2016 IEEE Innovative Smart Grid Technologies - Asia (ISGT-Asia)

Melbourne, Australia, Nov 28 - Dec 1, 2016

Distribution System Reliability Assessment Using

Sequential Multilevel Monte Carlo Method

A.S.N. Huda Rastko -!!/

School of Electrical & Electronic Engineering School of Electrical & Electronic Engineering
The University of Adelaide The University of Adelaide
Adelaide SA 5005, Australia Adelaide SA 5005, Australia
asnazmul.huda@adelaide.edu.au rastko.zivanovic@adelaide.edu.au

Abstract&The paper proposes a new time sequential multilevel improvement strategies in the distribution system are adding
Monte Carlo (MLMC) method for estimating distribution system protection devices, switching and reclosing devices,
reliability. Usually, the reliability indices of a distribution system accessibility of alternative power supply, speed up the repair
are accurately assessed by sequential Monte Carlo simulation process and system reconfiguration [3]. Therefore, fast
(MCS). A disadvantage of sequential MCS is the computational assessment of reliability indices is a primary need to speed up
burden which may be prohibitive for achieving a high accuracy. the process of reliability improvement.
The aim of the proposed method is to enhance the computational
efficiency of sequential MCS. The basic idea of MLMC is to use Distribution system reliability can be evaluated by two
the stochastic differential equations (SDE) based approximation approaches: analytical and simulation. Analytical approach is
of the reliability indices with multiple timesteps. In MLMC concerned only with the mean values of the indices regardless
method, a large number of simulations are performed on the their future performance analysis. On the other hand, Monte
coarse grid levels at a low cost and few on the fine grids at a high Carlo method [4] is widely utilised simulation approach to
cost in such a way which can reduce the overall variance of the overcome this deficiency of analytical approach. Monte Carlo
reliability indices estimators of MCS for the desired accuracy. simulation (MCS) can be classified into non-sequential and
Comparisons with the analytical and sequential MCS methods sequential modes [5]. Sequential MCS is an accurate tool to
are carried out on a small distribution system to demonstrate the simulate the chronological issues of the complex power
effectiveness of the proposed method. The results verify that the systems. A disadvantage of sequential MCS is the
proposed method gives acceptable values of reliability indices by computational burden when a large number of samples are
reducing their calculation time compared to sequential MCS.
necessary to obtain a high accuracy. To lessen the
computational burden of sequential MCS, several approaches
Keywords
distribution system; reliability; multilevel Monte have been proposed. The approaches include parallel
Carlo; computational efficiency. computation [6], state-space pruning [7], and variance
reduction techniques [8]. In this paper, a new variance
reduction approach based on multilevel Monte Carlo (MLMC)
I. INTRODUCTION method is proposed with application to power distribution
system reliability assessment.
Performance of a power distribution system is measured
by reliability indices. Customers demand a distribution system The idea of MLMC was first presented by Heinrich [9] for
with low interruptions. In the distribution systems, more than the computation of high-dimensional, parameter dependent
80 percent of the customer interruptions occur due to the integrals, and was later applied to stochastic differential
failure of distribution components [1]. Reliability assessment equations (SDE) by Giles [10]. MLMC method uses an
is a probabilistic method that measures past performance and optimal number of samples at different levels to estimate the
predicts future performance of a distribution system. The expected values of uncertainty parameters. In this method,
assessment techniques utilise a set of data including feeder more samples are simulated at the coarse levels with low
lengths, loading, customer data, failure rate and restoration computation cost and few samples are required at the more
time of components. Some common system performance expensive fine levels due to the smaller variances [11]. In this
indices are system average interruption frequency index way, the overall computation time is reduced compared to the
(SAIFI), system average interruption duration index (SAIDI), MCS, where all the samples are simulated on the
customer average interruption duration index (CAIDI), computationally most expensive finest level. Readers are
average service availability index (ASAI), average service encouraged to refer [10] for detailed mathematical proofs of
unavailability index (ASUI), and energy not supplied (ENS) MLMC method.
[2]. Based on these indices information, the distribution
This paper will examine the applicability of the proposed
system planners and designers can take necessary steps to
sequential MLMC method in distribution system reliability
improve the system reliability. Some of the reliability
evaluation. The impacts of each component failure on the load

