Professional Documents
Culture Documents
I. INTRODUCTION
Manuscript received November 3, 2003; revised October 6, 2004. This work (2)
was supported in part by NSERC and in part by the MITACS-NCE. Paper no.
TPWRS-00619-2003.
C. L. Anderson is with Department of Civil and Environmental Engineering,
The University of Western Ontario, London, ON N6A 5B7, Canada (e-mail: This standard form CDF is the whole state distribution, which
landerso@uwo.ca). assumes that the component has entered its current state at the
M. Davison is with the Department of Applied Mathematics, The instant of inspection. Since we assume that only one component
University of Western Ontario, London, ON N6A 5B7, Canada (e-mail:
mdavison@uwo.ca). can change state at a given time, we must consider the remaining
Digital Object Identifier 10.1109/TPWRS.2005.856992 state distribution for all other components. For example, as seen
0885-8950/$20.00 © 2005 IEEE
1784 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 4, NOVEMBER 2005
in [4], the “long run” remaining life distribution of a renewal crucial information about the system, while leaving some re-
process is maining questions. If we know the TTF of a system, can we
determine which component caused the state change? This is
(3) given by the probability of component failing first, conditioned
on the TTSC. In other words, given the failure at time , what is
where is the mean duration of the state. the probability that component is the causing component?
This is an important factor in the development of our model, The conditional distribution of the time to failure can be found
which is detailed in Section II. using Bayes Theorem
(5)
(9)
After completing the integration, we obtain
Equations (8) and (9) give us the conditional probability of
each component causing the state change. We can then use these
probabilities to select the causing component for the next state
of the system. Following this sequence, we can simulate the
(6) state, and thereby the generating capacity, of the system over
where represents the upper incomplete gamma function the duration of time we wish to model.
for , from to . This development clearly illustrates that our Weibull-based
Now that we have obtained the distribution, we can use it to model is efficient since it uses the same structure for system
simulate the times to state change of a system. This provides state. When the system changes state, the model responds by
ANDERSON AND DAVISON: AGGREGATE WEIBULL APPROACH FOR MODELING 1785
changing the parameters of the distribution to represent the new component: once with failure parameters and once with repair
configuration. Computationally, this framework is straightfor- parameters. These factors are used in each calculation for both
ward to implement. TTSC and determining the conditional probabilities but are only
The development of a framework for a system-wide Weibull calculated once.
model naturally opens the question of whether the additional A key element of this model is the estimation of Weibull pa-
work is worth the effort. Many authors agree that the exponen- rameters for failure and repair times. This is easily completed
tial model is sufficient to describe the useful life of a component with the availability of system data for various generating plants.
but that its application to component repair is an over-simplifica- The most common method is a maximum likelihood estimator
tion [5]. Some authors have suggested that although repair times technique. For example, detailed information about forced out-
are decidedly nonexponential, the use of an exponential model ages of California generators is publicly available. From this
for time to repair is acceptable given the complexity of other data, we can determine the set of TTFs and TTRs observed for
models [5]. However, the mathematical framework illustrated every plant in the system that experienced a forced outage over
here allows the appropriate distributional shape to be applied to the last three years. With this data set, MLE can be used to esti-
the appropriate task, without requiring any modifications to ac- mate the and values for both failure and repair of the plants
count for exponential or nonexponential durations. The use of in the system.
the Weibull framework allows the same distribution to be ap- For the immediate purpose of testing the model, we chose
plied and the shape of the distributions to be determined by the the IEEE Reliability Test System (RTS) described in [2]. In this
parameters used. These parameters are obtained from the system model, the generation system is described using exponential pa-
of interest and no a priori assumptions about the shape of the rameters for TTF and TTR of the generating units. We used
distributions are needed. the exponential parameters to test the validity of the aggregate
In the next sections, we will show that the modest addition of model by using for both failure and repair of all com-
computational cost is justified by the realism that can be gained. ponents in the system. We then tested the model behavior with
other Weibull parameters.
