NOTE: Other “modules” referenced in these notes may be found at http://home.engineering.iastate.edu/~jdm/ee653/ee653schedule.htm.

Also, in the below notes, text in blue will not be covered in class, though you should read them.
Module PE.PAS.U19.5 Generation adequacy evaluation 1
Updated:
2/14/2024

Module PE.PAS.U19.5
Generation adequacy evaluation

U19.1 Introduction
Probabilistic evaluation of generation adequacy is traditionally
performed for one of two classes of decision problems. The first
one is the generation capacity planning problem where one wants
to determine the long-range generation needs of the system. The
second one is the short-term operational problem where one wants
to determine the unit commitment over the next few days or weeks.
We may think of the problem of generation adequacy evaluation in
terms of Fig. U19.1.

Fig. U19.1: Evaluation of Generation Adequacy

In Fig. U19.1, we see that there are a number of generation units,
and there is a single lumped load. Significantly, we also observe
that all generation units are modeled as if they were connected
directly to the load, i.e., transmission is not modeled. The
implication of this is that, in generation adequacy evaluation,
transmission is assumed to be perfectly reliable.
We begin our treatment by first identifying the necessary modeling
requirements in terms of, in Section U19.2, the generation side,
Module PE.PAS.U19.5 Generation adequacy evaluation 2

and, in Section U19.3, the load side. Section U19.4 describes a

common computational approach associated with the generation
capacity planning problem, and Section U19.5 illustrates how this
approach is typically used for capacity planning. Section U19.6
provides an alternative way of computing generation capacity
planning indices. Section U19.7 briefly summarizes three
important issues central to a more extended treatment of the topic.
U19.2 Generator model
In the basic capacity planning study, each individual generation
unit is represented using the two-state Markov model illustrated in
Fig. U19.2.

Up λ=1/m D
=1/r

Fig. U19.2: Two-State Markov Model

This model was described in Sections U16.4-U16.5 and Section
U18.2.3 of modules U16 & U18, respectively (suggest reviewing
U16, U16.4-U16.5), using p (t ) = p(t ) A , where p(t) is a vector of
state probabilities and, with λjk and λj as transition intensities, A is
the transition intensity matrix given by (for the general case)
− 1 12  1n 
 − 2  2 n 
A =  21 
     
 
 n1   nn 

Important relations for this model, in terms of long-run availability

A and long-run unavailability U, are provided here again, for
convenience, where m=MTTF, r=MTTR, µ and λ are transition
rates (number of transitions per unit time) for repair (D to Up) and
for failure (Up to D), respectively; T is the mean cycle time, and f
is a “frequency” which gives expected number of direct transfers
between states per-unit time. In U16.6.2, we show that f= λpup
Module PE.PAS.U19.5 Generation adequacy evaluation 3

where pup is the long-run probability of being up, i.e., it is

p1(t→∞), p(t)=[p1(t), p2(t)]. We also show that
• whereas λ gives the expected number of transfers from Up to D per unit
time, given the system is Up,
• f gives the expected number of transfers from Up to D per unit time, with no
condition specified, and

μ m m f SH
A= pup = = = = = (U19.1)
λ + μ m+ r T λ FOH + SH
λ r r f FOH
U = pD = FOR = = = = = (U19.2)
λ + μ m+ r T μ FOH + SH
In (U19.2), the FOR is the forced outage rate. One should be
careful to note that the FOR is not a rate at all but rather an
estimator for a probability. The terms in the right-hand-expressions
of (U19.1) and (U19.2) are defined as follows:
• Forced outage hours (FOH) is the number of hours a unit was in
an unplanned outage state;
• Service hours (SH) is the number of hours a unit was in the in-
service state. It does not include reserve shutdown hours (a
reserve shutdown exists whenever a unit is available but is not
synchronized1).
Module U18 also describes how one can approximate the effects of
derating (the unit is operating but at reduced capacity due to, for
example, the outages of auxiliary equipment such as pulverizers,
water pumps, fans, or environmental constraints) and of reserve
shutdown (very important for peaking units), by using the
equivalent forced outage rate, EFOR, according to:
forced equivalent forced
+
outage hours derated hours
EFOR =
equivalent reserve
forced service
+ + shutdown forced
outage hours hours
derated hours

1
See Section 4.8.1, “Reserve Shutdowns,” in “Powergads User Manual,” Integ Enterprise Consulting, at
https://www.pjm.com/~/media/etools/egads/egads-user-guide.ashx.
Module PE.PAS.U19.5 Generation adequacy evaluation 4

FOH + EFDH
EFOR = (U19.3)
FOH + SH + ERSFDH
The basis for (U19.3) is not simple, and so we will not address it
here. But it is very well explained in Module U18 (see Fig. U18.3).
U19.2.1 Capacity outage table for identical units
A capacity table is simply a probabilistic description of the
possible capacity states of the system being evaluated. The
simplest case is that of the one-unit system, where there are two
possible capacity states: 0 and C, where C is the maximum
capacity of the unit. Table U19.1 shows capacities and
corresponding probabilities.
Table U19.1: Capacity Table for 1 Unit System

Capacity Probability

C A
0 U

We may also describe this system in terms of capacity outage

states. Such a description is generally given via a capacity outage
table, as in Table U19.2.
Table U19.2: Capacity Outage Table for 1 Unit System

Capacity Outage Probability

0 A
C U

Figure U19.3 provides the probability mass function (pmf)2 for this
one-unit system.
2
A PMF, for a discrete RV, provides for each value that the RV may assume, the probability of occurrence for the corresponding
outcome of an experimental trial. Notationally, we write f X ( x ) to denote the PMF of the RV X. It may be interpreted as P X = x , or,
in words, “the probability that X equals x”, where x is any specific value that X may assume.
Module PE.PAS.U19.5 Generation adequacy evaluation 5

Probability
U

0 C 2C
Capacity outaged

Fig. U19.3: pmf for Capacity Outage of One-Unit Example

Now consider a two-unit system, with both units of capacity C. We

can obtain the capacity outage table by basic reasoning, resulting
in Table U19.3.
Table U19.3: Capacity Outage Table for 2 Identical Units

Capacity Outage Probability

0 A2
C AU
C UA
2C U2

We may also more formally obtain Table U19.3 by considering the

fact that it provides the pmf of the sum of two random variables.
Define X1 as the capacity outage random variable (RV) for unit 1
and X2 as the capacity outage RV for unit 2, with pmfs fX1(x) and
fX2(x), each of which appear as in Fig. U19.3. We desire fY(y), the
pmf of Y, where Y=X1+X2. Recall from Module U13, Section
U13.3.2, that we obtain fY(y) by convolving fX1(x) with fX2(x), i.e.,
Module PE.PAS.U19.5 Generation adequacy evaluation 6


fY ( y ) =  f X 1 (t ) f X 2 ( y − t )dt (U19.4)
−
But, inspection of fX1(x) and fX2(x), as given by Fig. U19.3,
indicates that, since X1 and X2 are discrete random variables, their
pmfs are comprised of impulses. Convolution of any function with
an impulse function simply shifts and scales that function. The
shift moves the origin of the original function to the location of the
impulse, and the scale is by the value of the impulse. Fig. U19.4
illustrates this idea for the case at hand.

