You are on page 1of 8

IEEE Transactions on Power Apparatus and Systems, Vol. PAS-100, No.

2, February 1981 771


SENSITIVITY STUDY OF THE CUMULANT METHOD OF
CALCULATING GENERATION SYSTEM RELIABILITY

John P. Stremel, Member, IEEE


Tennessee Valley Authority
Chattanooga, Tennessee

Abstract - The cumulant method is a very fast the full load duration curve (LDC). Extension of the
method for the underlying calculations which support analysis to include the LDC (which would be more
production costing and reliability evaluations. This appropriate for production costing comparisons) can be
paper applies the method to a variety of system con- accomplished by following the logic developed in [1].
figurations and measures the accuracy against estab- Since most systems concentrate on a planning criterion
lished methods. A surprisingly wide range of of 0.1 to 1 days/year, this analysis will concentrate
applicability is established. The method proves to on probability levels as low as 3 x 10 5 as the region
be adequate for generation systems where the forced of interest. Loads which incur smaller probabilities
outage rates are -greater than 10 percent and the are assumed to be inconsequential for planning.
largest unit is less than 10 percent of installed
capacity. The EPRI test systems offer a convenient source
for varied applications. The full system descriptions
are available in [3]. System A is described in
INTRODUCTION table 1 as a representative example and for the
reader's convenience.
Since the introduction of the cumulant method for
production costing and reliability evaluations, there
has been considerable interest in extending the Table 1
analysis to smaller systems and multiple area studies.
However, it was pointed out in the discussion of [1] EPRI SYNTHETIC SYSTEM A
that the method will not be as appropriate for smaller
systems. Full System
Capacity: 53,350 MW
This paper explores the accuracy of the cumulant
method under a wide variety of applications. Forced
Number Unit Unit Outage
The cumulant method gets its name from the use of Units Size Description Rate
cumulants (which are related to the moments of a
random variable) of the generation system outage dis- 5 1200 Nuclear 0.15
tribution. The generation system outage cumulants 5 1000 Nuclear 0.15
appear as coefficients in a Gram-Charlier approxima- 3 800 Fossil 0.24
tion to the system MW outage distribution. 17 600 Fossil 0.21
21 400 Fossil 0.13
Since the method is an approximation, there will 74 200 Fossil 0.074
be conditions where it performs poorly. The sections 5 200 PS, Hydro 0.05
of this paper: benchmark the accuracy of the approxi- 12 50 Fossil 0.027
mation against some EPRI developed synthetic systems; 87 50 C. Turbines 0.24
benchmark the approximation against some hypothetical 12 50 Hydro 0.012
identical unit systems; and discuss some alternatives
which may be more fruitful for specialized
applications. Reduced System*
Capacity: 10,700 MW

EPRI SYNTHETIC SYSTEMS 1200 Nuclear 0.15


1 1000 Nuclear 0.15
This section tests the accuracy of the cumulant 4 600 Fossil 0.21
method for realistic systems. An EPRI project by 4 400 Fossil 0.13
Power Technologies [3] has developed a data base for 15 200 Fossil 0.074
six generation systems which vary from 21,900-MW to 4 50 Fossil 0.027
53,350-MW capacity. 19 50 C. Turbines 0.24
1 200 PS, Hydro 0.05
In order to maintain a relatively simple frame- 3 50 Hydro 0.012
work, the comparisons will be based upon the genera-
tion system MW outage curve, rather than incorporating
"System has been modified slightly to accommodate a
50-MW step Calabrese reliability model.

Each of the six EPRI synthetic systems were used


to produce a generation system outage curve comparison
as in figure 1. The error is measured by fixing a
probability level and reading the MW-outage level.
The MW difference is recorded as the error
80 SM 593-4 A paper recommended and approved by the attributable to the cumulant method. This comparison
IEEE Power Syste-m Engineering Committee of the IEEE is straight forward for large systems since both
Power Engineering Society for presentation at the models produce continuous looking outage curves. For
IEEE PES Summer Meeting, Minneapolis, Minnesota, smaller systems the discrete model would consist of a
July 13-18, 1980. Manuscript submitted January 2, series of steps. In those cases, the midpoints of
1980; made available for printing May 19, 1980. the steps have been joined to maintain a more
-realistic comparison as in figure 1.

