Professional Documents
Culture Documents
DepartmentqœSystemsEngineering,CaseWesternReserveUniversity,Cleveland,Ohio
BENJAMIN F. HoBBs
Departments
of SystemsEngineeringand Civil Engineering,
CaseWesternReserveUniversity,Cleveland,Ohio
forthesingle-step
Markov chaintransitionmatrixTt. Equa- w(t + At)= [c(t)- q(t)]At + w(t) (6a)
tion(5a)showsthat we can solvefor any P• in the same 0 < [c(t) -- q(t)]•t + w(t)< w•, x
manner
as for the time-invariantcase.If P• exists,then the
long-term
probability
that the system
is in system
statei at w(t+ At) = 0 [c(t)-- q(t)]At+ w(t)< 0 (6b)
:ho,.ur
t isPwFromthePt andTt, wecanfindexpected fre- [c(t)- q(t)]At q-w(t)> Wm•
•. (6c)
w(t q- At)= Wma
x
quent.es
anddurationsfor all capacity-demand-storage
states.
Thetwonumerical
methods
wejust described
for finding in which
st.eady
stateprobabilities
involvematrix operations.
For "real
world"water
supply
systems,
thematrices
Tt willbetoobigto w(t) storagelevelat time t'
::'be
handled
bymost
computers,
because
ofthehuge
number
of q(t) demand
levelthesystem
isfacing
at t'
ayetern c(t) available
exactstates.We avoidthe problemof unusuallylarge capacity
of thesystem
at t'
matrice•
byexploiting
independence components.w•,• reservoir
of capacity capacity.
1448 BEIMAND HOBBS'MARKOVMODELOF CAPACITYRELIABILITY
empty,
differ
fromtheD(S)andP•sin theworkbyHobbs
and inadequate
capacity
will be biaseddownwards
by no more
Beim[thisissue].
Thelatterreferonlyto capacitydeficits
ofat than 2%.
least
S,irrespectiveof theamountofwaterin storage.)
VERIFICATION
OFTHEMODIFIEDFREQUENCY-DURATION
INCLUDINGON-STREAMSTORAGEIN THE MODELS AND MARKOV MODELS
Our analytical simulation models, including the above Thissectionpresents
an applicationthat illustrates
andver-
Markovmodel,are mostappropriatefor systems whichhave ifiestheanalytical
simulation
methodologies describedin this
capacity
components
in paralleland/or series,random and in our companion
paper[HobbsandBeim,1986•.We
demand,
and moderateamountsof finishedwater storage. addressthefollowingquestions.
Among the"capacity
components"canbe streamflow,if there 1. Do the modelsindeedprovidesatisfactory estimates
of
isnosignificant
rawwaterstorage
in theformof reservoirs. reliabilityindices,whencomparedto Monte Carlosimulation
However,there are many water supply systemswhich store results ?
large
amounts of raw water.Somesystems
haveenoughstor- 2. In particular,do the simplifying assumptions madeby
agetomeetseveralmonths ofdemand.Representing
thewater each of the analytical methodsaffect their results?
rapplyas a component whosecapacityequalsthe natural 3. How do their computationaltimes compareto the
streamlow is clearly inadequatein such cases,sincethe reser- timesneededto obtain tight boundsfor indicesderivedfrom
voirservesto even out flows over time and releasescan be Monte Carlo simulation ?
alteredin responseto system demand and capacity avail-
ability.
We suggest two approaches for interfacingreliability The Nahal Oren Water SupplySystem
modelsof capacityand largereservoirs. The Nahal-OrenProjectsupplieswaterfor the city of Haifa
One approximationthat might be adequatefor planning in Israel.A reliabilityanalysisof this systemwasundertaken
purposes wouldbe to assume that the supplycapacityis the in Damelinet al. [1972] usingMonte Carlo simulation.In
outflowof the reservoir, based upon an operating policy essence, the systemcan be describedas having10 pumpsin
chosen
without consideration of outages of other systemcom- parallel which feed into a reservoir 4000 m3 in size, from
ponents or randomvariations in demand.Theprobabilitydis- whichwateris drawnto meetdemand.Two of thesepumps
tributionof reservoirreleasescan then be estimatedby simu- are boosterpumpswhichpumpfrom a second,
largersupply
latingthe operationof the reservoirwith that policy.Transi- reservoir,and the other eight are wells.
tionratesbetweenvarious levelsof releasesr (e.g.,from r < R We make the following assumptions,as do Damelin et al.
to r > R) can be calculatedfrom the resultsof the simulation. [1972].