A A

Probability
Probability

*
U U

0 C 2C 0 C 2C
Capacity outaged Capacity outaged

A2
+
Probability

Probability

AU AU
U2

0 C 2C 0 C 2C
Capacity outaged Capacity outaged

Fig. U19.4: Convolution of Generator Outage Capacity pmfs

Figure U19.5 shows the resultant pmf for the capacity outage for 2
identical units each of capacity C.
Module PE.PAS.U19.5 Generation adequacy evaluation 7

A2

Probability
2AU
U2

0 C 2C
Capacity outaged

Fig. U19.5: pmf for Capacity Outage of 2 Unit Example

We note that Fig. U19.5 indicates there are only 3 states, but in
Table U19.3, there are 4. One may reason from Table U19.3 that
there are two possible ways of seeing a capacity outage of C, either
unit 1 goes down or unit 2 goes down. Since these two states are
the same, we may combine their probabilities, resulting in Table
U19.4, which conforms to Fig. U19.5.
Table U19.4: Capacity Outage Table for 2 Identical Units

Capacity Outage Probability

0 A2
C 2AU
2C U2

In fact, we saw this same kind of problem in Section U10.2 of

module U10, where we showed that the probabilities can be
handled using a binomial distribution, since each unit may be
considered as a “trial” with only two possible outcomes (up or
down). We may then write the probability of having r units out of
service as:
n!
PX = r = Pr[ X = r , n,U ] = U r ( A) ( n − r ) (U19.5)
r!( n − r )!
Module PE.PAS.U19.5 Generation adequacy evaluation 8

where n is the number of units.

It is interesting to note we may also think about this problem via a
state-space model, as shown in U19.6 where we have indicated the
state of each unit together with the capacity outage level associated
with each system state. Note that we are not representing the
possibility of common mode or dependent failures.
1u, 2d
C out 

1u, 2u   1d, 2d
0 out 2C out

 
1d, 2u 
C out

Fig. U19.6: State Space Model for 2-Unit System

From Section U16.8 of module 16, since the two middle states of
Fig. U19.3 satisfy the merging condition (a group of (internal)
states can be merged if the transition intensities to any external
states are the same from each internal state) and they satisfy rule 3
(two states should be combined only if they are of the same state
classification – in this case, the same capacity), we may combine
them using rule 1 (when two (internal) states have transition rates
that are identical to common external states, those two states can
be merged into one; entry rates are added, exit rates remain the
same.) Therefore, Fig. U19.6 becomes Fig. U19.7.
1u, 2u 2 1d, 2u
 1d, 2d
0 out 1u, 2d 2C out
 C out 2

Fig. U19.7: Reduced State Space Model for 2 Unit System

The 2λ transition in Fig. U19.7 reflects the fact that the “0 out”
state may transition to the “C out” state because of unit 1 or
because of unit 2, but it does not reflect a common mode outage
since the middle state is a state in which only 1 unit is failed.
Module PE.PAS.U19.5 Generation adequacy evaluation 9

Similarly, the 2 transition in Fig. U19.6 reflects the fact the “2C
out” state may transition to the “C out” state because of repair to
unit 1 or repair to unit 2, but it does not reflect a common mode
repair since the middle state is a state in which only 1 unit is
repaired.
One may also compute frequency and duration for each state in
Fig. U19.7 according to (U16.32) and (U16.33) from module U16,
repeated here for convenience:

f j = p j ,   jk (U19.6)
k j

1
Tj =
  jk (U19.7)
k j
Table U19.5 tabulates all of the information.

Table U19.5: Capacity Outage Table for 2 Identical Units with

Frequencies and Durations

Capacity Probability Frequency Duration

Outage

0 A2 2λA2 1/2λ
C 2AU 2AUλ 1/λ
2C U2 2U2 1/2

U19.2.2 Capacity outage table for units having different capacities

Reference [1] provides a simple example for the more realistic case
of having multiple units with different capacities, which we adapt
and present here. Consider a system with two 3 MW units (units 1
Module PE.PAS.U19.5 Generation adequacy evaluation 10

and 2) and one 5 MW unit (unit 3), all of which have forced outage
rates (FOR) of 0.02. (The fact that all units have the same FOR
means that we could handle this using the binomial distribution,
which would not be applicable if any unit had a different FOR).
The pmfs of the two identical 3 MW units can be convolved as in
Section U19.2.1 to give the pmf of Fig. U19.8 and the
corresponding capacity outage table of Table U19.6.

0.9604
Probability

0.0392

0.0004

0 3 6 9 12
Capacity outaged

Fig. U19.8: pmf for Capacity Outage of Convolved 3 MW Units

Table U19.6: Capacity Outage Table for Convolved 3 MW Units

Capacity Outage Probability

0 0.982=0.9604
3 2(0.98)(0.02)=0.0392
6 0.022=0.0004

Now we want to convolve in the 5 MW unit. The pmf for this unit
is given by Fig. U19.9.
Module PE.PAS.U19.5 Generation adequacy evaluation 11

0.98
Probability
0.02

0 3 6 9 12
Capacity outaged

Fig. U19.9: pmf for 5 MW Capacity Outage

Convolving the pmf of Fig. U19.8 with the pmf of Fig. U19.9
results in the pmf illustrated in Fig. U19.10, with the
corresponding capacity outage table given in Table U19.7.

Table U19.7: Capacity Outage Table for Convolved 3 MW Units

and 5 MW Unit

Capacity Description Probability

Outage

0 All units up. 0.980.9604=0.941192

3 U3 up, U1 or U2 down. 0.980.0392=0.038416
5 U3 down, U1 and U2 up. 0.020.9604=0.019208
6 U3 up, U1 and U2 down. 0.980.0004=0.000392
8 U3 down, U1 or U2 down. 0.020.0392=0.000784
11 All units down 0.020.0004=0.000008
Module PE.PAS.U19.5 Generation adequacy evaluation 12

0.941192
Probability
Unit 3 “0 MW capacity outage”
convolved with two 3 MW units pmf
0.038416

0.000392

0 3 6 9 12
Capacity outaged

Unit 3 “5 MW capacity outage”

Probability

convolved with two 3 MW units pmf

0.019208

0.000784 0.000008

0 3 6 9 12
Capacity outaged
0.941192
Probability

Resultant final pmf accounting for all

three units
0.038416
0.019208
0.000392
0.000784 0.000008

0 3 6 9 12
Capacity outaged

Fig. U19.10: Procedure for convolving Two 3 MW units with 5

MW Unit (top two plots) and final 3 unit pmf
U19.2.3 Convolution algorithm
The procedure illustrated above can be expressed algorithmically,
which is advantageous in order to code it.
Module PE.PAS.U19.5 Generation adequacy evaluation 13