i) 1981 IEEE
772
Figure 1 Realistic application of the cumulant method for
the EPRI systems is very encouraging. Since the deci-
sion for capacity expansion will be driven by outcomes
i, in the third and fourth columns, the magnitude of the
EPRI SYSTEM A error will be less than 25 MW for a system which is
planning unit additions on the order of 1,000- to
2,000-MW per year.
Although the cumulant method appears to be
entirely adequate for these systems, they are large.
The approximation also needs to be applied to smaller
systems.
eni1
Vf)
0 Each of the EPRI synthetic systems was accompa-
0
nied by a reduced "equivalent." These systems were
0
0
not intended for use in reliability studies.
Reserve requirements to maintain a realistic planning
criterion would be on the order of 30 to 40 percent.
F- However, an analysis of these systems produces some
0
m interesting results.
m
0
(r

Table 3

EPRI SCALED DOWN SYNTHETIC SYSTEM


ERROR ANALYSIS (MW)
x
14000 15000 16000 17000 18000 19000
MW OR MORE ON OUTAGE

System Probability Level


(Capa-
3xlO -4
Large
-2 lO -3 3xlO -
1
city) 3xlO 3x 3x30 Units
Figure 1 clearly indicates that there is very
little error in the approximation. Table 2 summarizes A 0 0 -25 -125 250 1,200
the accuracy of the approximation for the six (10,700 MUI) 1, 000
synthetic systems.
B 0 -25 0 -150 -350 1,000
(10,000 MU1) 800
C 0 0 0 -175 250 1,100
Table 2 (10,950 MW) 1,100

EPRI SYNTHETIC SYSTEM D 0 0 0 -200 300 1,200


ERROR ANALYSIS (MW)* (10,050 MW) 1,200
E 0 0 0 -75 -100 800
System Probability Level (10,300 MW) 800
(Capa-
city) 3xlO 3xlO 3xlO 3xlO 3x1O-5 3xlO-6 F O 0 0 -150 225 1,200
(10,700 MW) 1,000
A 0 0 0 -50 75
(53,350) Average
Error 0 4 4 146 246
B 0 0 0 -25 -25 150
(46,100)
Table 3 indicates that the approximation is not
C O 0 0 -25 0 225 as good for these smaller systems. The increased
(21,900) error for these systems can be traced to the effects
of several large units which contribute a substantial
D O 0 0 -25 25 225 portion to the total system outages. Consequently,
(32,000) the generation system outage distribution contains
large outage impulses (due to the large units) which
E 0 0 0 0 -75 50 make it very difficult for a continuous approximation
(45,700) to represent the system outage distribution.

F O 0 0 -25 -25 175 Table 3 also indicates that the error from an
(31,800) analysis for such systems would be on the order of
100 MW for a system which would be expanding at about
Average 500-MW per year. However, in order to put this
Error 0 0 0 16 32 150 evaluation into perspective, a frequently used
approximation is the representation of generation
units as binary devices with an equivalent forced
'The synthettic system units are chosen to be in 50-MW outage rate derived from total outages and deratings.
increments, all errors are rounded to the nearest The EPRI data used here is also reported for a more
25-MW increement. detailed 3-state representation of each unit. The
773
error introduced by ignoring the 3-state representa- Figure 2
tion for reduced system A is 0, 125 MW, 200 MW, 225 MW
and 225 MW error for the probability levels in Relative Unit/System Size Sensitivity
table 3.
20% F.O.R
Consequently, even though the approximation is A: 50 units, 200 MW
less precise for the reduced systems, it may be B: 20 units, 500 MW
appropriate for some smaller systems and is certainly C: 10 units, 1,000 MW
appropriate in cases in which binary unit representa- c: cumulant function
tion is acceptable. d: discrete function
10-2
Deratings or multiple state unit representations
are easily introduced into either model. Table 4 I// I L
compares the cumulant and discrete models, and indi-
cates that similar results are obtained when deratings 0
-i
are included. A comparison of tables 3 and 4 shows LL.

