You are on page 1of 8

1118 IEEE Transactions on Power Systems, Vol. 8, No.

3,August 1993

A CONDITIONAL PROBABIUTY APPROACH TO THE CALCULATION OF


FREQUENCY AND DURATION INDICES IN COMPOSITE REUABILIlY EVALUATION
A.C.G.Melo M.V.F.Pereira A.M.Leite da Silva
Centro de Pesquisas de Energia Eletrica Power Systems Research Inc. Catholic University of Rio de Janeiro
Rio de Janeiro Rio de Janeiro/New York Rio de Janeiro
Brazil Brazil/USA Brazil

Abstract - This paper describes a new methodology for frequency The frequency of failure, LOLF, is evaluated as the expected value of
and duration (F&D) assessment in composite generation and the following test function [4]:
transmission reliability evaluation. The proposed approach uses the
concept of conditional probability to characterize the contribution of if x is a success state, then F(x) = 0;
each component to the frequency indices, and allows the calculation if x is a failure state, then F(x) = sum of transition rates
of F&D indices at both system and bus level. The algorithm is easy between x and all success states which can be reached from x
to implement, and requires the same computational effort as the in one transition
estimation of LOLP and EPNS indices. Case studies with
utility-derived systems are presented and discussed. Finally, the loss of load duration (LOLD) is defined as the ratio
LOLP/LOLF
Keywords Reliability, composite reliability, frequency and duration
indices, probability, Monte Carlo simulation, large scale sytems
Diffiarlties in the Computation of F&D Indices in
Composite Reliabilii

The computational effort in calculating reliability indices is directly


related to the difficulty of calculating F(x). As shown previously, the
Calculation of Reliability l n d i i
function associated to the calculation of LOLP and EPNS is
implemented through an adequacy assessment (power flow and
The problem of calculating reliability indices is equivalent to
remedial actions) for each selected state.
calculating the expected value of a given test function [l]:
In contrast, the F&D calculation requires the identification of all
success stat= which can be reached from x in one transition. In other
words, if x has rn components, each with two states, we have in
where: principle to carry out m + 1 adequacy assessments to update the
x m-vector representing the ystem state; each component frequency estimate. The number of evaluations increases further if
{ xk, k = 1, ... , m } in x represents the state of a syszem there are components with multiple states. Depending on the
element (e.g. generators, circuits or loads) characteristics and dimensions of the system being analyzed, this
X state space, i.e. the set of all possible states x arising from the approach may become computationally infeasible. For this reason,
combinations of component states F&D indices are usually calculated in a simplified way, either as
P(x) probability of state x ; transitions among component states are upper bounds [3,5], or by limiting the analysis to a small set of
usually represented by Markov models [2,3] outages in the system [6]. More recently, a technique for alleviating
F(x) test function; its objective is to verify whether that specific the computational effort of F&D has been proposed, based on the
configuration of generators and circuits is able to supply that use of sensitivity information associated to the remedial actions
specific load; because the system state vector is a random model [4].
variable, the test result is also a random variable.
This paper describes a new formulation of F(x) for F&D assessment
For example, the system loss of load probability (LOLP) corresponds in composite generation and transmission reliability evaluation. The
to the expected value of an indicator function F(x), where F(x) = 1 if proposed approach is based on the concept of conditional probability,
x is a failure state (i.e. if there is no load curtailment associated to and allows the calculation of F&D indices at both system and bus
that state); otherwise, x is a success state, and F(x) = 0. In turn, if level. The algorithm is easyly implemented through either state
F(x) is the amount of load curtailment associated to state x , E(F) enumeration or Monte Carlo simulation, and requires the same
represents the system expected power not supplied (EPNS). The computational effort as the estimation of LOLP and EPNS. The
most widely used indices, i.e., the loss of load expectation (LOLE) methodology is illustrated in case studies using nonsequential Monte
and the expected energy not supplied (EENS), can be respectively Carlo simulation. The utilityderived test systems comprise the IEEE
obtained multiplying the LOLP and EPNS indices by the study Modified Reliability Test System, the Brazilian
period. Southern/Southeastern system and the Boneville Power
Administration system.

92 SM 425-9 PWRS A paper recommended and approved


by the IEEE Power System Engineering Committee of A NEW F m FOR FEQUENCY CALCULATIONS
the IEEE Power Engineering Society for presentation
at the IEEE/PES 1992 Summer Meeting, Seattle, WA, Reference [7lshows that the LOLF can be expressed as:
July 12-16, 1992. Manuscript submitted January 31,
1992; made available for printing June 7, 1992. m
LOLF = LOLFk
k=l
where LOLFk is the contribution of component k to system
frequency of failure. Assuming that the component is represented by
a two-state model (a= 1, up; xk = 0, down) and that state of xk is
statistically independent from the rest of the system, we have:
0885-8950/93$03.000 1992 IEEE