978-1-5090-4303-3/16/$31.00 ©2016 IEEE 867

points are determined by average failure rate and annual  J _ 6l J # m \ 2l J # m



* ` ` 
(1)
unavailability indices of load point. To find the effects of load
points interruption frequency and duration on the overall where J is the value of component random variable, TTF at a
distribution system performance, system performance indices time . The solution of this SDE can be obtained by using a
are determined. The proposed method will calculate three discretization scheme. In this article, Euler-Maruyama (EM)
system performance indices: SAIFI, SAIDI and CAIDI. The discretization is used for an approximate numerical solution of
performances of the MLMC method are evaluated based on SDE [18]. An approximation of the solution of SDE is found
the accuracy and computational efficiency. The outcomes of by linear interpolation of 2 # % % # G [19]. The EM
the proposed method are then compared to the analytical and discretization with n timesteps, step size  _ - and
sequential MCS approaches in order to check the validity of Brownian increments 0G is expressed as follows:
the proposed technique.
II. METHODOLOGY OF RELIABILITY EVALUATION USING G83 _ G \ 6l G # G m \ 2l G # G m0G # (2)
SEQUENTIAL MLMC
where 0G _ G83 ] G with  _ *# % % #  ] + and
A. Modelling of Reliability Indices G _  with  _ *# % % # .

03 # 04 # % % # 0H93 are
In the current study, distribution components are assumed independent and normally distributed random variables.
to represent by two-state (up-down) models [12]. The time The artificial operating histories of each component are
during which the component remains in the up state is called generated in chronological order using a random sampling
time-to-failure (TTF). The time during which the component technique from the exponential probability distributions of
is in the down state through repairing, replacing of component component failure time, G83 from (2). Finally, KI is
or switching is called time-to-repair (TTR). Failure of a
component depends on several factors including component generated by converting the uniform distribution random
type, size, design, weather and geographical location [13]. The variable into an exponential distribution using the inverse
process of failure is a completely random event and there is no transform method [20].
such assurance that a component will fail according to the Assume that G83 _ MEF . Then, KI using SDE model of
failure history of similar components. Therefore, the failure *
!
-to-failure G83 can be expressed as
rate and repair time of a component are random variables and follows:
simply an approximation of the actual failure rate and repair
time. In MLMC method, SDEs are utilised for the KIEF _ ] MEF ^ l+m# (3)
approximation of the component actual failure rate and repair
time. A time sequential simulation process is usually used for
reliability assessment of distribution system. The key where + is a uniformly distributed random variable between
prerequisite in time sequential simulation is to generate the [0, 1].
artificial operating (KI m
and restoration (CH m histories of Likewise, following the above procedure, the SDE model
system components. Reliability indices can be evaluated using of another random variable, TTR can be determined. For TTR,
the values of KI and CH . the values of 
%

!"
2


 


TTF stochastic process. If "EF is a constant for component j
In the distribution system, reliability is generally assessed
with the useful life period of the component. Where the failure connected to  and "EF _ EF ^ )EF , then initial time-to-repair
rate of a component is assumed as constant and exponential for component j connected to  is 2 _ "EF -)EF . Therefore,
probability distribution is usually employed for modelling the like (3), CH can be expressed as follows:
randomness of TTF [14]. On the other hand, one of the
requirements to apply the proposed MLMC in reliability CHEF _ ]"EF ^ MEF ^ l,m# (4)
evaluation is to construct the SDE based modelling of TTF.
Therefore, a combination of SDE and exponential probability where ,
is a uniformly distributed random variable between
distribution [15] based modelling of random variables is [0, 1]. The values of average failure rate, (EF and average
proposed in this study. unavailability, EF for a component j connected to  can be
At first, we construct the SDE model of TTF for a calculated using the following expressions:
component j connected to load point i (). An average
failure rate (failures/year) and an average repair or switching 4EF _ 1T
>P
# (5)
time (hour/failure) for this component are considered as 5EF YQO ?[ZWXlYm

and EF , respectively. The randomness of variable TTF of this

1TYQO ?VYWXlYm
component is presented by the Brownian motion,  on the EF _ # (6)
?VYWXlYm
time interval [*# ] [16]. The rate of change of the average 1YQO ?[ZWXlYm 81T
T
YQO d
7652
value of stochastic failure process (drift), and the degree of
variation of stochastic failure process over time (volatility) are where 4 is the number of times component j failures during
considered as μ 
2
!"
Then, the SDE model of entire simulation period and N is the desired number of
TTF with given specific μ, 2 parameters and an initial time-to- simulated periods.
failure l 2 _ +-5EF m can be expressed as [17]:

868
 Now, average failure rate, 4E and annual unavailability, E 3
=; _ 1> SR lEm
E:3  A $ (13)
for  could be calculated as follows: >SR

93
HX The variance of this estimator is =; n A o, so the root
4E _ 1F:3 4EF # (7)
mean square error (RMSE) is
l+-j =; m. To achieve an
HX  "

1
 
=; _
l194 m samples. Therefore, for
E _ 1F:3 EF # (8) the =; estimator being a sufficiently accurate approximation
of no with small , a large number of samples need to be
where F is the number of components fails that affect the simulated. This results in huge computational cost.
service of . The proposed reliability indices can be defined The basic idea of MLMC method is simple. Instead of
as follows: estimating no by n A o on a single fine level, MLMC
SAIFI finds the average number of sustained interruptions method uses a number of levels of resolution, . _ *#+# % % # 
per customer during a year in the distribution system. The with . _ * being the coarsest level and . _  being the finest
index is determined by dividing the total annual number of level. Mathematically, the idea of MLMC can be formulated
customer interruptions by the total number of customers as follows:
served during this period. The unit of this index is
interruptions/system customer.year. n A o _ n 2 o \ 1<.:3 n . ]  .93 o _ 1<.:2 n. o# (14)
YW L >
1WQO W RW where for simplicity we have set 2 _  2 for . _ * and
 _ # (9) . _  . ]  .93 for 3 a *. In MLMC method, each of the
>U
expectations on the right-hand side of (14) is independently
where ; E is the number of customers at , ? is the total estimated by MCS using a different number of samples at
number of customers served and E is the number of load different levels in such a way which minimises the overall
points in the distribution system. variance for a given accuracy.

SAIDI is designed to provide information regarding the Let k2 is an unbiased MCS estimator for n 2 o using 2
average duration of interruption for each customer during a samples for . a * and k 3 for 3 c + is the MCS estimator for
year. It is calculated by dividing the sum of all durations of n 3 ]  393 o using 3 samples. Then we have
service interruptions to customers by the total number of
customers served in the distribution system. This index is 3 lEm
k2 _
 1>
E:3  2 #
N
(15)
measured in the unit of hours/system customer.year. >N

3
YW @ >
1WQO W RW and k
3 _ 1> lEm lEm
E:3h 3 ]  393 i#
3
(16)
 _ $ (10) >3
>U

The estimator for n 3 ]  393 o is computed in the form of

CAIDI determines the average outage duration or average D3] B393 . where D3 is a fine path estimator using timestep
restoration time that any given customers would experience. It size [22]
is measured in the unit of hours/customer interruption.
YW @ > D _ ,93 # (17)
1WQO W RW
 _ YW L >
$ (11)
1WQO W RW
and B393 is the corresponding coarse-path estimator using
timestep size [22]
B. MLMC Approach
Before introducing the MLMC method, MCS method is B _ ,9l393m . (18)
briefly reviewed in this section. Let n A o be the
approximation of no. MCS states that an expected value Therefore, on the coarsest level (3 _ *), the simulations
n A o of a function of simulated random variable can be use just 1 timestep, while on the finest level (3 _ ), the
approximated by averaging over a large number of samples on simulations use ,< uniform timesteps. Each level has twice as
a single fine level [21]. In this study, we can define  as many timesteps as the previous level. To avoid the
follows: introduction of an undesired bias, we require that
# 

! # fD3 g _ nB3 o. (19)
 _ e # 

! # (12)
# 

! $
Based on the expected value computed on the coarsest
lEm level, the expected difference from this level to the next finer
If A is the ith sample of  A and =; is the total number level n 3 ]  393 o is added, until the finest level

is reached.
of Monte Carlo samples. Then, the MCS estimator is As the level 3 increases and the grid resolution becomes finer,