that the cumulant method is more accurate when 0

deratings are included. Introduction of deratings has Un


apparently smoothed out the impulses which make it
difficult for the continuous approximation. Li.
10-4
0

Fn

Table 4 co
0

EPRI SCALED DOWN SYNTHETIC SYSTEM //C B A


WITH DERATINGS ERROR ANALYSIS (MW)
10 L
20 40 60 80 100
Probability Level LOAD AS % OF INSTALLED CAPACITY
System
(Capacitn v) 3x10
3l 3x10 -2
l
3x10 -3 3x10 -4 3xlO 5
The continuous approximation is within 50 MW of
A O 0 0 -100 250 the numerical probability for each case, indicating a
wide range of applicability for various system
B O 0 0 -150 300 compositions.

C O 0 0 -150 300 Effect of Forced Outage Rate

D O 0 0 -200 300 Consider a system of constant size (50 units


of 200 MW each) where unit forced outage rates are
E 0 0 0 -50 50 varied. Systems A, B, and C in Figure 3 have F.O.R.s
of 0.10, 0.20, and 0.30. Both generation models
F 0 0 -100 250 predict that as the forced outage rate falls, the
expected load loss decreases.
Average Error 0 0 0 125 242
Figure 3

IDENTICAL UNIT SYSTEMS Forced Outage Rate Sensitivity

It is difficult to perform a thorough investiga- 50 uitits, 200 MW


tion with practical systems because of the multitude A: 10% F.O.R.
of parameters that need to be taken into account. The B: 20% F.O.R.
most important parameters are system size, number of C: 30% F.O.R.
units and forced outage rates. This section uses c: cumulant function
easily verified identical unit systems to investigate d: discrete function
the effects of varying these parameters on the
accuracy of the cumulant method approximation to the
capacity outage distribution.
The comparisons which follow plot the probability 0

of load loss against the load as a percent of 3l io-


installed capacity. I.
0
,C B A-
Effect of Number of Units Un
U)

104
Consider a system of constant size (10,000 MW).
All units have a F.O.R. (forced outage rate) of 20
percent. F
=II A=~~
m o-
Figure 2 summarizes the effect of the number of 0
<

units on the approximation. Both the discrete and


cumulant models demonstrate the well-known relation- 0.

ship that system load carrying ability increases


with increasing number of units (installed capacity io'6
0 20 40 60 80 1CDO
held constant). This is due to the system being less LOAD AS % OF INSTALLED CAPACITY
affected by outages of individual units, as their
number increases.
774
A close inspection of figure 3 shows that the Figure 5
continuous approximation is for all practical purposes
identical to the discrete case and consequently quite Small System Sensitivity
acceptable.
10 Units, 1,000 MW
Effect of System Size A: 10% F.O.R.
B: 20% F.O.R.
In order to study the effect of system size, the C: 30% F.O.R.
number of units (50) and the forced outage rate (0.1) C: cumulant function
are kept constant. Systems A, B, and C in Figure 4 d: discrete function
are composed of 100-, 200-, and 300-MW units.

Figure 4
I D1// L
System Size Sensitivity 0

0
50 units, 10% F.O.R. 104
A: 100 MW units IL.
0
|//C //B A
B: 200 MW units
C: 300 MW units 0

c: cumulant function 104


d: discrete function 0
d d

m
co
ct
'I
a-A>
L=
7!!1/.
0
H

c]
.0
j 104 .LL.
0
20 40 60 80 100
LOAD AS % OF INSTALLED CAPACITY
in
0

outage rates fall.

