Anda di halaman 1dari 6

WATER RESOURCES RESEARCH, VOL. 24, NO.

8, PAGES 1277-1282, AUGUST 1988

A Method for LocatingWells in a GroundwaterMonitoring Network


Under Conditions of Uncertainty
PHILIP D. MEYER AND E. DOWNEY BRILL, JR.

Departmentof Civil Engineering,


Universityof Illinoisat Urbana-Champaign,
Urbana

A methodis presented
for locatingwellsin a monitoring
networkunderconditions of uncertainty.
The
methodcouples the useof a simulationmodelof contaminant transportand a facilitylocationmodel.
The MonteCarlotechnique is usedwiththesimulationmodelto translateuncertaintyin thesimulation
model parametersinto uncertaintyin the contaminantconcentrationdistribution.The simulationmodel
determines
whichwelllocations
woulddetecta givenrealizationof a contaminant
plumewith a con-
centrationabovea specifiedlimit. The facility locationmodelis then usedto selecta fixed numberof well
locationsso that a maximumnumberof suchplume realizationsare detected.The selectedwell network
maximizesthe probabilityof detection.
The methodis appliedto an exampleproblem.Althoughthe
technique
is computationally
intensive,
theresultsindicatethatpracticalproblems
aretractable.

INTRODUCTION concentration) is carried out using Monte Carlo simulation.


Protection of groundwater resources from contamination Massmann and Freeze [1987b] use limited trade-off infor-
hasbeenof increasingconcernthroughoutthe pastdecade.A mation to illustrate a landfill owner/operator'sresponseto
growing awareness of the environmental effects of waste dis- different monitoring alternatives.
posalpresentsa greatchallengeto thosemanagersresponsible The monitoring alternativesconsidered,however,do not
for groundwaterprotection. In a groundwatercontamination seemto be optimalwithrespect
to maximizingtheprobability
scenario a manager needs information about the con- of detection. The methoddeveloped in thispapercanbe used
centration distribution and movement of contaminants. to selectnetworksthat maximizethe probabilityof detection
Moni-
toring of the subsurfaceenvironmentthrough remote geo- in thefaceof uncertainty;
themethodcouldeasilybeintegrat-
physicaltechniquesor directsamplingfrom wellscanprovide ed into the frameworkof Massmann and Freeze[1987a] to
thisinformation.A methodis presentedfor locatingwellsin a generatethese optimal monitoring alternatives.In addition,
monitoring network to provide detection of contaminant the methodcan be usedto obtaina detaileddescriptionof the
plumes.Oncedetected,the plumecanbe properlymanaged. trade-offbetweendifferentnetworkdesignsand the probabil-
This paper begins by describinga framework within whichity of detection.
this method for monitoringnetwork designfits. The method An example problem, which is similar to that of Massmann
and Freeze [1987a], is used to illustrate the method. For sim-
utilizestwo mathematicalmodels,a simulationand an opti-
mization model. These models are developedand their cou- plicity, an analytical model is usedhere, and the uncertainties
pling is detailed. Finally, the attributes of the method are are assumed to be associatedwith the valuesof the dispersi-
illustratedthroughan applicationto an exampleproblem. vitiesand the direction
of the velocityvector.Conceptually,
the application of the method would remain the same if a
MODEL FRAMEWORK numericalmodelwereusedand if uncertaintyin the hydraulic
A comprehensive frameworkfor designof a landfill oper- conductivitywere considered.In a practicalsense,however,
ation has been detailed by Massmannand Freeze [1987a]. the methodwould be limitedby its computational require-
Their objectiveis to maximize the net present value of a ments.Further discussion
of these limitationsis given in the
concluding section.
stream of costs, benefits, and risks. The risks are associated
with the costs sustained in the event of failure, with failure
defined as a concentration measured in excess of a standard at MODEL DEVELOPMENT
the compliancesurface.Monitoring contributesto the objec- The method developedhere involvesthe use of two inde-
tive functionby reducingthe probabilityof failure,or equiva- pendentbut linked models,a groundwatercontaminanttrans-
lently, increasingthe probability of detection. This framework port simulationmodel and an optimizationmodel.The simu-
reflectsthe notion that in designinga monitoring network the lation modelof the exampleproblemutilizesan analytical
objective is not merely the acquisition of information but solutionto the solutetransportequation.Any simulation
rather the improvement of the overall decisionsthat are made.
modelcouldbe used,however,
andlinkedto the optimization
Massmannand Freeze [1987a] integrate monitoring into modelin an identicalmanner.The optimizationmodelusedis
theirframeworkbycalculating
theprobability
ofdetection aform
ofthep-median
facility
location
problem
known
asthe
given a particular
monitoring
network.
Uncertainties
in the maximal
covering
location
problem.
physicalsystemprecludea deterministicpredictionof detec-
tion. Massmannand Freeze [1987a] consideruncertaintyin SimulationModel
thehydraulic
conductivity
field.
Translating
parameter
uncer- Thegroundwater
system
considered
isdepicted
inFigure
1.
tainty
intouncertainty
inthemodeloutput
(hydraulic
head or Thecontaminant
source
islocated
attheorigin
ina homoge-
neous and isotropic aquifer of infinite areal extent and con-
Copyright
1988bytheAmerican
Geophysical
Union. stantdepth.
Alternatively,
theboundary
oftheareain Figure
Papernumber8W0311. 1 is locatedfar enoughfrom any geologicfeaturesto make the
0043-1397/88/008W-0311505.00 approximationvalid.This boundarymay be interpretedusing
1277
1278 MEYER AND BRILL: LOCATING WELLS IN A GROUNDWATER MONITORING NETWORK