Two-state model:
The algorithm is simplest if we assume that all units are
represented using two-state models.
Let k denote the kth unit to be convolved in, Ak and Uk its
availability and FOR, respectively, and Ck its capacity.
The composite capacity outage pmf before a convolution is
denoted by fYold(y), and after by fYnew(y), so that for unit k, the
capacity outage random variables are related by Ynew=Yold+Xk. We
assume that there are N units to be convolved.
The algorithm follows.
1. Let k=1.
2. Convolve in the next unit according to:
f Ynew ( y ) = Ak f Yold ( y ) + U k f Yold ( y − Ck ) (U19.8)
for all values of y for which fYold(y)0 and/or fYold(y-Ck) 0.
3. If k=N, stop, else k=k+1 and go to 2.
Note that in (U19.8) the influence of the argument in the last term
fYold(y-Ck) is to shift the function fYold(y) to the right by an amount
equal to Ck. This corresponds to the shift influence of the kth unit
pmf impulse at Xk=Ck.
U19.2.4 Deconvolution
An interesting situation frequently occurs, particularly in
operations, but also in production costing programs, when the
composite pmf has been computed for a large number of units, and
capacity outage probabilities are fully available. Then one of the
units is decommitted, and the existing composite pmf no longer
applies. How to obtain a new one?
One obvious approach is to simply start over and perform the
convolution for each and every unit. But this is time-consuming,
and besides, there is a much better way! We have a better approach
based on the following fact:
Module PE.PAS.U19.5 Generation adequacy evaluation 14

The computation of fYnew(y) is independent of

the order in which the units are convolved.
Consider, in (U19.8), the term fYold(y). This is the composite pmf
just before the “last” unit was convolved in.
Given we have fYnew(y), we assume that the “last” unit
convolved in was the unit that we would like to decommit.
It may not have been the last unit, in actuality, but because the
computation of fYnew(y) is independent of order, we can make this
assumption without loss of generality.
In that case, we may “convolve out” the decommited unit.
How to do that? Consider solving (U19.8) for fYold(y), resulting in:
f Ynew ( y ) − U k f Yold ( y − Ck )
f Yold ( y ) = (U19.9)
Ak
The problem with the above is that the function we want to
compute on the left-hand-side, fYold(y), is also on the right-hand-
side, as fYold(y-Ck).
There is a way out of this, however. It stems from two facts.
Fact 1: The probability of having capacity outage less than 0 is
zero, i.e., the “best” that we can do is that we have no capacity
outage, in which case the capacity outage is zero. Therefore any
valid capacity outage pmf must be zero to the left of the origin.
Fact 2: fYold(●) is a valid capacity outage pmf.
Implication: For values of y such that 0<y<Ck, fYold(y-Ck) evaluates
to the left of the origin and therefore, since fYold is a valid capacity
outage pmf, it MUST BE ZERO in this range. As a result,
f Ynew ( y )
f Yold ( y ) = , 0  y  Ck (U19.10)
Ak
But what about the case of Ck<y<IC, where IC is the total installed
capacity? Here, we must use (U19.9). But let’s assume that we
Module PE.PAS.U19.5 Generation adequacy evaluation 15

have already computed fYold(y) for 0<y<Ck. Then the first time we
use (U19.9) is when y=Ck. Then we have:
f Ynew (Ck ) − U k f Yold (0)
f Yold (Ck ) =
Ak
But we already have computed fYold(0) from (U19.10)!
And we will be able to use the values of fYold(y), 0<y<Ck, in
computing all values of fYold(y), Ck<y<2Ck. In fact, we will be able
to compute all of the remaining values of fYold(y) in this way!
As an example, try deconvolving one of the 3 MW units from the
capacity outage table of Table U19.7 (which is also illustrated at
the bottom of Fig. U19.10). In this case, C3=3, A3=0.98, U3=0.02.
The computations are given in Table U19.8.
Note that, since fYold(y-Ck)=0 for y<Ck, (U19.9) includes the case
of (U19.10), and we can express the algorithm using (U19.9) only.
The deconvolution algorithm is given below. There is just one step.
We assume that we are deconvolving unit k.
1. Compute:
f Ynew ( y ) − U k f Yold ( y − Ck )
f Yold ( y ) =
Ak
consecutively for y=0, ….,IC such that
fYnew(y)0 and/or fYold(y-Ck)0,
where IC is the installed capacity of the system before
deconvolution.
2. Stop.
Module PE.PAS.U19.5 Generation adequacy evaluation 16

Table U19.8: Computations for Deconvolution Example

Capacity fYnew(y) fYold(y)

Outage y

0 0.94119200 f Ynew (0) .941192

f Yold (0) = = = .9604
A3 .98
3 0.0384160 f Ynew (3) − U 3  f Yold (3 − 3)
f Yold (3) =
A3
.0384160 − 0.02  .9604
= = .0196
.98
5 0.019208 fYnew (5) − U 3  fYold (5 − 3)
fYold (5) =
A3
.019208 − .02  0
= = .0196
.98
6 0.00039200 f Ynew (6) − U 3  f Yold (6 − 3)
f Yold (6) =
A3
.000392 − .02  .0196
= =0
.98
8 0.00078400 f (8) = fYnew (8) − U 3  fYold (8 − 3)
Yold
A3
.000784 − .02  .0196
= = .0004
.98
11 0.00000800 fYnew (11) − U 3  fYold (11 − 3)
fYold (11) =
A3
.000008 − .02  .0004
= =0
.98
Module PE.PAS.U19.5 Generation adequacy evaluation 17

U19.2.5 Multi-state models

We have so far addressed only the case where all units are
represented by two-state models. It may be, however, that we
would like to account for derated units, in which case we need to
address the multi-state model as well. This situation presents no
additional conceptual difficulty relative to the two-state model, as
the pmf for each unit will still consist of only impulses, except
now, each unit will have a pmf consisting of as many impulses as it
has states, instead of only two.
We do, however, need to generalize the algorithms for convolution
and deconvolution.
Convolution algorithm for multi-state case:
With N the total number of units:
1. Let k=1.
2. Convolve in the next unit according to:
nk
f Ynew ( y ) =  pkj f Yold ( y − Ckj ) (U19.11)
j =1

for all values of y for which fYold(y) or fYold(y-Ckj) are non-zero.

Here, nk is the number of states for unit k; pkj is the jth state
probability for unit k; Ckj is the jth capacity outage for unit k.
3. If k=N, stop, else k=k+1 and go to 2.
Note (U19.11) is the same as (U19.8) if nk=2, with Ak=pk1, Uk=pk2.
f Ynew ( y ) = Ak f Yold ( y ) + U k f Yold ( y − Ck ) (U19.8)
Deconvolution algorithm for multi-state case:
We again assume that we are deconvolving unit k. To determine
the deconvolution equation for the multi-state case, rewrite
(U19.11) by extracting from the summation the first term,
according to:
Module PE.PAS.U19.5 Generation adequacy evaluation 18

nk
f Ynew ( y ) = pk 1 f Yold ( y ) +  pkj f Yold ( y − Ckj )
j =2

where we have assumed that the first capacity outage state for unit
k is zero, i.e., Ck1=0. Solving for fYold(y), we have:
nk
f Ynew ( y ) −  pkj f Yold ( y − C kj )
f Yold ( y ) = j =2

pk 1
We assume that we are deconvolving unit k. The algorithm is:
1. Compute:
nk
f Ynew ( y ) −  pkj f Yold ( y − C kj )
f Yold ( y ) = j =2
(U19.12)
pk 1
consecutively for y=0, ….,IC, and y such that fYnew(y)0,
fYold(y-Ckj)0, where IC is the installed capacity of the system
before deconvolution.
2. Stop.
U19.3 Load model
Consider Fig. U19.11, instantaneous demand as function of time.
Load (MW)

300

200

100

Time (days)

Fig. U19.11: Instantaneous demand vs. time

Module PE.PAS.U19.5 Generation adequacy evaluation 19

Although this curve is only illustrated for seven days, one could
easily imagine extending the curve to cover a full year.