Consequently, it would be dangerous to extend


m
this analysis to systems outs'ide the sensitivity range
co considered. A 50-unit hydro system with forced outage
0
rates on the order of 2 or 3 percent will not neces-
sarily be adequately represented. These traits follow
from the fact that the approximations are subject to
the law of large numbers. Although this method is
LOAD AS % OF I NSTA LLED CAPACITY significantly better than' one which would use the
Normal distribution, the specified parameters used
The results shown in figure 4 indicate that the here are incapable of modeling severe cases of non
systems are identical when compared in relative terms. Normality. Thus, as the equivalent forced outage rate
falls to near zero or as the number of units becomes
Investigation of Small Systems very small, the approximation becomes inadequate.
Fortunately for most practical examples, neither is
Figure 5 contains an analysis of three 10-unit the case.
systems consisting of identical 1,000-MW units.
Systems A, B, and C have unit forced outage rates of
0.10, 0.20, and 0.30. These systems are consistent IEEE TEST SYSTEM
with what was seen with the reduced EPRI systems,
the errors are larger. Systems B and C appear to be The IEEE test system [7] is the most realistic
accurate within 50 to 100 MW while System A misses at small synthetic system. The major system character-
one point in the range of major interest by 200 MW. istics are: an installed capacity of 3,405 MW, a
Larger errors below 1 x 10 4 will have very little large unit of 400 MW, a peak load of 2,850 MW and an
im,pact on effective capacity and load carrying ability annual capacity factor of 62 percent.
calculations.
Since the test system is published with a
Consequently, estimates obtained for using the capacity outage table, table 5 reports a similar
cumulant method for 10-unit systems are entirely ade- calculation for the Cumulant Method and translates
quate for most applications, the level of error is the error into a MW equivalent.
still less than 10 percent of the capacity on outage
and within the range of error that would be implied Table 5 shows large errors for this size of a
from ignoring specific deratings and modeling binary system. The worst errors occur in the extremes of
units with an equivalent forced outage rate. the MW outage distribution. However, for planning
purposes the range past 1,300 MW is not particularly
In conclusion, the synthetic system analysis disturbing since these loads have virtually no influ-
indicates that any system in which the largest unit is ence on the system planning criterion.
less than 10 percent of the total installed capacity
can be successfully modeled with the cumulant method. Although the Cumulant Method cannot be recom-
However, it is important to note that the method mended in this case, when a planning criterion is
improves in accuracy with the introduction of derat- calculated based upon all the hourly loads, the
ings and decreases in accuracy as equivalent forced accuracy improves dramatically. Figure 6 shows the
775
Table 5 SIMILAR METHODS

CAPACITY OUTAGE TABLE* Rau, Schenk, and Toy [4] have proposed a nearly
identical method which uses the Gram-Charlier series A
for IEEE Test System rather than the Edgeworth form of the series in [1].
Other forms of the Gram-Charlier approximation may
produce slightly different results (for example,
x P (x) MW Error Gram-Charlier series C, see [5], imposes nonnegativity
restrictions on the estimates). Experience at TVA
100 .604 50 indicates that these formulations are all approxi-
200 .435 30 mately equivalent but that the Edgeworth form of
300 .307 -50 series A maintains better properties for large
400 .204 -10 systems.
500 .120 0
600 .063 0 An inverse function developed by Cornish and
700 .035 -50 Fisher [6] allows for the reading of MW (see figure 7)
800 .022 -10 given the probability index. This is just as accurate
900 .013 40 as the usual formulation and numerically faster for
1000 .0068 40 repeated calculations and may hold some promise for
1100 .00274 30 specialized applications.
1200 .00085 10
1300 .00021 -50 The basis for the approximation in [1,4] is the
1400 .000041 -100 Normal distribution and variations about the Normal
1500 .0000063 -120 as a first approximation. It is also possible to
1600 .00000083 -130 base the approximation on other distributions such as
the Poisson (Gram-Charlier series B). The Poisson has
much better small sample properties and consequently
*Based on four cumulants in the Edgeworth form of the may be very helpful for representing small systems
Gram Charlier series. such as the IEEE test system.

Figure 7
accuracy for various load levels and expansion units.
In fact, for the IEEE test system, the timing of unit
additions would be nearly identical for the discrete INVERSE FUNCTION
model and the Cumulant Method of calculating LOLP.

I
x

0
Figiyure 6 J
1-
TEST
SYSTEM UNIT 2 UNITS 3 UNITS 4 UNITS 0
6
/ / / , 0X

PLOLP

:
lx

(In
cc

i CONCLUSIONS
-i
CD The analytical approximation discussed in this
paper is an accurate and easily computed method for
(D
2
a
4
reliability evaluations in capacity planning and
0
U.
maintenance scheduling models.
0
(n(f)
0 TVA's experience with this method is very
encouraging and the use of the method appears to be
satisfactory for most applications in which the
largest unit is less than 10 percent of total
installed capacity and system forced outage rates
are greater than 10 percent.