2. Determine the maximum concentration at the boundary


for plume i.
3. If this concentration exceedsthe standard, go to step 5,
otherwise continue.
4. Determine the maximum concentration, for all time, at
all well locations. Go to step 7.
5. Determine the time at which the concentration of plume
i first exceeds the standard at the boundary. Call this time
TBOUND i.
6. Find the maximum concentration at all well locations
o for all time previousto and including TBOUND i. Note that
}30 m
for a constant, continuous source input the maximum con-
centrationwill be at time TBOUND i.
7. If the concentration at any possiblewell location j ex-
possible
well ceeds the standard, then a well at location j is eligible to
__12atiøns provide detection for plume i.
8. Return to step 1 to continue Monte Carlo simulation if
500m the number of distinct plumes generated is less than I, a pre-
determined value judged to be sufficientto describethe uncer-
Fig. 1. Example problem layout.
tainty. Otherwise, terminate the algorithm.
The analytical solution of Wilson and Miller [1978] was
the concept of a containment-dispersal(c-d) threshold [Le- used to calculate the concentrations at the well locations with
Grand, 1982]. The c-d thresholdis a subjectivelydefinedgeo- the Hantush well function calculated using the approximation
graphical position that determines the action to be taken in given by Wilson and Miller [1979]. Note that the algorithm
the event of contamination. A contaminant detected on the effectivelyassumescontinuous monitoring.
containmentside is managedmore efficientlywith (expensive) The algorithm distinguishes between the detection of
remedial measures, whereas a contaminant detected on the plumes that exceedthe standard at the boundary (step 6) and
dispersalside can only be dealt with efficientlyby allowing the detection of plumes that exceed the standard inside but
dispersionand dilution into the environment,with consequent not at the boundary (step 4). In practice, a single well sample
harmful effects.The boundaryin Figure 1 representsone pos- that exceedsthe standard inside the boundary would not be
sible c-d threshold; for instance, it could be a compliance sufficient to determine if the plume would exceed the standard
surfaceset by a regulatory agency. Domenicoand Palciauskas at or outside the boundary. Once a plume is detected, the
[1982] discussalternative boundariesfor compliance.Failure monitoring strategy may be reevaluated with respect to new
is defined as the detection of a contaminant at or outside this objectives, such as determining the extent of the plume
boundary, with a concentrationhigher than a predetermined [Herzog et al., 1986]. New wells may be installed.
maximum value (a standard). Any plume that will cause a It would be feasible to design the monitoring network to
failure should be detectedinside the boundary before failure maximize the probability of detecting only those plumes that
occurs. Once detected, appropriate action can be taken to would exceed the standard at or outside the boundary. The
prevent failure.
The possible well locations are assumed to be known. For
the example they are evenly spaced 30 m apart along each
line. There are 50 m between lines with the first line of wells TABLE 1. Parameter Values for the Simulation Model
placed 150 m from the contaminant source.Judgement would
Description of Parameter Value
be required in a given application to select a finite number of
locations that represent, from a practical point of view, all
Porosity 0.35
locations. The wells are assumedto measurea depth-averaged Contaminantinjectionrate 2.4 kg m-1 d-1
concentration. Magnitude of average linear 0.46 m/d
The source is modeled as a continuous point input of a velocity
conservative contaminant with a constant concentration and Expected value of 2.13 m
volumetric flow rate. The contaminant is assumed to be com- longitudinal dispersivity
Expected value of 0.43 m
pletely mixed over the depth of the aquifer. Additional as- transverse dispersivity
sumptions include a steady state, uniform regional flow field, Standard deviation of 0.709 m
undisturbed by the source input. Molecular diffusion is ne- longitudinal dispersivity
Standard deviation of 0.088 m
glected.
transverse dispersivity
The longitudinal and transversedispersivitiesare uncertain Correlation coefficient of 0.273
parameters.They are assumedto be positively correlated and longitudinal and
to have normal marginal distributions. The values for the ve- transverse dispersivity
Normal distribution:
locity vector direction are selectedfrom an independent distri-
bution, assumedto be either normal or uniform. All parameter expected value of velocity 0.0 deg
vector direction
values for the example are given in Table 1. standard deviation of 10.0 deg
The algorithm for the method uses the simulation model as velocity vector
follows. direction
Uniform distribution:
1. Select values for the uncertain parameters from their
respective distributions. These values represent a distinct interval of velocity vector [-30,30] deg
direction
plume, plume i.
MEYER AND BRILL: LOCATING WELLS IN A GROUNDWATER MONITORING NETWORK 1279

