You are on page 1of 15

SPE-181685-MS

Estimation of Inter-Well Connections in Waterflood under Uncertainty for


Application to Continuous Waterflood Optimization of Large Middle-Eastern
Carbonate Reservoirs

Xiang Zhai, Tailai Wen, and Sebastien Matringe, Quantum Reservoir Impact

Copyright 2016, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in Dubai, UAE, 26-28 September 2016.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
The management of large and mature waterfloods is a notoriously challenging exercise. The vast amount
of data available usually cripples reservoir simulation efforts and operational teams usually revert to simple
classical engineering calculations, diagnostics plots and maps to make their decisions. Some powerful
technologies based on reduced-physics modeling have been developed over the past decade to address this
issue. In this paper, we present one such approach that was designed for the management of a large Middle-
East carbonate waterfloods.
The reservoir model used is based on the Surveillance Model proposed by Batycky et al. (2008) but
differs from it in two aspects: the inter-well allocation factors are computed through the solution of a tracer
equation rather than through streamline computations and the fractional flow behavior is estimated through
an empirical model rather than computed numerically. Using the tracer allows an improved treatment of
unstructured grids and dual-porosity systems, both features being important for the application of interest.
Modifying the fractional flow model allows for the automation of the history-matching step. The model can
thus integrate new data quickly and estimate the strength and efficiency of each inter-well connection.
An optimization algorithm is used to translate the reservoir management strategy of the asset team in
terms of an objective function and a series of constraints at the well, well-group or facility level. Constraints
such as voidage replacement ratios, surface facility limits, fracturing pressures can be integrated into the
optimization engine to control the field.
A new uncertainty modeling process uses a Markov-Chain Monte-Carlo algorithm to evaluate the
robustness of each recommended change. The less mature or less data-rich areas of the field are typically
harder to calibrate and more uncertain. Decisions to change the rate of a producer or injector in those
areas are more risky. The algorithm is able to quantify this risk to help the operator make a more informed
decision. As the field gains in maturity, the algorithm shows how the model learns with new data and how
the proposed decisions continuously gain in robustness.
The application of the methodology to giant Middle-East carbonate fields is discussed. The proposed
methodology was able to integrate all relevant facility, well group, individual well and reservoir constraints
but remains fast enough to be run daily as new data becomes available.
2 SPE-181685-MS

Introduction
The key to the successful exploitation of secondary recovery projects is the efficient management of the
injected and produced fluids. The oil production should be maximized while minimizing the production
of unwanted fluids (e.g. water and gas). Injection rates should be controlled to provide the right amount
of pressure support without driving fast breakthroughs or excessive water cycling. Asset teams routinely
review the current production and injection data in order to define adjustments to the secondary recovery
strategy that would increase the reservoir performance. The objective of the exercise is usually focused on
increasing production and reserves or minimizing operating expenses.
A myriad of approaches have been used in the past with different levels of success to support Production
Optimization decisions. Most asset teams rely on basic heuristics, simple classical engineering calculations,
diagnostics plots or maps to make their decisions. Such approaches usually require a significant amount of
experience to design and of manual work to perform. Some teams customize spreadsheets to help clarify
the workflows and accelerate the analyses.
A more rigorous solution relying on reservoir simulation is favored by a portion of the academic and
corporate research community. The advocates of the approach have performed a significant amount of
research to deliver new technologies to decrease the time and effort required to build, calibrate and optimize
a reservoir simulation model. However, in field applications, the size and complexity of the datasets and
the short time frames allocated to make decisions often preclude teams from using the simulation-based
approaches. For the moment, so-called closed-loop waterflood management projects remain at the pilot
stage for the most part.

Reduced-Physics Models
In recent years, a new class of approaches has been investigated that tries to strike a balance between the
data-driven and physics-based extremes of heuristics and simulation. Two main category of methods have
been proposed that have demonstrated positive results in field applications: the Capacitance-Resistance
models (CR) (e.g. Sayarpour, 2008) and the Surveillance models (SRV) (Thiele and Batycky, 2006). Other
ideas have been proposed but have remained at a more prototypical stage (Albertoni, 2002, Albertoni and
Lake, 2003, Gentil, 2005, Liu, 2007, Lee, 2010).
CR models rely on an analogy between the equations governing the flow of fluids in porous media and
those describing the behavior of an electrical system composed of a capacitor and resistor set up in series.
First introduced in the 1940's as an analogue reservoir simulation technology (Bruce, 1943), CR methods
have evolved until the 1960's where they were successfully used for history-matching and predictions (Wahl
et al., 1962). The technology was then phased out in favor of cheaper numerical methods until a revived
interest in the 2000's (Yousef, 2005, Yousef et al., 2006a and 2006b, Liang et al., 2007, Sayarpour, 2008,
Weber, 2009, Izgec and Kabir, 2009). CR models are designed as a quick way of modeling large secondary
projects as a network of wells. The wells are considered to be interacting through subsurface connections
that are viewed as a simplification of the flow paths between injectors and producers. Several other academic
and professional teams have since then investigated the subject further and are now using CR models as
an alternative to reservoir simulation models for complex reservoirs with long production and injection
histories. CR models have been used on clastic and carbonate reservoirs, fractured or not, to optimize a
variety of reservoir management strategies ranging from waterflood to gas injection and including steam
injection.
SRV models were proposed by Thiele and Batycky (2006) as a simplification of the streamline method
geared toward understanding and improving the efficiency of a waterflood project. The methodology post-
processes streamline information to represent the reservoir as a connected network of inter-well connections.
The efficiency of each connection is quantified and the information is used to optimize the production and
injection targets of each well in the reservoir. The technique has been applied successfully in a number of
SPE-181685-MS 3

