Вы находитесь на странице: 1из 23

SPE-173829-MS

Dynamic Well Conversion and Rate Optimisation Using Ensemble-Based


Method
Kurt R. Petvipusit, Imperial College London, UK; Yuqing Chang, International Research Institute of Stavanger,
Norway

Copyright 2015, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Bergen One Day Seminar held in Bergen, Norway, 22 April 2015.

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 acknowledgements of SPE copyright.

Abstract
This work describes the use of ensemble-based optimisation (EnOpt) approach to maximise a pre-defined
economic criterion, the Net Present Value (NPV), while considering two well parameters simultaneously,
the well type and the corresponding well control rates for a given subsurface model. EnOpt has previously
been used to optimise well control settings for a fixed number of producers and injectors. It relies on the
ensemble of well control settings from which the sensitivities of a suitably constructed objective function
are approximated. The sensitivities to the control settings are then used with a steepest descent algorithm
to determine optimal well controls. Our proposed approach relies, instead, on the selection of three
variables for each well: liquid production rate, injection rate, and well conversion time. Well conversion
time dictates the optimal time to switch a producer to an injector, which also plays a key role in
determining well type and the optimal number of injectors and producers. The superior performance of
our approach is indicated by the increase in NPV, cumulative oil production and sweep efficiency within
a predefined reservoir life. This work presents three case studies: only well-type optimisation, only well
rate control optimisation and simultaneous well conversion and well rate optimisation. We also discuss the
effect of localisation used in the EnOpt algorithm on well conversion and well control optimisation. Based
on numerical results of the reservoir PUNQ-S3 model, our approach revealed potential benefits of
dynamic well conversion and rate control to optimise waterflooding. Comparing the proposed approach
with the traditional EnOpt workflow, using a fixed number of producers and injectors, clearly indicates
improved project revenues with a comparable computational effort.

Introduction
Waterflooding has been a great success worldwide and typically results in additional recovery of about
10-20% of original oil in place (OOIP) over primary production. Waterflooding leads to incremental
recovery by maintaining reservoir pressure and by mechanically displacing residual oil. In 1988 Asheim
[4] pioneered the use of a gradient-based algorithm applied to a water-drive problem to include both
natural and induced water drives to optimise waterflooding. They maximised sweep efficiency by
allocating rates between injector and producer pairs. The optimised cases showed an improvement in NPV
compared to a simple rate allocation. Later, Yeten et al. [18] used a similar approach by approximating
2 SPE-173829-MS