optimal network for this objective,however, will likely be dif- S contaminant concentration standard;
ferent from the optimal network obtained by maximizing the do contaminant
concentration
of plumei asmeasured
at
probabilityof detectingall plumesthat exceedthe standard, well sitej at the time of failure TBOUNDi, or the
both inside and at or outside the boundary. The latter objec- maximum concentration, over all time, for those plumes
tive is the approachusedhere. For the exampleproblem,all that would not cause a failure;
plumes generated exceed the standard at or outside the x•-1 if a wellis installedat sitej;
boundary. x•= 0 otherwise;
The output from the simulation is, for each plume, the well y•= 1 if plume i is detectedby a well;
locations where the concentration exceeds the standard before y• = 0 otherwise;
the standard is exceeded at or outside the boundary. For a N,= {j • dido > S};
networkwith only one well the locationthat maximizesthe N•=0 if d•j _<S ¾j • J'
probability of detectionis the site that would detect the largest ai weight appliedto plume i, ai = 1 ¾ i in the example;
number of plumes generated by the I simulations. When two P number of wells to be installed (this number can be
or more wells are involved, however, the placement of wells varied to examine the trade-off between the number of
thatmaximizes theprobability of detectionismored{fficult
to wells and the percentage of plumes detected).
determine, since all combinations of locations must be con-
The objective is to maximize the detection of plumes that
sidered.The facility location model describedbelow provides
exceedthe concentrationstandard.The weightsai can be in-
a means for selectingthe optimal combination.
terpreted as the probability of any distinct plume occurringor,
Optimization Model alternatively, as the value of detecting any distinct plume.
The general facility location problem can be divided into Since all plumes are obtained using Monte Carlo simulation
two broad categories,the p-median and the m-centerprob- and since they all exceed the standard at the boundary, they
lems,distinguished
by theformof theobjective
functionaregiven
equal
weights
(arbitrarily
setatone)intheexample.
[Handler
andMirchandani,
1979].
Thep-median
problem
can Asdiscussed
above,
thismodelmaximizes
detection
ofplumes
beconcisely
stated
asfollows
[Church
andReVelle,
1976]'thatexceed
thestandard
bothatoroutside
theboundary
and
given
a network
ofdemand
nodes,
locate
a fixednumber
of alsoinside
theboundary.
Applying
equal
weights
tothese
two
facilities
onthenetwork,
whereeachdemand
nodeis servedgroups
ofplumes
implies
thatthevalueof detecting
a plume
bytheclosest facility,
such
thatthetotalweighted
travel
dis- in eithergroupis thesame. If thedetection
of plumes that
tance(theweighted sumofthedistance
fromeachnodetothe exceed thestandard at or outside
theboundaryis of higher
facility
thatserves
it)isminimized. value,thea•fortheseplumes would begreater
thantheai for
Theintroduction
of a constraint
onthemaximum
allowable the groupof plumesthat exceeds
the standardinsidethe
service