fields and has recently been complemented with a more rigorous set of optimization approaches (Wen et
al., 2012, Wen, 2014).

Proposed approach
The approach presented in this paper has been heavily influenced by both CR and SRV models but a few
key modifications were made to alleviate some of the limitations that the authors faced when trying to
apply these models to large Middle-East carbonate reservoirs. For example, classic CR models assume that
interwell connections and drainage volumes are not time dependent. This assumption was not applicable for
some of the reservoirs under study, which were under active development. This drove the team to choose
the SRV workflow as a basis, but specific elements had to be modified. The dual permeability nature of
the reservoirs of interest made it hard to justify using streamlines which only account for a single flowing
medium. In the proposed approach, streamlines have been replaced by numerical tracer calculations to
remove this limitation. This modification also allowed for unstructured grids to be leveraged to provide a
very refined grid around wells while maintaining a reasonable number of grid cells. Removing streamlines
from SRV models means that a new method was required to estimate the efficiency of connections. An
empirical fractional flow model is borrowed from CR models that allows the proposed approach to feature
a fully-automated history matching algorithm.
The model obtained remains a reduced-physics approach, where critical aspects of the reservoir flow and
transport problems were simplified to gain speed. The proposed methodology has been tested in several
fields. Models can be built from scratch in under a week, can be updated with fresh data in under an hour
and can be run in a few seconds. This acceleration is due to the fact that the model is heavily data-driven.
The history-matching is now performed automatically and with virtually no human intervention.
The model is completed with an optimization engine borrowed directly from Wen et al. (2012, Wen, 2014)
and designed to provide the optimal set of operational changes to be performed on a well by well basis and
considering the strategic objective of the reservoir management strategy and the current field constraints.
Finally, a new strategy is proposed to quantify the uncertainty range of calculated connection efficiencies
based on the amount and quality of data available. Understanding these uncertainties, allows the team to
only alter the Reservoir Management strategy for wells that are well "understood" by the model.
The remainder of this paper will be organized as follows. The proposed reservoir modeling approach
will first be described in details to provide its theoretical background and to highlight its similarities
and differences with existing methods. Next, the production optimization methodology will be presented,
followed by a section explaining how the model quantifies uncertainty. The paper will conclude with some
synthetic and applied examples.

Proposed Reservoir Modeling Approach


General Workflow
During waterflood operations, a variety of data is routinely collected to analyze, diagnose and optimize the
current performance of the field. Well rates, production ratios, pressures, logs, fluid samples are all examples
of measurements that are regularly collected to help the Reservoir Management team improve the behavior
of the field (Kikani, 2013). The modeling approach proposed in this paper makes use of this standard dataset
in the most direct way possible, to make data integration as fast and as seamless as possible.
The model requires two general categories of data. A reservoir characterization dataset should be
provided, in terms of a raw dataset (e.g. logs, cores, geologic surfaces), a predeveloped geologic model
or an existing simulation model. Production data should also be integrated (e.g. production and injection
rates, pressures, production logs, chromatography, fluid samples or tracer data). The geologic information
is integrated as part of a static model, similar to one that would be used in Reservoir Simulation while the
4 SPE-181685-MS

production data is typically expected to be updated monthly or weekly, although in some applications, it
has been updated daily.
The geologic model is represented as a numerical grid on which rock and fluid properties are defined.
The model is completed with well information, such as trajectories and completion data. The geologic
model represents the best current understanding of the underlying geology of the field. Depending on the
application, this could be a fully detailed three-dimensional geostatistical model, in other cases, simpler
models are being used that are two-dimensional or "2.5D" (stacked two-dimensional models). When
available, a probabilistic description of the geology can be incorporated in the process.
The grid used is either provided by the input model or generated during the process. Often, geologic or
simulation models do not have sufficient accuracy to properly model the complex well or reservoir geometry
and a new grid needs to be defined. When such a need arises, a well-centered unstructured gridding algorithm
similar to the one described in Houzé et al. (2013) is used to develop a grid that is refined around the
wells and coarsened away from them. Creating the model grid has been found to be a critical step for this
algorithm. The grid should be coarse enough to maintain speed but refined enough around the wells to
guarantee that the strength of inter-well connections is calculated accurately.
Once the grid created, the workflow starts a recurrent loop over a pre-defined set of time-step. The time-
step size should be chosen to be long enough for pressure to equilibrate within the reservoir. For most
application, the process is repeated on a monthly basis.
For each time-step, the algorithm starts by calculating the pressure distribution within the reservoir using
the stationary pressure equation. Next, a numerical tracer computation is used to determine the volume of
influence of each well and that information is post-process to compute a well-allocation factor between each
producer and injector. This can be understood as the strength of each inter-well connection.
Next, the efficiency, or oil-water ratio, of each connection is determined by calibrating a fractional flow
model to each connection. This defines the production water cut as a function of the cumulative water
injected in a connection. Each fractional flow function is parameterized with coefficients that are adjusted
to provide an accurate match to the production data.
The end result of the workflow is thus a determination of the strength and efficiency of each connection
between wells. The remainder of this section covers the details of each step.