the gradient of the objective function from a finite difference approximation. In addition to that, Brouwer
and Jansen [5] later developed the gradient-based algorithm coupled with adjoint-based approach in a
waterflood application. The goal of their work was to optimise the flood front in high permeability regions
so that they could maximise sweep efficiency by controlling the appropriate well pressure profiles. Sarma
et al. [14] developed the adjoint equations and later showed that it improved the efficiency of the adjoin
calculations. Lien et al. [11] considered a multiscale regularization method to group well control settings
both in space and time for waterflooding optimisation, resulting in an improvement of cumulative oil
recovery.
Alhuthali et al. [2] considered equalizing the arrival time of the flood front reaching producers. This
strategy had the benefit of accounting for reservoir heterogeneity as well as existing well patterns through
the use of streamlines for the computation of arrival times. In general, the authors indicated that such a
strategy could result in delayed water breakthrough, increasing in cumulative oil production and oil
recovery factor in comparison to a constant rate control. In 2009, Alhuthali [1] extended their approach
to account for geologic uncertainty via an optimisation framework using multiple realisations. However,
this approach was likely to delay or slow production at early stages of the reservoir life in order to allow
the flood front to move uniformly out from the injectors. Equalization of arrival times with a slightly
different approach was documented by Chen [6] where the authors incorporated a penalty cost term for
the amount of water production. Wang et al. [17] used gradient-based algorithms with the gradient
information derived from several approaches for their waterflooding studies. They compared the steepest
ascent approach with the gradient approximated by finite-differences, by single Simultaneous Perturbation
Stochastic Approximation (SPSA) and average SPSA while Nwaozo [12] implemented a EnKF-based
approach for optimization instead of data assimilation and showed very promising results. In a related
work, Asadollahi and Naevdal [3] investigated the impact of different control settings on the performance
of the Brugge field with 20 producers and 10 injectors. The authors showed that producers controlled by
liquid production rates and injectors controlled by water rate tended to provide the highest NPV. They also
investigated the convergence rate of the conjugate gradient and steepest descent methods with different
starting points and reported that the conjugate gradient algorithm was more efficient than the steepest
descent method when the initial point was not close to the optimum solution.
Ensemble based optimisation method is a relatively new technique compared to other optimisation
approaches applied to wateifllooding. The formulation is similar to the Ensemble Kalman Filter (EnKF)
approach for data assimilation except that the state vector comprises well control settings instead of
reservoir state variables and parameters. The ensemble is used to determine the sensitivities of the
objective function to various control settings. EnOpt was first proposed by Nwaozo [12] for waterflooding
optimisation. The Bottom-Hole Pressure (BHP) of producers were optimised to find a maximum NPV
while the total injection rate was maintained a constant. All wells were controlled with reactive control
at a water cut value of 0.93. After optimisation, an increase of 9 % was observed in cumulative oil
production. In contrast to the adjoint-based approach, the effort to incorporate EnOpt into existing
simulators would not necessarily be very difficult, and conveniently it does not require access to simulator
source codes.
EnOpt was further developed by Chen [6] for a closed-loop production optimisation where they
reported that EnOpt was essentially a gradient-based optimisation method with the gradient information
acquired from an ensemble. In 2010, EnOpt was compared to the adjoint-based approach by Leeuwen-
burgh et al. [8] in terms of ensemble size, perturbation of the ensemble, regularization, smoothing and the
quality of the solution. They considered the impact of control step sizes and showed that the gradient
estimation was poor for large time step intervals. Their results in terms of the smoothing was in line with
Chen [6] where the solution was shown to be superior when the measure of smoothing was considered.
However, both studies did not mention any guidelines as to the level of smoothing required to attain higher
values of the objective function. For complex problems with smart wells, Su and Oliver [15] used EnOpt
SPE-173829-MS 3

for downhole control valve settings (ICVs) by minimising water production while keeping liquid
production a constant. The results of the reactive control strategy and EnOpt were compared which
showed the improved performance of 50% reduction in cumulative water production.
Production optimisation approaches generally centre around one specific problem; for example, this
includes optimising well rates for a given well configuration or vice versa (optimising well locations for
a predefined well rate). While the traditional approaches have shown to improve oil recovery and increase
revenues, it is highly likely that the optimal solution is dependent on a given predefined constraints and
hence there exists a set of optimal solutions that corresponds to a set of predefined constraints.
Considerably optimisation of simultaneous predefined constrains and the well control parameters could
mitigate the dependency of the optimal solution and might result in a better optimal solution. As addressed
by Li and Jafarpour [9], optimisation of combined variables, such as rate allocations and well locations,
showed significant improvements of NPV.
Unfortunately, in recent years, a few studies [10, 13] have considered solving optimisation of the
combined well control parameters. For example, Li et al. [10] considered solving the optimisation of well
locations together with well rates simultaneously. Although optimal well placement is a key aspect of
production optimisation, it may be more cost-effective to instead focus on existing wells and then
determine optimal well conversion times to switch production to injection and simultaneously determine
their corresponding control settings. For this reason, we consider the optimal solution that involves
conversion of existing producers to injectors and their corresponding well rates to improve oil recovery
by waterflooding. The objectives of this study are given below:
1. Optimal well types: producers and/or injectors
2. Optimal number of wells of each type
3. Optimal conversion time from production to injection, if necessary
4. Optimal production and injection rates for particular wells