(Vogel[1987] analysesreservoir releasesin this manner.) 1. The pipe system interconnectingthe pumps, the tank,
These probabilities
and transitionratescan then be treatedby and demandis perfectlyreliable.
thefrequency-duration analysesin the same manner as any 2. The power supply to the pumps is perfectly reliable,
othercapacitycomponent. even though power interruptionshave causedfailures of this
Underwhat circumstances might suchan approachbe ade- water system.
quate? Independencebetween(1) releasesand (2) the statusof 3. The small reservoir is operated so that is never filled
othercapacitycomponentsand demand would be sufficient. beyondapproximately 3800m3, whichwe take as its effective
Thisassumptionis not unreasonableif (1) the reservoir size, capacity.
measured in days of demand that can be satisfied,is large 4. The large supply reservoir always has an adequate
ramparedto the length of possible capacity deficits due to supply.We will make no furtherreferenceto thisreservoir.
problems elsewherein the system;(2) capacitydeficitsdue to These assumptionsare not necessaryones. For example,
othercapacitycomponentsare infrequent; and (3) finished Monte Carlo simulation, the modified FD models, and the
waterstorageis used to even out demand, with releasesfrom Markov model can all easilyhandle pipe outages,particularly
thelargereservoirs beingrelativelyconstantover the day or if the pipesare in seriesor parallelbetweenthe pumpsand the
week.Undertheseconditions,capacitydeficitscausedby daily tank (see Hobbs eta!. [1987] or Duan and Mays [1987]).
.demand fluctuations
or capacityfailureswill not significantly Power supplycan be modeledas a componentin seriesbe-
affect
the operationof the reservoir,in termsof the level of tweenthe pumpsand the tank [Hobbset al., 1987].Finally, as
releases followingsuch a deficit. This is reasonable,since explainedin the previoussectionof this paper,thesemodels
rarely,if ever,do capacitydeficitsallow utilitiesto cut back can be interfaced with simulation models of large reservoir
reservoir releases
enoughso that later releasescan be signifi- systems.
cantlyincreased. Table 1 givescapacityand reliability data for the pumps.
However, if demands increase duringdroughtperiods,the Their capacities sumto 2398m3/hour.Damelinet aI. [1972]
above approach shouldbe modified.For example, separate assume that times to failure for individual pumps are ex-
.•analyses
couldbe donefor normaland droughtperiodsto ponential.Times to repair,by contrast,are assumedto be
accountfor correlations between demands and streamflows. lognormal.Figure1 showsthe probabilitydistribution of ca-
Another approach wouldbe to studycapacityreliability pacityon outageX for the pumpsand the transitionratesto
0•lyduring theperiodsfor whichreservoir releases are ade- higherandlowerlevelsof X usedby the modifiedFD method.
to meet demands,and to assumethat during times of Theseprobabilities
.q•uate andratesare calculated with the versions
mufficientreleases thecapacity ofothercomponents isnoless of (8)-(10)of HobbsandBeim[thisissue]whichuseround-off
t.M,n
Re reservoir
release.
Thisapproximation will be suf- proceduresto restrictstatesto multiplesof a chosendis-
f•qently
accurate
forevaluating
capacity components if re- cretization.
'eases
areadequate
mostofthetime.If, forinstance,
thereare The demandpatternis a simpleone, with a two levelsof
moa,.gh
reservoiroutflows to meet demandsfor 51 out of demandin each day: a peak level during midday and an
.e.•ry
52weeks,
estimates of failuredueto off-peak
of theprobability levelat othertimes.Thesizesanddurations
of peak
1450 Bœ!M
ANDHOB•S:MARKOVMODœL
OFCAPACITY
RELIABILITY
0.95 -
•••09999
_
0.9998
x 0.9-
o 0.9997
• 0.8- 0.9995
0.9994
_o 0.75 -
0.9995
E
o 0.7-
0.9992
0.65 -
0.9991
0.6 i i "1 i i i i i i i ] i i
0 I 2 3 4 5 $ 7 8 9 10 11 12 13 14 15
Capecity on Outego X (100 cubic m/hr)
0 I 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Capocilyon OutageX (100 cubicm/hr)
Fig.1. Probability
distributions
andtransition
rates
forcapacity
onoutage
X, NahalOrenSystem.