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.
1119

where:
B set of states in which the system is good if ~k = 1, and fails if
Xk = 0
pk repair rate of component .a

Expression (3) can be further developed when the system is coherent


where D' indexes the components which are down in failure state x'.
[8]: if a failed component is repaired, the system performance never
It is shown in the Appendix that expressions (9) and (10) are similar
gets worse; conversely, if a working component fails, the system
to the concept of incremental frequency used in generation capacity
performance never becomes better. The assumption of coherence is
reliability evaluation [11,12,13].
always valid for generators in either single-bus or composite
reliability analysis, when a linearized power flow model is used.
Equations (9) and (10) can be used to define a new test function
However, it may not be true for transmission circuits in some
F(x):
situations. This issue was discussed in Ref. [4], which showed that,
for the three systems analysed, network noncoherence had a
if x is a success state, then F(x) = 0
negligible effect on the overall system reliability indices.
(11)
Under the coherence assumption, it can be shown that [9,10]:

where Ox and di index the components which are down and up,
c P(x'(Xk=O) = P(SFk=O) - P(SFIXk=l) (4) respectively, in state x.
icB
With F(x) as in (ll),it is no longer necessary to carry out additional
where SF indicates system failure.
adequacy analysis to test the component transitions. In other words,
we can evaluate the F&D indices with the same computational effort
Substituting (3) and (4) into (2)' we obtain the expression of LOLF
as LOLP and EPNS.
in terms of conditional probabilities:
As the components are represented by a two-state Markov model, the
m
system is balanced in frequency, i.e the transition frequency between
LOLF = [P(SFIn=O) - P(SFIx?.=l)] P(xk=O) pk (5) two states is the same in both directions [12]. As a consequence,
k=1
expression (11) simplifies to:
Replacing { P ( a = 0)) by (1 - P(xk = 1)} in (S), and rearranging if x is a success state, then F(x) = 0,
terms. we have:
otheIwise,F(x) = pk - Xk
kcDX kcdi

The above concepts will be illustrated for the simple hvo-component


system shown in Figure 1. The transition and repair rates are
Note that P(SF) in (6) corresponds to the system LOLP. Therefore,
denoted by X and p.
it can also be expressed as the sum of probabilities of all failure
states:

P(SF) = LOLP = E P(x')


i 2
SUCCESS
where 9 is the set of failure states. The term P(SFbk = 1) in (6)
can also be represented as a summation over the up components in
failure states, that is, FAILURE

where rr' is the set of components which are up in the failure state xi. -
Figure 1 Two-Component System
Expression (8) can be rewritten using b y e s ' formula for conditional
probability By inspection, the system LOLF is:

m LOLF = p3 p2 + p4 pz = (p3 + p 4 ) p2
LOLF = c P(xj c p k - c c .[p(xj/p(a=1)1w
icy k=l icykcrr' Using expressions (12) and (l),we have:
(9)
. m LOLF = PI x 0 t ~2 x 0 t ~3 (p2 - XI) t p4 (/A + p ~ )
= P(X> [ pk - . pk/P(Xk=l)]
icy k=l kcli' = p3 p2 + p4 p2 + @4 pl -p3 Xl)

The above equation can also be expressed in terms of the up and As the components are represented by a two-state Markov model,
down components in each failure state: p4 pi = p3 xi. Hence,

LOLF = @3 + p4) p2
as obtained previously.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.
1120

Note that, although the transitions between failed states appear in if x is a success state, then F(x) = 0;
the above expressions, they cancel out in the summation. A more
complex example with multistate components is shown in the next m
section. otherwise, F(x) = [X i - Ai]
k= 1

D(TENSK)N OF F&J TO W L l l 4 T A T E COMPONENTS where:


mk
In the definition of a new test function F(x), we assumed that the I: Xju transition rate of component xk from j to
system components were represented by a twu-state Markov model. = u=j+l
?he methodology can be extended to represent multi-state "higher" states
components. Multi-state Markov models have been used, for j-1
example, to represent derated states of generating units and multiple A- = I: X j v transition rate of component xk from j to
load levels (see, e.g. [3,6,21,22]). v=l
"lower" states
For ease of presentation, equation (2) is reproduced below:

m The application of (17) will be illustrated for the system of Figure 2,


LOLF = LOLFk (13) with m = 2 components, each modeled by three states. The transition
k=l rates are denoted by A and 7 .

where LOLFk represents the contribution of component xk to system


frequency of failure. Assuming that xk is now modeled by mk states,
we have:
FAILURE SUCCESS

where:
j indexes the possible states of xk = (1, 2, .... mk}; the following
convention is used: if xk is a generator or circuit, xk = 1
corresponds to the state with the smallest capacity, if xlr is the
system load, XIS = 1 corresponds to the highest load level
Xjl is the transition rate of component xk from j to 1.
B c X is the set of states in which the system is failed when
j, but is good whenxk = 1.
xk =