Application to other generation systems is implemented by
III. PARAMETERS AND IMPLEMENTATION
development of the system generator “profile” and estimation
The model developed in Section II gives us a stable and effi- of failure and repair distribution parameters.
cient framework for implementation. The steps to implementa-
tion are as follows. IV. RESULTS
1) Draw a TTSC from the distribution given in (6). In order to assess the usefulness of our model for power plant
2) Calculate the probability, conditioned on the TTSC, that failure and repair, we compare it to two other techniques.
a given component caused the state change. The first method was to model the same system using expo-
3) Use this probability to simulate which component caused nential distributions for both failure and repair times. This is
the state change. equivalent to the homogenous Markov model (HMM). We also
4) Change the state of the causing component. use the purely Weibull model with the same distribution param-
The model will repeat steps 1 through 4 for the duration of eters and model the plants in independent sequential simulation.
the simulation. In theory this is simple although there exist some We then use these results of these methods as a comparison for
complications. Drawing a TTSC from the distribution in step 1 the aggregate Weibull model.
is more complicated than it might seem. There are three main The exponential model is developed using assumptions sim-
options for drawing the TTSC. These include analytical inver- ilar to the Weibull model described in Section II, but the expo-
sion, numerical approximation (such as acceptance–rejection) nential model simplifies quickly.
or numerical root-finding, such as bisection. For a single component indexed by
Although there are inversions for the incomplete gamma
function, the inversion of a product of incomplete gamma (10)
functions does not seem feasible. The presence of the incom-
plete gamma functions does provide the possibility of using For the entire system, we obtain
the properties of special functions to estimate upper and lower
bounds. If these bounds can be determined in a reasonably (11)
accurate manner, the acceptance–rejection technique could be
very efficient.
The most immediate method is a standard root finding tech- where .
nique, such as bisection. The inversion of a CDF in this manner The probability of component causing the change of state,
is straightforward as the function is monotone. Once the TTSC conditioned on the TTSC, is given by
can be simulated using one of these techniques, it is straight-
forward to calculate the conditional probabilities for each com- (12)
ponent and simulate the causing component of the subsequent
state. It should be noted that the stable structure of the distri- As a second comparative model, we also complete an inde-
bution provides additional efficiencies because the incomplete pendent sequential simulation of the plants of the system. This
gamma function need only be calculated twice for each method requires that each plant be simulated independently and
1786 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 4, NOVEMBER 2005
TABLE I TABLE II
COMPARISON OF LOLE FOR MODELS COMPARISON OF OUTAGE FREQUENCY FOR
its failure and repair times be recorded. At the end of the simula-
tion, capacity is calculated for each day in the simulation based
on which plants are operational and which are under repair. We
show that this simple method results in the same long run system
behavior as our proposed model.
TABLE III
Recalling that the purpose of the model is to simulate avail- COMPARISON OF OUTAGE DURATION FOR
able system capacity, an important criterion for assessing the
validity of a capacity model is its ability to model the system
generating capacity and to meet system load requirements.
We first compare the loss of load expectation (LOLE), a
common risk measure, for each of the three models. We then
examine the simulated system capacity of the IEEE RTS for
the aggregate model and compare the results to the traditional
exponential model.
A. LOLE
The LOLE is a common measure of the adequacy of system Tables II and III compare the frequency and duration of out-
capacity. LOLE is a measure of the expected proportion of time ages, for each plant type, as the shape parameter of the repair
that available system capacity is unable to meet the electrical distribution ( ) is varied.
load [6] and can be quoted as a proportion or in the form of The results provided in Tables II and III support the previous
days/year or hours/year. The LOLE is estimated for the RTS results. It is clear that, as expected, the frequency of outages is
based on each of the three simulation techniques described in not affected by the variation of the the shape parameter of the
the previous section. repair distribution ( ). However, the outage duration is very
The LOLE values calculated were then compared with the sensitive to this parameter.
LOLE values estimated for the exponential model of the system. The model described here was developed to simulate system
Two-sided t-tests of the LOLE values confirmed that there was capacity in an overall electricity spot price model. For this
no difference between the mean LOLE values obtained using reason, we are primarily interested in the ability of the model
the aggregate Weibull model with bisection, the sequential sim- to simulate overall system capacity.
ulation of Weibull plants and the exponential model, when ex-
ponential parameters ( ) are used in the Weibull models. B. Overall System Capacity
To test other scenarios, we considered various shape ( ) pa-
rameters for component repair and compared the results of the System capacity histograms for exponential parameters
three models. In all cases, the t-tests confirmed that the results ( ), and are presented in Figs. 2–4,
of the Weibull bisection and sequential simulation models were respectively.