(Top)Cumulative
distribution
ofX. (Bottom)
Transition
rates
tolower
andhigher
amounts
ofavailable
capacity.
O.g --
0.8-
0.7-
0.6-
0.5-
0.4-
0.3-
0.2-
0.!-
0.026
0.024
0.022 -
0.02 -
o. o18-
0.016 -
0.014 -
0.012-
0.01 -
0.008-
0.006 -
0.004 -
0.002 -
Fig.2. Probability
andtransition
ratedistributions
fordemand
q,NahalOrenSystem.
(Top)Exceedance
probabilities
fordemand.
(Bottom)
Transition
ratesfromq > Q to q < Q.
summerseason.Thepractical
significance
ofthedemand,
stor- it avoids
thesimplifications
justdescribed.
Weperform
Monte
age,andstateduration
approximations
areassessed
in the Carlosimulations
withbothexponential
andlognormal
times
nextsection. to failure
for thepumps to determine
whether
useofthe
former distribution is a source of bias.
Results Additional
results
areshown
in Figures
3 and4,which
displayP/ and E(UD) for a rangeof'reservoir
sizesforthe
Table
3displays
thecalculations
madebythemodified
FD MonteCarlo
model(exponential
failure
times)
andthem•-
approach
toobtain
theE(UD)andsystem
unavailabi!ity
Pfof fledFDapproach.
Duetothelongruntimes
oftheMonte
theNahalOren
system.
These
use
thedata
inFigures
1and2 Carlosimulation,
only30-year
runsweremade
forFigU•3
andtheprocedures
in theworkby HobbsandBeirn[1986]. and 4.
Table4 presentsthe reliabilityestimates
for the Nahal Oren
We obtain 99% percentconfidencelimits for the MOnte
projectas obtainedby the modifiedFD methods,the Markov
Carloestimates by assuming that statistics
for adjacent
yean
model,andour MonteCarlosimulations
of 60 yearsof oper- of the simulationare independentrandomvariables. Theva-
ation of the system.The Monte Carlo runs are the basisfor lidityof thisassumption
hasbeenchecked
by calculating
the
evaluating
theaccuracy
of the analyticalsimulation
models,
as lag-1autocorrelations
forthestatistics'
theydidnotd•fi'
BEIM AND HOBBS:MARKOV MODEL OF CAPACITYRELIABILITY 1453
TABLE 2. Nahal Oren Distributionof SummerDemand, obtained by using a larger time step and/or a larger dis-
Markov Model
cretizationinterval for the systemstates.
The results of Table 4 differ from those of Damelin et al.
Demand,
Period
of theDay m-¾hour Probability [1972], even for the Monte Carlo model. We are not able to
account
forthis.Theonlydifference
in terms
ofmodel
formu-
900 0.0275 lation is the operatingrule of the reservoir.Our Monte Carlo
I A.M. to 6 A.M. and 1000 0.1154 simulationinvokesan operatingrule identicalto the bne used
11P.M. to 1 P.M. 1300 0.1880
1400 0.6692 in the analytical simulation models, rather than the more
900 0.0275 complicatedrule based on set points used by Damelin et al..
6 A.M. to 9 A.M. and 1000 0.1154 However, we argued above that this differencewill not affect
8 P.M. to 11 P.M. 2000 0.0600 the reliability estimates. Since the estim:ites in Table 4 are
2100 0.5226
2200 0.2746 reasonablycloseto one another, we supposethat the differ-
1400 0.0716 ences between Damelin et al.'s results and ours are due either
9 A.M. to 8 P.M. 1500 0.0712 to typographicalerrorsor differencesin assumptionsnot men-
2000 0.0600 tioned in their paper.
