Вы находитесь на странице: 1из 11

A Generalized Wellbore and Surface

Facility Model, Fully Coupled to a


Reservoir Simulator
B.K. Coats, SPE, G.C. Fleming, SPE, J.W. Watts, SPE, and M. Ramé, Landmark Graphics, and G.S. Shiralkar, BP

Summary oil reservoir model for improved simulation of steam-assisted


The formulation of a black-oil or compositional fully coupled sur- gravity drainage with dual horizontal wells. Reservoir mass and
face and subsurface simulator is described. It is based on replacing energy balances; segmented wellbore energy, momentum, and
the well model in a conventional reservoir simulator with a gen- mass balances; and the perforation-rate equations were solved si-
eralized network model of the wells and facilities. This allows for multaneously through Newton iteration. Capacitance was repre-
representation of complex wellbore geometry and downhole sented in the wellbore control volumes, and simultaneous annular/
equipment. The method avoids the inefficiencies and/or inaccura- tubular flow was modeled for a given wellstream. The authors
cies of other coupled models, in which wells and facilities are found that simulator timestep size was impractically limited for
treated as separate domains or in which the global system is not field-scale problems when wellbore conditions were changing rap-
solved simultaneously. Example cases demonstrate the perfor- idly. More recently, Holmes et al.2 presented a black-oil model
mance of the model for cases with simple and segmented well- with an implicit configurable segmented wellbore. Four primary
bores (with and without facilities). variables were included for each wellbore segment—a weighted
Introduction and Background total flow rate, weighted fractional flows of water and gas, and
pressure. The corresponding equations were three component mass
The objective of our work is to develop an integrated reservoir and balances and a hydraulics relationship for each segment (repre-
facility simulator that is as efficient and general as possible. This senting pressure drop caused by gravity, friction, and acceleration)
involves many technologies. This paper focuses on the primary through correlations or hydraulics tables. The pressure equation for
aspect of our new model that differentiates it from previous work, the segment terminating at the well’s datum depth was replaced
specifically the manner in which domain decomposition is applied with the limiting boundary condition—either a rate constraint, a
to simulation of the coupled system of reservoirs, wells, and facilities. bottomhole pressure constraint, or, for wells on tubinghead pres-
sure control, a hydraulics relationship for the tubing represented by
Conventional Well Model. A basic example of domain decom- hydraulics tables. Tables also could be used for pressure drops
position is the well model in a reservoir simulator. A simple rep- across flow-control devices. Stone et al.3 then extended this work
resentation of a reservoir simulator is shown in Fig. 1. In the well to compositional and thermal applications. We note that these
model, within each simulator Newton iteration, each well is de- advanced wellbore models are very similar to the network models
coupled from the reservoir through reservoir boundary conditions of facilities.
consisting of fixed current iterate values of the perforated grid-cell
variables. Well rates and pressures are determined subject to the Coupled Surface/Subsurface Models. In reservoir models, flow
well boundary conditions, which apply at the boundary with the in the reservoir and flow between the reservoir and wellbores is
facilities. At the end of each simulator Newton iteration, the res- decoupled (either bottomhole or at the wellheads) from flow
ervoir and well equations are solved simultaneously. Most current through the remainder of the production and injection facilities by
reservoir models implicitly represent well rates in terms of an specifying pressure and/or rate constraints for each well. If indi-
additional variable (to the primary reservoir cell variables), well vidual-well rates and pressures are known from production history,
bottomhole pressure. Conventional well models treat the wellbore then the decoupled reservoir/well model is sufficient to match
as a mixed tank, but many include methods for explicit inclusion historic reservoir behavior by specifying and matching the ob-
of pressure losses caused by friction and acceleration through vari- served boundary conditions as a function of time. However, when
ous approximations and assumptions. used in predictive mode for reservoirs with gathering and distri-
bution networks, the proper decoupled well boundary conditions
Modeling of Advanced Wells. The advent in recent years of are in general variable in time and dependent on reservoir behav-
horizontal- and multilateral-well drilling technology and of com- ior, equipment performance, production strategy, hydraulics rela-
plex wellbore configurations resulting from implementation of in- tionships, and pressure, rate, and source composition constraints
telligent completion systems has resulted in the development of that may be applied within the surface network. When production
increasingly sophisticated well models for use in reservoir simu- is controlled in the surface facilities, it is in general necessary (or
lation. The complex geometries, increased lengths of perforated at least desirable) to include the facilities in a full-field model to
intervals, and potential presence of downhole-flow-control devices predict how the otherwise specified boundary conditions will vary
in these wells, along with the need to more accurately model in time.
low-potential recovery methods such as steam-assisted gravity Because this is an obvious limitation of decoupled reservoir
drainage, have resulted in the requirement of a much more detailed simulation, many authors have presented methods for simulta-
representation of wellbore composition, rate, and pressure distri- neous solution of the reservoir and facility equations. Most meth-
butions than is possible in conventional well models. ods are based on modification of a reservoir simulator to iteratively
Stone et al.1 presented a three-phase, thermal, segmented well- converge separate solutions of the well and facility domains
bore casing and tubing model implicitly coupled to a thermal dead- (sometimes referred to as an equilibration loop) before identifying
a conventional solution of the combined reservoir and well do-
mains, as shown in Fig. 2. These methods are referred to as par-
tially coupled because at no point are all domains solved simulta-
Copyright © 2004 Society of Petroleum Engineers
neously. The methods differ according to the frequency of equili-
This paper (SPE 87913) was revised for publication from paper SPE 79704, first presented bration and the definition of final timestep convergence. If
at the 2003 SPE Reservoir Simulation Symposium, Houston, 3–5 February. Original manu-
script received for review 3 February 2003. Revised manuscript received 20 January 2004.
equilibration is performed only on the first Newton iteration of
Paper peer approved 4 February 2004. each timestep, then the surface model is coupled at the timestep

132 April 2004 SPE Reservoir Evaluation & Engineering


Fig. 1—Reservoir model.