Observing that

mk
c P(xib=j) = c [P(sFIxk=j) - P(SF~=I)I (15)
icB I = j +1

and replacing (14) and (U) into (13), we have:


; j
, .
m mr-1 mk !. !
+.. .........................
!..........
LOLF = I: [P(SF(xk=j)-P(SF(xk=I)]P(xk=j)Xjl (16) !
k = l j =1I = j + 1 j
Expression (16) also allows the definition of a test function F(x) for
the calculation of frequency failure in systems with multiple-state Figure 2 - System with Two Multi-State Components
components.
2 3
Suppose that x , x , x6, x7 and x8 are failure states, and that the
It is shown in the Appendix that:
remainder are success. By inspection, the LOLF for this system is:
if x is a success state, then F(x) = 0; LOLF = pZ(A~P713)+ p3Xu + p 6 ( X \ p + ~ ~ +) p7X13 + ps(X~z+Au)
(17)
m mk j-1 (17,we have:
Using (1) and
otherwise, F(x) = I:[ C Xju - I:(P(xk=v)/P(xk=j)) Xvj]
k=lu=j+l v=l LOLF = PZ[AB - P(X1=1)/P(Xl=Z)X12t 712+713] i
p3[X12+A13+712+713] t
It should be observed that the evaluation of F(x) through (17) does p q X 2 3 - P(Xl=1)/P(Xl=2)X12+ 723 P(x2=1)/P(xz-2)712]+
not require additional adequacy assessments to test the component
~

p7[A12tX12+723 - P(Xz=l)/P(x2=2)712]t
transitions. In this way, the F&D indices require the same
pqX12+A13 P(XZ=l)/P(X2=3)713 P(XZ=2)/P(XZ=3)723]
computational effort as the methods for LOLP and EPNS.
~ ~

LOLF = p2(Aa+ 713) + ~3x13 + 723) + p7X13+ ps(Xl2+ XU) +


If the system is balanced in frequency, we have:
[p3 p2P(Xl=l)/P(X1=2)]A12+ (p7. p6P(xl=I)/P(x1=2)]A12 t
~

[p2 - psP(Xz=l)/P(x2=2)]712 t [p3 - p7P(Xz=l)/P(x2=2)]712 t


(P(xk=v)/F'(xk=j)) ~ v=
j ~ j v (18) [p3 p8P(x2=1)/P(x2=3)]
~ 713 t [p7 - p8P(x2=z)P(Xz=3)1723

And expression (17) can be further simplified to:


Cancelling terms, we obtain the desired LOLF expression. As in the
example of Figure 1, note that the internal transitions between failed
states cancel out in the summation.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.
1121

Load Lovoi (96)


CALCUIATK)N OF F&D INDICES AT Bus LML 100 r i

The extension of this approach to identify the boundary wall between


failure and success states associated with each bus, i.e., to calculate 90
F&D indices at bus level, is straightforward it is only necessary to
characterize as failure states those states in which there is load
curtailment at that bus [14]. 80

For example, if the components are represented by a -state


model, the frequency of failure at bus i is given by the expected value 70

of the following test function Fi(x):

if there is no load curtailment at bus i in state x, then 80

Fi(x) = 0,
(20)
otherwise,A(x) = E w k - Xk 60

kcD" kc#
where D" and v" index the components which are down and up, Figura 3 - Load Curvo of MRTS.
respectively, in state x.

Note that the bus F&D indices can be evaluated simultaneously with Load Lovoi (%I
the system F&D indices. 100

CASE STUDIES 90

1,
Test Wrns
80

The proposed methodology will be illustrated with three systems: the


Modified IEEE Reliability Test System (MRTS) [15], the Brazilian
70
Southern/Southeastem ( S E ) system [16] and the Boneville Power
Adminisration (BPA) system 111. The S E and BPA systems were
derived from utility data. The MRTS system is a modification of the
IEEERTS [18] with the objective of stressing the transmission 80 u i#uli,Ai;;;;
network. Table 1 summarizes the basic characteristics of the three
systems.
1 16 29 4$ 67 71 86 99 118 127 141 166

-
Table 1 Basic Characteristics of Test Systems
Flgura 4 -Load Curvo of SSE.
Characteristics MRTS SSE BPA

Number of buses 24 124 465


Number of circuits 38 272 679 Load Lovol(%)
100
Number of plants 14 106 104
Number of generators 32 393 561
Installed cap.(Mw) 6 810 56 842 45 214 DO
Peak load (MW) 5650 51158 39754
Number of load levels 23 25 24

Generators and circuits are represented by two-state models, and the


Eotuv v v U v v
load is represented by multi-state models unbalanced in frequency .
Figures 3 to 5 present the hourly load curves for each of these
systems, corresponding to a week, and expressed in p.u.