Mathematical Formulation of Ensemble-based Optimisation


EnOpt is an approximate gradient-based approach in which the sensitivity of the objective function with
respect to the control variables are approximate from the ensemble of control variables. Each variable
within the ensemble is assumed to follow a pseudo-Gaussian random field. Control variables are updated
based on the gradient information to maximise NPV the objective function. Additionally, to prevent
the abrupt change in control variables, the objective function is augmented by a regularization term to
impose a measure of smoothness in control variables. The form of the objective function used in this study
is given below:
(1)

where x is the state vector of control variables; Cx is the covariance matrix of the control variables,
describing the uncertainty of the control settings; f (x,yopt) is the Net Present Value (NPV). The objective
function is augmented by a penalty term for preventing the abrupt change in adjacent control variables,
the regularisation term on the right-hand side of equation; a is a weight factor between the regularisation
and the NPV term. The a value indicates the weight of smoothness of the updated x such that it is close
to the prior values xp. The a can also be described as the reciprocal of a step length for updating the
steepest descend algorithm (shown later). The choice of the initial a value in this study is subjected to the
scaling of step size. We simply select the initial a at the half of standard deviation of NPVs. The a value
will be increased overtime if NPV is not improved.
The objective function or the NPV function f in this study is formulated in the following:
4 SPE-173829-MS

(2)

where the model state vector y contains geophysical properties where the geology is presumably
known. Therefore, the term S(x) is specified for a presumably known reservoir model (i.e., reservoir rock,
fluid properties, geological structure, etc).
The gradient-based optimisation framework is derived from a basic formulation of Newtons method
in the following steps. First, the following objective function S(x) is approximated by using second-order
Taylers series expansion around the neighbourhood of x. Second, the maximisation of the approximated
quadratic function in a small neighbourhood, x around x defines the new updated value of x. This cycle
is repeated until the stopping criteria are reached.
(3)

where dS(x) and HS(x) denote the gradient and the Hessian of the objective function S(x). From Eq. 1,
we obtain the gradient and Hessian of the objective function in the following:
(4)

(5)

where the gradient and the Hessian of the NPV function are represented by g(x) and Hf (x),
respectively. Because we approximate the objective function with the first-order Taylors series expan-
sion, the Hessian is then negligible. Therefore, we can re-arrange Eq. 5 as follows:
(6)

The value of x that maximise the objective function S(x x) is given by solving the following set
of equations below:
For most applications, S(x) is the non-linear function of the control variables x. We hence use an
iterative approach to solve this equation with an iteration step given by index l.
(7)

Substituting dS(x) from Eq. 4 and HS(x) from Eq. 6 into Eq. 7, yields
(8)

Then, we obtain the updated state vector xl1 as:


(9)

Equation 9 is similar to the classical steepest ascent algorithm except for the additional pre-
conditioning term Cx for the gradient g(x). Chen [6] documented the approximation of the sensitivity
matrix g(x) of the objective function to the control variables x by using crosscovariance between the
function value and the control variables x, denoted as Cx, f(x) as follows:
(10)

The idea behind the pre-conditioning matrix Cx is to exploit physical knowledge of the system to
correct the gradient direction to a better (faster) direction. Also, the pre-conditioning term is used to
prevent the abrupt change in adjacent control variables. Moreover, without the regularization term, the
optimisation problems could become ill-posed [16]. The demonstration of how the cross-covariance
SPE-173829-MS 5

matrix between the objective function and the control variables are used to approximate the sensitivity of
the objective function is referred in [7, 13]. Then, the updated state vector x in Eq. 9 is explicitly given
in the following.
(11)

where Cx, f(x) is the cross-covariance between the function value and the control variables x.
For dynamic well-scheduling and well-type optimisation using the EnOpt algorithm, we define the list
of operational variables contained in the state vector x as: (i) time-dependent reservoir volume flow rates
for all wells and all time steps, (ii) time-dependent water injection rates for all wells and all time steps,
and (iii) conversion time for each well from production to injection. The size of state vector x is given as
2(Nt Nw) Nw, where Nt and Nw represent total time step and number of well respectively. The state
vector x is explicitly given in the following:
(12)

where vectors a, b, and c are reservoir volume rate, water injection rate, and conversion time vectors,
respectively. Each element of the vector a and b is the control rate for specific time step and its dimension
is Nt x Nw. Each well has only one conversion time, and therefore the vector c of dimension Nw comprises
the conversion time for each well. In Eq. 13, parameters vpr, inj, and t are the reservoir volume flow rate,
the water injection rate, and the conversion time. The superscript represents the number of wells while the
subscript denotes the specific time step.
(13)