level. If it is performed every Newton iteration, the coupling is at


the iteration level. Falling in between, a partially iterative coupling
performs the equilibration for some number of iterations. The fre-
quency of equilibration may also be controllable in time, with a
conventionally decoupled method used between equilibrated
timesteps. The coupling method is further classified as explicit,
partially implicit, or strictly implicit with respect to the facility
solution, on the basis of the final timestep-convergence criteria
and coupling level. If only convergence of the reservoir equations
is required, then the method is explicit if coupled at the time-
step level and partially implicit if coupled at the iteration level.
If convergence of the reservoir equations and the well/facility
boundary conditions is required, then the coupling is said to be
strictly implicit because it yields an effectively implicit facility
solution, regardless of the coupling level. The overall level of the
model’s implicitness depends on the implicitness of the coupling
and on the implicitness of the variables in the equations within
each domain.
Early models, beginning in 1971 with the work of Dempsey4
for gas/water systems, used timestep-level explicit couplings of the
well and surface domains, including simple surface models inte-
grated into the reservoir model code. Extensions to black-oil sys-
Fig. 2—Partially coupled models.
tems were presented by Startzman et al.5 and Emanuel and Ranney.6
More recently, Litvak and Darlow7 presented a coupled model
with the network model integrated within a commercial black-oil the reservoir grid surrounding each well, so that the boundary
and compositional simulator. This was the first reported coupled between the reservoir and the well domains is moved from the
compositional model. They described both fully coupled and par- sandface out into the reservoir, creating overlap of the well do-
tially coupled methods, but only the partially coupled method is mains with the reservoir. The idea is to move the boundary to a set
currently implemented. The implementation by default equilibrates of surfaces at which conditions are changing less rapidly while
for two iterations in each timestep, and timestep convergence is also including parts of the reservoir grid in the well/surface equili-
based on the reservoir equations alone, resulting in a partially bration loop. A fully coupled method was also investigated, but it
implicit method. Extensions of the model to provide integration was concluded that it was inefficient for complex facilities.
with production-optimization strategies applied to BP’s Prudhoe Byer et al.14–16 extended the work of Schiozer and Aziz13 to a
Bay field,8 tabular representation of compositional dependence for fully coupled method, in which the equations for all domains are
improved efficiency,9 automatic history matching,10 and more ad- solved simultaneously at the end of each simulator Newton itera-
vanced production-optimization strategies11,12 have been presented. tion while retaining the extended definition (into the reservoir) of
Schiozer and Aziz13 showed that two acceleration techniques the well subdomains. Also provided for were domain decomposi-
applied to the iteratively coupled implicit method can improve its tion and parallel treatment within the reservoir and the facilities,
efficiency. The first technique investigated was a preconditioner and an adaptive implicit reservoir formulation. Rather than using
applied at the beginning of the timestep. It provided estimates of the explicit preconditioning method of Schiozer, a coarse-grid so-
the new time-level reservoir boundary conditions for use in the lution was performed before equilibration of the well and facilities
well/surface equilibration loop on the first Newton iteration. The domains within each Newton iteration for improved estimates of
second technique extended the well subdomains to include part of reservoir boundary conditions. It was shown that some optimal

April 2004 SPE Reservoir Evaluation & Engineering 133


extent of the well subdomains into the reservoir minimized CPU duces no additional timestep size limitations over those in conven-
time for a given case. The coarse-grid solve and well/surface tional simulators.
equilibration were collectively referred to as “preconditioning.” It The network model represents the wellbores, pipelines, and
was claimed that the fully coupled method could be made efficient equipment in injection and production networks as a series of
by including this preconditioning, but it is difficult to determine nodes, connections between them, connections from nodes to res-
the overall utility of the method from the results presented because ervoir cells (perforations), and connections to sinks and from
example timings were given relative to the case with no precon- sources. The connections are discrete 1D representations of 3D
ditioning. An equivalent analysis for a conventional model with no flow paths, and nonperforation connections may be specified (or
facilities would be to remove the well model from the Newton loop implied through constraint specifications) as containing devices. A
(i.e., make the well model explicit), instead using the iterative set of nodes and connections representing a segmented linear or
change in bottomhole pressures to update well rates and to com- bilateral wellbore is by default automatically generated for each
pare the results with those obtained using some other precondi- well from the input perforation data, which describe each inter-
tioning method. section of perforated well casing with a reservoir grid cell. Flow
Several models have been presented in which advanced surface through each perforation is discretely represented as a connection
facility models developed for production engineering applications from the reservoir cell to a network node in a wellbore flow chan-
(stand-alone facilities modeling, or facilities modeling integrated nel. Fig. 3 shows segmented and lumped wellbore configurations
with simplified reservoir models) are used in place of simpler that can be generated automatically for a vertical well in a carte-
network models integrated into the reservoir code. Because the sian grid with four perforations. In a lumped wellbore, all perfo-
effort required for development of these advanced network models rations for a well are attached to the same node. Complex wellbore
is comparable to that of reservoir simulators, coupling them using geometries may be created through modification of the segmented
the method of Fig. 2 requires very little work compared to the configuration, such as the smart-well configuration also shown in
effort invested in each. Hepguler et al.17 and Tingas et al.18 Fig. 3. Well tubing connections and modifiable tree-structured
presented timestep-level implicit couplings of (black-oil) reser- networks can be generated automatically, or the well subnetworks
voir and network simulators. Trick19 extended the work of Hep- can be tied together with user-specified nodes and connections to
guler et al. (using a different network model) to the iteratively form production and injection networks. Rate and pressure con-
coupled implicit method and reported up to a 30-fold improvement straints may be specified anywhere in the network. The model
in efficiency. determines which constraints are limiting and applies them implicitly.

Approach Formulation
Our method can be represented by a logical extension of Fig. 1, The global system of equations includes the reservoir conservation
with the word “Well” replaced by “Network.” The well domain and constraint equations and the network equations. The network
has been redefined as a single domain containing the wells and equations consist of perforation-rate equations, component-rate
facilities, which we call “network”. Within each simulator Newton balances at network nodes, hydraulics equations, network con-
iteration, we solve the network through its own Newton iteration. straint equations, and composition equations. All quantities in the
Within the network Newton, individual-well subdomains are de- network equations are treated fully implicitly unless otherwise noted.
coupled from the network and solved, also through Newton itera- The total number of primary network equations is given by
tion. Our network model represents a decoupled facility/well simu-
共Nperfs + Ncons兲Nc + Nnodes, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
lator that operates at fixed reservoir boundary conditions. The
global network and well subnetworks are solved with the same corresponding to the primary variables of Nc component mass flow
code, except for a few minor differences. rates in the Nperfs perforations and in the other Ncons network
The main advantage of our fully coupled model over partially connections, as well as pressures at network nodes. The only pri-
coupled methods is that the latter do not implicitly represent the mary network equations that are always fixed in identity are the
facility equations in their conventional reservoir/well domain solve component mass rate balances at nodes and hydraulics relation-
at the end of the Newton iteration. As a result, if the partial cou- ships for connections representing flow paths that do not contain
pling is explicit, then errors and oscillations from timestep to adjustable devices. The remaining primary network equations are
timestep can occur, and if the partial coupling is implicit, iterative dynamic, and they apply as described in each subsection.
oscillations and/or reduced convergence rate can occur. If the par- The secondary network variables are device settings, one for
tial coupling is strictly implicit, a final convergence check is re- each adjustable device. In general, the number of secondary equa-
quired to ensure that rates and pressures at the interface of the well tions is equal to the number of constraints being imposed by the
and surface domains have not changed significantly; if they have, adjustable devices plus the number of fixed device settings. If the
extra simulator Newtons or timesteps are required. Otherwise, er- number of degrees of freedom (the number of adjustable devices
rors can also occur. minus the number of secondary equations) is greater than zero,
Our treatment of network convergence is exactly the same as there are three possible methods to obtain a solution:
that applied to the well model in a conventional simulator. By • Find the optimal solution that maximizes/minimizes some
default, timestep convergence is based only on the reservoir con- benefit/detriment function.
servation and constraint equations. However, because of the • Mathematically represent a production-optimization strategy
simultaneous global solve and prior resolution of the network, to supply the missing equations, resulting in implicit application of
the network equations will be converged within the accuracy of the strategy.
their linearizations, assuming continuity of active constraints. An • Explicitly apply a production-optimization strategy to deter-
option for specification of convergence tolerances for the perfora- mine a number of settings and/or constraints equal to the number
tion-rate equations is provided, mainly to test our assertion that of degrees of freedom.
using it is not necessary because it does not significantly affect the The type of optimization needed depends on the application. If the
overall result. goal is to determine which methodology would optimize produc-
Excluding the capability of representing facilities, our model is tion at a given time, then the first type is needed. If the goal is to
very similar to the advanced well models of Stone1,3 and Holmes.2 determine how production will behave given a methodology, one
The main difference is that our network model is steady-state, of the latter two types is needed. We have currently implemented
while the advanced well models include wellbore capacitance and only the last type, explicit application of predefined and/or user-
are thus transient. When wellbore conditions are changing, defined predictive control logic. Active production and injection
timestep size in the transient models is limited by the time scale targets on connections or groups of connections (group targets in
in which the nonlinearities of the wellbore flow equations can a conventional model) are explicitly allocated to the specified con-
be resolved through Newton iteration. Because our network trolling devices as constraints so that in the network model solu-
equations contain no time dependencies, the network model intro- tion, any active constraint is controlled by adjusting a single device.