indistinguishable and that both differed from the exponential We can see in Fig. 2 that the system capacity diagrams are in-
model. These results are presented in Table I. distinguishable for the Weibull and exponential models. This is
Table I also contains the mean time to repair (MTTR) asso- as expected, since the Weibull model with is the exponen-
ciated with each of the models. This is useful to compare the tial model. In Fig. 3, we show the histograms of system capacity
LOLE calculated in each case. For the RTS, the MTTR is 2.5 for the Weibull Model with and . With ,
days, but when , the MTTR for the system is slightly we approximate the assumption that repair times are better rep-
lower, translating to a lower LOLE for the Weibull models. Con- resented by a log-normal type of distribution than an exponen-
versely, the use of translated to a slightly longer tial. We can see that the Weibull model has higher system ca-
MTTR, meaning that failed plants take longer to return to ser- pacity more frequently than the exponential analog. This is sup-
vice, thereby increasing the LOLE for the system. ported by the fact that the new MTTR for this Weibull model is
The second row of Table I gives the LOLE calculated by each slightly faster than the MTTR for the exponential model. The
of the models using exponential parameters. This is included as shape parameter for the failure distribution is fixed at ,
a validation step to ensure that all methods are in agreement so the MTTF is the same in both cases.
when the exponential parameters are used. It is clear that this The example we provide here is the case of and
is the case. The fact that the values differ slightly at the fourth . It is clear that the selection of increases
decimal place is merely due to slow convergence of simulation the MTTR for failed components, and this is borne out in the ca-
approaches. pacity histograms. In Fig. 4, the Weibull model clearly provides
ANDERSON AND DAVISON: AGGREGATE WEIBULL APPROACH FOR MODELING 1787
V. CONCLUSIONS
The model presented here allows the simulation of a system of
generating units, without the need for a sequential Monte Carlo
simulation, or for simplifying the system with a homogenous
Markov model. Although slightly more intricate than the ho-
mogenous model, the aggregate model gives the advantages of
a much more flexible distribution. Instead of assuming a dis-
tributional shape, this method allows the system data to dictate
the shape of the distributions for TTF and TTR. The use of the
same distribution for failure and repair times, adjusted by pa-
rameters, also brings valuable efficiencies to computation im-
plementation. Many of the calculations can be completed “up
front” instead of being computed at each time step. Additional
advantages of this model, described in [3], are its applicability
for traditional reliability engineering calculations, such as mean
time to state change and steady-state transition frequencies. In
Fig. 2. Capacity histograms for exponential and Weibull model (with = 1). addition, the aggregate model allows for efficient determination
of steady-state system durations by recursion methods.
The aggregate model is compared to the traditional exponen-
tial model and a sequential Weibull simulation on the basis of the
IEEE RTS, detailed in [2]. Our results indicate that the Weibull
models do show different LOLE values for the RTS based on
varying Weibull shape parameters. This is to be expected given
that the Weibull model has the potential to allow more or fewer
extreme events by varying these parameters. It is possible that
the use of the Weibull model will give a more realistic measure
of short-term capacity risk, particularly when repair times are
decidedly nonexponential.
The development of such a model is particularly advanta-
geous for use in the hybrid model for electricity spot price de-
scribed in [1]. The use of a Monte Carlo simulation in this model
requires that all plants are simulated for the entire duration and
capacity data is stored until the simulation is complete. Alter-
natively, the aggregate model allows determination of system
capacity at each state change, eliminating the need for storing
Fig. 3. Capacity histograms for exponential and Weibull model (with = 2).
data past the previous change. This is a necessity when mod-
eling large generating systems even over shorter time periods.
While the computational expense of using Monte Carlo simu-
lation over long time horizons can be prohibitive, use of this
aggregate model will reduce this memory load.
APPENDIX
RELIABILITY BACKGROUND
The study of engineering reliability is of particular impor-
tance in power systems. One of the most unique aspects of power
generation is the inelastic demand and the significant costs, both
financial and social, of power shortages. The consumer of elec-
tric power does not expect to experience blackouts and events,
such as the August 2003 blackout in eastern North America,
carry heavy repercussions in both the energy sector and the po-
litical landscape. The need for sufficient power to meet whatever
demand is generated has lead to significant research in this area
Fig. 4. Capacity histograms for exponential and Weibull model (with = over the last 60 years.
0:65). The first major works on engineering reliability appear in the
literature in the 1940s. Key contributions include those of Cal-
lower system capacity more frequently than the traditional ex- abrese [7] and Lyman [8], whose work pioneered the application
ponential model. of probability to system reliability.