2100 0.5226
2200 0.2746
CONCLUSIONS
However,
that approachis mosteasilyappliedto systemsin Markov model as the basis of a management model which
whichcapacitycomponents are in parallelor seriesbetween definesoptimal reservoiroperatingpoliciesvia stochasticdy-
oneor more supplysourcesand demand.If it is desiredto namic programming (in a manner similar to Shoemaker
analyze
systems with multipledemandpoints,complexnet- ['1986]) shouldbe examined.To accomplishthis,however,the
works,and a large amount of storage,Monte Carlo simula- state spacewould have to be reducedin size,which would also
tionis still to be preferred.Analytical simulationmethodsfor improve the solutiontimesfor the Markov model.
handlingmore complex systemsare under development Finally, planning models for water supply systemswhich
[Gordon andHobbs,1987-1 and are computationallymoreef- explicitly account for random capacity and demand are
ficientthan Monte Carlo simulation. However, they have not needed.The modified FD method based on the exponential
yetbeenprovento be sufficiently accurate. approximationfor duration distributionsis attractivefor such
Further work is needed on several fronts. First, as just an application.This is becauseits quick run times make it
noted,extension of analytical simulation methods to more feasibleto imbed it in an optimization model in which numer-
complex systemsis desirable.Second,better informationis ous simulationsare required.Lanseyand Mays [!987] provide
neededon the reliability of capacitycomponents;current data a framework for conducting such analyses, in which they
are spotty and their applicability to any particular water interfacea hydraulic simulation model of a distribution system
systemwouldbe problematic[Mays and Cullinane,1986]. with a generalizedreducedgradient optimization package.In
Third,the approachwe proposedfor interfacinganalytical such a framework, the reliability model would be called upon
simulationmodelsof capacitywith simulation modelsof large to estimatesystemcostsand reliability and the reducedgradi-
surface
water reservoirsystemsneedsto be testedfor its accu- entswith respectto decisionvariables,suchas size of pumps
racy and usefulness.Fourth, the possibility of using the and reservoirs. Another possibility would be to imbed the
O. OS
Upper Bound, blodlfled Freq.--Durafion Method
nd, MCS
0.0..3
o
c MCS
0.02
0.01
10
0 0.5 I 1.5 2 2.5 $ $.5 4 4.5 5 6 8
Reservoir Size, 1000 Cubl• Me!ers
Lower Bnd,FD Melhod o Exp.Appr.,FD M•hoci
Fig.3. Probability
ofbeing
ina system
failure
state
asfunction
ofreservoir
size.
1456 BEIMAND HOS•S' MARKOVMODELOFCAPACITYRELIABILITY
0.01 -
0.009
0.008
-
-
., Modified
FD
Mefhod,
Upper
Bound
0.007 -
0.005 -
0.004 -
o.oo$
o. oo2
o.001
-
-
Bound,
MC:•
• ean,
MCS
o ! - !
APPENDIX: CALCULATION OF TRANSITION RAWESFOR To start the algorithm(n = 0), the followinginitial conditiom
PARALLEL CAPACITY COMPONENTS are imposed'
Thestatespace
oftheMarkov
model
canbesignificantly
2o.c•
= ;. 2C•,0= ].t px(O)'- 1 -- r px(C1)= r (14)
reducedif parallel capacitycomponents,each with its own
transitionrates,are reducedto an equivalent singlecompo- Unfortunately,if the parallelcomponentsdifferin size,the
nentwithonesetof rates.An example of sucha setofcompo. aboveproceduremay not significantlyreducethe number of
nentswouldbe a bank of pumpsdrawingfrom a delivering states.For realisticallysizedsystems,it is desirableto usea
waterto a singlereservoiror dentandpoint,asin theworkby round-offprocedurewhichlimits X to 0, I, 21, ..., whereI is
Dame!in et al. [1972].Thetransition
ratesarecalculatedusing somepredefinedincrement.To accomplishthis,we generalize
the following recursive formulae. These formulae can be
theprocedure of AllanandTakieddine[1977].Bythismethod,
viewedas generalizations of thosedeveloped by Allan and
addinga unit of capacityC•+• is equivalentto adding two
Takieddine[1977].