869
the required time increases, but in the meantime the required c) Desired accuracy, ;
number of samples decreases. This suggests that MLMC runs d) Drift value, μ; and
most of the iterations on the cheaper lower levels and just a e) Volatility, 2.
few samples on the computationally expensive higher levels.
In this way, the total computational time is significantly saved In the next step, for a component j connected to the ,
compared to MCS which spends all its effort on the SDE models of a random variable, TTF are constructed at
computationally most expensive finest grid. Since MLMC levels 3 _ * and 3 a * using (2), (17) and (18). Two random
considers the estimations on a sequence of grids so that the numbers between 0 and 1 are generated using the MATLAB
less accurate approximation on the coarsest grid is function rand. Then, these two numbers are converted into
sequentially corrected by the estimators on the following finer KIEF and CHEF of the component according to exponential
grids and thereby achieves the finest grid accuracy. Thus probability distributions using (3) and (4), respectively. After
MLMC can achieve the similar values of reliability indices as that, the values of average failure rate, (EF and average
MCS with less computation time. unavailability, EF
are calculated at level . _ * and . a *. In a
Finally, by using (15) and (16), the overall estimator of similar way, the values of average failure rate and
MLMC for each reliability index can be expressed as follows: unavailability for each component in the system are
determined. The load points affected by each component
k3 $ failure is found using the method in [4]. The values of average
=< _ 1<3:2  (20) load point failure rate, (E (failures/yr) and average annual
unavailability, E (hr/yr) are determined by accumulating the
Like MCS, MSE of the MLMC estimator also consists of individual component value connected to the relevant load
two terms: variance of the estimator 1<3:2 393 3 and point. The system performance indices SAIFI, SAIDI and
approximation error nl A ] mo4 : CAIDI are determined at level . _ * and . a *. Finally, the
sum of reliability indices values at level . _ * and . a * is
< 93
l=< 4 4
A m _ 13:2 3 3 \ nl A ] mo $ (21) calculated. The whole process is repeated until the number of
samples is reached to N.
In order to ensure the MSE of MLMC estimator is less
After estimating the values of reliability indices at level
than 14 , it is sufficient to confirm that both 1<3:2 393 3 and 3 _ * and 3 a *, the overall MLMC estimator for each of the
nl A ] mo4 are less than 14 &,. The value of 3 on each reliability indices is determined in order to achieve the target
level 3 ensures that the estimated variance of the combined accuracy. Initially, the minimum refinement level of MLMC
multilevel estimator is less than 14 &,. Therefore, it is essential method is set at  _ ,. Then, the number of samples Ns at
to choose optimally 3 to obtain the optimal MLMC levels 3
_ *# +# ,
is determined using initial number of
convergence. To make 1<3:2 393 3 b 14 -,, the optimal 3
is samples 2 . At the same time, the sum of reliability indices
chosen as [23] values is updated at level 3
_ *# +# ,. Then, absolute value of
average of reliability index, 3 _ 'n3 o' and variance
3 _ ,194 j3 -3 h1<3:2 j3 3 i$ (22) 3 _ n3 o 
 