134 April 2004 SPE Reservoir Evaluation & Engineering


Fig. 3—Wellbore configuration.

Assumptions made in the network include the following: i, Qin


i and Qi
out
refer to flow rate across reservoir cell faces, ai is
1. Flow is at steady state; there is no capacitance. the rate of accumulation, Gi is the rate of generation, and Qip is
2. Phases are in equilibrium in the connections. the perforation rate of flow (positive for production) from reser-
3. Phases flow concurrently in the connections. voir cell r into a wellbore through perforation p. The Qip are
4. Temperature may vary in space but not in time. summed over the perforations contained in cell r. The conserved
5. A rate constraint is controlled by a device in the connection quantities are all fluid components including water and (for
on which it is specified or allocated. The upstream node represents thermal models) energy.
the device inlet. In addition to the conservation equations, a constraint equation
6. A minimum pressure constraint at a node is controlled by a is solved for each reservoir cell. For the volume balance method of
device in the downstream connection, with the node representing our simulator, the constraint is that the pore volume is equal to the
the device inlet, and it cannot be applied when more than one fluid volume, written in residual form as
downstream connection exists.
7. A maximum pressure constraint at a node is controlled by a RNc+1,r = VPr − VFr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)
device in the upstream connection, with the node representing the It is the convention in reservoir simulation to refer to constraint
device outlet, and cannot be applied when more than one upstream equations as those equations that apply to variables in only one
connection exists. grid cell. In the network equations, we define constraints as those
8. A hydraulics table assigned to a device represents its oper- equations that specify limiting values.
ating setting (i.e., fully open for valves and chokes, maximum In our isothermal reservoir model, the reservoir variables are
operating setting for pumps and compressors), which is modified component masses and pressures for each reservoir cell.20
only to impose a constraint.
9. Inline phase separations are not yet allowed, but produced Perforation Equations. Neglecting capillary pressure, the mass
phase streams may be reinjected (directed to sources) through production rate of component i from reservoir cell r through each
explicit predictive control logic. producing perforation p to network node n is given by
We intend to relax assumptions 3 through 9 in future work.
Nphases

兺␭
Given these assumptions, the numbers of secondary network
variables and equations are always the same and are equal to the Qip = Cp⌬⌽p kr ␳kr zikr , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
k=1
number of adjustable devices. The identity of the secondary equa-
tions is presented later in the subsection on Hydraulics Equations. ⌬⌽p = ⌽rp − ⌽wp, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5)

Reservoir Equations. All discretized reservoir simulators solve ⌽rp = Pr + ␥r共Dp − Dr兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (6)
conservation equations of the general form and ⌽wp = Pn + ␥wp共Dp − Dn兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (7)
Rir = Qiin − Qiout − ai + Gi − 兺
p∈pr
Qip, . . . . . . . . . . . . . . . . . . . . . . . (2) where C is the wellbore constant, equal to the well index21 times
the permeability-thickness product; ␭kr is the volumetric mobility
where the residual, R, of component i for each reservoir cell r is (relative permeability/viscosity) of phase k in cell r; ␳kr is the
driven to zero at full convergence of the equations. For component density of phase k in cell r; and zikr is the mass fraction of com-