units,oneof capacitymI and oneof capacity(rn+ 1)I,ral•
Let X be the amountof capacityon outagefor a set of n
C•+t < (m+ 1)I. The mI unit has a probabilityof P• = [(•
components in parallel,
andlet 20' bethetransitionratefrom + 1)I- C•+•]/!, whilethe (m+ 1)I unit hasa probability
of
outagelevel X = i to X = j. For example,if thereare three
P,,+t =(C.+•-mI)/I. Thus P,,+P=+•= 1 and Pdn/
pumpsin parallel,each having a capacityof 0.2 m•/s, then
+ P,,+z(rn+ I)I = C,+ •. Equation(13) is appliedseparately
2o.,,,o.2'
wouldbe thetransition ratefrom0.4 m•/son outage to eachof thesetwo unitsyielding,as a preliminarystep,two
(twopumpsout)to 0.2m3/sonoutage(onepumpout).The
transitionrate2ufora setof n + 1 pumpsis
setsoftransition ratesbetweenrounded states2m•jand,•,+1,0'
In a secondstep,thosetwo setsof ratesare combined:
2o= {1/p•(i)}{Px'(i){I
- r)EJ.,.f
+ ,•d(j-i- C,+
+ px'(i- C, •)r[,•l_c,+•,j_c,t'
+ #d(i-j- C,+1)]} (13) 20=2,nijPmx(i)Pm
+/•m+ •.uP=+ i.x(i)Pm+
p•(i)Pm + p•+ i.•(i)P,,+•
• (15)
where
wherep=•(l)istheprobabilityofcapacityoutage levelX when
a unitof capacityrnI is added,using(8) of Hobbsand .Be#n
px'(X),p•(X) probabilityof outagelevelX beforeand after [thisissue].
Thedenominator
of (15)is thep•(i)thatres•'t
s
addingthe n + !st component (from fromtheround-offprocedure
of BillintonandAllan[1984].,
equation(8),HobbsandBeim[thisissue] Thesetransition
ratesarethenusedin (2)and(3)too• •
or Bil!intonandAllan [!984]); thetransition
probabilities
for thesetof parallelcornpoaen,m
BEIM
ANDHOSES:
MARKOV
MODEL
OFCArACIT¾
RELIABILITY 1457
In ourapplicationto the Nahal Oren system,the ten pumps transitionratefrom statei to statej,
yielda 25x 25 transition
probability
matrixwhenan in- component k.
crement
of 100m3/houris used. transitionratefrom statei to j usedin
round-off procedure.
mean repair rate for a capacitycomponent.
NOTATION Acknowledgments.
S. Eckert made valuable suggestions
con-
cerning solution of the Markov model. M. A. Effendi constructed the
MonteCarlosimulationmodel,andP. Ola-Karlsson
helpedprogram
c,cl, c2,c(t) availablecapacity. the modifiedfrequency-duration
method.Supportwasprovidedby
C•+• capacityof n + 1stcomponent. the NationalScience
FoundationgrantECE 85-52524.
d(z) 1 if z equals0; otherwise,d(z) equals0.
D.r(S
) durationof a system
failureof severity
S REFERENCES
or more.
Allan, R. N., and F. N. Takieddine, Generator-maintenancesched-
el, e2 exact states of the system.
uling using simpIifiedfrequency-and-duration
reliability criteria,
D(S) duration of capacitydeficit of severityS Proc. Inst. Electr. Eng., 124(10),1977.
or more. Ang, H. S., and W. H. Tang,ProbabilityConcepts
in Engineerinõ
e(t)--(w(t),q(t),c(t)•,the exactstateof the PlanningandDesign,562pp.,JohnWiley,New York, 1984.
system,time t. Arad, N., A method for the evaluation of the systemand cost-
effectiveness
in largemultistage
flashdesaltingplants,Rep.R-1142,
F•.(S) average
frequency
of failureof severity
S Plann.Res.Corp.,Washington,D.C., 1968.
or more.
Billinton,R., andR. N. Allan,Evaluation
of PowerSystem
Reliability,
G(el) state compatibility function,giving the Plenum, New York, 1984.
statesin the next period that can be reached Damelin,E., U. Shamir,and N. Arad, Engineeringand Economic
from state el.
Evaluationof the Reliabilityof Water Supply,Water Resour.Res.,
8(4), 861-881, I972.
i column vector of ones.