(hutline of Monte Carlo Simulation

The indices will be estimated by nonsequential Monte Carlo


simulation [19]. In this approach, states x c X are sampled from their 1 16 29 43

joint probability distributions. The expected value of F, E Q , is Hours


estimated as: Figura 6 - Load Curvo of BPA.

estimator, which is an average of NS observations of F(x), is also a


Because expression (21) is valid for any type of function F, it random variable. The uncertainty around the estimate is given by the
provides unbiased estimates of the same reliability indices as the variance of the estimator.
enumeration approach, including the F&D indices.
V ( m 7 > = V(fl/NS (22)
It is important to observe in (21) that E(F) is not the "true"
(population) expected value E(F) (which is usually unknown), but an where V(F) is the variance of test function.
estimate of this value. In other words, if the experiment is repeated
with a different random sample, a different value would be obtained Expression (22) indicates that the uncertainty of the estimate
depends on the variance of the test function, V(F), and is inversely
for the estimate E(E). Because x and F(x) are random variables, this proportional to the sample size. This confirms the intuitive notion

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.
1122

that accuracy of a Monte Carlo experiment increases with larger Although the results in Table 4 are system-dependent and, also,
sample size NS.This uncertainty is often represented as a coeffcknt depend on the degree of detail in the load curve representation, they
of variation: seem not to support the suggestion of reference [21] of not
considering the transitions between load levels in the computation of
reliability indices.
(23)
Table 4
The coefficient of variation is used as a measure of the accuracy of
Relative Contribution of G, T and L to System LOLF
the estimates. Further details can be found in [1,19].
Relative Contribution (%)
Basic Reliability l n d i
System G T L Total
The adequacy evaluation of each sampled state was carried out
MRTS 12.0 0.6 87.4 100.0
through a linearized power flow model, coupled to a LP-based
SSE 5.7 3.6 90.7 100.0
remedial actions model [20]. All results were calculated for single
BPA 9.4 32.2 58.4 100.0
precision in an IBM/4381 computer.

The reliability indices for the MRTS and SSE system were obtained
Bus Reliability Indices
for a sample of twenty thousand observations; for BPA system, a
Tables 5 through 7 present the estimated bus LOLP, EPNS (MW),
sample s u e of ten thousand was used. Table 2 presents the LOLP
and EPNS indices for the three systems. The relative accuracy of the LOLF (x occ./hour) and LOLD (hours) indices and the
estimates is given by the coefficient of variation of the estimator, B, associated relative uncertainty p (%) for the three systems. Due to
also shown in the Table. As mentioned in the Introduction, the space limitations, only the data for the five buses with highest EPNS
LOLE and EENS indices can be respectively evaluated multiplying are shown. From these results we observe that bus load indices have
the LOLP and EPNS indices by the study period, which is 168 hours higher uncertainties than system indices, and that the LOLF
in this case. Note, however; that the coefficients of variation of estimates have roughly then same uncertainties as the EPNS
LOLE and EENS estimates are the same as LOLP and EPNS, estimates. Therefore, the calculation of LOLF index does not require
respectively. additional samplings once an acceptable EPNS estimate is obtained.

-
Table 2 Basic Reliablity Indices
Table 5 - Bus Reliability Indices - MRTS
LOLP EPNS
System Value B(%) Value B(%) Index BUS Number

MRTS 0.043 3.4 6.29 5.7 8 14 4 6 5


SSE 0.047 3.2 26.32 5.2
BPA 0.027 6.0 5.39 10.9 LOLP 0.014 0.017 0.006 0.016 0.004
pblp 5.8 5.3 9.4 5.5 10.8
Frequency and Duration hdices EPNS 2.765 1.284 0.498 0.476 0.300
&ns 6.9 7.1 11.1 11.0 13.9
LOLF 2.735 3.514 1.080 3.983 0.737
Table 3 present the indices LOLF (x occ./hour) and LOLD
pbU 9.8 8.5 16.1 7.7 19.7
(hours) obtained with the proposed approach for the MRTS, SSE
LOLD 4.3 5.0 5.2 4.0 5.8
and BPA systems. The Table also shows the coefficients of variation
pbld 7.8 6.5 12.6 5.3 15.7
B of each estimate.