ACKNOWLEDGEMENT
2450 2850 3250 3650 4050 4450 4850 5250
This study is indebted to the work of Brian
ANNUAL PEAK LOAD
Godsy who willingly prepared and studied mounds of
sensitivity runs. I would also like to acknowledge
the helpful comments of an anonymous reviewer which
have improved the presentation in this paper.
776
REFERENCES series shown in Table I below are an improvement over the author's
values in Table 5 of this paper.
[1] J. P. Stremel, R. T. Jenkins, R. A. Babb, and It can be seen that the extension of the series gives a better fit in
W. D. Bayless, "Production Costing Using the the tail of the density which is generally of interest to planners. In this
Cumulant Method of Representing the Equivalent system of 3405 MW capacity, the reserve for a risk level of 1 day in
Load Curve," IEEE 1979 PES Summer Meeting, 10 years, no matter how one calculates it, is of the order of 1200 MW.
Paper F79 674-3. Thus, an error of 70 MW represents only 5.8% of the reserve capacity.
This, we submit, is very small compared to the standard unit sizes the
[2] J. P. Stremel and N. S. Rau, "The Cumulant planners might be constrained to pick, the errors in load forecasting
Method of Calculating LOLP," IEEE 1979 PES and other uncertainties.
Summer Meeting, Paper A79 506-7. Another aspect of this work appears to need clarification. Under
the section, "similar methods," the author indicates that the methods
[31 Synthetic Electric Utility Systems for Evaluat- of [4] are different from those of [ 11. This is not true. The confusion
ing Advanced Technologies, EPRI Project TPS appears to have stemmed from the different terminologies adopted in
75-615, Power Technologies, 1977. different books. Cramer calls the series used in [1] an Edgeworth
series. Although this series was obtained by Edgeworth from the ori-
[4] N. S. Rau, K. F. Schenk, and P. Toy, "Expected ginal Gram Charlier's expansion, some books call the Edgeworth
Energy Production Costs by the Method of series the Gram Charlier's series of Edgeworth form. There are other
Moments," IEEE 1979 PES Summer Meeting, Paper nomenclatures as Type A and Type B series. In any event, in our first
F79 692-5. publication [8], we have called the series (A-25) the Gram Charlier's
expansion. Subsequently, in our paper [4], we still call it G.C.E. The
[5] M. G. Kendall and Alan Stuart, The Advanced author calls his expansion in [I] a G.C.E. as well. However, a com-
Theory of Statistics, Vol. 1, Hafner Publishing parison of the series in references [81, [4], [1] and [3] will reveal
Company, New York, 1969. that they are all identical! All the series are generic extractions from
the series of Chapter 17 of Cramer [9], falling under the category of
[6] R. A. Fisher and E. A. Cornish, "The Percentile Edgeworth's expansion. We therefore suggest that all the methods are
Points of Distributions Having Known Cumulants," "identical" rather than being similar.
Technometrics, 1960, Vol. 2, No. 2, pp. 209-225. One final point concerning the names attributed to this method.
In [8] and in [4] we have used the termninology method of moments
(7] IEEE Reliability Test System, IEEE Transactions, or, for short, the moment method. This technique utilizes the statistical
PAS-98, pp. 2047-2054, November/December 1979. moments as well as cumulants in the Gram-Charlier expansion to evalu-
ate system reliability. The author refers to this approach as the cumu-
lant method. Although the cumulants are utilized directly in the Gram-
Discussion
Charlier series, these cumulants must be obtained from the moments.
Hence the name "method of moments" used by us is synonymous with
the author's preference to call it the cumulant method.
N. S. Rau, K. F. Schenk and P. Toy (National Energy Board, Canada): We do commend and appreciate the author's interest and con-
Since the publication of the work in [8] and [21, there have been tinued work in this fertile area of reliability analysis.
several questions regarding the accuracy and sensitivity of the new
method for calculating system reliability. To this important aspect,
this discussion has been addressed. TABLE 1
We do not agree with the author that the cumulant method can-
not be recommended for the case of the IEEE test system. We shall Obtained from Extended Series
substantiate this below.
Cramer [91 has discussed the question of accuracy of the so-called 1 2 3 4
Gram-Charlier and other expansions in Chapter 17. The author of this
paper has examined the accuracy of the expansion as given in [8] and X MW on Outage P(x) Probability MW from the Error, MW
[2] with only terms up to the fourth central moment and the sixth outage table Col. 1 - Col. 4
derivative of the Normal. The errors resulting therefrom are shown in corresponding
Table 5. We find, however, that under some circumstances, it is ad- to p(x) in
visable to extend the series to include higher order terms. The result- column 2
ing computational effort is minimal.
Elderson & Johnson [ 10] define a measure of skewness as 100 0.6047 50 50
200 0.3972 192.7 73
300 0.2706 376.6 -76.6
s -
4 l (B2+ 3) 400 0.2036 415.5 -15.5