Duan, N., and L. W. Mays, Reliabilityanalysisof pumpingstation
I incrementof capacityusedin round-off and storagefacilities,in Hydraulic Engineering,edited by R. M.
procedure. Ragan, pp. 600-605, American Society of Civil Engineers,New
York, 1987.
px(X) probabilityof X amount of capacitybeing Fiering,M. B., StreamflowSynthesis,
Harvard UniversityPress,Cam-
on outage. bridge,Mass., 1967.
P(e(t)) probability of being in system state e(t). Gordon, S., and B. F. Hobbs, Analytical simulation of bulk water
Ps(S) probabilityof thesystembeingin a failure supply network reliability, in Hydraulic Engineering,edited by
state exceedingseverityS. R. M. Ragam pp. 594-599, American Society of Civil Engineers,
New York, 1987.
P• row vectorof steadystateprobabilities Hobbs, B. F., and G. K. Beim, Verification of a supply reliability
for systemstatesi, time period t. model, in Water Forum '86: World Water Issues in Evolution,
Per rowvectorof steadystateprobabilities
for edited by M. Karamouz, G. R. Baumli, and W. J. Brick, vol. 2,
demand, time t. American Societyof Civil Engineers,New York, 1986.
Hobbs, B. F., and G. K. Beim, Analytical simulationof water system
q,ql, q2, q(t) quantity demanded. capacityreliability, I, Modified frequency-durationanalysis,Water
r unavailability of a capacity component. Resour. Res., this issue.
r•, rk+• probabilitiesusedin addingunit of capacity Hobbs, B. F., G. K. Beim, and A. Gleit, Reliability analysisof power
in round-off procedure. and water supplysystems,in StrategicPlanningin Energyand Natu-
ral Resources,edited by B. Lev et al., North-Holland, Amsterdam,
S, Sj severityof capacitydeficitq -- c. (Asan 1987.
argument in D(S) or F(S), S refers to a Houck, M. H., and J. L. Cohon, Sequentialexplicitlystochasticlinear
systemcapacity deficit of S or more.) programmingmodels:A proposedmethod for designand manage-
S* value of S from (21) of Hobbs and Beim ment of multipurposereservoirsystems,Water Resour.Res., 14(2),
161-169, 1978.
[this issue].
Jacoby, H. D., and D. P. Loucks,Combined use of optimization and
S(ei) severityof capacitydeficit associatedwith simulation modelsin river basin planning, Water Resour.Res.,8(6),
exact state ei. 1401-1414, 1972.
t time step of Markov chain. Kleiner, V., A two-step probabilistic model of storage reservoir with
T matrix of transition probabilities. correlatedinputs, Water Resour.Res.,6(3), 756-767, 1970.
Lansey, K. E., and L. W. Mays, Optimal design of large scale water
Tk matrixof transition
probabilities
T•j between distribution systems,in Hydraulic Engineering,edited by R. M.
statesi and j of component k. Ragan, pp. 475-480, American Society of Civil Engineers,New
T, matrix of transitionprobabilities,time York, 1987.
period t. Lloyd, E. H., Reservoirswith serially correlated inflows, Techno-
metrics, 5, 85-93, 1963.
Tqt matrixof demandtransitionprobabilities,
Lloyd, E. H., and S. D. Saleera,Waiting time to first achievementof
time period t.
specifiedlevelsin reservoirssubjectto seasonalMarkovian inflows,
UD unserved demand.
in Reliability in Water ResourcesManagement, edited by E. A.
wl,w2,w(t) quantityof waterin storage. McBean, K. W. Hipel, and T. E. Unny, Water ResourcesPub-
W storagecapacity. lications, Ft. Collins, Colo., !979.
X amountof capacityon outage. Loucks, D. P., J. R. Stedinger,and D. A. Haith, Water Resource
SystemsPlannin.qand Analysis,Prentice-Hall, EnglewoodCliffs,
;. meanfailurerate for a capacitycomponent. N.J., 1981.
;'o,J-if tr.ansition
ratefromcapacity
outage
level Mays, L. W., and M. J. Cullinane,A reviewand evaluationof reliabil-
X--itoX=j. ity conceptsfor designof water distribution systems,Rep. M.P.
1458 BEIMANDHOBBS:MARKOVMODELOFCAPACITYRELIABILITY