Table 3 - LOLF and LOLD Indices


Table 6 Bus Reliability Indices - SSE
LOLF LOLD
Index Bus Number
System Value a(%) Value p(%)

MRTS 8.80 5.3 4.8 4.0 200 246 234 1 70


SSE 26.85 3.8 1.7 21
LOLP 0.018 0.026 0.016 0.015 0.023
BPA 8.44 10.9 3.2 9.0
pblp 5.1 4.3 5.6 5.8 4.6
EPNS 5.787 2.708 1.970 1.529 1.149
These results indicate that the estimation of F&D indices considering
&pns 6.3 6.9 6.9 10.1 5.9
multi-state components unbalanced in frequency, such as the system
LOLF 11.61 16.00 9.296 9.233 14.17
load curve, is computationally feasible with the proposed
I methodology. pbu 5.8 4.9 6.5 6.5 5.3
LOLD 1.6 1.7 1.7 1.6 1.6
pbld 2.8 2.4 3.1 3.0 2.5
klative Contribution of Generation, Transmission and Load
Transitions to oarall system LOLF

It was shown previously that the LOLF can be obtained from the Table 7 Bus Reliability Indices - BPA
contribution of each system component to the system frequency of
Index BUS Number
failure. It was also seen that the system components can be divided
into generators, circuits and load. merefore, it is interesting to study
the relative contribution of each of these types of transition to the 257 356 395 425 308
global system LOLF.
LOLP 0.001 0.005 0.008 0.002 0.003
pb~p 27.7 14.5 10.8 20.4 19.6
Table 4 presents the relative contribution of generation (G),
transmission (T) and load (L) transitions to the MRTS, SSE and EPNS 0.652 0.525 0.510 0.491 0.413
BPA systems. Results indicate that, for the systems analysed, the &pns 41.0 17.3 18.2 23.4 23.8
major contribution comes from load transitions. ?he next most LOLF 0511 1.896 3.117 1.286 0.875
relevant contribution comes from generation in the MRTS and SSE pbu 43.0 23.0 17.7 28.0 33.3
systems, and from circuit transitions, in the case of BPA. LOLD 2.5 2.5 2.7 1.9 3.0
pbid 19.9 16.2 13.3 15.0 22.4
1123

[2] J.Endrenyi, Reliability Modeling in Electric Power Systems, John


Wiley and Sons,New York, 1978.
One popular method of calculating the LOLF index is the simple [3] RBillinton, R N A I a n , Reliability Evaluation of Power Systems,
sum of frequencies of occurence of failure states. In this approximate Pitman Advanced Publishing, New York, 1984.
method the internal transitions between failure states are also [4] A.C.G.Melo, M.V.F.Pereira, A.M.Leite da Silva, "Frequency
included, leading to an upper bound to the frequency indices. and Duration Calculations in Composite Generation and
Ref. [4] proposed the use of Lagrange multipliers to identify the Transmission Reliability Evaluation", 1991 IEEE PES Summer
boundary wall between failure and success states. This method was Meeting, Paper 91 SM 44@8 PWRS, San Diego, USA, August
shown to be accurate and computationally efficient, but required an 1991.
optimization-based adequacy evaluation model. [SI RBillinton, T.KP.Medicherla, "Overall Approach to the
Reliability Evaluation of Composite Generation and
Table 8 compares the computational effort, measured by the required Transmission Systems", IEE Prmeedings Pt. C, Vol. 127, No. 2,
number of adequacy assessments, and the acurracy of LOLF 1980.
(x 10E-3 occ./hour) estimate using the three methods. The three test [6] A.P.Meliopoulos, "Bulk Power Reliability Assessment with the
systems are the same, but with a single load level (MRTS 5130 - RECS Program", Proceedings 1985 PICA Conference, pp 3846
Mw, SSE - 48316 Mw; BPA -
39 754 MW). Data for the [7] C.?gh, "Calculating the Frequency of Boolean Expression
approximate method and Lagrange approach were extracted from Beine 1". Correspondence IEEE Trans. on Reliability, Vol. R-26,
Ref. [4], for methods M1 and h45,respectively. No. &ember 197.
M.N.Fardis, C.A.Cornel1, "Analysis of Coherent Multistate
Systems", IEEE Trans. on Reliability, Vol. R-30, N0.2, June
-
Table 8 LOLF Indices 1981.
C.Singh, "Calculating Time-Specific Frequency of System
System Method Value p(%) Adequacy Asses. Failure", IEEE Trans, on ReIiability, Vol. R-28, No. 2, June
1979.
Approximate 5.74 2.7 20,000 C.Singh, "Rules for Calculating the Xme-Specific Frequency of
MRTS Lagrange Multip. 1.94 3.3 20,616 System Failure", IEEE Trans. on Reliability, Vol. R-30, NO. 4,
Conditional Prob. 1.97 4.1 20,000 October 1981.
X.Wang, C.Potle, "A Concise Frequency and Duration
Approximate 69.94 4.0 10,000
Approach to Generating System Reliability Studies", IEEE
SSE Lagrange Multip. 11.48 5.6 10,721
Trans. on PAS, Vol. PAS-102, No. 8, August 1983.
Conditional Prob. 10.52 6.9 10,000
C.Singh, "Forced Frequency Balancing Technique for Discrete
Approximate 46.60 3.1 10,000 Capacity System", IEEE Trans. on Reliability, Vol. R-32,
BPA Lagrange Multip. 9.06 3.7 10,191 October 1983.
Conditional Prob. 9.75 4.9 10,000 A.M.Leite da Silva, A.C.G.Melo, S.H.F.Cunha, "A Frequency
and Duration Method for Reliability Evaluation of Large Scale
Hidrothermal Generating Systems", IEE Procceedings Pt. C,
As discussed in Ref. [4], the approximate method had fairly high
Vol. 138, No. 1, January 1991.
relative e r r o s for these test systems. The Lagrange approach was
A.C.G.Melo, Frequency and Duration Calculations in
accurate with the overhead of performing 616, 721 and 191 additional
Composite Generation and Transmission Reliability Evaluation
adequacy evaluations for the MRTS, SSE and BPA systems. The
for LaTe Scale Systems, DSc Thesis, Catholic University of Rio
conditional probability approach had the same estimated indices as
de Janeiro, October 1990 (in Portuguese).
the Lagrange approach within the accuracy limits, with the same
EPRI, Composite System Reliability Evaluation Metho&, Report
computational effort as the approximate method. The conditional
EL-5178, June 1987
probability approach is also more flexible as it does not require an
S.H.F.Cunha, G.C.Oliveira, M.V.F.Pereira, V.LAnenti,
optimization-based adequacy evaluation model.
A.C.G.Melo, "Composite Generationflransmission Reliability
of the Brazilian Southern/Southeastern System", 2nd PUAPS,
San Francisco, USA, September, 1988.
CONCLUSlONS
EPRI, Development of a Monte-Carlo Based Composite
Reliability Evaluation Program - the CREAM Model, Report
This paper described a new methodology for F&D assessment in
EL6926,August 1990.
composite generation and transmission reliability evaluation. The
Subcommittee Report, "IEEE Reliability Test System", IEEE
proposed approach is based on the concept of conditional
Transactions on PAS, Vol. PAS-98, November/December 1979.
probability, and allows the calculation of F&D indices at both system
J.Kleijnen, Statistcs Techniques in Simulation - Part I, Marcel
and bus level. The algorithm was shown to be easy to implement, and
Decker, New York, 1974.
to require the same computational effort as the estimation of LOU'
B.Stott, J.L. Marinho, "Linear Programming for Power System
and EPNS. The approach was illustrated in case studies with
Network Applications", IEEE Transactions on PAS, Vol.
utilityderived systems.
PAS-98, May/June 1979
EPRI, Reliability Evaluation for Large Scale Bulk Transmission
Systems, Report EL5291,1988.
RBillinton, C.Singh, "System Load Representation in
Generating Capacity Reliability Studies - Part II: Applications
Part of this work had the technical and financial support of the
and Extensions", paper T 72 072-2, presented at the IEEE
Transmission Dept. of Eletrobras and of the multi-utility Reliability
Winter Meeting, New York, NY,January %February 4, 1972
Working Group (project "2). The authors gratefully acknowiedge
the suggestions and contributions of G.C. Oliveira of Cepel.
APPENDIX

