Vous êtes sur la page 1sur 12

WATERRESOURCESRESEARCH,VOL. 27, NO.

8, PAGES1873-1884,
AUGUST 1991

The Laplace Transform Finite Difference Method


for Simulationof Flow ThroughPorousMedia
GEORGE J. MORIDIS1

Environmental
andWaterResources
Engineering,
CivilEngineering
Department,
TexasA&M University,
CollegeStation

DONALD L. REDDELL

Departmentof AgriculturalEngineering,
TexasA&M University,CollegeStation

A newnumerical method,theLaplacetransform finitedifference


(LTFD)method,wasdeveloped to
solvethe partialdifferentialequation(PDE) of transientflowthroughporousmedia.LTFD provides
a solutionwhichis semianalyticalin timeandnumericalin spaceby solvingthediscretizedPDE in the
Laplace spaceand numericallyinvertingthe transformedsolutionvector. The effectsof the traditional
treatment of the time derivative on accuracyand stabilityare renderedirrelevantbecausetime is no
longera consideration.For a singletimestep,LTFD requiresno morethaneightmatrixsolutions and
an executiontime eighttimeslongerthanthe analogous finitedifference(FD) requirementwithoutan
increasein storage.This disadvantage is outweighedby an unlimitedtime stepsizewithout any loss
of accuracy,a superioraccuracy,and a stable,nonincreasing roundoff error. Thus, a problemin
standardFD format may require several hundredtime stepsand matrix inversionsbetween the initial
conditionandthe desiredsolutiontime, but LTFD requiresonlyonetime stepandno more thaneight
matrix inversions to achieve a more accurate result.

INTRODUCTION where R K is the remainder term usually referred to as the


"truncation error."
The basicconcept of any numerical solution of a partial Parabolic type PDEs, such as the ones encountered in
differentialequation (PDE) is the substitution of a set of fluid flow throughporousmedia, contain the time derivative
algebraicequations for the original continuous PDE. These 0 U/Ot, of which the numericalapproximationis consistently
algebraicequationsare valid within constant-propertysub- the most important source of instability and error in numer-
domainsof the solution domain, are easier to solve, and ical models. The treatment of the' time derivative is usually
providea solutionarithmetically "close" to the true solution identical in both the finite difference (FD) and finite element
of the originalproblem. In transient flow through porous (FE) methods that are used to analyze flow through porous
media,the PDE to be solved is [Aziz and $ettari, 1979] media. The continuous time coordinate is discretized into
time steps At, and numerical solutions are sought at the
U(x•, x2, x3, t) = O, (•) discrete time levels

whereU isthedependent
variable,(x•, x2, X3)arethe space
coordinates,and t is the time coordinate. t0=0, t•=Atl, t2=At•+At2,-'', tn= Z Atl,''',
In thefinitedifference(FD) method,insteadof solvingfor l--1

a continuous smoothfunction U(x•, x2, x3, t), Taylor


seriesapproximations (r•) of the solutionare obtainedat a where the dependentvariable U K(t) is approximatedby a set
finite
setofpredetermined N points(x•, x2, x3, t) •, (x•, x2, of values•t•, l = 1, 2, --- . A Taylorseries
approximation
x3,t)2, "', (x•, x2, x3, t)N withinthe solutiondomain.A of the time derivative yields
PDEproblemwith a continuoussmoothsolutionsurfaceis
thusreduced
to a set of algebraicequations
relatingthe
discrete
approximate
values• to eachpoint(Xl, x2, x3, = (Ug - U2)+ R•,n, (4)
t)• atallpointsin thedomain.Equation(1)is thenreplaced
by where

2
•(x•, x2, x3, t)K=0; K= 1, 2, "' , N. (2) Atn + 1 Atn + 1
= '" ----+ ... (5)
R•,n U'} 2 U• 6
AtanygridpointK, U and• arerelatedby
UK = U• + R•, representsthe truncationerror. For v = n, (4) representsa
(3)
forward differenceapproximationand resultsin the explicit
formulation of the FD, which is not unconditionally stable;
1Now
stGroundwater
Research
Program,
Department
ofAgri- for v = n + 1, (4) represents a backward difference
cultural
Engineering,
TexasA&MUniversity,
College
Station.
approximationand results in the unconditionallystable
Copyright
1991
bytheAmerican
Geophysical
Union. implicit formulationof the FD. The aboveapproximation
Paper
number
91WR0 ! 190. introduces an error of order O(At). Accuracy (and, in the
0043-1397/91/91WR-01190505.00 case of the explicit formulation, stability) considerations
1873
1874 MORIDISANDREDDELL:
LAPLACE
TRANSFORM
FINITE DIFFERENCE
METHOD

precludethe useof a largeAt. Minimizationof R •,n dictates RC a subscriptdenotingreservoir conditions;


theuseof a largenumberof smallAt valuesbetweendesired STC a subscriptdenoting standard conditions;
observationtimes, at the expenseof longerexecutiontimes • reservoir(aquifer)particle(grain)compressibility
and larger round off errors. (M-iLT2);
A new numerical simulationtechnique,the Laplace trans- •' reservoir(aquifer) skeletoncompressibility
form finite difference (LTFD) numerical method, is de- (M-1LT2);
scribedin thispaper.It eliminatestheproblemswith stabil- e fluidcompressibility
(M- 1LT2);
ity and accuracy posedby the FD treatmentof the time 0 a subscriptdenoting a reference state;
derivative, and allows an unlimited At without any loss of p fluiddensity (ML-3);
accuracyor stability.In the presentpaper,the mathematical t7 massflow rate of sink or sourceper unit volume
basisof the methodis presentedfor the parabolic-type PDE (ML-3T-I).
encounteredin fluid flow through porous media, and the In developingthe partial differential equationfor flowin
performanceof the methodis evaluatedagainstresultsfrom porousmedia, isothermalflow and negligiblelateraldefor-
analytical solutions and standard FD models. mation of the aquifer skeletonwere assumed.For liquid
phaseflow, the followingsimplificationscan be madewith-
THE FLOW EQUATION out loss of accuracy:
1. For near-incompressible
flow, the valueof e is very
The general nonlinear equation of transient flow through small and the effect of the fluid compressibilityon the fluid
porousmediais obtainedby combiningappropriateformsof properties(p,, p and, consequently,/3)can be neglected with
Darcy's law and the equationof massconservation,yielding impunity.This is a valid assumptionfor water (e = 4.47x
10-lø 1/Pa)andmostliquidphases.
Then
Vß Apo
VP- 3'Vz SsPo
'•-+q, Po
P=PPo /• = exp [e(p - P0)] •-- 1 + e(p - Psrc) = 1, (12)
(6)
where •= e exp [e(p- PSTC)]-----
e, (13)
Op
,k transmissibility of the fluid;
3' specific
weightof thefluid(ML-2T-2); k• -/•0, P • P0, A - A0, (14)
p fluidpressure (ML-1T-2); 2. The values of • and •' being very small, the following
P0 a referencefluidpressure(ML -iT-2); Taylor series approximations can be made with negligible
P relative fluid pressureas a fraction of P0 effects on accuracy:
(dimensionless);
z coordinate in the vertical downward direction (L); •b = 1 + ((b0- 1) exp {(• + •')(P0- P)}
t time (T);
q volumetric flow rate of source or sink per unit volume = 1 + (4>0- 1)[1 + (•: + •r)(p 0 -p)] = •b0, (15)
(T-l).
0o
The above formulation utilizes relative pressuresP, as •= (•b0 - 1)(• + •r) exp {(• + sr)(p0- p)}
opposedto actual pressuresp, to minimize round off errors. Op
The usual referencepressureP0 is the initial pressure.The
terms in (6) were defined by Moridis [1987] as --(! - •b0)(• + •'). (16)
Then the accumulationterm Ss becomes
x = --, (7)
Ss •- Cboe
+ (1 - •b0)(• q- •') (17)
and is a constant.
1 [Vf]aC
(8)
/• B exp{e(p
Psrc)};
B [Vf]sr
c
THE LTFD NUMERICAL METHOD