!
3




number of samples 3 at each level 3
_
*# +# $ $ $ $ $ #  is
where 3 is the cost of an individual sample on level 3. The determined based on (22). Next, the optimal 3 is compared to
test for weak convergence tries to ensure that nl A ] mo4 b the already calculated number of samples Ns at that level. If
14 -,. If the convergence rate of l A ] m with 3 for some the optimal 3 is larger, additional samples at each level as
constant 3 is measured by a positive value
[23]. Then, needed are evaluated. The values of the mean and variance at
each level are then updated. The aim to determine the optimal
n 3 ]  393 o b 3 ,903 # (23) 3 is to keep the variance of the estimator 1<3:2 393 3 less
than 14 &,. The weak convergence of MLMC estimator is
and the remaining error is tested using (25). This ensures that the remaining bias error is
less than 1-/,. If remaining bias error is greater than 1-/,,
n A ] o _ n < ]  <93 o-,0 ] +# (24) then  _  \ + is set. Then, the whole process is repeated
again until reaching the target accuracy level. Finally, the
overall multilevel estimator for each reliability index is
This leads to the convergence test
computed using (20).
n < ]  <93 o-,0 ] + ` 1-/,. (25) III. CASE STUDY
A. Test System
C. MLMC Simulation Procedure
Three system performance indices; SAIFI, SAIDI and
At first, failure rate and repair time of each component of CAIDI are calculated by applying the proposed method to a
the distribution system is defined. Some MLMC simulation simple radial distribution system [24] as shown in Fig. 1. In
parameters are also defined: this case study, it is assumed that the feeder breaker and fuses
a) Number of samples for convergence tests, N; are completely reliable because of their regular maintenance.
b) Initial number of samples on each level 3, 2 ; Disconnect switches S1 and S2 are normally closed and S3 is
normally open. Some data are assumed for this study. Average
870
switching time for switches S1, S2 and S3 are considered as MLMC is about 20 times faster than the MCS. Similarly, for
0.5, 0.5 and 1hour, respectively. For the main feeder and SAIFI and CAIDI calculation, MLMC method saves huge
lateral line, the interruption rates are 0.1 and 0.25 computation time compared to the MCS. From the Table, it is
interruptions/km/yr, respectively and the average time to seen that the maximum and minimum TS using MLMC
repair are 3 hours and 1 hour, respectively. Number of approach are 96.9419% (=0.001) and 93.6351% (=0.0005)
customers at load points LP1, LP2 and LP3 are 250, 100 and in SAIFI and SAIDI calculation, respectively. On average
50, respectively. Pre-defined parameters for all reliability sequential MLMC can reduce the simulation running time
indices calculation are
= 10000, 2
= 500, μ = 0.01 and 2 = about 95% compared to the sequential MCS.
0.4. The drift and volatility values of stochastic failure and
repair process are generally determined by using a time series TABLE I. SAIFI, SAIDI AND CAIDI VALUES
of TTF and TTR values [25]. These values are usually not  MCS MLMC DMC (%) DMLMC (%)
available in distribution system reliability analysis. Therefore, SAIFI (interruptions/system customer. yr)
the values of drift and volatility are determined by adjusting 0.0001 1.2129 1.2293 1.3902 0.0569
based on the accuracies for an index calculation and assumed 0.0005 1.2131 1.2302 1.3739 0.0162
as constant for rest of the indices calculation. Using analytical 0.001 1.2154 1.2252 1.1869 0.3902
technique, the system performance indices; SAIFI, SAIDI and SAIDI (hr/system customer. yr)
0.0001 1.5125 1.5318 0.1655 1.4430
CAIDI can be evaluated [24]. In this study, the results from 1.5130 1.5237 0.1986 0.9072
0.0005
analytical technique are considered as the base case. The 0.001 1.5134 1.5262 0.2251 1.0728
methodology was implemented using MATLAB and all CAIDI (hr/customer interruption)
computations were performed using an Intel Core i7-4790 0.0001 1.2469 1.2470 1.3739 1.3821
3.60-GHz processor. 0.0005 1.2472 1.2471 1.3983 1.3902
0.001 1.2474 1.2485 1.4146 1.5040

TABLE II. COMPUTATION TIME FOR SAIFI, SAIDI AND CAIDI ASSESSMENT

 TMC (s) TMLMC (s) TS (%)