Strength of Inter-Well Connections


Stationary Pressure Equation. To determine the strength of interwell connections, the approach proposed
by Shahvali et al. (2011) is used. For a single-phase system, the pressure equation reads:

(1)

where ρ and μ are the fluid density and viscosity, φ and K are the rock porosity and permeability, p is the
fluid pressure and q is a volumetric source or sink term, which essentially represents the production or
injection rate from wells. The equation is usually completed with boundary conditions in either pressure or
rate. Extending this equation to multi-phase problems is a standard Reservoir Simulation exercise, which is
used in the IMPES formulation (Aziz and Settari, 1979). The single-phase notation is maintained for clarity
but the procedure described in this paper does not have to assume a single-phase problem.
The stationary form of this equation describes the flow problem in a steady-state situation. The stationary
equation is obtained by neglecting the time-dependent accumulation term in the pressure equation to yield:

(2)
SPE-181685-MS 5

This pressure equation is solved using a classic finite volume discretization with a two-point flux
approximation. For grids with complex geometries or when the permeability anisotropy is not aligned with
the principal grid directions, a multi-point flux approximation can be used to improve the accuracy of the
calculation. Such algorithms are sometimes required in practice when an aggressive upscaling is performed
in some portions of the reservoir with an underlying anisotropic permeability field.
The stationary pressure equation is solved for by enforcing the average liquid production and injection
rates at each producer and injector over a given time-step as source terms. The equation assumes an
incompressible system and requires compatible boundary conditions. Most of the time, the production and
injection rates in the field are not equal. During that time, the pressure support is modeled as being provided
by aquifer support. The boundary can either be set up as an analytical or numerical aquifer or a series of
source terms can be distributed along the aquifer boundary with a rate corresponding to the net production-
injection voidage rate out of the reservoir.
The approach proposed delivers an estimated pressure distribution in the field at each time-step, without
the need of a pressure matching exercise. In practice, the reservoir and aquifer were modeled together so
that the aquifer influence was modeled through distributed source terms. The pressure field was thus only
defined up to a constant. The pressure distribution reflects the current production and injection rates as
well as the assumptions made to create the geologic model. For most production optimization applications,
knowing the pressure distribution up to a constant is sufficient. For other applications, such as pressure
mapping or if pressure data should be matched, the pressure constant can be related to the average pressure
of the reservoir, which can be determine by material balance (Dake, 1978). The proposed approach thus
makes it possible to adapt the level of effort placed in the history-matching process to the required objective,
i.e. matching or not the reservoir pressure information.
Numerical Tracer Equation. To understand the transport characteristics of the pressure field, the numerical
tracer equation is solved. The stationary tracer concentration equation for a single-phase incompressible
system reads
(3)
where u is the Darcy velocity of the fluid, and C is the tracer concentration. The Darcy velocity is simply
obtained by post-processing the pressure solution using the discretized version of Darcy's law. So the
determination of the tracer concentration for any set of boundary conditions is a relatively inexpensive
computation.
For each well in the reservoir, the tracer equation is solved with boundary conditions set to understand
the volume of influence of that well. A tracer concentration of 1 is set at the well of interest and a value of
0 is enforced in all other wells of the same type. For injectors, the tracer concentration is solved directly.
For producers, the velocity field is first inverted, for the tracer to follow an upstream direction.
In this and following steps, the aquifer is treated as an injector when the net voidage rate is positive
and as a producer otherwise. In the remainder of this paper, the treatment of the aquifer will not be further
discussed, as it is understood that mathematically, in our formulation it is no different than any other well.
The solution of the tracer equation provides an estimate of the reservoir volume that is hydraulically
connected to each well, i.e. the volume receiving pressure support from a given injector or the drainage
volume of a given producer. Combining the tracer concentrations from an injector and a producer yields
the volume swept by the well pair.
The tracer concentrations directly provide the well allocation factor defined by Thiele and Batycky (2006)
in the context of streamlines or referred to as connection factors in CR models

(4)
6 SPE-181685-MS

where the subscript i and j refer to an injector and a producer, respectively and q designates a volumetric
rate. The well allocation factor is thus the ratio of liquid rate from an injector to a producer to the total liquid
rate of that injector. A similar definition of the well allocation factor can be obtained for each producer by
computing the ratio between the rate of a connection to the total rate of the producer. The well allocation
factor can be understood as the strength of the connection between two wells.
Figure 1 shows an example visualization of the well allocation factor obtained for a well in a giant Middle-
East carbonate reservoir. The scales have been changed for confidentiality. In the early stages of production
the reservoir is under primary depletion and most of the pressure support comes from the aquifer. Over
time, neighboring wells are converted to injectors and the well starts to receive support from these nearby
injectors.