distance
leadsto theproblemknownasthelocationboundary.
setcovering
problem.
Thesolutiontothisproblem
locates
the Constraints
oftype1alloweach
plume
i tobedetected
(i.e.,
minimumnumberof facilities
to ensure
thatall demandnodes allowy•-- 1) onlyif a well is installedat oneor moresites
areserved
bya facility
withinthemaximalservice
distancecapable ofdetecting
it (i.e.,onlyif atleastonexj = 1forthose
[Church
andReVelle,
1976]. sites
j thatwoulddetect plume i).Ni isthesetofwells eligible
The maximal
covering location
problem (MCLP)arisesfordetecting
plume i. Forexample, if wellsat sites11,12,34,
whentheavailable
facilities
areinsufficient
toguarantee
that and35areeligibletodetect plume 23,thenN23-- {11,12,34,
all demandis servedwithinthemaximalservice
distance.
In 35}. N•- 0 only if no wellsare eligibleto detectplumei.
thatcasetheproblembecomes
as follows:
maximize
the Constraint
2 ensuresthatP wells
areinstalled.
Constraints
of
demandserved
withinthemaximal
servicedistance
givena types
3 and4 limitthevalueofeachvariable
tozeroorone.
fixed number of facilities [Church and ReVelle, 1976].
The groundwater monitoring network design problem can
EXAMPLE PROBLEM
be formulated as an MCLP by consideringthat (1) each of the
I plumes,examinedin the simulation model and obtained as a The integer optimization formulation was solved for the
distinct combination of uncertain parameters, represents a example problems on an Apollo Domain DN3000 using the
demand node, and (2) each possiblewell location representsa dual-simplex method in the linear programming package
facility site. The MCLP formulation given below is taken from XMP/ZOOM. These solutions required from 5-15 min of
Church and Re Velle [1974]. The notation remains identical, CPU time for the smaller problems and from 45 min to 2
but the definition of each symbol is restated in terms of the hours for the larger problems. A given size problem was more
groundwater problem. easily solved with more wells required in the network. These
maxz = • aiYi times are for problems with the velocity vector direction nor-
ieI mally distributed. When the direction is uniformly distributed
subject to the solution times are over twice these values.
A first step in solving this integer problem is to solve the
• x•_•y• ¾ieI, Ni•O (1) relaxed linear problem. That is, instead of constraining the
j•Ni
decision variables to be zero or one they are constrained to lie
Z x• = P (2) in the interval [0, 1]. Church and ReVelle [1974] reported that
jeJ
approximately 80% of the 27 relaxed integer MCLP problems
xj = (0, 1) ¾j e J (3) they solved terminated with an optimal solution that was also
y,=(O, 1) ¾i•I (4) an optimal integer solution. When the relaxed formulation
was solved for the groundwatermonitoring problems,the
where
code terminated with an optimal integer solution for each of
I set of distinct plumes' the more than 70 problemssolved.Thus although this method
J set of well sites; results in very large integer problems, it appears that they can
1280 MEYER AND BRILL: LOCATINGWELLS IN A GROUNDWATERMONITORING NETWORK