TheLaplacetransform
finitedifference
(LTFD)numerical
Ss= +13op (9) methodeliminatesthe accuracyand stability problems
•b = 1 + (•b0 - 1)exp ((• + •r)(p0-p)} (10) causedby the treatmentof the time derivativein (6) by
providing
a numerical
solution
in the Laplacespace.
This
approach
renderstimeconsiderations
andlimitations
irrele-
q = vant. The vectorof the unknownsin time is thenobtained
by
(p)src
numerically
inverting
theLaplacespacesolution
using
the
where Stehfest [1970] algorithm. LTFD consistsof four steps
k
describedin detail in the followingsections.
absolute permeability tensor of the porous medium
(L2);
Step 1' The Laplace Transform of the PDE
/.• fluidviscosity
(ML - 1T- 1);
B formation volume factor of the fluid (dimensionless); Because of the propertiesindicatedby (12)-(17),
the
•b porousmedium porosity (dimensionless); Laplacetransformof (6) expanded
in cartesian
coordinates
[Vœ]volume
occupied
bya fixedmass
offluid(L3); yields
MORIDISANDREDDELL:LAPLACE
TRANSFORM
FINITEDIFFERENCE
METHOD !875

(1/2)
= «(AXiq-•Xi +_1)'
(21)
Grid-block (i-l)

(&i4)+ Entirely analogousexpressionsare obtainedfor discretiza-


tion in the y andz direction.The subscripti refersto the
positionin the three-dimensional
grid definedas the ith,jth
(axi-:O-
and kth grid block center in the x, y and z direction
respectively,withi = 1,..., MX, j = 1, ..., MY and
k = 1, ßßß, MZ. The parametersMX, M Y andMZ are the
numberof subdivisions of the x, y and z coordinate.This
discretization
dividesthe solutionspaceinto MX x MY x
MZ locally uniform and homogeneousgrid blocks. For
Fig.1. Construction
of a block-centered
grid systemin the x simplicity,the subscriptsi, j and k are omitted and only
direction.
those needed to illustrate communicationbetween grid
blocks are mentioned.
Using a Taylor series,the dø derivative (dø --- x, y, z) on
the left sideof equation(18) is approximatedby

0 (xx)p0
Ox spo/

ado po
;7 * - *po
/,
ay(XY)p0
+-- •yy•- spo]
+--
o ad• ad
ø •- SPO g+(1/2)
oz (•Z)Pøaz spo
q
Xd* ß -
ad ø
= poSsS,tr- S sp(O) + -, (18)
s

whereAd*, d* --- X, Y, Z are the transmissibilities in the x,


y andz directions
respectively,
hx andhy aretheelevations spo /
alongthe x and y directionsrespectively,p(0) is the initial P0 Ad*
Ad• • +(1/2) Ad• + (1/2)
pressuredistribution,s is the Laplace spacevariable, and

= A •_hao -
with•{
braces.
} denotingthe Laplacetransformof the quantityin + Ad*
( :)-
• - (1/2) Ad•_(1/2)
sp
, (22)

Applyingthe same proceduresto a three-dimensional


cylindrical
coordinatesystemyields the followingequation where
intheLaplacespace:
A •II- = •- - ql_ , (23)
spo/ spo •_+ l spo /K
---
r Or r(AR)p0 +•-•-• (AO)
p0
K -= i d* -- X and d ø-- x
+ -- (AZ)po • -
Oz • spo K --j d* --- Y and dø -• y
q
• -= k d* = Z and d ø = z
= poSsS'P- S,p(O) + -, (20)
s

where
Ad*,d* ---R, ©, Z arethetransmissibilities
in ther, The subscript• -+ (1/2) refers to the interblock value of the
0andz directions
respectively. quantity
involved.
If Ad*•+_(1/2)
is approximated
as
Xd*K-+(1/2)--- •l(Xd*•+Ad*•-+ 1), (24)
Step
2:TheFD Scheme
in theLaplace
Space
Cartesian
coordinates.
A "block-centered"
gridmethod hdoT
isemployedinthespacediscretization
of thetransformed • ao= q• - ----, (25)
spo
PDE.Thegeometricalquantities
involvedintheconstruc-
tion
oftheblockcentered
gridsystemareshown inFigure
1.
It isevidentthat the local truncation error is [Aziz and Senari, 1979]
1876 MORIDIS AND REDDELL: LAPLACE TRANSFORMFINITE DIFFERENCE METHOD

Ad•o+ l - 2Ad• + Ad; _1 blocksresultin an algebraic


formof the spacederivative
Rd*• = (AdcI>ao)K identical to the one in (30), i.e.,
4Ad,•

(1/2))
2+ (Ad•_(1/2))
2
[2Ad*cI>bg Po
7•rr
r(AR)
•rr+•'•-[)tOO0
] +• AZw]
]
= AXi(•i- 1- •i) + CXi( •i+ 1- •i) + RXi
+ 3(Ad*'•}o)']• + O[(adø)2] (26)

The first term of Rd*• is generallyof zero order, which + AYj(•j_ • - •j) + CYj(•j + 1-- (I)j)q-RYj
impliesthat the schemeis an inconsistent
approximation of
+ AZk(cI>k
- 1-- (I)kq-CZk(CYDk
+ 1-- (I)k)q-RZk, (31)
order O[(Adø)ø].For regularspacing
the Rd*• is of
O[(Adø)2]. where•r = • and •0 = •. In the caseof cylindrical
Calculationof the exactpressuredropbetweengridblocks coordinates,
the radialintrablocktransmissibility
is obtained
,: and ,c _+ 1 (,c = i, j, k) in the FD schemeleads to the as
determinationof the transmissibilities at g +_(1/2) as