April 2004 SPE Reservoir Evaluation & Engineering 135


ponent i in phase k in cell r. ⌽rp is the perforation reservoir the primary network equation that applies is the rate or pressure
potential, obtained by correcting the reservoir cell pressure to per- constraint, Eq. 10 becomes a secondary equation, and the system
foration depth Dp using an explicit hydrostatic pressure gradient in is subject to an additional constraint of the form
the reservoir cell, ␥r. ⌽wp is the perforation wellbore potential,
obtained by correcting the node pressure Pn to perforation depth Pjin − Pjout ⱖ f 共qជ j,Pjin,Pjout,Vjmax兲. . . . . . . . . . . . . . . . . . . . . . . . . . (11)
with the wellbore gradient ␥wp.
For injecting perforation p contained in reservoir cell r, the
If a device is actively constraining a rate or pressure and Eq. 11
component mass rate equations are not independent, and the single
becomes violated in any network iteration, the rate or pressure
equation representing the total mass perforation rate (negative for
constraint is deactivated and replaced as a primary equation by Eq.
injection) is
10 at V=Vmax. Also, when a rate or pressure constraint is being
QTp = Cp␭Tr ␳inj
p ⌬⌽p, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (8)
applied, the device setting required to achieve the constraint may
be back-calculated from the secondary equation, Eq. 10, and the
where ␭Tr is the total reservoir cell mobility (sum of ␭kr) and ␳inj
p pressure and rate solution. The addition of device-specific models
is the perforation-injected fluid density. Endpoint mobility ap- for Eq. 10 will allow our model to predict device settings.
proximations are not used in our model because the injected fluid Currently, the hydraulics function f can be represented only
may be multiphase owing to multiple sources, crossflow, or rein- through hydrostatic pressure gradients or hydraulics tables, but
jection of produced streams. The injected fluid density is treated more rigorous options are being implemented. Hydrostatic gradi-
implicitly or explicitly and is computed from phase saturations and ents in connections are computed from the sum of phase densities
densities obtained by flashing the perforation fluid composition at weighted by their saturations, which are obtained from a flash of
perforation node pressure. the connection composition to average (of the inlet and outlet
In the automated construction of segmented wellbores, the node node) pressures. The treatment of composition and pressure de-
to which each perforation connects is placed at perforation depth, pendence can be either explicit or implicit. We default to explicit
so the wellbore gradient correction appearing in Eq. 7 is zero. treatment of hydrostatic gradients in the wellbore connections. We
Optionally lumping the wellbore (Fig. 3) in effect creates a simple have been completely successful in avoiding convergence prob-
mixed wellbore in the section of the wellbore along the perfora- lems with fully implicit treatment of hydrostatic wellbore gradients
tions. An approximate explicit mixed-wellbore gradient is com- and injected fluid densities only when we shut in crossflowing
puted for each perforation from the sum of phase densities perforations. The problems with implicit treatment arise because of
weighted by their saturations, which are obtained from a flash of the discontinuities in the governing perforation and composition
an average perforated grid-cell composition (weighted by the prod- equations on flow reversals, and also because functions of com-
uct of phase mobilities and wellbore constant) at the node pressure. position are not continuously differentiable in rate as rate changes
Our formulation of the perforation-rate equations offers several sign. Implicit treatment has been achieved in the reviewed ad-
advantages over that in conventional well models. Crossflow oc- vanced transient-well models, but at the expense of timestep
curs automatically, depending on the perforation potential differ- size limitations.
ence, without the need for additional considerations or relation-
ships. Apart from the use of mixed wellbore gradients in lumped Rate and Pressure Constraints. Rate constraints may be speci-
wellbores, wellbore hydraulics is represented as a separate rela- fied on any network connection, excluding the perforations. Pres-
tionship for each connection in the (configurable) wellbore and sure constraints may be specified on any network node. The only
is a function of connection composition and pressure. This pro- restrictions on network constraints, configuration, and flow pat-
vides a relatively simple framework for the inclusion of pressure terns are:
losses caused by friction and acceleration and for representation of • Minimum pressure must be specified at nodes connected to
wellbore phase behavior, complex wellbore geometry, and down- sinks (these are “terminal” nodes), controlled in the sink connection.
hole devices. • Maximum pressure must be specified on nodes connected to
The perforation-rate equations (Eqs. 4 and 8) couple the reser- sources (these are terminal nodes), controlled in the source connection.
voir equations to the rest of the network equations. • Terminal nodes must have exactly one inflow and one out-
flow connection (one to a source or a sink).
Network Rate Balances. The component mass rate balances for • All automatically generated wellbore connections between
each component i at each network node are perforation nodes are allowed to backflow; by default, all others

兺q + 兺Q = 兺q , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (9)
are not. The default may be overridden, except that sink connec-
ij ip ij tions are not allowed to backflow.
j∈In p∈pn j∈On
The terminal-node pressure constraints are always active and
where In and On are lists of the defined inflow and outflow non- applied in the source or sink connection, unless a rate constraint
perforation connections, respectively, at node n. Rates are negative controlled in the source or sink connection would be violated.
for flow in the reverse of the defined direction. The perforation The general form of a volumetric rate constraint controlled in
rates Q are summed over all perforations attached to the node. The connection j is
network rate balances are linear in the primary variables. All other
network equations, except for pressure constraints, surface volu- Qjmax ⱖ Qf 共qជ j,Pjin,Pjout兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (12)
metric rate constraints in black-oil systems, and explicitly treated
wellbore hydraulics equations (gradients), are nonlinear. where Qf is the result of a flash to either surface conditions
(through a specified fixed-condition separator, making the pressure
Hydraulics Equations. The general form of the hydraulics equa- dependence zero) or to in-situ conditions.
tion, applying to each connection j except for sinks, sources, and A minimum pressure constraint controlled in connection j takes
perforations, is the form
Pjin − Pjout = f 共qជ j,P jin,P jout,Vj兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . (10)
Pjin ⱖ Pmin, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (13)
where f is some function of rates, inflow and outflow node pres-
sures, and, if representing an adjustable device, its setting V. Given
and a maximum pressure constraint takes the form
our assumptions, it is not necessary to know the dependence of f on
V. When a device is not being adjusted to achieve a constraint
(when no constraint controlled by the device is violated), Eq. 10 Pjout ⱕ Pmax. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (14)
applies as a primary network equation, with the device at a fixed
operating setting (V⳱Vmax is the secondary equation). Otherwise, Active constraints apply as equalities.

136 April 2004 SPE Reservoir Evaluation & Engineering


Composition Equations. Composition specifications are required
for each network source. For each connection j that is a flow-
ing source, the independent source composition equations that ap-
ply are
冋 Aff Afp
Apf App Apr
Arp Arr
册冋 册 冋 册
␦xf
␦xp
␦xr
=−
Rf
Rp
Rr
, . . . . . . . . . . . . . . . . . . . (19)

qij = zij qTj . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (15a) where r denotes reservoir, and where the terms appearing in Eq. 18
are the values from the network domain solution. The reservoir cell
and 1 ⱕ i ⱕ Nc − 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (15b) mass and pressure coefficients of the perforation-rate equations,
Apr, are added to the perforation-rate equations. These include
where qTj is the total mass rate in connection j. coefficients caused by the implicit treatment of all perforated res-
When a node has more than one outflowing connection, we ervoir cell variables.
assume, pending the addition of inline phase separations, that each The reservoir conservation equation (Eq. 2) residuals (Rr) and
outflowing connection has the same composition. For each pair of coefficients (Arr, Arp) are then built. Arr are the coefficients caused
network connections (including perforations) j1 and j2 flowing out by intercell flow and accumulation. Arp are the coefficients of the
of the same node, the equal composition equations are perforation-rate terms (Qip), and for the rows of component con-
servation equations for a cell, these are identity submatrices for
qij1 qij2
⳱ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (16a) (columns corresponding to) each perforation contained in the cell.
qTj1 qTj2 At this point, the global system of equations has been built and
is ready for the elimination of the secondary reservoir equations
and 1 ⱕ i ⱕ Nc − 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (16b) and variables. For reservoir cells using the IMPES (implicit in
If a node has Nout outflowing connections, there are Nout–1 inde- pressure only) reservoir formulation, the conservation equations
pendent sets of equations of this form. (the secondary equations) are used to eliminate the mass coeffi-
cients of the cell volume constraint, Eq. 3, creating the pressure
Closure of the Network Equations. The number of network equation, and they are also used to eliminate the cell mass coef-
equations and unknowns balance exactly for physically possible ficients of the perforation-rate equations. For fully implicit reser-
flowing systems. This can be shown through careful accounting voir cells, the volume constraint (the secondary equation) is used
and the use of the observation that to eliminate the mass coefficients of the last component (water) in
the reservoir conservation equations and, for cells containing per-
Ninj + Ncons − Nnodes = Ncomp, . . . . . . . . . . . . . . . . . . . . . . . . . . . . (17) forations, in the perforation-rate equations. These linearized and
reduced reservoir, modified perforation, and other network equa-
where Ninj is the number of injecting perforations, Ncons is the tions can also be represented by Eq. 19, with the reduction having
number of nonperforation connections, and Ncomp is the number modified the dimensions and values of Rr, Rp, Arr, Arp, and Apr and
of independent sets of Nc–1 composition equations. “Physically the values of App. The equations are solved at the end of the
possible” means that each node has at least one inflowing and simulator Newton iteration with an unstructured solver, which is
one outflowing connection. When parts of the network are shut described in the Appendix. The reservoir domain (Rr, Arr, xr) may
in, a zero-rate constraint that is otherwise written as a single equa- be divided into subdomains, while the network system is treated as
tion must sometimes be expanded to Nc equations of zero compo- a single domain. Because of the form of the network equations (as
nent rate. we order them), standard iterative solution techniques cannot be
applied to the network domain. We currently apply a direct solve
Solution Method to the network domain using sparse elimination with partial pivoting.
At the beginning of each simulator Newton iteration, mobilities
and densities are computed for each reservoir cell either containing Results
a perforation or being treated implicitly in the reservoir. These are The first example case consists of several runs based on the SPE9
the variables, in addition to the reservoir grid-cell pressures and test case.22 We also present run statistics for a field application. In
compositions, that couple the network equations to the reservoir all cases, we compare results from our new model (Model A) and
equations. The network equations (Eqs. 4 through 16) are then results from a commercial model (Model B). Model B cases,
solved by Newton iteration, holding fixed the current iterate values which include production networks, use the partially coupled
of the reservoir variables. The form of the decoupled primary method of Litvak et al.8
system of network equations is Statistics for all example case runs are given in Table 1, including:

冋 册冋 册 冋 册
• Number of timesteps.
Aff Afp ␦xf Rf
=− , . . . . . . . . . . . . . . . . . . . . . . . . . . . . (18) • Number of global (outer) simulator Newton iterations.
Apf App ␦xp Rp • Number of network domain Newton iterations; these are zero
in Model A for cases decoupled bottomhole.
where f and p denote facility and perforation (facility designates
• Number of well subdomain Newton iterations.
network equations other than the perforation-rate equations), R
• Number of global linear solver iterations.
denotes the residuals, and A (the Jacobian) contains the deriva-
• Total CPU time. All Case 1 runs were made on a 3.056-GHz
tives of the residuals with respect to the variables x. Of those
PC with an Intel Xeon* processor.
equations designated by p, Ninj × Nc–1 of them are network com-
• CPU time in the network domain.
position equations.
• CPU time in the well subdomains (part of network). For
At the start of each network iteration, the variable network
Model B cases, this includes both the reservoir well model and the
equations (Eqs. 10 through 16) are checked to determine which are
network model and is comparable to network domain time for
active. For connections representing adjustable devices, the limit-
Model A.
ing constraint for a given iteration is taken as the constraint that is
• CPU time for linear solves in network (includes well solves,
most violated (i.e., hydraulics, rate, or pressure). Additional con-
also a part of the network domain time).
straint checking is required to detect and prevent overconstrained
• CPU time for global linear solves.
systems resulting from combinations of rate and pressure con-
• Cumulative oil, gas, and water production.
straints. When an overconstrained system is detected, constraints
All Model A runs use explicit wellbore gradients in the well-
are eliminated using estimates of which constraints are limiting.
bore connections and explicit injected-fluid densities. Hydraulics
These estimates must become accurate as convergence of the ap-
tables used in the test cases are available on request.
plied constraints is iteratively approached.
At convergence of the network equations, the fully coupled
system of network and reservoir equations is assembled, which is
represented by * Trademark of Intel, Santa Clara, California.

April 2004 SPE Reservoir Evaluation & Engineering 137


Case 1. We did not attempt to individually optimize run controls heads to gathering nodes), while Model B in effect decouples wells
in the various SPE9 runs. All runs for a given model use the same on rate constraint at bottomhole locations. Shown in Table 1 are a
controls as in Case 1a. lumped Model A run and two runs for Model B, one with default
Case 1a is the standard SPE9 test case, in which the 25 pro- two-iteration network coupling and one with coupling performed
ducing wells and single injection well are controlled by rate and every iteration (referred to here as implicit coupling).
bottomhole-pressure constraints. This is a difficult problem owing Agreement between the models is good in all cases (using the
to its high degree of heterogeneity, highly nonlinear relative per- implicitly coupled Model B network case results), with a maxi-
meability curves, and a step function in capillary pressure. The mum difference of 1.7% in water production in the wellhead case
reservoir formulation is fully implicit. Two runs are shown in (1b). A significant difference in Model B production is observed
Table 1 for Model A, one with segmented wellbores and one with between different levels of coupling in the network case (1c). The
lumped wellbores. implicitly coupled run gives 4.3% more gas and 9% more water
Case 1b modifies Case 1a to include tubing hydraulics in the than the run with two-iteration coupling. This demonstrates the
form of tables from the bottomhole datum location to the well- accuracy of Model A, the potential error with partially itera-
heads for producers. The same table is applied to all production tive partial coupling, and that multiple runs may be required
wells. The injector and all rate constraints are unchanged, while when using partially coupled models to determine the level of
the minimum bottomhole-pressure constraints for producers in coupling needed.
Case 1a are replaced with minimum tubinghead-pressure con- In all cases, Model A efficiency in terms of total Newtons,
straints of 200 psia. Again, we present both lumped- and seg- Newtons per step, and CPU time is significantly better than model
mented-wellbore runs for Model A. B. This is not related to the nature of the domain couplings but is
Case 1c modifies Case 1b to include a gathering network for primarily caused by differences in the reservoir domain formula-
the producing wells, with an identical connection (hydraulics tions, which are not the subject of this paper.
table) from each wellhead to its gathering node. The same table is Relative performance as tubings and gathering networks are
used for all the other surface connections. The four gathering added is similar and is very good for both models. The partial
nodes are connected to a field node, which in turn is connected to coupling in Model B causes no significant degradation of conver-
a terminal node representing a separator. A sink connection is gence rate in this problem.
attached to the terminal node (in Model A only; the Model B For Model A, the addition of well tubings (1b) and gathering
network model terminates with nodes, not connections). The pro- networks (1c) to Case 1a results in network domain Newton itera-
duction network is shown schematically in Fig. 4. The terminal tions, but the well subdomain Newton efficiency (total well do-
node is constrained to a minimum pressure of 100 psia (in Model main iterations per global Newton iteration per well) is not sig-
A, by a device in the sink connection), and all other pressure nificantly affected (1.2 bottomhole, 1.3 wellhead, 1.4 network).
constraints are eliminated. In Model A, the well rate constraints are This is because the decoupled well subdomain equations are linear,
applied at the well choke locations (in the connections from well- except for any composition equations that may apply as the result