1.0 1.0

,,-, .97 .98 .997 .998


.84 • 0.8
• 0.8
-o. Normal,1000
E 0.6 E 0.6
-*-' Normal,6000 -.u- Uniform,1000
-•- Uniform,6000
'• 0.4 '• 0.4
._o

.• 0.•

, I , I •
0.0 0.0
0 2 4 2 4 6
Number of Wells in Network Number of Wells in Network

b c
b

Normal,1000 Normal, 6000 Uniform,1000 Uniform, 6000

x x x
x

Fig. 2. Results for velocity vector direction normally distributed. Fig. 3. Results for velocity vector direction uniformly distributed.
(a) Tradeoff. (b) Optimal well locations for 1000 simulations.(c) Opti- (a) Tradeoff. (b) Optimal well locations for 1000 simulations. (c) Opti-
mal well locations for 6000 simulations. mal well locations for 6000 simulations.

computational whether the optimal solutionsfor the 1000 problem are good
usuallybe solvedoptimallywith a reasonable
effort. when applied to the 6000 problem. The resultsare presented
Problems of two sizes were solved on the basis of either in Table 2. For a normally distributed velocity direction the
1000 or 6000 Monte Carlo simulations. These will be referred well locations obtained using only 1000 simulations (1000
to as the 1000 problemand the 6000 problem,respectively. problem solution from Figure 3b) detect nearly the optimal
Sincesomeparametersof the simulationmodelare probabil- fraction of plumeswhen the physicalsituation is represented
runningmore simulationsallowsthe pa- by 6000 plumes.For instance,five wellslocatedas specifiedby
isticallydescribed,
rameter distributions to be better defined. This results in a the solution to the 1000 problem would detect 99.6% of the
greaterconfidence
in the solutions of wells plumesthat exceedthe standard,while five wells located by
for the placement
obtained from the facility location model. the solution to the 6000 problem would detect 99.8%. When
The implicationsof usingmore simulationscan be seenin the velocity is uniformly distributed, the solutionsfrom the
Figure2, whichgivesresultswhenthe velocityvectordirec- 1000 problem do not perform quite as well, yet they are still
tion is normallydistributed.Figure2a showsthat the trade-off quite close to optimal for the 6000 problem. It may be desir-
between detection and the number of wells in the network is able, in such a case, to obtain additional simulation results to
nearlyidenticalfor boththe 1000problemandthe 6000prob- better characterize the uncertainty. These results agree with
lem.The optimalwell locationsfor the two problemsmay be the trade-off curves, which suggestthat for the uniformly dis-
significantlydifferent,however,as can be seenin Figures2b tributed problem (Figure 3a) more simulationsshould be ob-
and 2c, whichgivethe optimalsolutionsfor a five-wellmoni- tained, since there is a large difference in the curves for the
toring network(P--5). Althoughadditionalresultsare not 1000 problem and the 6000 problem. When the required
displayed, the differences in well locationsfor the two prob- number of wells in the network is less than four, the optimal
lemsincreaseas the percentageof plumesbeingdetected(or solution to the 1000 problem is also optimal for the 6000
the number of wells in the network) increases. problem.
The results for the normal distribution can be contrasted to The problem structure suggests,and the computer output
those obtained when the velocity vector direction is uniformly confirms, that alternative solutions to the M CLP exist that
distributed(Figure3). In thiscase,the trade-offbetweendetec- are optimal or very closeto optimal with respectto the objec-
tion and the number of wells in the network is not indepen- tive function (number of plumes detected)but different in de-
dent of the number of Monte Carlo simulations (Figure 3a). cision space(location of wells). A complete analysisshould
The solutionsobtained for the 1000 problem overestimatethe examine these alternative solutions, since they may be su-
fractionof plumesdetected,as comparedto the 6000 problem. perior to the optimal solutionwhen an unmodeledobjectiveis
The well locationsfor the 1000 problem(Figure 3b) exhibit the considered.For instance, the optimal solution presented in
symmetrylacking in the earlier solution(compareto Figure Figure 2c locatesthe wells closeto the contaminant source.It
2b) yet differ from the optimal well locationsfor the 6000 may be believed,however, that the contaminant is not fully
problem (Figure 3c). mixed over the aquifer depth so closeto its source,thus rais-
The question of how many simulations are necessaryto ing somedoubt in the validity of the analyticalsolutionused
characterize the uncertainty fully can be addressed by asking to model contaminanttransport.It would be valuable,in such
MEYER AND BRILL' LOCATING WELLS IN A GROUNDWATER MONITORING NETWORK 1281

TABLE 2. Well Location Solutions of the 1000 Problem Applied to the 6000 Problem

Fraction of the 6000 Plumes Detected

Velocity Vector 4 5 6
Distribution Solution Wells Wells Wells

Normal 6000 problem 0.984 0.998 0.999


1000 problem 0.981 0.996 0.996
Uniform 6000 problem 0.725 0.807 0.866
1000 problem 0.685 0.709 0.813

The 6000 problem solution is optimal. The 1000 problem solution is not optimal for the 6000 plumes.