(32)
Ad*
$d• + $d• x 1
....... (27) 'kRi-
(1/2)
=In(rint/ri)
In(ri_+
ilrint)
'
•c-+(1/2)6d• $d•o_.1 ARi
+
XRi_+1
Xd* 1
inwhich
rint-=ri - • Ari.Thentheflowcoefficients
AXand
where CX are

Po(XRi_(1/2)) Po(ARi+(1/2))
AX i = CX i --- .
riAri riAri
(33)
Therefore, Ad*•+_.(1/2
) is theharmonic meanvalueof trans-
missibilitiesof the adjacentgridblocksg and g _+1 whenthe The flow coefficientsfor flow along the 0 directionin a
cylindrical system are obtained from
grid transmissibilities
are )td*• and )td*•_.
1. Equation(27)
minimizes the mass balance error.
Equation (22) is then written as
P0(• Oj_(1/2)) P0()[Oj+(1/2))
, - (34)
AYj=ri2A
OiA
Oj_(!/2
) CYj
- ri2AOjAO
j+(!/2)
Od
ø (Ad*)p0
• •aø = Ad*•(•,•_
1- •,•) and•'•}j-(1/2)is computedfrom (27). The flow coefficients
AZ•, and CZ• alongthe vertical direction are the sameasf0r
cartesiancoordinates.The quantities Adø -- r, 0, z are
(28) obtairiedin the mannerdepictedfor /3x in Figure1. The
where termsRXi, RYj and RZ•, representthe truncation errors
resultingfrom the discretization of the solution domaininr,
P0(Ad*•_ (1/2)) P0(Ad*•+ (1/2)) 0, andz respectively
and are generallyof orderO(zXrø),
Ad*• = Cd*• = . (29) O(A0ø) andO(Azø).
Ad,•Ad• _ (1/2) Ad• Ad,•+ (1/2) Thefinalform. Assumingthat appropriatespacediscret-
Theleft-handsidein thetransformedequationof flow(! 8) izationresultsin negligibletruncationerrors,andsubstitut-
is thusreducedto the followingalgebraicform' ingfor •, the FD equationin the Laplacespacebecomes
AX(•i-1 - • + GXi) + CX(xlti + 1 -- xlt-- GXi + 1)
p0 xx ,xj xr + xz + A Y(•j_ 1-- nitq-GYj) + CY(•j + 1- •
= AX•(•_ • - •i) + CX•(• + • - •i) + RX• -GYj + 1)+ AZ(•_ 1 -- xttq- GZk) q-CZ(Xltk
+1
+ AYj(•j_ 1- •j) + CYj(cypj+
I -- (I)j)q-gYj - a•- GZ• + •)
+ AZk(•tc- 1-- CI•k)
q-CZk(•k + 1-- C•k)q'RZt:. q
= po$ss't•- Ssp(O)+ -. (35)
(30)

The terms Ad* and cd* (d* --=X, Y, Z) are called "flow ThetermsGX, GY andGZ representthe Laplacetrans-
coefficients." the formsof the gravitationalpressuredifferentials
The termsRXi, R Yj andRZ&represent between
consecutive
truncation errors resulting from the discretization of the grid pointsin the x, y and z directionrespec-
solution andaregenerally tively.Theyaregivenby theexpressions'
domainin x, y, andz respectively
of orderO(/•xø),O(Ayø) andO(Azø).
Cylindricalcoordinates.An entirelyanalogous
analysis 3'Ax in (x, y, z)
of the spacediscretizationin cylindricalcoordinatesand a GXi=2--•p
o 3'Ax i-1 i
calculation
of theexactpressure
dropbetween
adjacent
grid (36)
MORIDIS AND REDDELL: LAPLACETRANSFORMFINITE DIFFERENCEMETHOD 1877

Step 4: The Numerical Inversion of the Laplace


yAy in (x, y, z) Solution
J
(37) The vector of the unknown relative pressures[P] at any
time t is obtainedby usingthe Stehfest [1970] algorithmto
! numericallyinvert the Laplace solutions[•]•. The proce-
dure is describedby the following equations:
=2--p0 +(ZXz)A. (38)
In2 Ns
Collecting
andrearranging
terms,the finalform of the FD [e(t)]
=-T- [p(t)] = po[P(t)] (47)
equation
intheLaplace
space
is
AXil-1 + A Y•j -1 + AZ XIr
k-1 + B• + CX xIri + 1 V•,= (-1) (Ns/2)
+•
+ CY•j+ 1+ CZXltk+l =D, (39) rnin{ •,,Ns/2}
•c(Ns/2)(2•C)
!
where K = (1/2)(• + 1)
((Ns/2)- K)!K!(K- 1)!(V- K)!(2• -- V)!

B= -(AX + A Y + AZ + CX + CY + CZ + poSss), (40) (48)


Although the accuracy of the method is theoretically
q expected to improve with increasingNs, Stehfest [1970]
D = -Ssp(O) + -s + GG, (41) showed that with increasing Ns the number of correct
significantfiguresincreaseslinearly at first and then, due to
round off errors, decreases linearly. Testing his algorithm
GG= CX(GXi+ l) + CY(GYj+ l) + CZ(GZi•+ 1) against 50 equations with known inverse Laplace trans-
forms, he determined the optimum Ns = 10 for single
- AX(GXi) - A Y(GYi) - AZ(GZ•). (42)
precisionvariables(eightsignificantfigures)andNs = 18for
doubleprecisionvariables(16 significantfigures).However,
as described in detail below, our experience indicated no
Step3: TheSolutionin the Laplace Space
significantdifferencesin the performanceof LTFD for Ns
between 8 and 20.
TheFD approximationof the partial differentialequations
overthe solutiondomain (x, y, z) or (r, 0, z) in the Laplace The solutionin the Laplace spaceeliminatesthe stability
spaceresultsin and accuracyproblemscausedby the treatmentof the time
derivative in standard FD simulators, thus allowing an
Nr = (MX)(MY)(MZ) (43) unlimitedtime stepsize. The truncationerror of the method
is limited to the truncation error caused by the space
simultaneous
equations.Written in matrix notation, the FD discretization because the domain is not discretized in time,
systemof simultaneousequations (39) becomes and the methodprovidesa solutioninherentlymore accurate
than the standardFD method for the same grid system. The
(M}[•] = [D], (44)
ability to useunlimitedtime stepsize boundsthe accumula-
tion of roundoff error by an upperlimit definedas the round
where{M} is the flow coefficientmatrix, [•] is the vector of off error accumulatedafter the Ns solutionsrequiredby the
theunknowntransformedpressures,and [D] the "known" method. Thus, LTFD offers a stable, nonincreasinground
fight-handside vector. off error irrespectiveof the time of observationt ohsbecause
Thecomputationof {M} and [D] necessitatesarithmetic calculationshave to be performedat this time only usinga
valuesof the s variableof the Laplace space.For a desired At = robs;onthe otherhand,calculations in a standard FD
observation time t they are provided by the first part of methodhave to be performedat all intermediatetimes
Stehfest's
[1970]algorithmas
l L