Use of the Concept of Incremental Frequency


[l] M.V.F.Pereira, LM.V.G.Pinto, S.H.F.Cunha, G.C.Oliveira,
"Monte Carlo Based Composite Reliabiulity Evaluation - The index LOLF in (2) and (13) was defined as the sum of failure
frequencies due to each system component. From (2), after algebraic
Modeling Aspects and Computational Results", in Reliability
manipulations, we arrived at (lo), where LOLF is expressed as the
Assessment of Composite Generation and Transmission Systems,
sum of contributions of each failure state to system frequency of
IEEE Tutorial Course 9OEHO311-1-PWR, February 1990.
failure.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.
1124

It should be observed that this last definition of LOLF is similar to


p(xi) = ~(x'lxk= j) ~ ( x k= j)
that used in generation capacity reliability, i.e.
P(xn) = P(xnl xk = v ) P(xk = v )
. .
LOLF = r(x')
icx' Noting that the states xi and xn differ only in component xk,

where the term r ( x 5 is known as incremental frequency [U]. In P(X' Ixk = J3 f P(X" Ixk = V)
references [12,14], the definition of incremental frequency was
extended to multi-state components. From (27) and (28) we conclude that

In generation reliability, the incremental frequency was defined as the


difference of frequency between two accumulated states. Therefore, it
is related to the process of accumulation (combination) of states, i.e.
Substituting (29) into (U),we have:
it depends on the order in which states are combined. Thus, the
incremental frequency associated to the state xi being combined is m mk
given by [14]: y(xj = P(X~ E c Xju -
k = l u= j +1
= c ifu" - m j-1
~(x') c c
~ ( x k= v)/~(xk = j)
mas+ ncS-
k=l v=l
= P(x5 c A; - c P(xn) A;
m mk i -1
mcS+ ncS-
k=lu=j+l v=l
where:
Finally, from (30) and (24), we amve at the following expression for
LOLF:
f i = frequency and transition rate from state xi to state x"

fni, X ', = frequency and transition rate from state xp to state xi


m LOLF = P(xi)
m
[
mk
Xju
j-1
-
P(Xk = V)/P(xk = j ) Xvj ]
icx' k=l u = j + l v=l
S i = set of states not yet combined
(31)
S! = set of states already combined
This last expression leads to the definition of F(x) in (17).