a case, to examine alternatives that locate the wells farther solutions with four, five, and six wells are, as above, all within
from the source while maintaining near optimality. Brill 0.2% of the 6000 problem optimal solutions.
[1979] has addressed the issue of generating alternatives that The ad hoc methods consist of producing an alternative
are different, yet good with respect to the modeled objectives. network using special knowledge of the problem, evaluating
Church and Huber [1979] present several techniques for find- the objective function for that network, and comparing the
ing close to optimal solutions of various location problems, objective function value with the optimal value. The MCLP
including the MCLP. Chang et al. [1982] also present formal does not need to be solved to do this. A simple computer
techniques for modeling to generate alternatives (MGA) that program was used to determine the total number of plumes
could be applied to obtain alternative solutions for mathemat- detected by a given network by searchingthe output from the
ical models of the type presented here. For the groundwater simulation model (for each plume, the set of sites where an
monitoring network problem, however, ad hoc methods can installed well would detect the plume). This search required
also be used. less than a minute of CPU time. It is thus quite easy to
One such method has been discussed above. As a by- determine the objective function value for a large number of
product of asking how many simulations are required to de- selected alternatives.
scribe the problem uncertainty adequately, it was found that The methods described above represent a simple, heuristic
the solutions for the 1000 problem are good alternatives to the approach. Although formal MGA techniques could be used,
optimal solutions for the 6000 problem. the idealized symmetry of the problem allows good alternative
Good alternatives can also be found by taking advantage of solutions to be produced and examined readily with very little
the symmetry of the problem. Since the problem is approxi- computational effort. In contrast, a formal technique would
mately symmetric about the x axis, solutions obtained for only require the solution of a modified form of the MCLP with its
one half of the domain should be good solutions. The asym- accompanying computational costs.Formal MGA techniques,
metric solution for the 1000 problem with five wells in the however, may be necessaryto produce widely different alter-
network, depicted in Figure 2b, translates into the two sym- natives for a lessidealized problem.
metric alternatives presented in Figure 4. When applied to the
CONCLUSIONS AND FINAL REMARKS
6000 problem, each of these two symmetric solutions detects a
fraction of plumes within 0.2% of the optimal solution. Simi- A method has been developed for locating the optimal
larly good results were obtained when symmetric alternatives placement of wells in a monitoring network using linked simu-
to the optimal four- and six-well network solutions were lation and optimization models. Contaminant transport simu-
examined. In those cases the symmetric alternatives may not lation provides information about the detection of individual
contain the same number of wells as the optimal solution. plumes while the use of the Monte Carlo technique accounts
This method could not be used for the uniformly distributed for uncertainty in transport parameters. The optimization
problem, sincethe solutionsfor 1000 simulationswere all sym- model locates a given number of wells to maximize the prob-
metric. ability of detection. The application of the method to an ex-
The well locations that are the mirror image of an asymmet- ample problem illustrates the use of the method and demon-
ric solution also form a good alternative (Figure 5). The objec- stratesthe type of information the method can produce.
tive function values for the mirror images of the 1000 problem The strengths of the method include its conceptual sim-
plicity as well as the separability of the simulation and opti-

o
o
x

Fig. 4. Alternative symmetric solutions. Velocity direction nor-


1000 Problem Solution, P=5 Mirror Image Alternative
mally distributed.The X indicatesa well in the original solution and
O indicates a well added to create symmetric solution. Fig. 5. Mirror image alternative for normally distributed velocity.
1282 MEYER AND BRILL: LOCATING WELLS IN A GROUNDWATER MONITORING NETWORK