sv=
In 2
• v- 1, .-- , N s, (45)
(tint)i=
• Ati l= 1,2,''', L, • = robs,
t i= 1 i= 1
(49)
where
Ns isthenumberof summation
termsin thealgorithm continuouslyaccumulatinground off error in the process.
andNs is an evennumber.Optimumvaluesfor Ns are
discussed
inthefollowing
section.
Solution
of (44)returns
a
setofN s vectorsof the transformedpressures[•]• Simulationof SlightlyCompressible
Fluid Flow
[*]•: [•(s•)] = (M(s•)}-l[D(s•)] v=l,'",Ns. Slightly compressible fluids, as opposed to near-
incompressible
fluids,resultin weak nonlinearities
in the
(46)
equationof flow (6), in which

Toobtain
asolution
atatimet, allvectors
[•] •, v = 1, ßßß, 13= 13(p(t)), dp= c)(p(t)), ,k= X(p(t)),
Nsareneeded,
i.e.,thesystem of simultaneous equations
has
tobesolved
Ns times. Ss = Ss(P(t)).
1878 MORIDISANDREDDELL',
LAPLACE
TRANSFORM
FINITE DIFFERENCE
METHOD

,I 50
(b)
:2.

40 ß-o- LTFD,
Nsu10
/'-'meisso]•'øn I ß-A- FD,1 time step
-,- FD,2 time steps
!- I.,,•,..FD,1timestep I r- ß-a-- FD,4 time steps
9- • I...A...
FD, 2timesteps
[ 3O -x- FD,5 lime steps
- -o- FD,6 time steps
e- k I..a.,FD,4timesteps[
,'• 7- % [.,,x,.,
FD,5timesteps[ -+- FD,15t!,,me,,st,epsl
vE 6-N
k [..•..FD,
6time
steps
I 20-

10-.':
-

'

-,o-':-
0.1 Ill•me
n5days,!
I ....... .20x!1S
a.
I I I I I !
0 1000 2000 3000 4000 5000 6000
Distance from the well (m) Distance from the well (m)

Fig. 2. Comparison
of theLTFD solutionfor testcase1 at t = 5 daysto the analyticalTheissolutionandtheFD
solutionsfor variousnumbersof time stepswith respectto (a) drawdowndistributionand (b) drawdown difference
distribution.

The LTFD method is easily extendedto cover the simulation FD methods. Double-precision variables with 20 significant
of weakly nonlinear flow through an iteration procedure figureswere used in all simulations.
entirely analogousto the treatment of such nonlinearitiesin
FD or FE. The procedure is describedas follows: Verification and Test Case 1
1. An initial solution [P]l is obtainedwith/3, •b, it and,
Ss values computedfor p = p(0). Test case 1 was a test problem investigated by Bouwer
2. The updatedvaluesof/3•, •b•, (0t3/Op)•,and (Oc)/Op)i [1978],and representsthe one-dimensionalradial flowprob-
are obtained for [p]• using the unapproximatedpart of lem toward a well of radius r w --> 0 in a homogeneous
(!2)-(16). Then the updated it and Ss are computedas circular aquifer with infinite boundaries.The aquiferhada
transmissivity
t = 1000m3/dayanda storagecoefficient
$=
0.0001,
andwaterwaspumped
ata rateof Q = 1000
m3/day
Ss $ (50) [Bouwer,1978].Two observationswere madeat t• = 5 days
]'/' 1 I and t 2 = 10 days.
where The aquiferwas simulatedby a one-dimensional radial
-- 1 -- 1
gridsystemof 60 grid blocksof unequalsize, anda uniform
fi = •(fi0 + fi•), • = 2-('/'0
+ •)- thickness of 50 m was assumed. The initial aquifer pressure
3. An improved solution [P]2 is obtained using the up- wastakenasP0 = 6 x 105Pa. The well radiuswasrw'-
dated it and Ss. 0.001 m, and the spacediscretizationin r was as follows:
4. Steps 2 and 3 are repeated until a desiredconvergence 0.01,0.02,0.04,0.08, 0.16, 0.32, 0.5, 1, 2, 3, 4, 5, 6, 8, 9, 10,
criterion is met. 15,20, 30, 40, 50, 60, 70, 80, 90, 100, 120, 140,160,180,
200,
220,240,260,280,300,320,340,360,380,400,420,440,460,
480,500,550, 600,650, 700, 750, 800, 850, 900,950,1000,
VERIFICATION AND TEST CASES
1200,1500,2000,and3000m. In the termsof ourformula-
The LTFD numerical method was tested in four test tion,theaquiferproperties andparameters correspondedto
problems of groundwaterflow which representedincreasing kr = 2.35965 x 10-ll m2 Ss = 2.03874 x 10-•ø 1/Pa,
levels of complexity. An analytical solution, the Theis solu- Ix= 1.0x 10-3 kg/rn/s,
and• = 1.15741
x 10-2m3/s.ANs
tion, exists for the first test problem of radial flow to a well = 10 was used in the LTFD simulator. A variable At was
in an aquifer with infinite boundaries.The LTFD solution usedfor all testcasesin the FD simulator,givenbythe
recursive formula
was verified through comparison to the analytical solution.
No analytical solutionsexisted for the other three test cases.
In all four cases the results obtained from the implementa-
Atl= min{1.5x AtI _ 1,Atmax} l = 1,2,"- (52)
tion of LTFD were tested against results obtained from a withAtma
x -- 8.64x 104S= 1day.Thenumberoftimesteps
standardimplicit FD simulator for the same spacediscreti- wasthereforea functionof the originaltime stepAt0.
zation. A direct banded-matrix solver was used to solve the The analytical
solutionof the drawdownat t l = 5 days
system of simultaneousequations arising in the LTFD and along
ther axiswasgivenbytheTheissolution.
InLTFD
a
MORIDISANDREDDELL.'
LAPLACE
TRANSFORM
FINITEDIFFERENCE
METHOD 1879

1 I I I. . I.. I I

(a) (b)

• 40?I' '4- am,


I "'•- FD,1 time step I
--- Theissolution J-•- FD,2 timesteps J
ß..o...
LTFD,
NS-10 I .-a..FD,3 timesteps I
...A-..
FD, 1 time step I .-x- FD,4 timesteps I
...&...FD, 2 time steps I .-o-.FD,6 timesteps I
..o. FD,3 timesteps
...x...FD,4 timesteps .• 20.•I,-+-?D.7.me stepsI
ß..o...FD,6 timesteps ß