From such a yearly curve, we may identify the percent of time for
which the demand exceeds a given value. If we assume that the
curve is a forecasted curve for the next year, then this percentage is
equivalent to the probability that the demand will exceed the given
value in that year.
The procedure for obtaining the percent of time for which the
demand exceeds a given value is as follows.
1. Discretize the curve into N equal time segments, so that the
value of the discretized curve in each segment takes on the
maximum value of the continuous curve in that segment.
2. The percentage of time the demand exceeds a value d is
obtained by counting the number of segments having a value
greater than d and dividing by N.
3. Plot the demand d against the percent of time the demand
exceeds a value d. A typical such plot is illustrated in Fig.
U19.12.
Demand, d (MW)

Percent of time 100

Fig. U19.12: Load duration curve

Module PE.PAS.U19.5 Generation adequacy evaluation 20

Fig. U19.12 is often generically referred to as a load duration

curve (LDC). However, one should be aware that there is a
significant difference between LDCs based on hourly segments
and LDCs based on daily segments.
• Hourly: the load duration curve indicates the percentage of
hours through the year that the hourly peak exceeds a value d.
• Daily: the load duration curve indicates the percentage of days
through the year that the daily peak exceeds a value d.
Thus, one must realize that the load duration curve gives the
percentage of time through the year that the load exceeds a value d,
only under the assumption that
• Hourly: the load is constant throughout the hour at the hourly
peak.
• Daily: the load is constant throughout the day at the daily peak.
Clearly, the smaller the segment, the better approximation is given
by the LDC to the actual percentage of time through the year that
the load exceeds a value d. Nonetheless, both daily and hourly-
based LDCs are used in practice.

The LDC may also be drawn in another way that is convenient for
computation. Consider first normalizing the abscissa (x-
coordinate) by dividing all values by 100, so that we obtain all
abscissa values in the range of 0 to 1.
The abscissa then represents the probability that the demand
exceeds the corresponding value d. We denote this probability
using the notation for a cumulative distribution function (cdf),
FD(d). However, one should realize that it is actually the
complement of a true cdf, i.e.,
FD (d ) = P( D  d ) = 1 − P( D  d )
Here, D is a random variable and d are the values it may take.

Finally, we can switch the axes of the LDC so that we plot FD(d) as
a function of d. Figure U19.12 illustrates the curve, which we refer
Module PE.PAS.U19.5 Generation adequacy evaluation 21

to as the load shape curve or the load model for the given time
period.

1
FD(d)

Demand, d (MW)

Fig. U19.12: Load shape

Note that Chanan Singh in his notes on “Load Modeling” gives an
algorithm for getting the load model from one scan of the hourly
load data [12].
U19.4 Calculation by Capacity Outage Tables
Module U17 identifies the loss of load probability (LOLP) and the
loss of load expectation (LOLE) as two indices for characterizing
generation adequacy risk. The LOLP is the probability of losing
load throughout the time interval (year). LOLE is the expected
number of time units (hours or days) per time interval (year) for
which the load will exceed the demand.
Fig. U19.13 illustrates a typical load-capacity relationship [1]
where the load model is shown as a continuous curve for a period
of 365 days. The capacity outage state, Ck, is shown so that one
observes that load interruption only occurs under the condition that
the load exceeds the installed capacity less the capacity outage, i.e.,
d>IC-Ck. The minimum demand for which this is the case is
dk=IC-Ck. Thus, the probability of having an outage of capacity C k
and of having the demand exceed dk is given by the capacity
Module PE.PAS.U19.5 Generation adequacy evaluation 22

outage pmf and FD(dk), i.e., fY(Ck)FD(dk)= fY(Ck)FD(IC-Ck). (This

assumes independence between outage events and demand).

t FD(d)

365 1

Observe
IC>dmax

Ck
FD(dk) or tk
0 0
dk IC
Demand, d (MW)

Fig. U19.13: Relationship between capacity outage, load model [1]

The LOLP is computed as the sum over all capacity outage states:
N
LOLP =  f Y (Ck ) FD ( IC − Ck ) (U19.13)
k =1

and the LOLE as:

N N
LOLE =  f Y (C k ) FD ( IC − C k ) * 365 =  f Y (C k )t k (U19.14)
k =1 k =1

where N is the total number of capacity outage states.

Module PE.PAS.U19.5 Generation adequacy evaluation 23

Example: Compute the LOLP and the LOLE for the capacity
outage table of Table U19.7, for the (daily) load shape curve given
by Fig. U19.14. Table U19.7 is repeated below for convenience.
Table U19.7: Capacity Outage Table for Convolved 3 MW Units
and 5 MW Unit

Capacity Description Probability

Outage

0 All units up. 0.980.9604=0.941192

3 U3 up, U1 or U2 down. 0.980.0392=0.038416
5 U3 down, U1 and U2 up. 0.020.9604=0.019208
6 U3 up, U1 and U2 down. 0.980.0004=0.000392
8 U3 down, U1 or U2 down. 0.020.0392=0.000784
11 All units down 0.020.0004=0.000008

t FD(d)
The values of FD(d) for the
365 1 possible capacity outage states.
0.875

0.375
0.25

0.0625
0 0
IC=11
d=3 d=5 d=6 d=8
Demand, d (MW)

Fig. U19.14: Load shape curve for example

Module PE.PAS.U19.5 Generation adequacy evaluation 24

From (U19.13), we then have:

N
LOLP =  f Y (Ck ) FD ( IC − Ck )
k =1

= f Y (0) F D(11) + f Y (3) F D(8) + f Y (5) F D(6)

+ f Y (6) FD (5) + f Y (8) F D(3) + f Y (11) F D(0) =
= .941192 * 0 + .038416 * .0625 + .019208 * .25
+ .000392 * .375 + .000784 * .875 + .000008 *1
= 0.008044 / year
We could compute LOLE using (U19.14),
N N
LOLE =  f Y (C k ) FD ( IC − C k ) * 365 =  f Y (C k )t k (U19.14)
k =1 k =1

but it is easier to just recognize that

LOLE=LOLP*365=0.008044*365 =2.93606 days/year.
This means that we can expect to see 2.93606 complete days of
load interruption each year, assuming that the peak load per day
lasts all day. Another index often cited is the years/day, in this
case, 1/2.93606=0.3406 years/day. This is the number of years that
must pass before we see a full day of load interruption. A well-
known metric is “1 day in 10” indicating LOLE=0.1 day/year, or
years/day should be 10.
Two important qualifiers should be emphasized:
• This LOLE is the load outage time expected as a result of
generation unavailability and does not include the effects of
transmission or distribution system components unavailability.
• This amount of outage time would correspond to the long-run
average of this system only if
o all 3 units are always committed, i.e., no reserve shutdown,
and there is no maintenance;
o the demand remains at its daily peak throughout each day.
Module PE.PAS.U19.5 Generation adequacy evaluation 25