SAIFI
0.0001 69.2342 2.4362 96.4812
0.0005 3.1049 0.1023 96.7052
0.001 1.0039 0.0307 96.9419
SAIDI
0.0001 177.842 9.0342 94.9201
Fig. 1. A simple radial distribution system 0.0005 10.2312 0.6512 93.6351
0.001 3.6482 0.1136 96.8861
B. Test Results CAIDI
0.0001 146.7123 7.6191 94.8067
The values of system performance indices computed using 0.0005 9.4883 0.3779 96.0164
sequential MCS and MLMC methods are shown in Table I. 0.001 3.0238 0.1326 95.6134
Three different target accuracies 0.0001, 0.0005 and 0.001 are
used to approximate these indices. DMC and DMLMC are
expressed in percentage. These are the absolute values of the
difference of reliability indices values of sequential MCS and
MLMC with respect to the base case value. It is found that the
results using MLMC method compared to both analytical
approach and MCS are very close. The maximum deviation of
the results computed by the proposed approach is 1.5% which
occurs in CAIDI calculation (=0.001). For SAIFI and SAIDI
calculation, the maximum DMLMC are 0.3902% (=0.001)
and 1.4430% (=0.0001), respectively. These results are (a) (b)
acceptable for distribution system reliability evaluation which
Fig. 2. Required number of samples for convergence [SAIDI] (a) at each level
depends on uncertainty quantifications. This proves the of MLMC and (b) MCS
accuracy level of the proposed MLMC approach.
Table II presents the calculation time of reliability indices. Fig. 2(a) shows the required number of levels and samples
TMC and TMLMC are the CPU time in second (s), taken by the at each level of MLMC for system index SAIDI calculation to
MCS and MLMC methods, respectively to reach the target achieve  = 0.0001, 0.0005 and 0.001. For the desired
accuracy. TS is the percentage of computation time saving by accuracy, the value of . 


!
3
increases.
MLMC approach over MCS simulation. Results show that the For example, when  = 0.001, three levels are needed and the
time taken by sequential MCS is significantly more than that . at the highest level . = 3 is just 41; when  = 0.0005, the
taken by MLMC to achieve the same accuracy. From the . at the highest level . = 3 increases significantly to 148.
results, the maximum CPU time using MLMC method is Fig. 2(b) displays the number of required samples for MCS.
9.0342s which is needed for SAIDI calculation (=0.0001). To Comparing between Fig. 2(a) and 2(b) indicates that in the
reach this accuracy level, MCS needs 177.842s. In this case, case of MCS, the number of required samples to reach the

871
desired accuracy is higher than the number of optimal samples [4] R. Billinton and P. Wang, 'Teaching distribution system reliability
at the highest level of MLMC. For example, when  = 0.0001, evaluation using Monte Carlo simulation,) IEEE Trans. Power Syst.,
vol. 14, pp. 397-403, 1999.
14503 samples are required for MCS convergence. Where
[5] W. Li, Risk Assessment of Power Systems: Models, Methods, and
only 3452 samples are required at the highest level . = 3 for Applications, 2005, IEEE Press/Wiley.
MLMC method. This is how MLMC saves computation time [6] C. L. T. Borges, D. M. Falcao, J. C. O. Mello and A. C. G. Melo,
by performing most of the simulations at the cheaper lower 'Composite reliability evaluation by sequential Monte Carlo simulation
levels. Similarly, MLMC simulation-based calculations of on parallel and distributed processing environments,) IEEE Trans.
SAIFI and CAIDI reduce computational cost compared to Power Syst., vol. 16, no. 2, pp. 203-209, 2001.
MCS technique. [7] C. Singh and J. Mitra, 'Composite system reliability evaluation using
state space pruning,) IEEE Trans. Power Syst., vol. 12, no. 1, pp. 471-
479, 1997.
[8] B. Zhaohong and W. Xifan, 'Studies on variance reduction technique of
IV. CONCLUSIONS Monte Carlo simulation in composite system reliability evaluation,)
Elect. Power Syst. Res., vol. 63, no. 1, pp. 59-64, 2002.
In this paper, a new sequential multilevel Monte Carlo
[9] S. Heinrich, 'Multilevel monte carlo methods,) in Large-scale scientific
method has been presented for reducing the calculation time of computing, vol. 2179, Lecture Notes in Computer Science, Springer-
distribution system reliability indices. MLMC is a variance Verlag, 2001, pp. 58-67.
reduction technique for MCS and improves the computational [10] M. B. Giles, 'Multilevel Monte Carlo methods,) Acta Numerica, vol. 24,
efficiency of sequential MCS by conducting the reliability pp. 259-328, 2015.
indices calculations on a geometric sequence of grids. Where a [11] 
$
'An introduction to algebraic multigrid)