The selection of well type to be either injector or producer depends on the conversion time variables.
In this example, a well is initially a producer and then it will be converted to an injector later based on
the conversion time ti in the state vector. Conversion time is demonstrated in Fig. 1, where the conversion
time equal to 3 means that the well operates as a producer for the first two time steps and then it is
converted to an injector at the 3nd time steps for a total production of 5 simulation time steps. The
corresponding rates are selected from the green and the blue circle for production and injection,
respectively.

Figure 1The selection of well type by conversion time parameters.


6 SPE-173829-MS

Reservoir Descriptions
The proposed procedures are tested with a 3-D Cartesian synthetic reservoir model defined on a 24 x 24
x 3 grid lattice. The dimension of each grid are 30 ft x 30 ft x 100 ft. The area of the reservoir is 12 acres
(720 ft x 720 ft). The field has five vertical wells which are completed in all three layers and are arranged
in a 5-spot pattern as shown in Fig. 2(b). All case studies are restricted to two phase flow and
closed-boundary conditions. The logarithmic permeability distribution for heterogeneity case is shown in
Fig. 3. For reservoir structure, our examples include a non-dipping reservoir and an inclined reservoir. The
descriptions in the models are given as follows:

Figure 2(a) The inclined reservoir structure with water oil contact(WOC) shown in blue at depth 7025 ft. The WOC is at the layer 2 and
3 of wells 2 and 4. (b) Wells 1 and 3 are located at the top and the bottom left corner of the figure while wells 2 and 4 are located at the
top and the bottom right corner.

Figure 3Top view of the 2-D cartesian synthetic reservoir with heterogeneous permeability distribution. Black dots show the well
locations.

1. For the inclined reservoir, the datum depth is 6035 ft with a corresponding pressure at 2500 psia.
Water-oil contact (WOC) is at 7250 ft while gas-oil contact is located at 1500ft. The top of
reservoir is located at 6000 ft below the surface.
SPE-173829-MS 7

2. For the non-dipping reservoir, the datum depth is 9035 ft with a corresponding pressure at 3600
psia. Water-oil contact (WOC) is at 11000 ft while gas-oil contact is located at 8500 ft. The top
of reservoir is located at 9000 ft below the surface.
Figure 2 shows the structure and well locations while Table 1 summarises the model descriptions,
which will be used to test our proposed EnOpt algorithm.

Table 1Summary of reservoir properties


Description properties

Number of grid blocks 24243


Number of layers 3
Grid block size 30 ft 30 ft 100 ft
Porosity 0:30
Permeability (homogeneous case) 10 mD
Initial reservoir pressure 3604 psia
Reservoir Temperature 212 F
Production time 5 years
Well control step 6 months

Results and Discussions


First, we present case studies documenting only well-type optimisation. Second, we investigate only well
rate control optimisation. Next, case studies including simultaneous well-type and well-scheduling
optimisation are demonstrated. Last, we present the effect of localisation of the gradient for this problem.