'"•: FD,17timesteps II
ß 10
u

Tim
e=..
10daYsllJ [[,Tim
e=;'
10,,,daYs•J
0.1
0 1000 2000 3000 4000 5000 6000
Distance from the well (m) Distance from the well (rn)

Fig. 3. Same as Figure 2 for t = 10 days.

singleobservationwas madefor t = t •. In the FD simulator, tended toward the LTFD and the Theis solutions with an
fora•t0 = 5 days, 2.5 days, I day, 0.5 day, 0.25 day, and 10 increasingnumber of time stepscorrespondingto smaller zkt
rain,the numberof required time-stepsto reach t• was 1, 2, values.
4, 5, 6, and 15 respectively. The drawdown results for both An identical pattern was observed for t 2 = 10 days in
the analyticaland the numerical solutions (LTFD and FD), Figures 3a and 3b. The FD simulator requires 1, 2, 3, 4, 6,
as well as the difference between the Theis solution and the and 17 time stepsto reach t 2 = 10 days for zXt0= 10 days,
numericalsolutions, are presented in Figures 2a and 2b 5 days, 2.5 days, 1.125days, 0.5 day, and 10 min, and tended
respectively. The Theis solution and the LTFD solution toward the Theis and LTFD solutions which practically
practicallycoincided, having a maximum difference of 3.0 x coincided.
10-3 matr = 0.01m (wherethedrawdown
is 2.3487m)and The effectof Ns on the accuracy of the LTFD schemeis
a differencedecreasingwith increasing r. The FD solution shown in Figure 4. The differencebetween the Theis and the
LTFD solutionswas negligiblefor an Ns rangingbetween 6
and 20. The minimum differencewas observedfor N s = 6,
and increasedwith an increasing N s until Ns = 12. The
solutionsfor an Ns between 12 and 20 were identical and
exhibited arithmetic differences in the 10th decimal place.
The implicationsof the results in Figure 4 are that (1) the
! accuracy of LTFD for this one-dimensional problem is
practicallyinsensitiveto the value of Ns, and (2) the number
-o- •D, .;=• I of summationterms Ns for an accurate solutionmay be 6,
_-•_
,m),
.,=,II far smallerthan the Ns = 18 which Stehfest [1970] deter-
LTFD,
Ns=10
mined for double-precisionvariables. This drastically re-
LTFD,
Ns-1• duces the execution time and makes the LTFD method even

]••,•_•
LTFD,
Ns=12
I
LTFD,
Ns=16•
LTFD,
Ns=18•
LTFD,
Ns=2,01
more efficient than theoretically predicted.

Test Case 2

The secondtest problem involved flow toward a single


well in a two-dimensional cylindrical (r, z) system. The
geometry,boundaries,and propertiesof the simulatedaqui-
I II fer (i.e., the permeabilities
kr, k•. in (r, z) and the porosities
Time
=•idaY•11 4>)are shownin Figure5, and describean extremelyaniso-
tropic and inhomogeneous layered system.Sinceno analyt-
ical solution is available, a comparison was made between
0, the LTFD and the FD solutions. The initial pressure and the
Distance from the well (m) boundarypressureat the constantheadboundarywerep0 =
Fig.4. Effect
ofNs ontheaccuracy intest 6.0 x 105Pa, andtheSs wasasin testcase1.
oftheLTFDmethod
case 1. The domain was discretized in 500 = 25 x 20 grid blocks
1880 MORIDISANDREDDELL:LAPLACETRANSFORM
FINITE DIFFERENCEMETHOD

: " ' R=1277m, 25 varlablysized subdivisions•

51m
I
!•"•"
• kr-5.0E-12
m2,
kz=5.0E-13
m2, • •' o

• [•,•,•.• • •
ß• ........... • ..... •-,•
_ •.•.,•.,•.•.•.•.,•.•.•.•.,•.•.•.•.,•.•.,•./
kr = 3.0E-13m2 , •=10E-13m2,
3ml ......................
..... kr=3.0E-13 1•m••½-0• ß
m2,• =1.0E ß ...... ..... .....................
.•.•.•.• ..... •.•.• ......... • ..... •.•
.....,.,.,.........................
ß- *.•:•'•::(:
•.•:::::•::•':
•;::'
•::
•":•
5•L'•:.•::
:':5:.
•5:::'
:•:•:•:••'•::
:.:':•:
' •:.'•::
•}:.:::::7
•:•:5
:'• •:•:;:
:•:•• •:'•::
•:'••5:•:.":'i:•-•
:.':•::
•:•::::
;•::
:::'
:::::
• :•::•:
::•::"•i:•
F• •.::•5•.'•.:
••:•:'::•:::.
::::':.•:•::
::•
•:•:•::•
•:::5:•?
5d:•:•':•::
•-:::•:"•:
:'::•'•:
•:•::
:•::
:•:' ".•::'•
::5:.
•:.•:
....
ß.::•55•::•-•:'.:Z•
:":•-:
•::•'•5
:•:.5:'::
......
•:•:.•":'
:.:.•::
::.:.:::•:•
::•':.5•:.•::•:.•:.h:.•.•::
::5:"•:•:'?:•:
95:::•:•:.•::
:::•
'•::::
•:•:
::::::::::::::::::::::::::
:•:::•:::.•:
::::?"
•:.:::.
•::"•'-'•'•:.
:.•?::
:::5:•'
•:•: •:::5:::•.•:.'.?
::•:•
•:.•
::•'::•
:'5:.::::5•5'
:•::•:):;:
:•:::•
:::::::.:.:•
•:.•.:•5
•:•::•::•:.
Z•:5•'•:
•-••: . ?::::::
• :•:•:
•:• :..•-:•:
.......................
:........ :.•½•:..'.::•:.•:5.:.:..:•.•:::
•. . . •::::
• ::.-••:•;:
.:.:::::•
::::•:•.::
::::.:•:
::.:•..:•:5:•::... .::5,:•:••::5
:.•; :.:.•"•:.•:
:.::•5:
•::::•5.•::5:..
':::
:½5:.::5".:5.:.•:.'.
•:.•::•:•::.::•;•:.'::'..5
•::.::::'::5'
'" ' '"'"" ':':::•:::•:.:L..:•:•
::':F'.:::::::
::-..:.'.
ß...."?•.::•::'.':::::::
?::•.:
:-:'::•
'•:'::.•
•:•:'3..:
?'::•:.•:.:'
•:::.
:::•'.:..:.: :::::•-:
":..'
.::::::•'
':':.:.3'..
-"':':
":':: :':
...........:'' :""•:''"......
:'?•:':":•:
:::•'
:•:":::'-::
::'::'::•::•:'""•:•:
:::•":":' O
.......................
..: ::•:•:.:•::•::•:.:.:•::•..

