differentiated with respect to voltage.
The complex conjugate of A.3 is added to A.3 and substitution made
PIv(2) = E Vi- V i
from the differential form of the injection equations. Finally, terms are
matched with corresponding terms from A.5 to give sufficient equa- To calculate the performance function we would need to carry out two
tions to solve for the adjoint voltages. The solution is not a second full
a.c. load flow but simply a solution of 2n-2 simultaneous sparse equa- calculations of the adjoint voltages.
tions. Specific Comments to Further Points Raised by Discussers:
3) Several of the discussers mention the possibility of using line current Drs. Puttgen and Sullivan correctly point out that we do not at-
flow or MVA flow rather than MW flow in setting up the line overload
performance index. This was tried in some early experiments and was tempt to obtain an evaluation of the system's state and that their sen-
not found to result in any significant difference when compared to a sitivity coefficients are used "...to actually evaluate the state of the
MW flow performance index. system under contingencies." It is our opinion that first order sensitivi-
4) The issue of the choice of performance index also appears in several ty coefficients, while perhaps providing adequate accuracy to predict
discussions. Indeed, the choice of performance index is undoubtedly a line MW flows, cannot provide adequate accuracy for voltage
key consideration in the use of a method such as reported in our paper-. magnitudes. Major EHV transmission systems may require that voltage
An ideal performance index would have the following properties: drop be held to as little as 3% under contingency conditions. Any
a) It would always produce a perfect ordering, and method whose error is in the range of 1-5% in voltage magnitude would
b) The terms of the normalized gradient of the performance index be unacceptable in planning such systems.
would be accurate enough to indicate directly whether a branch Mr. Ramarao makes the important observation that contingencies
overload or bus voltage problem existed. involving multi-terminal lines are the major portion of a realistic con-
Obviously, the performance index used in the paper is far from having tingency set. We model multi-terminal lines as described in Ref. [81 of
these ideal properties. Of the two properties mentioned, the second is the paper.
by far the most difficult to attain. Dr. Fischl is the first researcher to apply Tellegen's theorem to
Our performance index was chosen to give a global measure of the power systems analysis and the detailed comments in his discussion
system's performance. The measure was obtained by summing penalty reflect his extensive knowledge of the problems involved. We trust that
functions over all lines and buses in the system. The idea behind this ap- most of his points were addressed in our general comments. Dr. Fischl
proach is to obtain a scalar quantity whose value is a measure of system asks if the use of a security region would not be better than APIMW. The
performance only. However, by summing all penalty functions we lose answer is dependent upon the level of system trouble one is dealing
the identity of which elements are out of limit. The chief advantage of with. In our planning program, Ref. [8], we use APIMW to identify cases
this approach is its simplicity-its disadvantages are many as clearly in trouble. When a problem case is found, we apply a gradient
pointed out by the discussers. By adding the penalty functions together, redispatch procedure to evaluate whether a secure region exists or not,
the method may suffer from a masking effect where, for example, many and if not, how much load needs to be shed to relieve the overloads.
lines at moderate loading give the same performance index value as one Under normal conditions the redispatch and possible load shed are only
line with an overload and the others at light loading. To overcome the invoked on a small number of cases. When the system is under a great
masking, one can try to raise the exponent used in the penalty func- deal of stress, however, redispatch and load shed are needed for each
tions. Our experience with the MW flow performance index shows that case.
although the higher exponents do accentuate the overloaded cir- Drs. Van Slyck, Irisarri, and Sasson provide many useful com-
cuits-the increased non-linearity causes the final ordering to become ments relevant to use of the contingency selector in real time applica-
worse. The example reported in the paper shows results for the MW tions. Particularly, they ask about what conditions should be used to
performance index with n = 3. When run with an exponent of n = 1, trigger the selector. Obviously, any significant change in system
the results were quite similar; however tests on a different network topology should trigger the selection since this is where it really helps by
reported in [RI], show a worse ordering as n is increased. identifying the next most troublesome cases. When no switching has oc-
The authors do not claim to have all the answers to the discussers' curred, we suspect that a periodic re-execution of the selector followed
comments concerning performance index structure. However, some of by a series of ac load flows will be the most trustworthy approach. Any
our own thoughts will be offered to indicate what type of future reliability function approach is subject to the same "masking" effects
research might be profitable. Our index was a scalar function of system as the performance index functions discussed earlier, thus making the
performance. Better results might be obtained if the performance were reliability function inaccurate.
measured with a vector function. That is, two or more attributes would These discussers further point out that the stopping criterion when
be calculated by the performance function and the gradient replaced by running load flows can have multiple attributes. This is an excellent sug-
a matrix of first derivatives. Selection of cases to be run now becomes gestion since the attributes can be accurately determined and should
much more complicated. One could simply order within each attribute enhance the selector's ability to stop correctly.
and run full load flows on each ordered list, but an analysis of the entire In comparing our method of that of Carpentier, we would note
matrix looking at combinations at attributes would perhaps prove first of all that we do not take economics into account in the redispatch
superior. strategy used (see response to Fischl above). Otherwise our approach is
The performance index need not be global, but could be restricted quite similar since we cycle between a redispatch and the contingency
to only certain important subsets of system lines or buses. This may overload checking logic.
have the effect of decreasing some of the "noise" in the ordering. Drs. Stott, Alsac, and Thome questions were covered in our
Several iterations would be required to cover the entire network using general comments on performance indices. An additional insight pro-
separate performance indices each covering a different part of the net- vided in their discussion is that correct ranking can still be achieved if
work. The results presented in the paper reflect this approach in that we the errors in API are consistent. This is a very useful point to remember
did not try to mix line MW flow, bus voltage, and generator outage con- in doing further research on performance indices.
tingencies together. Mr. Stadlin is quite correct in noting that direct differentiation
The problem of monotonicity of the performance index was men- achieves the same result as Tellegen's theorem. This is described quite
tioned by several discussers. It is quite correct to note that Equation (1) succinctly in Ref. [R2]. We trust that our expanded comments on the
of the paper does not present a monotonic increasing or decreasing adjoint voltages answers the question about solving a second load flow.
function of voltage magnitude. A possible remedy would be to form a The adjoint voltages involve only a network solution and do not require
two attribute voltage performance index. Each attribute would be "one a load flow.