Figure 1—Well allocation factor vs. time for a producer, showing the contribution of the
aquifer during the primary depletion phase and of the various reservoir injectors after that.

At this point, the algorithm proposed follows closely the workflow of Thiele and Batycky (2006) but
replaces the streamline-based computation of interwell connections with a tracer-based computation as
described in Shahvali et al. (2011). This offers two key advantages that derive from the fact that the tracer
equation is solved using a similar finite volume method as for the pressure equation. First, the tracer
calculations can be performed on complex grids with anisotropic permeability tensors, which is a possible
but cumbersome exercise for streamline methods. This improvement was important as the large mature
fields under study contained hundreds of wells of various geometry. A grid was needed that had sufficient
accuracy around vertical and horizontal wells and that remained of a manageable size so that it could be run
in seconds. A second advantage of the tracer formulation is that it is readily extensible for dual permeability
systems. In the fields of application, the current reservoir understanding called for a dual-permeability
description of the carbonate rock. Including such effects is an area of current research in reservoir simulation
but is a straight forward exercise for tracer calculations.

Efficiency of Inter-Well Connections


The most naïve approach to waterflood management, especially popular in pattern floods, consists in
balancing each pattern so that each injector delivers the expected amount of pressure support to each
SPE-181685-MS 7

producer. This approach is valid when no further information is provided but leaves behind a lot of
opportunity.
Going a step further, the efficiency of each connection should be taken into account to decide how to
adjust the injection rates of a waterflood. Thiele and Batycky (2006) define the connection efficiency as the
ratio of oil produced to water injected in a given connection and propose to compute the ratio by solving
the transport problem along the streamlines. Lake et al. (2007) and the CR community instead rely on
an empirical fractional flow model. The model parameters of each connection are matched based on the
production history.
The proposed methodology uses the approach proposed by the CR models to compute the efficiency of
interwell connections. Two main reasons drove this decision. First, since the strength of connections was
computed using tracers and not streamlines, the more rigorous approach of SRV models could not be applied
directly. Second, the use of empirical models made it possible to simplify dramatically the history-matching
exercise and reduced the problem to a simple parameter estimation exercise.
For each interwell connection, a fractional flow function is thus defined as a parameterized function Gnij.
Several functions have been tested and good results were obtained with the functions presented in Sayarpour
(2008). The function currently used in most models reads:

(5)

where aij and dij are the fractional flow model parameters and WICnij is the cumulative water injection from
injector i to producer j from t0 to tn:

(6)

Physically speaking, dij can be understood as the cumulative volume of water injected to reach a
connection efficiency of 50% and 1/aij is proportional to the volume of water that should be injected to
fully sweep the connection. Other definitions of Gnij have been used successfully in field applications. The
functions can be adapted to properly field a specific reservoir. Since these functions are parameterized,
they need to be calibrated against historical production data before being used. To do so, the contributions
are combined to yield a calculated fractional flow at each producer. By summing the contribution of each
injector, the oil fraction fno,j of producer j at the th time step is expressed as

(7)

where the notation that i = 0 is used to describe the aquifer. The oil fraction can be used to evaluate the oil
production rate through a simple multiplication with the total fluid production rate of a well:

(8)

where qno,j and qnj are the oil and total fluid rates of producer j and the th time step, respectively.
The history-matching process is then performed by applying an optimization algorithm to minimize
the mismatch between the model and historical production. Depending on the fractional flow model used,
several optimization approaches have been used, ranging from basic pattern search to gradient-based
8 SPE-181685-MS

methods. The adjoint formulation usually outperforms other approaches in terms of efficiency but can
require cumbersome derivations.
Figure 2 proposes an example of the efficiency of a connection computed between a producer and an
injector. The efficiency decreases over time as the amount of water delivered to the connection by the
injector increases.

Figure 2—Efficiency of a producer-injector connection vs. time, showing how the


efficiency deteriorates over time as more liquid is injected through the connection.

Once calibrated, the model has essentially simplified the subsurface model into a series of wells
interacting through connections described in terms of their strength, or well allocation factor and efficiency,
or fractional flow model. Figure 3 proposes a view of the various injectors connected to a producer of
interest. The display describes the connection to each injector with a bubble plot. The diameter of the bubble
describes the relative strength of the connection while the ratio displayed by the blue and green pie chart
represents the efficiency of the connection, or water and oil fractional flow, respectively.