138 April 2004 SPE Reservoir Evaluation & Engineering


Fig. 4—SPE9 network configuration.

of crossflow, when explicit gradients and injected-fluid densities wellbores. The Model B network case was run with the default
are used. The well subdomain iterations do not increase by very two-iteration coupling. Of particular interest in this study was the
much when network domain iterations are performed because successful modeling of well crossflow and numerical performance.
within most network domain iterations, the well subnetworks are Cases 2b and 2c are nearly physically identical problems because
converged and do not require iteration. of low pressure drops between the wellhead nodes and the manifold.
For the Model A wellhead case (1b), network domain Newton Table 1 gives cumulative production predictions for the three
efficiency (network Newtons per simulator Newton) is approxi- sets of runs, normalized to the Model B Case 2a results. Differ-
mately 1.5. This is significant because in this case, the network ences between the models are less than 1% in the bottomhole case
domain includes implicit nonlinear tubing hydraulics equations. and approximately 4% in the tubinghead case. While Model A
The addition of a gathering network (1c) causes this figure to rise gives exactly the same production in the wellhead and network
to more than 3, along with a corresponding rise in network-domain cases as expected, Model B shows about a 6% discrepancy in oil
CPU time. This is because of the increased difficulty in solving and gas and a 9% discrepancy in water. Model A results fall in
systems with multiple hydraulics functions in series. between the Model B wellhead and network results. Differences in
In these cases, the Model A network model is generally less the model results are attributed to explicit treatment of tubinghead-
efficient than the conventional well model with or without partially pressure constraints in Model B Case 2b and to the two-iteration
coupled facilities in Model B. While network domain CPU time in partial coupling in Model B Case 2c.
the network case (1c) is approximately the same as that in Model Table 1 clearly shows the improvement in efficiency gained by
B with implicit coupling (2.22 seconds for Model A, 2.27 for the new formulation. While improvements in global Newton effi-
Model B), it is 40% higher as a fraction of total run time. This is ciency will be the subject of a future paper, what is of most interest
primarily because of Model A’s use of the Newton method on a here is the change in outer iterations and CPU time between the
larger system of equations. wellhead case (2b) and the network case (2c). While adding the
Use of a segmented wellbore in Model A results in no differ- network to Model B causes Newton iterations to increase by more
ence in production and virtually no difference in performance over than 40% and CPU time to increase by 50%, these figures are
the lumped case for these problems. Differences in production are virtually unchanged for the new formulation. This is directly at-
expected only when perforation-composition differences affect the tributable to the fully coupled method.
wellbore composition and pressure profile sufficiently for perfo-
ration potentials (buildup or drawdown) to be significantly af- Conclusions
fected, which is not the case here. The well subdomain Newton 1. Complete integration of a facilities model with a conventional
efficiency is almost unaffected by the wellbore treatment. well model yields a fully coupled formulation with configurable
gridded wellbore capabilities.
Case 2. The integrated model was benchmarked on a deep-sea oil 2. The generalized formulation is a highly efficient and accurate
reservoir, modeled with two separate grids coupled through a pro- method for coupling surface facility and reservoir models, and
duction network. Wells in each grid flow into a common back- for conventional simulation of decoupled systems.
pressure manifold node. Hydraulic tables are used to model the 3. Complex preconditioners are not required for fully coupled ef-
pressure drops in each of 25 producers. Maximum liquid rates are ficiency, contrary to previous assertions.
specified at the wells, and a minimum manifold pressure is also 4. Segmented wellbores can be modeled with very little perfor-
specified for the network case. The system is black-oil and fully mance penalty over cases with simple mixed wellbores.
implicit with a total of 56,995 active cells. Pressure maintenance is
Nomenclature
achieved through water injection.
Three sets of runs were made: Case 2a, with minimum bottom- a ⳱ rate of accumulation
hole-pressure constraints specified for individual producers; Case A ⳱ Jacobian matrix
2b, with minimum well tubinghead-pressure constraints specified C ⳱ wellbore constant
for individual producers; and Case 2c, with minimum pressure D ⳱ depth
specified at the manifold node. Model A was run with lumped e ⳱ vector of ones