These qualifiers are obviously pointing towards inaccuracies in the

model and as a result, indicate that the indices computed should
not be perceived as accurate in an absolute sense. However, the
indices do serve well for comparative purposes.
U19.5 A capacity planning example
Reference [1] provides an illustrative example showing how the
generation adequacy calculation procedure in the previous section
can be applied to the capacity planning problem. We adapt that
example here.
Consider a system containing five 40 MW units each with a
FOR=0.01, so that the installed capacity is 200 MW. We can go
through a convolution process for this, or else we can observe that
there are six distinct capacity outage levels: outage of 0, 40, 80,
120, 160, and 200 MW, and that for each of these capacity outage
levels, there are a number of states equal to the combination of 5
things taken r at a time, where r is 0, 1, 2, 3, 4, and 5, respectively.
It is a combination, where order does not matter, instead of a
permutation, where order does, because, for example, the status
vector {U1, U2, U3, U4, U5} characterizes the same state (and
corresponding probability) as the status vector {U2, U1, U3, U4,
U5}. Therefore we may identify the number of each specific
combination of states with 0, 1, 2, …,5 outages as Crn = n! / (n - r)!r! .
For example, the number of states with a capacity outage level of
80 MW (2 units out) will be
C25 = 5! / (5 - 2)!(2!)= 5  4 / 2 = 10
The probability of any one such state will be
(0.01)(0.01)(0.99)(0.99)(0.99)=0.00009703, and so the probability
of this state will be 10*(0.00009703)=0.0009703.
Proceeding in this way (or alternatively, performing the
convolution), the capacity outage table for the system is developed
and is shown in Table U19.9, where capacity outage states having
probabilities less than 10-6 have been neglected.
Module PE.PAS.U19.5 Generation adequacy evaluation 26

Table U19.9: Capacity outage table for example [1]

Capacity Description Probability

Outage

0 All units up 0.950991

40 1 unit down 0.048029
80 2 units down 0.000970
120 3 units down 0.000009

The next year’s system load model is represented by the load shape
curve of Fig. U19.15a, which is a linear approximation of an actual
load shape curve. Note that the forecasted annual system peak load
is 120 MW.

t FD(d)

365 1

Ppeak

0 0

20 40 60 80 100 120 140 160 180 200

Demand, d (MW)

Fig U19.15a: Load shape curve for example [1]

The procedure of the previous section was applied, i.e., the
following expression was evaluated,
Module PE.PAS.U19.5 Generation adequacy evaluation 27

N
LOLE =  fY (Ck )FD (IC - Ck )* 365 (U19.14)
k=1

and the LOLE and years/day were computed as 0.002005

days/year and 498 years/day, respectively. Certainly this is a very
reliable system! The reason for the high reliability is, of course,
that the installed capacity is so much greater than the system
annual peak.
However, the load will grow in the future, so it is of interest to see
how these indices vary as peak load increases. Table U19.10
summarizes LOLE and years/day for the system peak beginning at
120 MW and increasing to 200 MW in units of 10 (this is to just
illustrate the effect on the indices; the 10 MW increment should
not be interpreted as an annual load growth). This is imposed on
the load model by shifting it to the right by 10 MW for each 10
MW increment, as illustrated by the dark line in Fig. U19.15b for a
160 MW system peak. A similar thing is done for 130, 140, 150,
170, 180, 190, and 200 MW system peaks, and indices are
recomputed for each.

t FD(d)

365 1

Ppeak

0 0

20 40 60 80 100 120 140 160 180 200

Demand, d (MW)

Fig U19.15b: Load shape curve for example [1]

Module PE.PAS.U19.5 Generation adequacy evaluation 28

Table U19.10: Variation in LOLE with System Annual Peak [1]

System annual peak Indices
(MW)
LOLE, days/year Years/day
120 0.002005 498
130 0.04772 20.96
140 0.08687 11.51
150 0.1208 8.28
160 0.1506 6.64
170 1.895 0.53
180 3.447 0.29
190 4.837 0.21
200 6.083 0.16
The LOLE (days/year) is plotted on semi-log scale in Fig. U19.16.

LOLE=6 days/yr

Fig. U19.16: LOLE as a function of system annual peak load [1]

Module PE.PAS.U19.5 Generation adequacy evaluation 29

Obviously, we must add some capacity before we reach an annual

peak demand of 200 MW, as the LOLE is about 6 days/year at that
level (very high!). But at what peak demand level should capacity
be added?
The answer to this question can be identified if we select a
threshold risk level beyond which we will not allow. This is
basically a management decision, but of course, all management
decisions can be facilitated by quantitative analysis. We will
forego such analysis here and instead arbitrarily select 0.15
days/year as the threshold risk level.
Assume:
• we have forecasted a 10% per year load growth (very high) and
so the peak loads will be, for years 0, 1, 2, 3, 4, 5, 6, 7, 8, given
by 120, 132, 145, 160, 176. 193, 213, 234, and 257 MW,
respectively;
• we have decided to add one 50 MW unit at a time, each with
FOR=0.01, as the load grows, in order to ensure the system
satisfies the identified threshold risk level.
The question is: when do we add the units?
To answer this question, we will repeat the analysis of Table
U19.10, except for four different installed capacities: 200 MW,
250 MW, 300 MW, and 350 MW, corresponding to additional
units of 0, 1, 2, and 3, respectively.
Table U19.11 summarizes the calculations. Fig. U19.17 illustrates
the variation in LOLE with peak load for each case, together with
vertical lines indicating the peak load value for each year.
Module PE.PAS.U19.5 Generation adequacy evaluation 30

Table U19.11: LOLE Calculations for Example [1]

System annual LOLE (days/year)
peak (MW)
200 MW 250 MW 300 MW 350 MW

100 0.001210 - - -
120 0.002005 - - -
140 0.08687 0.001301 - -
160 0.1506 0.002625 - -
180 3.447 0.06858 - -
200 6.083 0.1505 0.002996 -
220 - 2.058 0.03615 -
240 - 4.853 0.1361 0.002980
250 - 6.083 0.1800 0.004034
260 - - 0.6610 0.01175
280 - - 3.566 0.1075
300 - - 6.082 0.2904
320 - - - 2.248
340 - - - 4.880
350 - - - 6.083
Module PE.PAS.U19.5 Generation adequacy evaluation 31

3 4 5 6 7 9
8

Fig. U19.17: Capacity planning example [1]