2 (5B 6B1 - 9) 500 0.1430 451.8 48.2
2 2 600 0.08013 568.8 31.2
700 0.03594 747.5 -47.5
B (-. 3)
2 800 0.016290 834.9 -34.9
2 ~ (2B 2 -3B 1 -6)(4B 2 -3B)
1 900 0.009496 922.6 -22.6
1000 .005772 973.6 26.4
1100 .002847 1060.1 39.9
1200 .001066 1171.9 28.1
where B1 = G2 1300 .0003039 1320.0 -20.0
1400 .00006674 1445.7 -45.7
1500 .0000151 1569.9 -69.9
B = G + 3
2 2
REFERENCES
SI is always positive while S2 can be positive or negative. If SI < 0.5 [8] N. S. Rau and K. F. Schenk, "Application of Fourier Methods of
and IS21 < 0.5, the density has a small skew. We have found that for Capacity Outage Probabilities," Paper A 79 W3-3, IEEE Winter
densities which have SI and IS21 > 0.5, extension of the general series Power Meeting, February 1979.
in Chapter 17 of Cramer gives better accuracy [11]. We use terms up [9] H. Cramer, Mathematical Methods of Statistics, Princeton Uni-
to and including the eighth derivative of the Normal and the eighth versity Press, Princeton, 1946.
cumulant. For the IEEE test system discussed in the paper, SI = [10] W. P. Elderton and N. L. Johnson, Systems of Frequency Curves,
1.4812 and S2 = -0.6574. Thus, the values obtained by our extended University Press, Cambridge, 1969.
777
[ 11 ] K. Schenk, "Analysis of Reliability Criteria for Generation Plan- R. Billinton and G. Hamoud (University of Saskatchewan, Canada):
ning - Addendum," National Energy Board Report, January We would like to complement the author on an interesting and timely
1979. paper. The author has used a variety of system configurations in order
to examine the accuracy of the cumulant technique. As expected, the
Manuscript received July 14, 1980. accuracy is acceptable when a system is composed of identical units
with relatively large forced outage rates. The latter aspect is extremely
important. The accuracy of the technique degrades considerably when
G. T. Heydt (Purdue University, West Lafayette, IN): This paper is a units with low forced outage rates are included in the model. In apply-
further extension of a recent series of papers on the use of the Gram ing the cumulant method to the IEEE Reliability Test System it was
Charlier series for the purpose of calculating generation system reliabil- found that the results are reasonably accurate over a certain load range.
ity. The salient contribution of this paper is a discussion of the accu- If the forced outage rate of each unit is doubled, the cumulant method
racy of this technique over a variety of applications. becomes more accurate and the valid range is increased by approxi-
The Gram Charlier series type A is essentially the same series as mately 25%. The actual risk is of course much higher than before.
that of Edgeworth. This series is an infinite series which, when con- The accuracy of the cumulant method has been examined in detail
vergent, is an exact representation of a probability density function in when applied to practical systems. Three existing systems were used
tenns of the statistics of a random variable. The approximation de- for this purpose. These are the Saskatchewan Power Corporation (SPC)
scribed in the paper occurs due to the truncation of the infinite series. system, the Manitoba Hydro (MH) system and the combined four
The accuracy of the truncated series depends on the values of the high western province system (Saskatchewan, Manitoba, British Columbia,
order cumulants of the random variable being modeled. When these Alberta). The SPC system has 39 units and the unit size varies between
high order cumulants are small, the truncated finite sum approximation 10 MW and 280 MW which makes the system outage distribution free
is a good representation of the density function. Stremel has apparently from large outage impulses. The MH system has 101 units and their
found the practical conditions under which the latter occurs. capacities vary between 3 MW and 117 MW in a manner that the dis-
Viviani and Heydt [1,2] have recently described an alternative crete distribution of the system capacity outages looks as a continuous
approach to the use of the Gram Charlier series type A. The approach distribution. The combined four western province system is relatively
is to transform the random variable being modeled. If T(x) is the trans- large and has 31 1 units with a total installed capacity of 23,376 MW.
formed variable and x is the original random variable, the parameters The LOLE index obtained on a daily basis for each of the three systems
of the transformation T are selected to force the cumulants of T(x) was calculated using both the cumulant and recursive methods and the
to those of a normal variate. By this means, the truncated, finite series results are illustrated in the figures on the following pages.
is an accurate representation of the probability density function. Al-
ternative forms of T and other details are found in the references
cited here. The comments of Stremel would be appreciated. SPC System
The interested reader should also note that Sauer and Heydt [3] - Recursive Method
have recently described both the Gram Charlier series type A and its 1.0 -- Cumulant Method