April 2004 SPE Reservoir Evaluation & Engineering 139


G ⳱ rate of generation 2. Holmes, J.A., Barkve, T., and Lund, O.: “Application of a Multiseg-
I ⳱ defined inflow connections ment Well Model to Simulate Flow in Advanced Wells,” paper SPE
N ⳱ number of 50646 presented at the 1998 SPE European Petroleum Conference, The
O ⳱ defined outflow connections Hague, 20–22 October.
3. Stone, T.W. et al.: “Thermal Simulation with Multisegment Wells,”
P ⳱ pressure
paper SPE 66373 presented at the 2001 SPE Reservoir Simulation
q ⳱ mass flow rate Symposium, Houston, 11–14 February.
Q ⳱ mass or volumetric flow rate 4. Dempsey, J.R. et al.: “An Efficient Method for Evaluating Gas Field
R ⳱ residual Gathering System Design,” JPT (September 1971) 1067.
V ⳱ volume 5. Startzman, R.A. et al.: “Computer Combines Offshore Facilities and
x ⳱ vector of unknowns Reservoir Forecasts,” Petroleum Engineer (May 1977) 65.
z ⳱ mass fraction 6. Emanuel, A.S. and Ranney, J.C.: “Studies of Offshore Reservoir With
␥ ⳱ pressure gradient an Interfaced Reservoir/Piping Network Simulator,” JPT (March 1981)
␭ ⳱ mobility 399.
␳ ⳱ mass density 7. Litvak, M.L. and Darlow, B.L.: “Surface Network and Well Tubing-
head Pressure Constraints in Compositional Simulation,” paper SPE
⌽ ⳱ potential
29125 presented at the 1995 SPE Symposium on Reservoir Simulation,
San Antonio, Texas, 12–15 February.
Subscripts
8. Litvak, M.L. et al.: “Integration of Prudhoe Bay Surface Pipeline Net-
B ⳱ black work and Full Field Reservoir Models,” paper SPE 38885 presented at
c ⳱ components the 1997 SPE Annual Technical Conference and Exhibition, San An-
comp ⳱ composition equations tonio, Texas, 5–8 October.
cons ⳱ nonperforation connections 9. Litvak, M.L and Wang, C.H.: “Integrated Reservoir and Surface Pipe-
f ⳱ facility equations line Network Compositional Simulations,” paper SPE 48859 presented
F ⳱ fluid at the 1998 SPE International Oil and Gas Conference and Exhibition
G ⳱ global in China, Beijing, 2–6 November.
10. Litvak, M.L., Macdonald, C.J., and Darlow, B.L.: “Validation and
i ⳱ component
Automatic Tuning of Integrated Reservoir and Surface Pipeline Net-
inj ⳱ injecting perforations
work Models,” paper SPE 56621 presented at the 1999 SPE Annual
j ⳱ connection Technical Conference and Exhibition, Houston, 3–6 October.
k ⳱ phase 11. Litvak, M.L. et al.: “Prudhoe Bay E-Field Production Optimization
max ⳱ maximum System Based on Integrated Reservoir and Facility Simulation,” paper
min ⳱ minimum SPE 77643 presented at the 2002 SPE Annual Technical Conference
n ⳱ node, network and Exhibition, San Antonio, Texas, 29 September–2 October.
nodes ⳱ nodes 12. Wang, P., Litvak, M.L., and Aziz, K.: “Optimization of Production
out ⳱ outflowing connections Operations in Petroleum Fields,” paper SPE 77658 presented at the
p ⳱ perforation, perforation equations 2002 SPE Annual Technical Conference and Exhibition, San Antonio,
Texas, 29 September–2 October.
P ⳱ pore
13. Schiozer, D.J. and Aziz, K.: “Use of Domain Decomposition for Si-
perfs ⳱ perforations
multaneous Simulation of Reservoir and Surface Facilities,” paper SPE
phases ⳱ phases 27876 presented at the 1994 SPE Western Regional Meeting, Long
r ⳱ reservoir grid cell, or reservoir domain Beach, California, 23–25 March.
R ⳱ red 14. Byer, T.J., Edwards, M.G., and Aziz, K.: “Preconditioned Newton
Schur ⳱ Schur complement Methods for Fully Coupled Reservoir and Surface Facility Models,”
T ⳱ total paper SPE 49001 prepared for presentation at the 1998 SPE Annual
w ⳱ well Technical Conference and Exhibition, New Orleans, 27–30 September.
15. Byer, T.J., Edwards, M.G., and Aziz, K: “A Preconditioned Adaptive
Superscripts Implicit Method for Reservoirs With Surface Facilities,” paper SPE
in ⳱ inflow 51895 presented at the 1999 SPE Reservoir Simulation Symposium,
Houston, 14–17 February.
inj ⳱ injected
16. Byer, T.J.: “Preconditioned Newton Methods for Simulation of Reser-
min ⳱ minimum voirs with Surface Facilities,” PhD dissertation, Stanford U. (2000).
max ⳱ maximum 17. Hepguler, G., Barua, S., and Bard, W.: “Integration of a Field Surface
out ⳱ outflow and Production Network With a Reservoir Simulator,” SPECA (June
0 ⳱ initial 1997) 88.
18. Tingas, J., Frimpong, R., and Liou, J.: “Integrated Reservoir and Sur-
Overbars face Network Simulation in Reservoir Management of Southern North
A single tilde, as in Ã, denotes an array or vector after elimination Sea Gas Reservoirs,” paper SPE 50635 presented at the 1998 SPE
of the red unknowns. A double tilde, as in Ø, denotes an array or European Petroleum Conference, The Hague, 20–22 October.
vector after elimination of the red and network unknowns. 19. Trick, M.D.: “A Different Approach to Coupling a Reservoir Simulator
With a Surface Facilities Model,” paper SPE 40001 presented at the
Acknowledgments 1998 SPE Gas Technology Symposium, Calgary, 15–18 March.
We gratefully acknowledge significant contributions made by 20. Shiralkar, G.S. et al.: “Falcon: A Production Quality Distributed
Terry Wong, Dominic Camilleri, Richard Hinkley, Ron Moss- Memory Reservoir Simulator,” SPEREE (October 1998) 400.
barger, and Shahin Abasov. We also thank Landmark Graphics for 21. Peaceman, D.W.: “Interpretation of Well-Block Pressures in Numerical
permission to publish this work and BP for allowing us to present Reservoir Simulation with Nonsquare Grid Blocks and Anisotropic
results of their field application. Permeability,” SPEJ (June 1983) 531.
22. Killough, J.E.: “Ninth SPE Comparative Solution Project: A Re-
References examination of Black Oil Simulation,” paper SPE 29110 presented at
1. Stone, T.W., Edmunds, N.R., and Kristoff, B.J.: “A Comprehensive the 1995 SPE Symposium on Reservoir Simulation, San Antonio,
Wellbore/Reservoir Simulator,” paper SPE 18419 presented at the 1989 Texas, 12–15 February.
SPE Symposium on Reservoir Simulation, Houston, 6–8 February. 23. Appleyard, J.R.: “Nested Factorization,” paper SPE 12264 presented at

140 April 2004 SPE Reservoir Evaluation & Engineering


the 1983 SPE Reservoir Simulation Symposium, San Francisco, 15–18 We next eliminate the red unknowns from Eq. A-4, yielding the
November. reduced matrix equation, which is of the form

冋 册冋 册 冋 册
24. Wallis, J.R.: “Incomplete Gaussian Elimination as a Preconditioning
for Generalized Conjugate Gradient Acceleration,” paper SPE 12265 Ãnn ÃnB xn −R̃0n
= , . . . . . . . . . . . . . . . . . . . . . . . . . . (A-5)
presented at the 1983 SPE Reservoir Simulation Symposium, San Fran- ÃBn ÃBB xB −R̃B0
cisco, 15–18 November.
25. D’Azevedo, E., Forsyth, P.A., and Tang, W.-P.: “An Automatic Or- where
dering Method for Incomplete Factorization Iterative Solvers,” paper −1
Ãnn = Ann − AnR ARR ARn, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6a)
SPE 21226 presented at the 1991 SPE Reservoir Simulation Sympo-
−1
sium, Anaheim, California, 17–20 February. ÃBn = ABn − ABR ARR ARn, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6b)
26. D’Azevedo, E.F., Forsyth, P.A., and Tang, W.-P.: “Towards a Cost-
−1
Effective ILU Preconditioner with High Level Fill,” BIT (1992) 32, No. ÃnB = AnB − AnR ARR ARB, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6c)
3, 442. −1
27. Saad, Jousef: Iterative Methods for Sparse Linear Systems, second ÃBB = ABB − ABR ARR ARB, . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6d)
edition, Soc. of Industrial and Applied Mathematics, Philadelphia, −1 0
R̃0n = R0n − AnR ARR RR, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6e)
Pennsylvania (2003).
−1 0
and R̃B0 = RB0 − ABR ARR RR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-6f)
Appendix—The Linear Equation Solver
At each Newton iteration, it is necessary to solve the following We actually solve the reduced matrix equation, Eq. A-5, rather
global matrix equation. than the original global matrix equation, Eq. A-1. We obtain xR

冋 册冋 册 冋 册
by backsolving.
Ann Anr xn −R0n To solve Eq. A-5, we construct an approximate factorization of
= . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-1) it using the nested factorization column sum principle.23 The form
Arn Arr xr −Rr0 of the factorization is
Ann and Anr contain the network equation coefficients that multiply
network and reservoir unknowns, respectively. Arn and Arr contain
the reservoir equation coefficients that multiply these same un-
knowns; xn and xr are the network and reservoir unknown vectors,
冋 册 冋 册冋
Ãnn ÃnB
ÃBn ÃBB