The unit additions would need to be made in years 3, 5, and 7. The
dotted line tracks the year-by-year risk variation.
This approach ensures that the stated reliability criteria are met;
however, the other influence to the decision-making process is, as
always, economic. Recall that we assumed that we would solve our
capacity problem by adding capacity at increments of 50 MW at a
time. It would be quite atypical if this were the only solution
Module PE.PAS.U19.5 Generation adequacy evaluation 32

approach considered. For example, one might consider larger or

smaller increments, or more or less reliable units (different FOR).
Different decisions would have different influence on the system
risk; they would also have different present worth values. The
influence on risk and present worth would need be weighed one
against another in order to arrive a “good” decision.
Question:
Why would you want to perform this kind of calculation for a
system in which generators are built by electricity market
participants rather than a centralized vertically integrated utility
company?
U19.6 The effective load approach
Most of what we have seen in sections U19.1-U19.5 characterize
the view taken by [1]. We now provide another view, based on [2].
U19.6.1 Preliminary Definitions
Let’s characterize the load shape curve with t=g(d), as illustrated in
Fig. U19.18. It is important to note that the load shape curve
characterizes the (forecasted) future time period and is therefore a
probabilistic characterization of the demand.
t

t=g(d)

dmax

Demand, d (MW)

Fig. U19.18: Load shape t=g(d)

Module PE.PAS.U19.5 Generation adequacy evaluation 33

Here:
• d is the system load
• t is the number of time units in the interval T for which the load
is greater than d and is most typically given in hours or days
• t=g(d) expresses the functional dependence of t on d
• T represents, most typically, a day, week, month, or year
The cumulative distribution function (cdf) introduced in Section
U19.3 is given by
t g (d )
FD (d ) = P( D  d ) = = (U19.15)
T T
One may also compute the total energy ET consumed in the period
T as the area under the curve, i.e.,
dmax
ET =  g( )dλ
0 (U19.16)
The average demand (a power quantity) in the period T is obtained
from
1 1d d
= ET =  g ( )d =  FD ( )d
max max

d avg (U19.17)
T T 0 0

Now let’s assume that the planned system generation capacity, i.e.,
the installed capacity, is CT, and that CT<dmax. This is an
undesirable situation, since we will not be able to serve some
demands, even when there is no capacity outage! Nonetheless, it
serves well to understand the relation of the load duration curve to
several useful indices. The situation is illustrated in Fig. U19.19.
Module PE.PAS.U19.5 Generation adequacy evaluation 34

t=g(d)

tC

CT dmax

Demand, d (MW)

Fig. U19.19: Illustration of Unserved Demand

Then, under the assumption that the given capacity CT is perfectly
reliable, we may express three useful reliability indices:
• Loss of load expectation, LOLE: the number of time units that
the load will exceed the capacity,
LOLE = tC = g (CT )
T
(U19.18)

• Loss of load probability, LOLP: the probability that the load

will be interrupted during the time period T
LOLP = P( D  CT ) = FD (CT ) (U19.19)
One may think that, given CT<dmax, then LOLP=1, i.e., the event
“load interruption during T” is certain. The reason why it is not
certain is because the load model is probabilistic. So LOLP is
simply reflecting the uncertainty associated with demand, i.e.,
the demand may or may not exceed CT, according to FD(CT).
• Expected demand not served, EDNS: If the average (or
expected) demand is given by (U19.17), then it follows that the
expected demand not served would be:
Module PE.PAS.U19.5 Generation adequacy evaluation 35

d max
EDNS =  FD ( )d (U19.20)
CT

which would be the same area as in Fig. U19.19 when the

ordinate is normalized to provide FD(d) instead of t. Reference
[2] provides a rigorous derivation for (U19.20).
• Expected energy not served, EENS: This is the total amount of
time multiplied by the expected demand not served, i.e.,
d max d max
EENS = T  FD ( )d =  g ( )d (U19.21)
CT CT

which is the area shown in Fig. U19.19.

U19.6.2 Effective load
The notion of effective load is used to account for the unreliability
of the generation, and it is essential for understanding the view
taken by [2].
The basic idea is that the total system capacity is always CT, and
the effect of capacity outages are accounted for by changing the
load model in an appropriate fashion, and then the different indices
are computed as given in (U19.18), (U19.19), and (U19.20).
A capacity outage of Ci is therefore modeled as an increase in the
demand, not as a decrease in capacity!
We have already defined D as the random variable characterizing
the demand. Now we define two more random variables:
• Dj is the random increase in load for outage of unit i.
• De is the random load accounting for outage of all units and
represents the effective load.
Thus, the random variables D, De, and Dj are related according to:
N
De = D +  D j (U19.21)
j =1
Module PE.PAS.U19.5 Generation adequacy evaluation 36

It is important to realize that, whereas Cj represents the capacity of

unit j and is a deterministic value, Dj represents the increase in load
corresponding to outage of unit j and is a random variable. The
probability mass function (pmf) for Dj is assumed to be as given in
Fig. U19.20, i.e., a two-state model. We denote the pmf for Dj as
fDj(dj)
fDj(dj)
Aj

Uj

0 Cj Outage load, dj

Fig. U19.20: Two state generator outage model

Recall from module U13 that the pdf of the sum of two random
variables is the convolution of their individual pdfs. In addition, it
is true that the cdf of two random variables can be found by
convolving the cdf of one of them with the pdf (or pmf) of the
other, that is, for random variables X and Y, with Z=X+Y, that

FZ ( z ) =

 F
= −
X ( z −  ) f Y ( )d (U19.22)

Let’s consider the case for only one unit, i.e., from (U19.21),
De = D + D j (U19.23)
Then, by (U19.22), we have that:

F (d e ) =  FD( 0 ) (d e −  ) f D ( )d
(1)
De e j
(U19.24)
 = −
Module PE.PAS.U19.5 Generation adequacy evaluation 37

where the notation FD () indicates the cdf after the jth unit is
( j)

convolved in. Under this notation, then, (U19.23) becomes

De( j ) = De( j −1) + D j (U19.25)
and the general case for (U19.24) is:

F (d e ) =  FD( j −1) (d e −  ) f D ( )d
( j)
De e j
(U19.26)
 = −

which expresses the equivalent load after the jth unit is convolved
in.
Since fDj(dj) is discrete (i.e., a pmf), we may rewrite (U19.26) as

FD( j ) (d e ) =  FD( j −1) (d e − d j ) f D (d j ) (U19.27)

d j 0
e e j

From an intuitive perspective, (U19.27) is providing the

convolution of the load shape FD () with the set of impulse
( j −1)

functions comprising fDj(dj). When using a 2-state model for each

generator, fDj(dj) is comprised of only 2 impulse functions, one at 0
and one at Cj. Recalling that the convolution of a function with an
impulse function simply shifts and scales that function, (U19.27)
can be expressed for the 2-state generator model as:
FD( j ) (d e ) = Aj FD( j −1) (d e ) + U j FD( j −1) (d e − C j )
e e e
(U19.28)
So the cdf for the effective load following convolution with
capacity outage pmf of the jth unit, is the sum of
• the original cdf, scaled by Aj and
• the original cdf, scaled by Uj and right-shifted by Cj.
Example: Fig. U19.21 illustrates the convolution process for a
single unit C1=4 MW supplying a system having peak demand
dmax=4 MW, with demand cdf given as in plot (a) based on a total
time interval of T=1 year.
Module PE.PAS.U19.5 Generation adequacy evaluation 38