Figure 3—Display of the various injectors connected to a producer of interest. The producer is displayed
as the red ang green dot and the injectors are designated with bubble plots. The size of the bubble
indicates the strength of the connection and the efficiency is represented by the colors in the pie charts.
SPE-181685-MS 9

Production Optimization
The reservoir model presented in the previous section characterizes the reservoir as a set of interwell
connections described by their strength and efficiency. The strength of each connection changes over time
and depends on the underlying geology as well as the production and injection rates of neighboring wells.
The efficiency of each connection is also a function of the underlying geology and varies over time as the
volume of water injected in the connection increases. The model was designed to be built and matched
quickly so that it could be used to support Production Optimization efforts.
The waterflood optimization exercise requires a forecast of the reservoir behavior subject to operational
changes. To perform a prediction, the calibrated model is run forward for the next time step. The liquid
production or injection rate of each well is defined and enforced to solve the stationary pressure and tracer
equations to determine the new well allocation factors. The fractional flow models are then updated to
determine the expected oil and water production rates at each producer.
The reservoir model can thus estimate the oil and water production behavior for each producer given the
liquid rate enforce in each well. The decision of setting the liquid rates of wells as control parameters was
driven by the mature fields under study. In these reservoirs, the surface and wellbore production systems
allowed the operators to define their target liquid rates for each well. Similarly, the rate of each injector
could be adjusted.
The Production Optimization problem is then formulated in a form similar to that proposed by Wen and
collaborators (Wen et al., 2012, 2014, Wen, 2014). The control variables are the target liquid production
and injection rates of each well. The constraints are set to limit these control variables and the objective
function can be designed to affect either the oil or water production rates.
A variety of constraints can be enforced on these control variables. For producers, a maximum liquid rate
can be set to avoid coning or to reflect the limitations of the artificial lift system. For injectors, a maximum
rate can be defined to avoid formation fracturing. Surface facilities have limitations that can be enforced on
groups of wells, such as maximum liquid, oil, gas or water handling or maximum/minimum water injection
volumes. More subtle constraints can also be developed to imprint a Reservoir Management logic into the
optimization algorithm. For example a voidage replacement ratio can be set as a target for a group of wells,
a fault block, a geologic layer or the whole reservoir. Additional constraints that were used in practice
were related to the amount of change that was allowed on a monthly basis. For example, the maximum
liquid rate of a producer can be set to increase only up to 10% of its past value, this essentially dampens
the operational changes that are recommended by the algorithm. Similarly, limits can be set in terms of
how many operational changes are allowed in a month; for example, if a reservoir contains 100 wells, the
operator might only want to perform 10 operational changes (producer or injector adjustments) instead of
altering the rate of every well. When naïve constraints are first enforced, surprising or unreasonable cases
tend to be proposed by the optimization engine. The team then needs to think about what constraints are
missing that should be enforced to yield a more realistic or reasonable case.
In terms of objective functions, two main classes have been tested: increase oil production and decrease
water production. Since the optimization is mainly geared toward improving the efficiency of a waterflood,
the liquid rates are pre-defined and the optimization focuses on how to best distribute the injected water to
either maximize the recovery or reach a recovery target while minimizing the water production.

Solution of the Optimization Problem


The optimization problem is solved directly using a commercial numerical optimization algorithm. During
the solution, the time required to evaluate the gradient of the problem is the main computational bottleneck.
To accelerate the process, at each optimization iteration, the reservoir model is linearized by approximating
the gradients in a fast manner. Small perturbations of the solution are assumed so that the strength and
efficiency of the interwell connections can be fixed. The change in oil or water production is therefore
10 SPE-181685-MS

written as the change in water injection at a connected well multiplied by the associated well allocation
factor and connection efficiency. This approach was initially proposed by Wen and collaborators (Wen et
al., 2012, 2014, Wen, 2014) in the context of streamline-based production optimization but can be used
directly in this approach. Details on the formulation and solution of the optimization problem can be found
in the referenced literature.

Uncertainty Quantification
The calibration of the empirical fractional flow model is based on the observed production history. When a
give connection has received a limited volume of cumulative injection, a significant amount of uncertainty
might remain on the efficiency of that connection. This occurs either for fairly new connections or for
connections that have existed for a longer period of time but that have been characterized by a low strength.
In either case, the total volume of water injected within that connection is small and the amount of water
produced by that connection might be hard to detect at a given producer, as it is typically masked by the
contribution of stronger connections.
In such a case, the fractional flow model of the connection should be described probabilistically to reflect
the range of possible efficiencies to be considered. This is especially critical during production optimization,
as the algorithm tries to push more water in the efficient connection and tries to reduce the water injected
in the inefficient ones. Understanding that a connection efficiency is still uncertain is thus critical to limit
the risk taken when implementing an operational change.
To quantify the uncertainty of connection efficiencies, an optimization-based method was first tested
(Wen et al. 2016) that was inspired by Van Essen et al. (2010). The method investigated the lower and upper
bound of the efficiency that would still yield an acceptable match of the production data. The bounds are
obtained by solving a nonlinear constrained maximization and minimization problem. The approach starts
with the regular history-matching procedure described earlier. The fractional flow model parameters and
corresponding matching error are first recorded. The minimization/maximization problem is then set up to
determine the minimum/maximum parameters that lead to a matching error within a predefined tolerance
of the best model. This algorithm quickly determines the range of possible parameters and corresponding
fractional flow models that lead to a history-match.
The previous approach was applied successfully to field studies and is relatively fast: a full uncertainty
analysis study is usually performed in a few hours. However, only the lower and upper bounds of the
uncertainty range are determined. In order to deliver a more complete probabilistic description of the
uncertainty parameters aij and dij defined in (5), a new Markov-Chain Monte-Carlo (MCMC) approach is
proposed. For any producer j, the history-matching error up to time step n is defined as