6m :•:•...:
.....
Well
sere,
e_n.J

Fig. 5. Geometry,boundaries,and propertiesof the cylindricalaquifer in test case2.

in (r, z). The grid block size was uniform in z, with zXzk= 1the percent drawdown difference between the LTFD andthe
m, k = 1, - --, 20. Discretization in r was nonuniform with FD solutions were made along the z axis at r = 0.15 m. The
resultsof the comparisonappear in Figures 6. The accuracy
the following Ar distribution: 0.1,0.2, 0.3, 0.5, 0.8, 1, 1.5, 2,
of the LTFD solution was indicated by the fact that (1)the
3, 5, 7.5,10, 15, 20, 30, 50, 60, 80, 100, 120, 150, 170, 200,
150, and 100 m. The well radius r,. = 0.1 m. The well FD solution for a decreasing At size (which results in smaller
pumping rate was300m3/day= 3.4722x 10-3 m3/s,and truncation errors, more accurate solutions, and larger num-
was equally divided among the six grid blocks corresponding bers of time step) tended to the LTFD solution, and conse-
to the well screen (Figure 5). A single observation was made quently (2) the difference between the two solutionsde-
at t ohs = 10 hours. The LTFD method was applied with a creased and tended to zero. The power of LTFD is
Ns = 10. In the FD simulator,for a At0 = 10 hours,5 hours, demonstratedin Figure 6a, in which it captured in detailthe
1 hour, $ min, and 6 s, the number of required time steps to significant variations in drawdown due to the presenceof
reach tot•. was 1, 2, 5, 15, and 22 respectively. wells and zones of drastically different permeability. The FD
Comparisons of drawdown (computed as -Ap/-y) and of solution showed significant deviations and insufficientaccu-

.... I . , ,. , I , , , , I ß , , , .... ! .... I .... I


20 - -
.
15 - -

I/•Time"
10
hoUrS
i •, . (b)-
•; -x-FD,
2lime-stePs
I
• ---F_D,
5_fime-ste•
I •
•FD; 51tirne-st•psl '
-e- LI'FD,
Ns-10 i
-a- FD,1 time-step I
•c10- i-/-FD,
22time-steps
I '
-x- FD,2 lime-steps I
-,- FD,5 lime-slopsI "'"
. -o--FD,
15time-steps
I - ":'
-/- FD,22time-steps
I
.
õ 5-
•o10 .
a) :

' '
(a) . ,. .
6 43 0 '
"' " ' I " " ' " i ' ' "'' ! "'" ' "' ' " " ''1 " " " ' I ß ' " ' ! ....
0 5 10 15 20 0 5 10 15 20
Depth(m) Depth(rn)
Fig. 6. Comparison of the LTFD solutionfor test case 2 at t = 10 hours to the FD solutions for various numbers
of time stepswith respectto (a) drawdowndistribution,
and(b) percentdrawdowndifferencedistributionalongz at
r = 0.15 m.
MORIDIS
ANDREDDELL'
LAPLACE
TRANSFORM
FINITEDIFFERENCE
METHOD 1881

7,200 m, MX = 48 equally-sized subdivisions kx=9.0E- 13


ky=7.0E- 13
4=0.43
kx= 1.0E- 11
ky--6.0E- 12
O.38
kx---3.5E- 11
ky= 1.0E-11
0.33
kx=5.0E-12
ky=2.0E-12
4=0.43
kx=5.0E-13
ky=3.0E-13
4=0.45
kx=8.0E-12
ky=3.0E-12
41)=0.40
kx= 1.0E- 16
ky= 1.0E-16
Initial pressure: Boundaries: 41)=0.01
6.0E5 Pascals
Pressures@ B1 & B2:
6.0E5 Pascals
Constant-head
Bi and B2,
at

No-flow elsewhere Wells


at
(x,y)/(IJ):
!
(33,23) - Q = 400 cu.m/day
(16,15) - Q = 500 cu.m/day
,,

Fig.7. Geometry,
boundaries,
andproperties
of theaquifer
intestcase3. Here,k.•andkyarein meters
squared.

racycausedby the averaging effect of the treatment of the difference between the two solutions decreased and tended
timederivative for larger At values, deviations which de- to zero.
creased with smaller zXt values. The power of the LTFD method is further demonstrated in
Figure 8b. The capability of LTFD and the inability of FD
Test Case 3 (caused by the averaging effect of the treatment of the time
derivative for larger At values) to accurately describe the
Test case 3 represented an extremely anisotropic and effects of the presence of wells and zones of significantly
inhomogeneous two-dimensionalcartesian(x, y) systemof different permeability are reflected in the difference between
flowto two wells.The geometry,boundariesand properties the two solutions. The location of wells and permeability
oftheaquifer,as well as the locationand pumpingratesof zones can be identified by the existence of peaks and sharp
thetwowells,are shownin Figure 7. Initial and boundary variations in drawdown, variations which decrease in mag-
pressures, as well as the Ss, remained the same as in test nitude with a decreasing At size in the FD solution.
case2. The aquifer had a uniform thickness Az = 50 m. No
solutionis possiblefor this problem;the measure Test Case 4
analytical
ofthepower,validityand accuracyof the LTFD method Test case 4 was the extension in three dimensions of test
wasprovidedby comparingthe LTFD and the FD solutions. case 3, and represented an extremely anisotropic and inho-
The solutiondomain was discretized in 1536 = 48 x 32 mogeneous three-dimensional cartesian (x, y, z) system of
grid
blocks
of uniform
sizein (x, y), withAXi = Ayj -- flow to two wells. The geometry, properties, and parameters
150m,i = 1,-.- , 48 andj = 1, --- , 32. A singleobserva- of the aquifer remained as in Figure 7. A single observation
tionwasmadeat tot,s = 2 days.The LTFD methodwas was made at tot,s = 2 days. The LTFD method was applied
appliedwitha Ns = 10.In theFD simulator, for a At0 = 2 with a N s = 10. In the FD simulator, for a At 0 = 2 days, 0.5
days,I day,0.5day,2 hours,10min,and6 s, thenumber of day, 10 min, and 6 s, the number of required time steps to
requiredtimestepsto reachtobswas1, 2, 3, 7, 13,and24 reach tooswas 1, 3, 13, and 24 respectively. The differences
respectively. between test cases 3 and 4 were the following:
TheLTFD methodwas evaluatedby comparing the 1. The thickness of the aquifer was divided into four
distribution
ofthedrawdown andofthepercent drawdownequally sized subdivisionswith Azk = 12.5 m, k = 1, ---, 4.
difference
between theLTFDandtheFDsolutions alongthe This resulted in a very large system with a total of 6144 grid
xaxis
aty - 2175mandpassing
bythewellgridblockat (i, blocks.
J)= (16,15).Thedistributions
of drawdownandpercent 2. The vertical permeability was taken as kz = 0.05kx
drawdowndifference
areshown inFigures
8aand8brespec-over the whole aquifer.
tively.
Thesame pattern
observedinallprevious
testcases 3. The wells were placed at the deepest aquifer layer,
isobvious-
LTFDproduced anaccurate solution,
a fact i.e., k =4.
indicated
bytherealization
that(1)theFDsolution
fora 4. The boundary conditions at the first layer in the z
decreasing
At size(whichresultsin smaller
truncationdirection (k = I) were the same as depicted in Figure 7. For
errors,
moreaccurate
solutions,
andlargernumbers
of time k = 2, 3, 4, no-flow boundaries were assumed everywhere
steps)
tended
totheLTFDsolution,andconsequently
(2)the along the perimeter of the aquifer.
1882 MORIDIS
ANDREDDELL:
LAPLACE
TRANSFORM
FINITEDIFFERENCE
METHOD

I , I •,. I [111 I IlJllllllll I Illl Ill I Ill I I [11111•.