It should be observed that, by defining the incremental frequency of


failure on the basis of the coherence property, the order of the HOGRAPHIES
accumulation process does not matter. Therefore, in expression (U)
Albefi C Gtbcr de Melo (M'91) was born in Manaus, Amazonas,
the sets S f e S! associated to the failure state xi are defined as
Brazil, in 1962. He received the BSc degree from Federal University
follows: of Pernambuco in 1983, and MSc and DSc degrees from PUC/RI in
1986 and 1990 respectively, all in Electrical Engineering. Since 1985
he has been with CEPEL, the Brazilian Electric Power Research
S i set of states coming from transitions which potentially can
Center, working in the areas of power system reliability and
cross the boundary wall, e.g. transitions of generators from expansion planning.
smaller to higher capacities, transition from a higher lower
load level to a lower load level Mario Veiga F. Pexeira has a BSc degree in Electrical Engineering
from PUC/W and MSc and DSc degrees in Systems Engineering
S! set of states coming from transitions which, due to coherence,
(optimization) from COPPE/UFRI. From 1976 to 1991 he was with
do no cross the boundary wall CEPEL, where he coordinated projects in the areas of power system
planning and operation. From 1983 to 1985 he was a project
manager at EPRI, where he initiated and managed several research
Consider now a failure state xi where component xk is at state j , i.e.,
projects in the planning and operation areas. From 1988 to 1990 he
xk = j. The states xm c S\ coming from transitions of component xlr, was with the EE Dept. of PUC/RI on sabbatical leave, where he
developed research in large scale and stochastic optimization. He is
have all components in the same state as in x', except component xk, now with PSRI and is a consultant for several Brazilian and
which is in one of the states xk = (i+l, j + 2 , ..., mk}. Analogously, American utilities. He has authored and coauthored about 200
technical papers, 60 technical reports, and textbooks in power system
the difference between states xn c S! , coming from transitions of
applications and optimizations. In 1985, he was one of the recipients
of the Franz Edelman Award for Management Science Achievement,
component xk, is that in xi, component xk is in one of the states
granted by ORSAmMS, for his work on stochastic optimization
xk = (1, 2, ...,j-1). In this way, (U)can be rewritten as:
applied to hydro scheduling.
m mk m j-1
Armando hk k i t e da Silva (S'V,M'78,SM'91) was born in Rio de
f"(xi> = p(xi) c c Xju - P(X") E c ~ v j (26) Janeiro, RI, Brazil, in 1954. He obtained his BSc degree from
k=l u=j+l k=lv=l
PUC/RI in 1975, his MSc degree from the Federal University of Rio
de Janeiro (COPPE/RJ) in 1977, and his PhD degree from the
where:
University of Manchester - UMIST, England, in 1980. He is a
Xju = transition rate of component xk from state j to state U
Professor in the Electrical Engineering Dept. of PUC/RI. He worked
Xvj = transition rate of component xk from state v to state j
as a visiting researcher at the Ontario Hydro Research Division,
Canada, from 1990 to 1991.
As the components are statistically independent,
1125