Ãnn 0
ÃBn ØBB
I Ã−1
0
nn ÃnB
I

, . . . . . . . . . . . . . . . (A-7)

respectively. The network unknowns are pressures at network where ØBB = ÃBB − diag共ÃBnÃ−1
nn ÃnBe 兲; . . . . . . . . . . . . . . . . . . . . . (A-8)
nodes and component rates on network connections, including
perforations. The reservoir unknowns are cell pressures if IMPES e is the column vector having all of its entries equal to one, and
is being used and cell pressures and masses if the calculations are diag(x) is the diagonal matrix having on its diagonal the entries of
fully implicit. R0n and Rr0 are the initial residual vectors. the column vector x.
To solve the global matrix equation, it is necessary to solve We need a preconditioner for ØBB. If the reservoir equations are
subproblems of the form IMPES, we need a preconditioner for the IMPES pressure equa-
tion. If they are fully implicit, we use Wallis’ constrained pressure
Ann␦xn = −Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-2) residual (CPR) method,24 which also requires a preconditioner for
the pressure equation. In either event, we need a preconditioner for
and Arr␦xr = −Rr , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (A-3)
the pressure equation; for this, we use incomplete lower triangular-
where Rn and Rr are residual vectors. Eq. A-2 is equivalent to Eq. upper triangular (ILU) factorization. To improve performance of
19. To solve Eq. A-2, we use sparse Gaussian elimination with row the preconditioner, we do two things:
pivoting. To decrease the cost of solving Eq. A-3, we perform a • Reorder the unknowns and equations in a way that reduces
red/black reduction. We determine the cell colors as follows. the amount of infill. Our procedure is similar to but simpler than
1. Designate all cells to be uncolored. that described by D’Azevedo et al.25,26 In deciding which un-
2. Designate the first cell as red. This creates the first group of known to eliminate next, we determine the infill that would be
red cells. generated by each possibility and eliminate the unknown that gen-
3. Find all uncolored cells connected to the latest group of red erates the smallest infill. Here, infill is computed as the sum of the
cells. Designate these to be black. magnitudes of all infill coefficients that would be introduced into
4. Find all uncolored cells connected to the latest group of the matrix during the elimination of this particular unknown.
black cells (from Step 3). These are candidates to be red. Designate • Determine which infill coefficients to retain on the basis of
the first one to be red. Cell by cell within this group, determine their sizes. The procedure is similar to Saad’s ILUT (ILU that
whether each cell is connected to any red cell. If it is not, designate discards infill larger than a specified threshold).27 We retain a
it to be red. If it is connected to a red cell, designate it to be black. possible infill coefficient if its magnitude is larger than 0.007 times
The red cells identified in this step become the next group of the magnitude of the diagonal coefficient in the row in which it falls.
red cells. The process of reordering the unknowns and determining
5. Repeat Steps 3 and 4 until no connected uncolored cells can which potential infill to retain is quite expensive computationally.
be found. As a result, we create a template containing the information needed
6. Search all cells for an uncolored cell. If one is found, des- to perform and use the factorization. The template is constructed at
ignate it to be red, and return to Step 3. the beginning of a run on the basis of the matrix coefficients for the
On a structured grid, this procedure yields the normal check- first Newton iteration of the first timestep. It can be updated from
erboard pattern. The procedure results in two lists, one containing time to time later in the run if needed.
the red cells and the other containing the black cells. Given these
lists, we can partition Arr and rewrite Eq. A-1 as follows.
SI Metric Conversion Factors

冤 冥冤 冥 冤 冥
ARR ARn ARB xR −RR0
bbl × 1.589 873 E – 01 ⳱ m3
AnR Ann AnB xn = −R0n . . . . . . . . . . . . . . . . . . . . (A-4)
ft3 × 2.831 685 E – 02 ⳱ m3
ABR ABn ABB xB −RB0 psi × 6.894 757 E + 00 ⳱ kPa
The R and B subscripts denote red and black cells, respectively.
ARR is diagonal for the IMPES pressure equation and block diago- Brian K. Coats is a senior technical advisor with Landmark
nal for the fully implicit set of equations. As a result, it can be Graphics, where he continues his work in reservoir simulation
inverted economically. (which began at J.S. Nolen and Assocs. and continued at

April 2004 SPE Reservoir Evaluation & Engineering 141


Western Atlas Software) in development of the VIP family of Editorial Committee, the SPE Annual Meeting Program Com-
reservoir simulators. e-mail: bcoats@lgc.com. Coats is a princi- mittee, the Cedric K. Ferguson Medal Award Committee, and
pal author of VIP-THERM and has authored several technical the SPE Journal Editorial Board. In addition, he has served as a
papers. He holds BS and MS degrees in chemical engineering member and chairman of program committees for SPE Reser-
from the U. of Texas at Austin. Graham C. Fleming is a senior voir Simulation symposia; a member of the Monograph Com-
technical advisor for Landmark Graphics Corp. in Houston. e- mittee; and a Technical Editor and Review Chairman for SPE
mail: gfleming@lgc.com. He is the development team leader Computer Applications. He gave the keynote address at the
for Landmark’s next-generation simulator project. Before join- 1997 SPE Reservoir Simulation Symposium. Marcelo Ramé is a
ing Landmark, Fleming worked 18 years in various positions for technical advisor with Landmark Graphics, where he works in
Arco Exploration and Production Technology, most recently as the area of iterative linear solvers for reservoir simulation. e-
Director of Reservoir Simulator Development. He holds a BE mail: mrame@lgc.com. Previously, he worked for the Center
degree in engineering science from the U. of Auckland and a for Research in Parallel Computing, a Natl. Science Founda-
PhD degree in applied mechanics from the California Inst. of tion Center located on the Rice U. campus. Ramé holds a BS
Technology. J.W. Watts is a consultant working under contract degree from the U. of Buenos Aires, an MS degree from Lehigh
to Landmark Graphics and ExxonMobil Upstream Research U., and a PhD degree from the U. of Houston, all in chemical
Co. in Houston. e-mail: bwatts@lgc.com. Before becoming a engineering. Gautam S. Shiralkar is a consulting engineer with
consultant, he was with Mobil Research and Development BP’s Technology group in Houston. e-mail: shiralgs@bp.com.
Corp., Scientific Software Corp., and Exxon Production Re- He holds a B.Tech degree from the Indian Inst. of Technology,
search Co. Watts holds BS and PhD degrees in chemical en- an MS degree from MIT, and a PhD degree from the U. of
gineering from Tulane U. He has been a member of the JPT California, Berkeley, all in mechanical engineering.

142 April 2004 SPE Reservoir Evaluation & Engineering

Вам также может понравиться