 


large number of simulations are performed on the 




$





computationally cheaper coarse grids with low accuracy at a London, 2001, pp. 413-532.
low cost and relatively few simulations are performed on the [12] B. Retterath, A. A. Chowdhury and S. S. Venkata, 'Decoupled
computationally expensive fine grids at high accuracy and a substation reliability assessment,) Int. J. Elect. Power Energy Syst, vol.
27, pp. 662-668, 2005.
high cost. Thus, the overall computation time is reduced using
[13] B. Retterath , S. S. Venkata and A. A. Chowdhury, 'Impact of time-
MLMC method when compared to MCS which calculates the varying failure rates on distribution reliability,) in Proc. Int. Conf.
reliability indices only on the computationally most expensive Probabilistic Meth. Appl. Power Syst., pp. 953-958, 2004.
finest grid. [14] A. Chowdhury and D. Koval, Power Distribution System Reliability:
Practical Methods and Applications, 2009, Wiley-IEEE.
The proposed MLMC approach has been tested on a small
[15] R. Zárate-Miñano and F. Milano, 'Construction of SDE-based wind
radial distribution system. Three system performance indices speed models with exponentially decaying autocorrelation,) Renew.
(SAIFI, SAIDI and CAIDI) have been computed at different Energy, vol. 94, pp. 186-196, 2016.
accuracies. Two basic random variables time-to-failure and [16] J. M. Harrison, Brownian motion and stochastic flow systems, 1985,
time-to-repair of each component are modelled by jointly Wiley New York.
using stochastic differential equations and exponential [17] B. Oksendal, Stochastic differential equations: an introduction with
probability distributions. Comparisons with the sequential applications, 2013, Springer Science & Business Media.
MCS method and base analytical approach demonstrated the [18] M. B. Giles, 'Multilevel Monte Carlo for basket options,) in Proc. of
accuracy level and computational efficiency of the proposed the 2009 Winter Simulation Conference (WSC), Proceedings of the 2009
method. The results show that there is no significant Winter, 2009, pp. 1283-1290.
difference between the outcomes of comparison studies. The [19] P. E. Kloeden and E. Platen, 'Higher-order implicit strong numerical
schemes for stochastic differential equations,) J. Stat. Phys., vol. 66, pp.
differences in reliability indices calculated data using MLMC 283-314, 1992.
are within 1.5% of values using an analytical approach. The [20] R. Billinton and S. Jonnavithula, 'A test system for teaching overall
MLMC approach speeds up the reliability indices computation power system reliability assessment,) IEEE Trans. Power Syst., vol. 11,
up to 97% compared to sequential MCS. The only feature that pp. 1670-1676, 1996.
the proposed method lacks in comparison to the sequential [21] M. H. Kalos and P. A. Whitlock, Monte Carlo methods, 2008, John
MCS is the ability to determine the probability distributions of Wiley & Sons.
these reliability indices. [22] M. Giles, 'Improved multilevel Monte Carlo convergence using the
Milstein scheme,) in Monte Carlo and quasi-Monte Carlo methods 2006,
In future, the proposed MLMC method will be tested on Springer, 2008, pp. 343-358.
the complex distribution networks of a large number of [23] M. B. Giles, 'Multilevel monte carlo path simulation,) Operations
components. Probability distributions of system performance Research, vol. 56, pp. 607-617, 2008.
indices will be evaluated to illustrate the capability of the [24] R. Billinton and W. Li, 'Distribution System and Station Adequacy
proposed approach compared to the sequential MCS. Assessment,) in Reliability assessment of electric power systems using
Monte Carlo methods, Springer, 1994, pp. 209-254.
[25] K. Wang and M. L. Crow, 'Fokker-Planck equation application to
analysis of a simplified wind turbine model,) in Proc. of North
REFERENCES American Power Symposium (NAPS), 2012, pp. 1-5.
[1] R. Billinton and J. E. Billinton, 'Distribution system reliability indices,)
IEEE Trans. Power Delivery, vol. 4, pp. 561-586, 1989.
[2] R. Billinton and R.N. Allan, Reliability Evaluation of Power Systems,
1996, Plenum Press, N.Y. and London.
[3] A. Kavousi-Fard and M. R. Akbari-Zadeh, 'Reliability enhancement
using optimal distribution feeder reconfiguration,) Neurocomput., vol.
106, pp. 1-11, 2013.

872