DISCUSSION AC.G. Melo, M.V.F. Pereira and AM. b i t e da Sihm We thank the
discussers for their interest in our work, and would like to offer the
C. Singh 8 N. Gubbala (Texas A & M University): The discussers following comments.
would like to congratulate the authors on this interesting and
important paper. We have some suggestions and questions. First, With respect to the first paragraph of comment # 1, we feel that the
however, we would like to put some of the statements in a proper references were adequate and correctly placed in the section
perspective. describing the main equations. Because the core of our work was
derived independently, we tried very hard to track all references to
This paper is based on the application of equation (10) for earlier work in this area, such as Ref. 17, published in 1977.
two state and equation (31) for multi-state components. Therefore, we do not share the discussers' impressions on the
Although the relevant papers are given in the Reference insufficient acknowledgement of related work. With respect to the
section, this paper however, seems to leave an impression that second paragraph, we think that one of the most attractive aspects of
equation (10) is new. We would like to note that this equation the proposed approach is the simplicity of the final expressions.
is identical to equation (13) in Ref [lo]. Similarly equation Therefore, not only the multi-state expressions, but all expressions in
(12) is the same as equation (U) in Ref [lo]. the paper, are straight forward, once the underlying concepts are
understood.
The extension of equation (10) to multi-state components is
straight forward once the structure of equation (10) is
With respect to comment # 2, the second author (Pereira) first came
understood. The positive portion gives the frequency due to
all repairs. The negative portion cancels those expected
across the concept of test functions - and their actual name in a-
paper by Professor Felix Wu, part of a 1977 EPRI Conference
transitions which d o not exit the boundary. For a multi-state
Proceedings. This concept has been used - with due
component in state j, the states j + 1 to mk are repaired states
with respect to j and states 1 to j - 1 are failed states with
-
acknowledgement to Professor Wu's work in several of Pereira's
earlier works on reliability. Therefore, we could not have possibly
respect to j. Using these concepts equation (31) directly
claimed that the concept of a test function is new. What we have
follows from equation (10). Here again the positive portion
proposed is the use of the efficient test function given by equations
gives the frequency due to the repair from j where as the
(11) and (17) to calculate F&D indices, rather than the conceptual
negative portion cancels those expected transitions that do not
definition of the test function for frequency, based on the analysis of
cross the boundaly.
all transitions from a given state.
To our knowledge, the concept of test function for frequency
With respect to load modeling, we completely agree with the
is not something new. It is described in Ref [Al, page 121 and
discussers that the proportional variation of load at all buses ensures
Ref [A2,page 6 7 , although it is called a discrete random
that load transitions are coherent. With other load models, a
variable instead of test function.
noncoherent behavior is theoretically possible. We have been
carrying out evaluations of the possible impact of load noncoherence,
These comments are made to provide clarifications and perspective
using a similar approach as our investigation of the noncoherence of
and are not meant to detract from the contributions of this paper.
circuit outages, reported in Ref. [4]. This work is being carried out
This paper makes a significant contribution to composite system
in cooperation with the Brazilian utilities, and in particular with
reliability based on the rules for converting probability into
SGC, which is a multi-utility working group on reliability issues.
frequency [lo]. As the authors are aware, however, these equations
are valid only when the system is coherent. This naturally leads to a
A very important part of this work is modeling the system load at
question on load modeling. The authors have shown in Table 4 that
bus level. We have been investigating a aggregation/decomposition
load transitions are significant contributors to frequency. The system
approach, borrowed from earlier stochastic streamflow modeling
remains coherent if at any given time the loads at all buses move in
work described in Ref. [Bl]. The idea is to model the stochastic
the same direction. The authors have assumed in their studies that
process of some "key" aggregate loads. The aggregation may range
the loads do move in the same direction at all buses. This is,
from a single bus, to a group of buses, to whole regions. Each
however, not true in reality. Have the authors studied the impact of
aggregate load is then disaggregated in each time stage, so as to
violation of this assumption. A combination of analytical
preserve spatial correlation. This aggregation/decomposition scheme
decomposition (first phase) and non-sequential simulation (second
has some similarities with the cluster based model proposed in Ref.
phase) have been used in [A3,A41 for calculating frequency and
duration indices in multi-area reliability studies. Reference A4 [A4]. Other load models are also been considered [B2].
(Tables 9 & 12) has shown that significant savings in CPU time can
Once again, we would like to thank the discussers for their interest in
be achieved for calculating the F & D indices if loads are perfectly
our work.
correlated. It has, however, further shown that this assumption could
lead to significant error in F & D indices if the area peaks are
shifted by significant duration of time. Also, have the authors References
thought of modeling the correlation of bus loads using cluster based
model [A4].
[sl] M.V.F. Pereira, G.C. Oliveira, C.G. Costa, J. Kelman,
"Stochastic Streamflow Models for Hydroelectric Systems",
The discussers would again like to congratulate the authors for this Water Resources Research, Vol. 20, No. 3, March 1984.
very interesting paper.
[B2] A.M. k i t e da Silva, S.M.P. Ribeiro, V.L. Arienti, RN. Allan,
M.B. Do Coutto Filho, "Probabilistic Load Flow Techniques
References Applied to Power Systems Expansion Planning", IEEE
Transactionson Power System, Vol. 5, No. 4, November 1990.
[All C. Singh, "Reliability Modeling and Evaluation in Electric
Power Systems", Ph.D. thesis, University of Saskatchewan, M a n u s c r i p t r e c e i v e d O c t o b e r 2 , 1992.
Canada, (1972).
[A21 C. Singh and R Billinton, System Reliability Modeling and
Evaluation, Hutchinson Education, London, 1977.
[A31 EPRI Report, Reliability Modeling of Interconnected Systems
Recognizing Operating Considerations, Report EL-4603, Vol.
1, December 1985.
[A41 C. Singh and A. Lago-Gonzales, "Improved Algorithm for
Multi-Area Reliability Evaluation Using the Decomposition
Simulation Approach", IEEE Transactions on Power Systems,
Vol. 4, No. 1, pp. 321-328, Feb. 1989.
M a n u s c r i p t r e c e i v e d August 6 , 1992.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE SAO JOAO DEL REI. Downloaded on November 3, 2009 at 14:07 from IEEE Xplore. Restrictions apply.

You might also like