(9)

where θ is the parameter vector , and the model and historical oil rate from producer j at
time step τ are respectively qoj(τ, θ) and .
The information provided by the historical oil production rates up to time tn limit the range of θ that would
lead to an acceptable history-match. Let the probability distribution of θ conditional to the historical oil rate
of producer j up to time tn be noted p(θ|qoj(τ = 1,…,n)). The objective of the MCMC approach is to estimate
the distribution of θ at tn+1: p(θ|qoj (τ = 1,…, n, n + 1)) from the distribution of θ at tn: p(θ|qoj (τ = 1,…, n)).
In the proposed approach, we take the above probability distribution as:
SPE-181685-MS 11

(10)

where γ is a normalization factor and tol is a pre-defined tolerance on the model mismatch. More stringent
distributions were tested but were found to have too strong a rejection rate in the presence of noisy data. This
distribution, despite an explicit expression, can be challenging to evaluate because of the nonlinearity of
the fractional flow model. To perform this sampling, an irreducible, ergodic, and aperiodic Markov-Chain
is assumed with the stationary distribution p(θ|qoj (τ = 1,…, n)).
The Metropolis–Hastings algorithm is then used to evaluate the distribution, through the following steps:
1. Definition of parameter distributions:
a. Assume that aij follows a Gaussian distribution .
b. Assume that dij follows a Gaussian distribution .
c. Define the joint distribution of and for all injectors i connected to producer j as Γnj(θ).
2. Initialization: For all injectors i and producers j, initialize the distributions and by an a priori
estimate of . These estimates can be performed by linking the parameters anij and
dnij to physical parameters such as cumulative water injected at breakthrough or at any arbitrary water
cut level.
3. Transition between time-step n and n+1 for each producer j:
a. Generate a random sample by sampling anij and dnij in the known distributions
of and
b. Compute the corresponding efficiency for all injectors and the
corresponding model oil production qij(τ,θo) for τ = 1,…, n + 1; and the historical mismatch j(n
+ 1, θ0).
c. Update:
i. Generate a random sample θ*1 as in step 3.a. and compute j(n + 1, θ*1) as in 3.b.
ii. Compute the Metropolis choice

(11)

iii. Generate a random number p ∊ [0,1] and define θ1 = θ*1 if p < A(θ*1|θ0) and θ1 =θ0 otherwise.
iv. Repeat 3.c.i through 3.c.iv to generate a distribution of θ, which according to the Metropolis-
Hasting theory relaxes to p(θ|qoj (τ = 1,…,n,n + 1)) as the number of samples increases.
d. Compute the mean and standard deviation of the retained parameters θ to establish the posterior
distribution Γn+1j(θ) at time τ = n + 1, which will become the prior at time τ = n + 2.
Figure 4 shows a top view of a part of the reservoir with every connection represented as an arrow between
injectors and producers or between a well and the aquifer (represented graphically at the reservoir boundary).
The color of each arrow represents the uncertainty range of each connection in terms of the computed
12 SPE-181685-MS

efficiency. Connections that have been more significant in terms of instantaneous rate and cumulative
volume tend to show less uncertainty to the weaker or newer connections.

Figure 4—Visualization of the uncertainty of inter-well and well-aquifer connections.

Validation with Reservoir Simulation


To demonstrate the validity of the proposed approach, the workflow was tested by using a reservoir
simulation model as a reference field. The reservoir simulation model selected was representative of
a portion of a large Middle-East carbonate reservoir. The scales have been normalized to preserve
confidentiality. The model contained 52 producers and 41 injectors and the reservoir history was relatively
short. In this instance, the reservoir was under active development and all the wells simulated were drilled
over a period of 4 years. The fluid system was represented with a black-oil formulation and the rock
properties were geostatistically distributed. The simulation model was history-match by the asset team and
was assessed to be a good representation of the reservoir and individual well behavior.
The workflow proposed in this paper was applied to the simulation model taken as the true reservoir. The
reduced-physics connection-based model was created and automatically calibrated. The simulation model
was then run forward in time as a do nothing scenario to establish a baseline of how the field would behave
without further adjustments.
The connection-based model was then used to optimize the field operational constraints over a six
months period. For every forecasted month, the data from the simulation model was retrieved and fed
to the optimization engine. The connection-based model was recalibrated using the new information and
new operational settings were recommended. The operational settings were then translated back into the
simulation model that was run forward for another month.
The objective of the test was two-fold. It was first used to verify that the automatic calibration could be
performed quickly and accurately on this model. Second, the test was designed to check that the operational
changes proposed by the optimization engine would lead to the expected behavior. Thus indicating that the
reduced-physics model captured enough of the critical reservoir and well behavior for its forecast to be
reliable and to be used in optimization calculations.
The average mismatch error in terms of oil and water rate at every well between the reduced-physics
and the "true", simulation-based values were found to be below 1%. The simple, two-parameter empirical
fractional flow model used is therefore sufficiently accurate when combined with the knowledge of the
connection strength between wells on a recurring basis.
SPE-181685-MS 13