14-

Time
- '2days]l • FD,1time-step
l
-c>-FD,2 time-steps
I
..•- FD,3 time-stepsI
-a- FD,7 time-steps
I
-e- LTFD,Ns '10 -x- FD,13time-stepsl
....-o-
FD•24time-stepq
.•. FD,
FD, 21time-steps
time-step
ee 10 FD, 3 time-steps
FD, 7 time-steps
,,era FD, 13 time-steps
o
,FD:,,
24time-steps

,--,

(a) (b)

I ..... '' I ' ''l ' •' ' ' I' ' ' ' [ ' ' '' I [,"'i"'i",I ' ' ' ' I ' '"'" ' [ ""
o 20O0 4000 6000 0 2000 4000 6000
x (m) x (m)

Fig. 8. ComparisonoftheLTFDsolution
fortestcase3 at t = 2 daysto theFD solutionsforvariousnumbers of
timesteps withrespect
to (a)drawdown
distribution,
and(b)percent drawdown difference
distribution
alongx aty =
2175 m.

The LTFD method was evaluated by comparing the for a decreasingAt size, thus causingthe differencebetween
distribution of the drawdown and of the percent difference the two solutions to decrease and tend to zero.
between the LTFD and the FD solutionsalong the x axis at The power of the LTFD method is further attestedtoin
y -- 2175 m and z = 43.75 m, and passingby the well grid Figure 9b. The LTFD methodis better equippedto accu-
block at (i,j, k) = (16, 15, 4). The distributionsofdrawdown ratelydescribethe effectsof the presenceof wellsandzones
and percentdrawdowndifferenceare shownin Figures9a of significantlydifferentpermeability,as the difference
be-
and9b respectively.The resultsindicateda patternidentical tween the two solutions demonstrates. The location of wells
to the pattern in test case 3: LTFD producedan accurate andpermeability
zonesis identifiedby the existence
ofpeaks
solution, and the FD solution tended to the LTFD solution and sharpvariationsin drawdown,variationswhichde-

i I .... ! i , I • ,

-a-FD, 1 time-step
. IITime
='2daYSlJ
•_•_
FD,
3time-steps
I-
FD,13time-steps •

II
II
--e-LTFD,
N$.-10
--ca-FD,1 time-step I
I FD,
24,,time-steps

•l -A- FD,3 time-steps
I
_i ! ..x- FD,13tlme-stepsl
, ,,,,, ,,, , ,,

, 10

b
2
(a) !• :
• _

I i I ' I ' I ' .... ' I ' ' I '" ' i


0 2000 4000 6000 2000 4000 6000
x (m) x (m)

Fig. 9. Comparison
of theLTFD solution
fortestcase4 at t = 2 daysto theFD solutions
forvarious
numbers
of
time stepswith respectto (a) drawdowndistribution,and (b) percentdrawdowndifferencedistributionalongx at y =
2175 m and z = 43.75 m.
MORIDISANDREDDELL:LAPLACE
TRANSFORM
FINITEDIFFERENCE
METHOD 1883

ß ß ., , I .... I . , , , ! , , , , I , .

ua

(in%oftheoriginal
fluidma,,ss),

--o-Test
caseI, 'Et ='1.770E-5
%1
--o-TestcaseI!, Et = 4.950E-5
%•
-x- Test
caseII!,Et = 2.276E-3
%•
E,,L=LTFD
(1,
n% Mass
Balance
ofthe Error
original for
fluid Ns=10
mass) ! -/•-Test
caseIV,EL."'
3.044E-3

-o- Testcase !1, E L= 4.950E-5


c:
-x- Testcase III, E, = 2.276E-3
-•- Testcase IV, E L'" 3.044E-3

ß ' .... '" ' ' ß " ' =' ' ' ' ' I ' ' *

Number of time-steps in the FD scheme N s = Number of summationterms in $tehfest'salgorithm

Fig. 10. Relativemassbalanceerror of the standardFD methodas Fig. 11. Relative mass balance error of the LTFD method for a
a fraction of the error of the LTFD method. varying Ns as a fraction of the error of the LTFD method for a
N s = 10.