Authorized licensed use limited to: PONTIFICIA UNIVERSIDADE CATOLICA DO RIO DE JANEIRO. Downloaded on November 27, 2009 at 10:57 from IEEE Xplore. Restrictions apply.
109
Drs. Sachdev and Ibrahim ask about the weighting factors WVi.
The WVi terms could be chosen so that division by 2n is avoided, but REFERENCES
then a multiplication by 2n appears in the gradient calculations. If [1] Clements, K. A. and Ringlee, R. J., "Final Report-Research on
voltage limits are not symmetrical about VSP, one would have to resort to Adequacy Assessment of Interconnected Electric Power
the one-sided approach mentioned in our general remarks on perfor- Systems," published by Worcester Polytechnic Institute, Dept. of
mance indices. Electrical Engineering, Worcester, Mass. 01609. Copies available
We have no doubt that the operator's or planning engineer's ex-
perience can be eliminated as a factor in the stopping criteria. However, by direct request to K. A. Clements.
it is doubtful that the resulting criteria would be any more accurate. In [2] Valtonen, Martti E. "Equivalence in Sensitivity Calculations Bet-
our opinion, experience should be employed until advanced stopping ween Direct Differentiation and The Method Based on Tellegen's
criteria are developed. As to using weighted sums of voltage squared or Theorem," Proceedings of the IEEE, Vol. 65, No. II, November
just weighted sums of voltage as sensitivity functions, we agree that 1977.
some simplification in calculations exists. We are somewhat skeptical,
however, as to whether the same ordering would result. Manuscript received June 28, 1978.

Bruce F. Wollenberg (M'67, SM'75) was born in


Buffalo, New York on June 14, 1942. He
graduated from Rensselaer Polytechnic Institute
Gabriel C. Ejebe (M'76) was born in Nimo, wth a BEE degree in 1964 and an M.Eng. degree
Nigeria on March 22, 1944. He received the in 1966, and from the University of Penn-
Bachelor of Engineering degree (First Class sylvania with a Ph.D. in Systems Engineering in
Honors) from Ahmadu Bello University, Zaria,
Nigeria in 1971 and the M.S. and Ph.D. degrees From 1966 to 1974, he was employed by
from Drexel University, Philadelphia, PA in Leeds & Northrup Co. in the Corporate
1974 and 1976 respectively both in Electrical Research & Development Division and in the
Engineering. Systems Analysis Department of the Systems
In July 1976, he joined Power Division. At L&N he was responsible for development of dispatcher's
Technologies, Inc., where he was involved in the load flow software as well as research into security dispatch, optimiza-
development and implementation of computer tion, and contingency analysis software.
programs for bulk power system reliability assessment. His research in- In 1974, he joined Power Technologies, Inc. and was given respon-
terests include the application of advanced system theory and optimiza- sibility for development of real time interactive power system
tion techniques to the planning and operational problems of intercon- simulators. In addition, he has contributed to researach in contingency
nected electric power systems. Presently, Dr. Ejebe is a lecturer at the selection techniques and load shedding algorithms using linear pro-
Department of Electrical Engineering at the University of Nigeria, gramming.
Nsukka, Nigeria. Dr. Wollenberg is a member of Tau Beta Pi and Sigma Xi.

You might also like