(a)
1 FD( 0 ) (d e )
r
1 fDj(dj) (b)
0.8 0.8
0.6
0.4
0.2
* 0,6
0.4
0.2
C1=4

1 2 3 4 5 6 7 8 de 1 2 3 4 5 6 7 8

1. (c) 1. (d)
0.8 0.8
0.6
0.4
+ 0.6
0.4
0.2 0.2

1 2 3 4 5 6 7 8 de 1 2 3 4 5 6 7 8 de

1.0
FD(1) (d e )
r

= 0.8
0.6
0.4
(e)

0.2

1 2 3 4 5 6 7 8 de

Fig. U19.21: Convolving in the first unit

Plots (c) and (d) represent the intermediate steps of the convolution
(0)
where the original cdf FD (d e ) was scaled by A1=0.8 and
e

U1=0.2, respectively, and right-shifted by 0 and C1=4, respectively.

Note the effect of convolution is to spread the original cdf.
Plot (d) may raise some question since it appears that the constant
part of the original cdf has been extended too far to the left. The
reason for this apparent discrepancy is that all of the original cdf,
in plot (a), was not shown. The complete cdf is illustrated in Fig.
Module PE.PAS.U19.5 Generation adequacy evaluation 39

U19.22 below, which shows clearly that FD( 0 ) (d e ) = 1 for de<0,

e

reflecting the fact that P(De>de)=1 for de<0.

1.0 FD( 0 ) (d e )
r
0.8
0.6
0.4
0.2

1 2 3 4 5 6 7 8 de

Fig. U19.22: Complete cdf including values for de<0

Let’s consider that the “first” unit we just convolved in is actually
the only unit. If that unit were perfectly reliable, then, because
C1=4 and dmax=4, our system would never have loss of load. This
would be the situation if we applied the ideas of Fig. U19.19 to
Fig. U19.21, plot (a).
However, Fig. U19.21, plot (e) tells a different story. Fig. U19.23
applies the ideas of Fig. U19.19 to Fig. U19.21, plot (e) to show
how the cdf on the equivalent load indicates that, for a total
capacity of CT=4, we do in fact have some chance of losing load.

1.0
FD(1) (d e )
r

0.8
0.6 CT=4
0.4
0.2

1 2 3 4 5 6 7 8 de

Fig. U19.23: Illustration of loss of load region

The desired indices are obtained from (U19.18), (U19.19), and
(U19.20) as:
Module PE.PAS.U19.5 Generation adequacy evaluation 40

LOLE = tC = g e (CT ) = T  FD (CT = 4) = 1 0.2 = 0.2 years

T r

A LOLE of 0.2 years is 73 days, a very poor reliability level that

reflects the fact we have only a single unit with a high FOR=0.2.
The LOLP is given by:
LOLP = P( De  CT ) = FDe (CT ) = 0.2
and the EDNS is given by:
d e , max
EDNS =  FDe ( )d
CT

which is just the shaded area in Fig. U19.23, most easily computed
using the basic geometry of the figure, according to:
1
0.2(1) + (3)(0.2) = 0.5MW
2
The EENS is given by
d e , max d e , max
EENS = T  FDe ( )d =  g e ( )d
CT CT

or TEDNS=1(0.5)=0.5MW-years, or 8760(0.5)=4380MWhrs.
U19.7 Four additional issues
A more extended treatment of generation adequacy evaluation
would treat a number of additional issues. Here, we just point to
these issues with a brief overview of each so that the interested
reader may follow up on them as desired. The main issues are
model uncertainty (U16.7.1), maintenance (U16.7.2), convolution
techniques (U16.7.3), and frequency and duration approach
(U16.7.4).
U19.7.1 Model uncertainty
We have modeled uncertainty in our analysis of generation
adequacy. However, we have assumed that our uncertainty models
Module PE.PAS.U19.5 Generation adequacy evaluation 41

are precise, i.e., the unit FORs and the load forecast used to obtain
the load duration curves are both perfectly accurate. The fact of the
matter is that the unit FORs and the load forecast are estimates of
the “true” parameters, and they will always be estimates no matter
how much data is collected! Therefore, it is of interest to model
uncertainty in the model parameters and then identify the influence
of these uncertainties on the resulting adequacy indices.
One method of modeling parameter uncertainty is to represent each
parameter with a numerical distribution. Then repeatedly draw
values from each distribution, and calculate the reliability indices
using those values. If the parameter values are drawn as a function
of their probabilities, as indicated by the distribution, then the
computed reliability indices will also form a distribution, from
which we may compute their statistics, e.g., mean, variance, etc.
For example, if the peak load is normally distributed, then the
distribution may be discretized, and each interval of the
distribution can be assigned to an interval on (0,1) in proportion to
its area under the normal curve. Then a random draw on (0,1),
which is then converted to the peak load value through the
assignment, will reflect the desired normal distribution. Figure
U19.24 illustrates the process.

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1

Fig. U19.24: Monte Carlo Simulation

Module PE.PAS.U19.5 Generation adequacy evaluation 42

This process is called Monte Carlo simulation (MCS) and is almost

always an available option for computing reliability indices under
parameter uncertainty. The advantage to MCS is that it is
conceptually simple to implement. The disadvantage is that it is
computationally intensive.
Load forecast uncertainty:
There are two basic methods. The first, well articulated in [1], is
the most computational but the easiest to understand. The approach
is to model the peak load using a discretized normal distribution, as
shown in Fig. U19.24, where the mean of the distribution
corresponds to the forecasted load.
0.382

0.242 0.242

0.061 0.061
0.006 0.006
-3 -2 -1 0 1 2 3

Standard deviations from mean

Fig. U19.25: Modeling of load uncertainty [1]

The load shape curve is adjusted for each of the load values
corresponding to the seven standard deviations from the mean
(-3, -2, -1, 0, 1, 2, 3), where 1 standard deviation is estimated based
on the load forecasting program used and the amount of time over
which the forecast is being done. A reasonable value could be 2%,
for example.
Then the indices are computed for each different load shape and
composite indices are computed as a weighted function of the
individual indices, where the weights are the probabilities given in
Fig. U19.25.
Module PE.PAS.U19.5 Generation adequacy evaluation 43

A second method is given in [1] but perhaps more thoroughly

described in [2]. The basic idea is that a single cdf is constructed
that reflects the uncertainty of the peak load forecast, using
FD( 0 ) (d ) =  FD( 0 ) (d |  ) f  ( ) (U19.29)