creasein magnitude with a decreasing At size in the FD


solution.
EL,Ns
Mass Balance Error Considerations ER'•Vs=
EL' (54)
A very important measure of the validity and accuracy of whereEL,•Vs
istheerrorcalculated
witha variableNs. It can
the LTFD numerical method was provided by the determi- be seen that the accuracy of the LTFD method, indicated by
nation
ofthemass balance errorEr, whichwas1.770x 10-5 decreasing valuesof ER.•v
s, improvedas Ns increased
percent,
4.950x 10-5 percent,2.276x 10-3 percent,and between 8 and 20. The improvement was more pronounced
3.044x 10-3 percentof theoriginalfluidmassin testcases in the two most inhomogeneousand anisotropic cases, i.e.,
1,2, 3, and4 respectively.In all casesa Ns = 10 was used. cases2 and 4. The greatestaccuracy was observed for Ns
Theextremelysmall magnitudeof the error further testified values between 16 and 20. However, the improvement over
to the power of the method. the rangebetweenNs = 8 and Ns = 20 was marginal,and
Therelativemassbalanceerror E R•u of the standardFD in no case exceeded 8%.
methodversusthe LTFD method was computed as The implications of these observations are that (1) for
Ns <- 20 the accuracyof LTFD is practicallyinsensitiveto
EFD the value of Ns, and (2) the number of summationtermsNs
ER = •. (53)
Eœ for an accuratesolutionmay be far fewer than the Ns = 18
which Stehfest [1970] determined for double-precisionvari-
EFz> isthe massbalanceerror (as a percentageof the initial ables. In light of the extremely small mass balance error, a
fluidmass) whentheFD isimplemented, andE•; isthemass value of Ns -- 8 is suggestedfor LTFD simulations,with
balance errorforLTFD discussed in thepreviousparagraph. little lossin accuracy.This drastically reducesthe execution
AsFigure10depicts,for the samedomaindiscretization the time requirementsand makes the LTFD method even more
LTFDscheme wasin all casesconsistently
moreaccurate efficient than theoretically predicted.
thanthe FD scheme.This was expectedbecauseof two As shownin Figure 11, the accuracy of the LTFD method
reasons:
(1) a smaller truncation error, equal to the trunca- may deterioratefor a Ns > 20. This is probably due to
tionerrorofthespacediscretization
only,and(2)in thecase limitations imposedby the computer accuracy in the deter-
oflarger
numbers of timesteps,a smallerroundofferrordue minationof the weightingfactors V• in (48). The calculations
tofeweroperations.Althoughtheimprovement in accuracy involve operationswith factorials,which may becomeinac-
wasnot pronouncedin test cases 1, 3, and 4, it was dramatic curatefor large numbers.Although this deteriorationmay be
intheinhomogeneous,
anisotropic,
two-dimensional
cylin- eliminatedin computerswith a greaternumberof significant
dricalproblemin test case2. This indicatesthat the LTFD figuresin double-precision
variablesthan was usedin this
formulation
is numerically
moreaccurate
andbettersuited study,this wouldbe pointlessand undesirablebecause(44)
tohandle
thedifference
in themagnitude
of theflowcoeffi- would have to be solved more times and would require
cients
acrossthegridblockboundarysurfaces,a conditionlonger executiontimes.
encounteredin suchproblemsdue to the increasingarea of
flowasr increases.
SUMMARY AND DISCUSSION
Theeffectof thenumber
of summation
termsNs in the
Stehfest
algorithm
onaccuracy
isdemonstrated
inFigure
11. A new numerical method, the Laplace transform finite
The
relative
massbalance
error
ER,•v
swascomputed
as difference(LTFD) method,was developedto eliminatethe
1884 MOlUDISAND REDDELL:LAPLACETRANSFORM
FINITE DIFFERENCEMETHOD

adverse effects of the treatment of the time derivative in the torswitha minimum
of effort.Inspection
of theLTFD fl0w
numericalapproximation of the parabolicpartialdifferential coefficients
Ad* andCd* (d* ---X, Y, Z) orR, O, Z reveals
equation(PDE) of transientflow throughporousmedia.The that they are identicalto the flow coefficientsin an FD
LTFD method is insensitive to the time step size and scheme.The only differencebetweenthe two numerical
consistsof four steps:(1) A Laplacetransformis performed methods liesinthevalues
of(1)theflowcoefficient B and(2)
on the PDE, (2) the transformedPDE is approximatedusing thefight-hand
sideD (equations
(40)and(41)respectively).
a finite difference(FD) method, (3) the resultingsystemof More specifically,(1) the term (poSs/At) in the FD formu.
simultaneousequationsis solvedandthe transformedvector lationis replaced
by theterm(poSss)in equation
(40);(2)
of the unknown pressuresis determinedin the Laplace thegravitational
termsin theFD simulatoraredivided bythe
space, and (4) a numerical inversion is performedon the Laplacespacevariables (equations
(36), (37), and(38));and
transformedvector of unknown pressuresobtainedin step 3 (3)theright-hand
side(-Ssp øIf/At+ q + GG*)intheFD
using the Stehf'estalgorithm. The solutionin the Laplace formulation
is replacedby (-Ssp(O) + q/s + GG) in (41)
space renders the effects of the time derivative on accuracy (poldisthepressure
attheendoftheprevious
timestep
and
and stability irrelevant becausetime is no longer a consid- GG = GG*/s).
eration. The disadvantageof having to solve the system of simul-
Four test cases were investigatedto evaluate the LTFD taneousequationseight times for a singletime stepis
method. Due to its formulation, LTFD requires solutionof outweighed by (1) an unlimitedtime stepsizewithoutany
the resulting system of simultaneousequationsNs times, lossof accuracy,(2) a superioraccuracy,and (3) a stable,
one for each of the N s different approximationsof the nonincreasinground off error. Therefore, calculationsin a
Laplace spacevariable s. Ns theoreticallyrangesbetween LTFD scheme are necessary only at the desired observation
10 and 18. Combinationof the resultingNs solutionsusing times, thus allowing "snapshots" in time. On the other
Stehfest's algorithm returns the actual solution in time. hand, in a standard FD method, calculations are neededat
Compared to the standardimplicit FD method, LTFD does all the intermediate times defined by the sequenceof small
not increase the storagerequirement because (1) the values time steps between observation times.
of the unknowns at the previous time step are not needed,
sinceno intermediate
time stepsare necessary,
andthetlme
Acknowledgments. This work was supported by the TexasAd-
stepsizecoincideswith the observationtime, and(2) the Ns vanced Research Program under grant 4473/87 and by the Texas
sets of unknowns can be stored and summed in a single Agricultural Experiment Station under grant 6993.
array. In FD this array is that of the unknowns at the end of
the previous time step. REFERENCES
It was determinedthat a Ns = 8 is sufficientto providean
extremely accurate solution. In essence,this means that the Aziz, K., and A. Settari, Petroleum Reservoir Simulation,476pp.,
Elsevier Science, New York, 1979.
solution at any time is obtained by solving the system of Bouwer, H., GroundwaterHydrology, 480 pp., McGraw-Hill, New
simultaneousequationsin LTFD eight times and algebra- York, 1978.
ically combining the solutions. Although the accuracy in- Moridis, G. J., An implicit two-phasenumerical simulatorfor mod-
creaseswith increasingNs for Ns < 20, the improvementis eling secondarywater recovery by air injection, Ph.D. disserta-
tion, Dep. of Agric. Eng., Texas A&M Univ., College Station,
marginal and insufficient to justify the additional execution 1987.
time.
Stehfest,H., Algorithm 368, Numerical inversion of Laplacetrans-
With a smaller time step size and more time steps,the FD forms, Commun.Assoc. Cornput. Mach., 13(1), 47-49, 1970.
solution tends to the LTFD solution. LTFD provides a
solution consistently more accurate than the FD solutionfor G.J. Moridis, GroundwaterResearchProgram, Department of
AgriculturalEngineering,Texas A&M University, CollegeStation,
the same space discretization. This is expectedbecausethe TX 77843.
truncation error was reduced to the truncation error of the D.L. Reddell, Department of Agricultural Engineering,Texas
space discretization, and the round off error was reduced A&M University, College Station, TX 77843.
due to the limited number of operations needed. An unlim-
ited time step size with a stable, nonincreasingerror is thus
(Received June 20, 1990;
possible. revised April 25, 1991;
Existing FD simulators can be changedto LTFD simula- accepted April 25, 1991.)

Vous aimerez peut-être aussi