multivariate counterpart.
10o1
REFERENCES
[1] Viviani, G., "Stochastic Optimal Energy Dispatch," Ph.D. thesis,
May 1980, Purdue University.
[21 Viviani, G., Heydt, G., "Stochastic Optimal Power Dispatch," IC - 2457 MW
submitted for publication, IEEE Winter Power Meeting, Atlanta, PL - 1895 MW
GA, January 1981. 104 N - 39 Units
[3] P. Sauer and G. Heydt, "A Convenient Multivariate Gram-Charlier
.I5

lo~
Type A Series," IEEE Trans. on Communications, Volume CM-27, II
I
No. 1, January 1979, pp. 247-248. I
II
I
f
Manuscript received July 28, 1980. I
I

.1
35-6 ",

Len L. Garver (General Electric Company, Schenectady, NY): When


.f
.1
.f
I
absolute values of reliability are important, the older recursive method I

of calculating capacity outage tables will continue to be used. This 1200 1400 1600 1800 2000 2200 2400 Peak Load
conclusion was reached after studying the results presented by Stremel, Fig. 1. Variation of LOLE With System Peak Load.
especially Table 5 based on the IEEE Reliability Test System. Both the
magnitude and the inconsistency in sign of the MW errors in the calcula-
tion are disturbing. However, when the capacity outage table is part of Manitoba-System
1.0
a larger model such as a production cost estimation in a generation - Recursive Method
Cumulanlt Method
expansion planning method, then the cumulants method appears
--

worth investigating. lo-I


Adding one column to Table 5 will help the reader understand
the meaning of "MW-error" and make comparisons easier. For example,
noting that the probability 0.604 is associated with 50 MW in the IEEE
table, reference 7, and 100 MW in the Cumulant method will help IC - 4038 MW
clarify the "MW-error" numbers. PL - 3311 MW
N - 101 Units
0
I

P(x) x IEEE-x MW-error


0.604 100 MW 50 MW 50 MW 10-
0.435 200 170 30
0.307 300 350 -50 10-6
0.204 400 410 -10
etc. 3200 3400 3600 3800 4000 4200 Peak Load MW

Manuscript received August 5, 1980. Fig. 2. Risk as a Function of Peak Load.