Once this cdf is obtained, the indices are computed using one of
our standard approaches.
It is important to realize that modeling of uncertainty in load
forecast always results in indices reflecting poorer reliability
because the rate of increase of the indices is nonlinear with peak
load, in that it is higher at higher load levels than at lower load
levels.
FOR uncertainty:
References [1, 4] address inclusion of FOR uncertainty using a
covariance matrix corresponding to the capacity outage table. The
method is based on [5]. One important conclusion from this work
is that although FOR uncertainty certainly affects the distribution
of the reliability indices, it does not affect their expected values.
This is in contrast to load forecast uncertainty.
U19.7.2 Maintenance
The conceptually simplest method for including unit maintenance
is through the capacity outage approach according to the
following:
1. Compute a “full” capacity outage table.
2. Divide the year into Ny intervals and obtain a unique load shape
cdf FDp(d) for each period p.
3. For each interval p=1, Ny
a. Identify the units out on maintenance in this interval
b. Deconvolve each outaged unit from the capacity outage
table to get a capacity outage table for period p, using the
Module PE.PAS.U19.5 Generation adequacy evaluation 44

algorithm of Section U19.2.4. Denote the resulting

capacity outage pmf as fYp(y).
c. Compute the LOLE for period p as (similar to (U19.14)):
Np N
LOLE p =  f Yp (Ck ) FDp ( IC − Ck ) * N days =  f Yp (Ck )t kp
k =1 k =1

(U19.30)
where Np is the total number of capacity outage states for
period p and Ndays are the number of days in period p.
4. The annual LOLE is then given as the sum of the LOLEp, i.e.,
NY
LOLE =  LOLE p (U19.31)
p =1

U19.7.3 Convolution techniques

We have seen that convolution plays a major role in both the
capacity table approach and the effective load approach. The
convolution method illustrated for both approaches is called the
recursive method. One drawback of this method is that it is quite
computationally intensive and can require significant computer
resources when it is used for systems having a large number of
units and/or units with a large number of derated states.
As a result, there has been a great deal of research effort into
developing faster convolution methods. This work has resulted in,
in addition to the recursive method, the following methods [3]:
• Fourier transform [6]
• Method of cumulants [7]
• Segmentation method [8, 9, 10]
• Energy function method [3]
Of these, the method of cumulants is very fast, and the recursive
method very is accurate. The segmentation method is said to
achieve a good tradeoff between speed and accuracy. Chanan
Singh summarizes the method of cumulants in his notes [12].
Module PE.PAS.U19.5 Generation adequacy evaluation 45

U19.7.4 Frequency and duration approach

The methods presented in this module so far provide the ability to
compute LOLP, LOLE, EDNS, and EENS, but they do not provide
the ability to compute
• Frequency of occurrence of an insufficient capacity condition
• The duration for which an insufficient capacity condition is
likely to exist.
A competing method which provides these latter quantities goes,
quite naturally, under the name of the frequency and duration
(F&D) approach. The F&D approach is based on state space
diagrams and Markov models. We touched on this at the end of
Section U19.2.1 above by showing that we may represent a 2
generator system via a Markov model and then compute state
probabilities, frequencies, and durations for each of the states.
The underlying steps for the F&D approach, outlined in chapter 10
of [11], are:
1. Develop the Markov model and corresponding state transition
matrix, A for the system.
2. Use the state transition matrix to solve for the long-run
probabilities from 0=pA and ∑pj=1 (note that we have dropped
the subscript  for brevity, but it should be understood that all
probabilities in this section are long-run probabilities).
3. Evaluate the frequency of encountering the individual states
from (U16.31), repeated here for convenience:

f j =   jk p j , = p j ,   jk (U19.32)
k j k j
which can be expressed as:
fj=pj,[total rate of departure from state j]
Module PE.PAS.U19.5 Generation adequacy evaluation 46

4. Evaluate the mean duration of each state, i.e., the mean time of
residing in each state, from (U16.33), repeated here for
convenience:

1 p j ,
Tj = =
  jk fj (U19.33)
k j
(Note that [11] uses mj to denote the duration for state j and uses
Tj to denote the cycle time for state j, which is the reciprocal of
the state j frequency fj. One should carefully distinguish
between the cycle time and the mean duration.
• The cycle time is the mean time between entering a given
state to next entering that same state.
• The duration is the mean time of remaining in a given state.)
5. Identify the states corresponding to failure, lumped into a
cumulative state denoted as J.
6. Compute the cumulative probability of the failure states pJ as
the sum of the individual state probabilities:
pJ =  p j (U19.34)
jJ

7. Compute the cumulative frequency fJ of the failure states (see

section U16.8.2) as the total of the frequencies leaving a failure
state j for an non-failure state k:
fJ =   f jk (U19.35)
kJ jJ
Because (see (U16.29)) fjk=λjk pj,, (U19.35) can be expressed as
Module PE.PAS.U19.5 Generation adequacy evaluation 47

fJ =    jk p j , =    jk p j ,
kJ jJ jJ kJ

=  p j ,   jk (U19.36)
jJ kJ
8. Compute the cumulative duration for the failure states, as:
pJ
TJ = (U19.37)
fJ
The above approach is quite convenient for a system of just a very
few states, and it is important for our purposes because it lays out
the underlying principles on which the F&D is based.
However, for a large system, the above approach is not very useful
because of step 1 where we must develop the Markov model. This
difficulty is circumvented by building the capacity outage table
using recursive relations for the capacity outage (e.g. state)
probabilities together with additional recursive relations for state
transitions and state frequencies [1, 2, 4, 11].
References
[1] R. Billinton and R. Allan, “Reliability Evaluation of Power Systems, 2 nd edition, Plenum Press, 1996.
[2] R. Sullivan, “Power System Planning,” McGraw Hill, 1977.
[3] X. Wang and J. McDonald, “Modern Power System Planning,” McGraw Hill, 1994.
[4] J. Endrenyi, “Reliability modeling in electric power systems,” Wiley and Sons, New York, 1978.
[5] A. Patton and A. Stasinos, “Variance and approximate confidence limits on LOLP for a single-area
system,” IEEE Trans. on Power Apparatus and Systems, Vol. 94. pp. 1326-1336, July/August 1975.
[6] N. Rau and K. Schenk, “Application of Fourier Methods for the Evaluation of Capacity Outage
Probabilities,” IEEE PES 1979 Winter Power Meeting, paper A-79-103-3.
[7] N. Rau, P. Toy, and K. Schenk, “Expected energy production costs by the method of moments,” IEEE
Trans. on Power Apparatus and Systems, vol. PAS-99, no. 5, pp 1908-1917, Sep/Oct., 1980.
[8] K. F. Schenk, R. B. Misra, S. Vassos and W. Wen, ‘A New Method for the Evaluation of Expected
Energy Generation and Loss of Load Probability’, IEEE Transaction on Power Apparatus and
Systems, Vol. PAS-103, No. 2, Feb. 1984.
[9] Y. Dai, J. McCalley, and V. Vittal, “Annual Risk Assessment for Thermal Overload,” Proceedings of
the 1998 American Power Conference, Chicago, Illinios, April, 1998.
[10] Y. Dai, “…,” Ph.D. Dissertation, Iowa State University, 1999.
[11] R. Billinton and R. Allan, “Reliability evaluation of engineering systems,” 2nd edition, Plenum Press,
New York, 1992.
[12] C. Singh, “Electric Power System Reliability – Course Notes.”