Well Type Optimisation


We used EnOpt to determine optimal well type settings for fixed well controls. Therefore all wells are
controlled by constant BHPs. Consequently, the only variable of interest is the well conversion time that
dictates the optimal time to initiate water- flooding. The dimension of the problem reduces to a single
unknown variable for each well. Wells are controlled by BHP at 8,000 psi for injector and 500 psi for the
producer. All BHP controls are kept constant throughout 5 years and these are compared to an inverted
5-spot pattern with fixed well configuration. The inverted 5-spot means that the injector is located at the
centre of the reservoir and the producers are located at the corner. For well type optimisation, we consider
the following case studies.
Case 1: Heterogeneous permeability with the non-dipping structure reservoir.
Case 2: Homogeneous permeability with the inclined structure reservoir.
Case 1: Heterogeneous Permeability Field with the Non-dipping Structure Reservoir. We present
two base case scenarios: the inverted 5-spot, and the 5-spot flood pattern in comparison to EnOpt in this
section. Figures 4 displays the NPVs versus the iteration of optimisation. Using the conversion times in
EnOpt clearly outperforms the two base cases which optimise only rate controls with fixed well types.
Numerical results of the figure are summarised in Table 2. We see that EnOpt provides the highest NPV
with 65% improvement in comparison to the 5-spot pattern flood. This indicates the value of well-type
optimisation even for a very simple and contrived reservoir.
8 SPE-173829-MS

Figure 4 The Net Present Value (NPV) is plotted versus the iteration of optimisation. EnOpt approach is shown by the red curve with
the circle-marker, while the triangle-marker and the cross-marker curves represent the inverted 5-spot and the 5-spot pattern,
respectively.

Table 2The comparison of well type optimisation for 3 scenarios


NPV increase(%) from 5-spot
No Scenario NPV ($USD)x106 pattern

1 EnOpt 214.6 65.2


2 inverted 5-spot 174.3 34.2
3 5-spot 129.9 0

Case 2: Homogeneous Permeability Field With the Inclined Structure Reservoir The inclined struc-
ture in Fig. 2 with homogeneous permeability field is demonstrated. Wells 2 and 4 were completed in the
water bearing zone while wells 1, 3, and 5 were completed in the oil bearing zone. It is therefore
intuitively reasonable that wells 2 and 4 are likely to be injectors to reduce the cost of water production.
The following two scenarios are considered:
1. We set the initial well configurations such that wells 1 and 3 are likely to be producers for the
entire life of the field, and wells 2, 4, and 5 operate as producers and injectors for equal lengths
of time. This is illustrated in Fig. 5 in a spider chart, where the spoke represents the percentage
of production of a particular well for the entire period of field. For example, the value of the spoke
at 0 indicates that the well operates as an injector whereas 100 indicates that the well mostly
operates as a producer. The value at 60 means that the well is operated as a producer for the first
60 percent of the field life and it is subsequently operated as an injector for the last 40 percent. The
value of the spoke is basically the conversion time of the particular well. Figure 5(a) shows all
iterations of the optimisation algorithm while Fig. 5(b) only displays the initial and the final well
configuration. From both figures, we see that the algorithm shows satisfactory results since well
2 and well 4 located in the water bearing zones at the bottom of the inclined reservoir move
towards the longer injection periods, as expected. Additionally, the plot of the NPV is shown in
Fig. 6, where EnOpt is compared with the fixed well type case that uses wells 2 and 4 for injectors
and wells 1, 3, and 5 for producers for the entire period of the field.
SPE-173829-MS 9

Figure 5Mean of the conversion time for 50 realisations for five wells is shown in the spider chart. The percentage on each spoke
shows the percentage contribution of the production throughout the entire period of the field. The value on the spoke also represents
time that wells are converted from producer to injector.

Figure 6 The Net Present Value (NPV) is plotted versus the iteration of optimisation for two cases: EnOpt (close circle marker) and
the fixed well type (open circle marker).

2. The second example illustrates a somewhat different outcome for a different starting configuration.
For this exercise, we assume that there is no prior information about the type of well. Therefore,
each well has a conversion time of 50 indicating equal injection and production periods. Our
results are shown in Fig. 7 and indicate that with no prior knowledge, all corner wells switch to
injection early in the life of the field while the central well remains a producer. Well 2 and well
4 move towards the injection like as expected. Since well 5 is likely to be a producer, wells 1 and
3 are converted to injectors to flood more oil to well 5. These results indicate the sensitivity of the
outcome to the initial configuration. Since EnOpt relies on a gradient-based approach to seek a
solution, there is a possibility of having a local maximum in this case.
10 SPE-173829-MS

Figure 7Mean of the conversion time for 50 realisations for five wells is shown in the spider chart. The percentage on each spoke
shows the percentage contribution of the production throughout the entire period of the field. The value on the spoke also represents
time that wells are converted from producer to injector. We used a different starting point from the one in Fig. 5.