Figure 5 shows the comparison between the baseline reservoir simulation forecast run and the run using
the optimized recommendations from the proposed workflow. The optimization engine was asked to deliver
a stable oil production while minimizing the water cut. In the results, a sharp decrease in the water cut is
observed immediately after the beginning of the forecast in the optimized scenario while the oil production
remains constant up to a tolerance. Over the next months, the water cut increases in both models but the
optimized scenario remains superior to the baseline as the oil production is maintained and the water cut
remains around 40% below the baseline.

Figure 5—Comparison of the baseline cade (dotted lines) with the optimized case (solid lines),
showing how the methodology proposed is able to improve the reservoir performance over time.

Discussion on Field Applications


The test case presented in the previous section shows the magnitude of the potential improvements that can
be achieved by using the methodology proposed. In practice however, many other factors can hinder the
direct application of the approach.
First and foremost, the dataset required to make decisions should be made available quickly to the asset
team quickly. It is often the case that the total field production is accurately monitored on a continuous
basis but that individual well rates are determined through a back-allocation exercises from their routine test
cases. For this methodology to work effectively, both the water and oil rate estimations should be sufficiently
accurate and recent to be used for decision making. This supposes that producers are tested at the very
least once during an iteration and that the allocation exercise is performed in a timely fashion. A similar
consideration should be made for injectors. In some fields, the injection rate in each well is not always
accurately measured.
Another obvious practical consideration is the ability to match the proposed rate target defined in the
workflow. This can be done in different ways operationally. Some operators adjust well-head chokes
while others have the ability to alter their artificial lift systems to reach a more precise target rate. The
recommended changes should also be implemented quickly, which requires a fast mobilization of resources.
Considering these limitations, the methodology has been applied to three reservoirs in North America
and three carbonate reservoirs in the Middle-East. The various reservoirs were located in four different
countries and operated by various international and national oil companies. The reservoirs selected had good
data management and well controllability for the approach to be applicable. The various reservoirs ranged
14 SPE-181685-MS

in size, complexity and maturity. Fluid properties were always modeled through a black-oil model. Rock
property models ranged in complexity from single-porosity single-permeability in clastic systems to dual-
porosity dual-permeability in some carbonate reservoirs.
Changing the existing Reservoir Management strategy of an asset is always challenging endeavor. The
biggest contribution delivered by the methodology presented here was on building alignment within asset
teams. The connection-based model allowed the interpretations from the geomodeling and simulation team
members to be accounted for by the operational teams. Conversations tend to evolve from trying to assess
the connections between wells within the reservoir with various groups within the asset team presenting
opposing arguments to a dialogue on what constraints should be considered and what strategy should be
devised to improve the management of the reservoir.
The optimization engine delivers a complete set of operational changes that should be implemented as
a whole. Some of these changes were intuitive to the team and were approved immediately while others
were deemed too risky. In such cases, some of the proposed changes were removed from the scenario and
the model forecast was updated to estimate the resulting field behavior. The platform was not used blindly
but was rather used as a scenario ranking tool to help with decision making. In particular, knowledge of
the connection strength and efficiencies over time, even when subject to uncertainty help the team better
understand inter-well behavior.

Conclusions
A complete workflow to assist in waterflood management was presented. The approach starts with the
creation of a reservoir model based on simplified physics that represents the subsurface problem as a network
of interwell connections characterized by their strength and efficiency. The connection-based model can
integrate an existing geologic or simulation model to describe the underlying geology of the field and is
equipped with an automated calibration algorithm to quickly to match production data. The calibration
algorithm can be matched to new production data within minutes with tolerances below 1%.
An uncertainty quantification strategy was also presented to help understand the limits of the model on
a connection by connection basis. More mature interwell interactions are well calibrated while newer ones
remain more uncertain.
Finally, the connection-based model is used by an optimization engine that uses field and well constraints
to describe the current reservoir management strategy of the field and helps define the optimal well liquid
rate targets under this strategy to either maximize recovery or minimize water production.
The accuracy of the connection-based model and the validity of the optimization approach were validated
against a reservoir simulation model and field applications were discussed. The methodology has proven
to be successful in helping asset teams understand and better manage their waterflooded reservoirs by
improving their operational decisions.

Acknowledgements
The authors would like to thank Quantum Reservoir Impact for permission to publish this work.