mization models.The optimization model resultsin an integer oping efficient ways to incorporate both the uncertainties as-
problem of considerablesize yet remains tractable, sinceit can sociated with more realistic problems and numerical models.
usually be solved as a linear program. Although in all cases
considered optimal integer solutions were obtained using Acknowledgments. The authors would like to thank Albert J. Va-
1occhifor time spent in discussionwith the first author during the
linear programming, Church and ReVelle [1974] discussheu-
early stages of this research and for his comments regarding this
ristic methodsto obtain a solutionif the linear program termi- paper. Computer support from the Civil Engineering Department of
nates noninteger. the University of Illinois at Urbana-Champaign is also acknowledged.
The method is also flexible and could be expanded to in-
REFERENCES
clude such factors as existing wells and discontinuousmoni-
toring schedules.Existing wells could be incorporated as ad- Brill, E. D., Jr., The use of optimization models in public sector
ditional constraintsin the optimization model. A monitoring planning, Manage. Sci., 25(5), 413-422, 1979.
schedule could be added to the simulation model. In that case Chang, S. Y., E. D. Bri.11,Jr., and L. D. Hopkins, Use of mathematical
models to generate alternative solutions to water resources plan-
a plume would be detected only if the concentration exceeds ning problems, Water Resour.Res., 18(1), 58-64, 1982.
the standard at the same time that the well location is sam- Church, R. L., and D. L. Huber, On determining many close to
pied. optimal configurations for single and multiple objective location
The method also has severallimitations. The objectivefunc- problems,Res. Set. 34, Dep. of Civ. Eng., Univ. of Tenn., Knoxville,
July 1979.
tion of the maximal covering location problem representsde-
Church, R. L., and C. ReVelle, The maximal covering location prob-
tection only. Other concernsin designinga monitoring net- lem, Pap. Reg. Sci. Assoc.,32, 101-118, 1974.
work might include the costsof constructionand samplingor Church,R. L., and C. ReVelle,Theoreticaland computationallinks
the effect of the random nature of water quality data. The between the p-median, location set covering and the maximal
method presented is not intended to stand alone. Rather, it is coveringlocation problem, Geogr.Anal., 8(4), 406-415, 1976.
Domenico, P. A., and V. V. Palciauskas, Alternative boundaries in
intended to be a part of a larger processof designor analysis, solid waste management, Ground Water, 20(3), 303-311, 1982.
such as that of Massmannand Freeze [1987a]. Handler, G. Y., and P. B. Mirchandani, Location on Networks, MIT
A major limitation of the method is that computational Press,Cambridge, Mass., 1979.
Herzog, B. L., B. R. Hensel, E. Mehnert, J. R. Miller, and T. M.
requirementsmay be significantlygreaterthan impliedby the
Johnson,Evolution and adequacyof ground water monitoring net-
simple example consideredhere. In practical situations, pa- works at hazardous waste disposal facilities in Illinois, in Pro-
rameters other than the velocity direction and the dispersi- ceedingsof the Sixth National Symposiumand Expositionon Aquifer
vities are also uncertain. Such parameters may include the Restoration and Ground Water Monitoring, pp. 98-119, National
hydraulic conductivity and the size of the contaminant source Water Well Association, Dublin, Ohio, 1986.
LeGrand, H. E., Hydrogeologic issuesin hazardous waste manage-
(as consideredby Massmannand Freeze [1987a]), the location ment, Ground Water, 20(4), 488-491, 1982.
of the contaminant source, and the reactive nature of the con- Massmann, J., and R. A. Freeze, Groundwater contamination from
taminant. These more realistic parameterscan be included in waste management sites: The interaction between risk-based engi-
the uncertainty analysis given the general conceptual ap- neering design and regulatory policy, 1, Methodology, Water
Resour. Res., 23(2), 351-367, 1987a.
proach presented here. Their inclusion, however, would in-
Massmann, J., and R. A. Freeze, Groundwater contamination from
crease the computational burden. Computational require- waste management sites: The interaction between risk-based engi-
ments would also increase if a numerical model were used neering design and regulatory policy, 2, Results, Water Resour.
insteadof an analyticalmodel of contaminanttransport.Ana- Res., 23(2), 368-380, 1987b.
Wilson, J. L., and P. J. Miller, Two-dimensional plume in uniform
lytical models generally require significantsimplification of ground-water flow, J. Hydraul. Div. Am. Soc. Civ. Eng., 104(HY4),
physical processes,making it difficult to incorporate impor- 503-514, 1978.
tant featuresof the groundwatersystem,including small-scale Wilson, J. L., and P. J. Miller, Two-dimensional plume in uniform
heterogeneity and complex boundary conditions. Numerical ground-water flow, Closure, J. Hydraul. Div. Am. Soc. Civ. Eng.,
105(HY12), 1567-1570, 1979.
models can better capture the complexitiesof practical prob-
lems. Their use, however, would substantially increase the E. D. Brill, Jr., and P. D. Meyer, Department of Civil Engineering,
computational effort since Monte Carlo simulation requires University of Illinois at Urbana-Champaign, Urbana, IL 61801.
the repeated solution of the transport problem. Future re- (Received October 27, 1987;
searchshould addressthesecomputational issuesby focusing revised March 16, 1988;
on an investigation of the computational limits and by devel- accepted April 7, 1988.)

Anda mungkin juga menyukai