1788 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 4, NOVEMBER 2005
Over the last 50 years, the study of engineering reliability spends in a given state. If is the state of the system at time
has developed into the main approaches of basic probability and represents the time of the th transition, then
methods, Markov and semi-Markov models, and simulation. is also a random variable whose distribution we can
It is important to note that these are not mutually exclusive choose to meet our requirements. Like the embedded Markov
methods, and some developments (frequency-duration ap- chain, the depends only on the states switched between
proaches, for example) have led to improvements of more than and and is independent of for .
one of these methods. We can determine the stationary state probabilities of each
In reliability theory, the most common choice of distribution of the states in a semi-Markov model. This method allows us
for TTF and TTR is the exponential distribution. The advan- to use the appropriate distribution for the appropriate task. For
tage of applying an exponential distribution is that the failure example, one can apply a more realistic bell-shaped distribution
rate and the repair rate are constant, allowing the application of to the duration of repair times.
homogeneous Markov techniques to the component model [9]. The general approach developed by [3] begins with the homo-
This model is very tractable, even for complex generating sys- geneous Markov chain for the system states. Beginning with the
tems. Due to its simplicity, the exponential distribution is often transition probability matrix, obtained from the transition rates,
applied to the repair of components as well. However, there is we can determine the stationary transition probability vector
some indication that repair times may be more suited to a log- based on the exponential distribution
normal distribution on an individual basis [3], as significantly
longer repair result from the need for parts or local expertise [6].
The lognormal distribution, while mathematically less tractable,
can still be utilized in a simulation approach to the problem. Here, represents the embedded chain probabilities. Using
It should be noted that the exponential distribution has some these steady-state probabilities, we can determine the steady-
very strong implications for the model. First, the exponential state probabilities for a Weibull–Markov model as follows:
distribution is memoryless. This implies that the expectation of
the time remaining in the current state is independent of the time
already spent in the state. Although counterintuitive, this model
has been shown to be somewhat successful for a component in
its useful life. However, the tractability of the model often out- where is the expected duration for state [13].
weighs its disadvantages, and as a result, the exponential distri- The expected duration of the transition can be found from the
bution has become the most common of the distributions applied expected value of the distribution, which is given in [14] by
to reliability modeling [10].
In some cases, it is possible to make a non-Markovian system
Markovian using supplementary variables. In theory, this allows
the use of other distributions in the Markov framework. In prac- The steady-state probabilities obtained in this manner will now
tice however, the use of more than one supplementary variable include the durations that follow the Weibull distribution, al-
for any state of the system results in a model too complex for lowing the inclusion of nonexponential repair times.
practical use[11]. The normal distribution, while simple, is not often used in
An additional approach for nonexponential durations is the reliability applications, although it has been used to model the
“device of stages” technique [11], [12]. This technique involves failed capacity of an entire system. As previously discussed, this
the representation of each state of the system as a number of approximation is only possible when the system contains a very
“stages,” either in series or in parallel. Each of these stages is large number of units or components.
represented by its own exponential distribution. The composi- Monte Carlo simulation (MCS) is another popular technique
tion of these stages allows most applicable distributions to be for modeling generating plant failures. The main advantage of
approximated. Systems modeled in this fashion may be solved MCS is the flexibility and realism it provides to the model. The
analytically, but if the complexity is great, numerical solutions disadvantage of such an approach is its computational expense,
are relatively straightforward. slow convergence, and lack of intuitive support. In some cases,
Another distribution that has been applied to generation the flexibility of the approach outweighs its disadvantages, in
reliability is the Weibull distribution. We use the two-parameter particular for the inclusion of such things as managerial deci-
version of the Weibull distribution, which allows a great deal sion making. This advantage is illustrated in works by Allan and
of flexibility in fitting data. The exponential distribution is Roman [15] and Li and Billinton [16].
merely a special case of the Weibull distribution. Although use MCS has been used extensively for complex multiarea sys-
of the Weibull distribution further complicates the problem, it tems. In multiarea systems, a supporting policy, which deter-
does allow the use of semi-Markov techniques [3]. The state mines the power serving priorities in the overall region, is an
transition of the system is a standard Markov chain, called integral part of the system. The sensitivity of reliability indices
the embedded Markov chain. The transition probabilities of to such management policies is addressed in [16], as the authors
the embedded chain are obtained using standard methods for use MCS to incorporate management decisions in the generation
homogeneous Markov models [6]. model. The simulation of generation plants is based on the work
The distinguishing feature of the semi-Markov model is the presented in [17], wherein a switching model is used along with
random variable that describes the duration that the system the forced outage rate (FOR) for each unit.
ANDERSON AND DAVISON: AGGREGATE WEIBULL APPROACH FOR MODELING 1789