References
Albertoni, A. 2002. Inferring Interwell Connectivity Only From Well-Rate Fluctuations in Waterfloods. M.S. Dissertation,
University of Texas at Austin.
Albertoni, A., Lake, L. 2003. Inferring Interwell Connectivity Only From Well-Rate Fluctuations in Waterfloods. SPEREE
6 (1): 6 –16.
Aziz, K. and Settari, A. 1979. Petroleum Reservoir Simulation. Applied Science Publishers.
Batycky, R. P., Thiele, M. R., Baker, R. O., and Chugh, S. 2008. Revisiting Reservoir Flood-Surveillance Methods Using
Streamlines. SPEREE, 11 (2) 387 –394. doi: 10.2118/95402-PA
Bruce, W.A. 1943. An Electrical Device for Analyzing Oil-Reservoir Behavior, Pet. Trans. AIME, 151: 112 –124.
SPE-181685-MS 15

Dake, L. P. 1978. Fundamentals of Reservoir Engineering. Developments in Petroleum Science, Elsevier, Amsterdam, NL.
Gentil, P. H. 2005. The Use of Multilinear Regression Models in Patterned Waterfloods: Physical Meaning of the
Regression Coefficients. M.S. Dissertation, University of Texas at Austin.
Houzé, O., Viturat, D., Fjaere, O. S. 2013. Dynamic Data Analysis. Kappa, Paris, France.
Izgec, O. and Kabir, C. S. 2009. Establishing Injector/Producer Connectivity before Breakthrough during Fluid Injection.
In SPE Western Regional Meeting, San Jose, CA.
Kikani, J. 2013. Reservoir Surveillance. Society of Petroleum Engineers.
Lake, L. W., Liang, X., Edgar, T. F., Al-Yousef, A., Sayarpour, M. and Weber, D. 2007. Optimization of oil production
in a reservoir based on capacitance model of production and injection rates. SPE-107713-MS presented at the SPE
Hydrocarbon Economics and Evaluation Symposium, Dallas, TX.
Liu, F., Mendel, J. M. and Nejad, A. M. 2007. Forecasting Injector-Producer Relationship from Production and Injection
Rates Using an Extended Kalman Filter. SPE-110520-PA, presented at the SPE Annual Technical Conference and
Exhibition, Anaheim, CA.
Lee, K.-H. 2010. Investigating Statistical Modeling Approaches for Reservoir Characterization in Waterfloods from Rates
Fluctuations. PhD. Dissertation. University of Southern California.
Shahvali, M., Mallison, B. T., Wei, K., Gross, H., 2011. An Alternative to Streamlines for Flow Diagnostics on Structured
and Unstructured Grids. SPE Journal 17 (03). DOI: 10.2118/146446-MS
Sayarpour, M. 2008. Development and Application of Capacitance-Resistive Models to Water/CO2 Floods. PhD
Dissertation, University of Texas at Austin.
Thiele, M. R., Batycky, R. P. 2006. Using Streamline-Derived Injection Efficiencies for Improved Waterflood
Management. SPEREE, 9 (2): 187 –196.
Van Essen, G. M., Jansen, J. D. and Van Den Hof, P. M. J. 2010. Determination of lower and upper bounds of predicted
production from history-matched models. Presented at the EAGE European Conference on the Mathematics of Oil
Recovery, Oxford, UK.
Wahl, W. L., Mullins, L. D., Barham, R. H., and Bartlett, W. R. 1962. Matching the Performance of Saudi Arabian Oil
Fields with an Electrical Model. JPT, 14 (12): 1275 –1282. SPE-414.
Weber, D. 2009. The Use of Capacitance-Resistance Models to Optimize Injection Allocation and Well Location in Water
Floods. PhD Dissertation, University of Texas at Austin.
Wen, T., Thiele, M. R., Echeverría Ciaurri, D., Aziz, K. and Ye Y. 2012. Reservoir management using two-stage
optimization with streamline simulation. In The 13th European Conference on the Mathematics of Oil Recovery,
Biarritz, France.
Wen, T., Thiele, M. R., Echeverría Ciaurri, D., Aziz, K., and Ye, Y. 2014. Waterflood management using two-stage
optimization with streamline simulation. Computational Geosciences, 18 (3-4), 483 –504.
Wen, T. 2014. Waterflood Optimization using Streamlines and Reservoir Management Risk Analysis with Market
Uncertainty. PhD Dissertation, Stanford University.
Wen, T., Zhai, X. and Matringe, S. F. 2016. Inter-well Connectivity in Waterfloods - Modelling, Uncertainty Quantification,
and Production Optimization. Presented at the EAGE European Conference on the Mathematics of Oil Recovery,
Amsterdam, NL.
Yousef, A. A. 2005. Investigating Statistical Techniques to Infer Interwell Connectivity from Production and Injection
Rate Fluctuations. Ph.D. Thesis, University of Texas at Austin.
Yousef, A. A., Gentil, P. H., Jensen, J. L., Lake, L. 2006a. A Capacitance Model to Infer Interwell Connectivity from
Production- and Injection-Rate Fluctuations. SPEREE 9 (5): 630 –646.
Yousef, A. A., Jensen, J. L., Lake, L. 2006b. A Capacitance Model to Infer Interwell Connectivity from Production and
Injection Rate Fluctuations. SPE 99998, presented at the SPE IOR Symposium, Tulsa, OK.

You might also like