778
capacity. I am still reluctant to make any more than a guarded recom-
mendation to apply the cumulant method for reliability analysis on
Combined System systems the size of the IEEE Test System.
- Recursive Method The discussors present an interesting concept for determining the
-- Comulart Method need for higher order cumulants. Considerably more work needs to be
completed to confirm using a simple cutoff rule for Sl and S2, how-
IC - 23376 KW
PL - 17250 MW
ever. For instance, Elderton and Johnson (referenced by the discugsors)
u . 'all 11^.4' show that for a related problem (Chapter 6) the values for SI and S2
interact to determine the appropriate region. This implies the need for
a more detailed test.
The discussors have also commented on the equation forms of the
Edgeworth Series and Gram Charlier Series. Cramer [1] shows that
there are several methods of developing an orthogonal expansion de-
10-4 /
rived from the normal distribution in order to represent an arbitrary
103
0- / frequency function (see sections 17.6 and 17.7). One of these
-6/
expansions in orthogonal polynomials became known as the Gram
Charlier Series of type A while ano-ther was developed by Edgeworth
/ [21 from an entirely different approach through the theory of ele-
mentary errors. For practical distributions, both will converge to the
true distribution with an infinite number of terms. In that- sense they
10-7 are identical. However, when only a finite number of terms are con-
20000 21000 22000 23000 Peek Load MW sidered, the expansions do differ. This has been the motivation for dis-
Fig. 3. System Risk as a Function of Peak Load. tinguishing between the two.
Mr. Heydt has pointed out an interesting possibility of introducing
a transformation of the random variable (i.e., a unit's outages) in order
It is seen from these figures that the cumulant method may or to obtain better approximation characteristics. This may prove bene-
may not give adequate results when compared with the recursive ficial if the ease of convolution can be maintained and a consistent
method depending on the system under consideration. It should be interpretation of the system outage distribution preserved.
noted that, in situations where the cumulant method is applicable, Messrs. Billinton and Hamoud have added to the presentation by
the results will be satisfactory only over a specified range of capacity including some predominantly hydro systems in the analysis. Hydro
outages. The accuracy of the cumulant method is not dependent upon systems will have unit forced outage rates of from 2 to 5 percent. This
a single parameter such as the system size, unit size, number of units, is considerably outside the range considered in the paper and not sur-
unit forced outage rate, but a function of all these parameters. It gen- prisingly indicates larger approximation errors than shown.
erally improves with increasing system size, average forced outage rate However, I have two reservations about drawing conclusions from
and number of generators. the discussors' analysis. First, for a predominantly hydro system a key
assumption for the reliability analysis (independent unit operation and
Manuscript received August 22, 1980. availability) is violated by the river flow dependency (which may be-
come critical during a dry year). My feeling is that the errors introduced
by the approximation are inconsequential compared to the errors intro-
John P. Stremel: The degree of interest in this topic underscores the duced by the operation assumption.
need for new methods in generation system reliability analysis. I would Second, an analysis of the efficacy of the approximation would
like to thank the discussors for their valuable contributions. require an analysis of the capacity addition decision as in Figure 6. The
I would like to thank Len Garver for clarifying the interpretation outage curve fit was so good for the EPRI Synthetic Systems that the
of Table 5. The changing sign of the entries in Table 5 is not an incon- additional analysis was unnecessary. However, the experience with the
sistency, however. The entries are a reflection of the approximation IEEE Test System indicates that the approximation may be quite
which is wrapping itself around the true values. The approximation adequate even with a rough approximation of the outage curve.
is high for a while, then low for a while and high again. The reason for As with any new method, the user should make a number of
the rather large errors is due to the fact that the IEEE Test System is a comparisons with existing procedures before incorporating the new
3,405-MW system dominated by two 400-MW units - nearly 25 percent method as a planning tool.
of the installed capacity. The two relatively large units cause the outage
curve to contain several large discrete jumps which makes it difficult for
the continuous approximation to form a smooth representation. When REFERENCES
the IEEE Test System is allowed to expand with two or three addi-
tional 400-MW units, the accuracy increases dramatically. [1] Cramer, Harold, Mathematical Methods of Statistics, Princeton
Messrs. Rau, Schenk, and Toy show that increasing the number of University Press, 1974.
terms in the approximation increases the accuracy. Although Figure 6 [2] Edgeworth, F. Y., "The Law of Error," Proceedings of Cambridge
supports the contention that the cumulant method would be appropri- Philosophical Society, Volume 20, 1905.
ate for expansion planning, the entries in column four of the discussors'
table indicate errors which could be as large as 2 percent of installed Manuscript received September 5, 1980.

You might also like