Well Control Optimisation


Chen [6] documented the use of EnOpt for production optimisation by allocating well controls. In this
section, we present the fixed rate compared with the optimised rate controls to illustrate the benefits of
well control optimisation from fixing well-types configuration. We used the 5-spot pattern in this case, as
shown in Fig. 3.
With a constant well rate control in a dipping reservoir, one should expect the unbalance of injected
waterfront at the centre of the field advances downhill towards producers due to the gravity effect. Water
saturation distribution for different period of injection are shown in Fig. 8, where the top row is the result
from the constant rate control while the bottom row represents the water distribution of the EnOpt method.
We see from Fig. 8 that the using the constant rate control results in poor sweep efficiency, high water
cycling, and premature water breakthrough. Since the water has created a flow channel in the subsurface,
controlling water front to the unswept area would be difficult and expensive. This asymmetrical flood
pattern can often be observed from the flow through a heterogeneous permeability field and/or the
imbalance controls between the injection and production. Without using the EnOpt algorithm, injection
and production rates may need to be fine-tuned with trial and error to balance flood pattern for a better
sweep efficiency. As expected, optimisation of well controls using EnOpt reveals the excellent flooding
performance as it balances the rates of injectors and producers for the uniform sweep pattern, improving
oil recovery of the field.
SPE-173829-MS 11

Figure 8 Water saturation distributions are shown between the constant control rates (the top row) and optimised control rates by
EnOpt (the bottom row) for 1.5, 3.5, and 5 years.

Simultaneous Well Type and Well Rate Optimisation


In the previous two sections, we have presented the benefits of well type and well rate optimisation. The
results show that naively categorising each well as a producer or an injector is not sufficient, and it may
be better to re-evaluate the well type and well configurations simultaneously. EnOpt is compared with the
traditional well type .i.e. 5-spot and inverted 5-spot flood pattern in this study, where two cases are
presented. For case A, two types of permeability fields, homogeneous and heterogeneous permeability
fields displayed in Fig. 3, are used with the 5-spot pattern for optimisation, while the localisation of
different correlation lengths are presented in case B. Both cases are summarised in Table 3.

Table 3Case studies of simultaneously well type and well scheduling optimisation by EnOpt.
Reservoir Description

Case Structure Permeability localisation

A1 non-dipping Homogeneous No
A2 non-dipping Heterogeneous No
B inclined Homogeneous Yes

Case A1: Homogeneous Permeability Field of a Non-dipping Reservoir Without Localisation W e


consider the following case studies in this section.
1. Inverted 5-spot with rate optimisation; this case will later refer to the inverted 5-spot case.
However, it should be noted that we use EnOpt for well rate optimisation in this case.
2. Simultaneous well type and well rate optimisation using EnOpt with random initial configurations.
3. Simultaneous well type and well rate optimisation with prior knowledge on well configurations
12 SPE-173829-MS

using EnOpt. We chose all the corner wells as producers for the initial well configuration, and
there is no preference of the well configuration for the well in the centre of the field.
4. Simultaneous well type and well rate optimisation using EnOpt with a selected initial configura-
tion. The selected initial configuration is chosen from the optimised well pattern in case 3 above,
where wells 1 and 5 are injectors while other wells are producers.
In this section, we present the impact of different initial well configurations on the NPV, where only
the well type is considered for the initial preferred well configuration. This means that we have a prior
knowledge about type of wells before implementing EnOpt. The NPVs of the four scenarios discussed
above are shown in Fig. 9 and the corresponding numerical NPV are given in Table 4. We see that EnOpt
with the selected initial points provides the higher NPV, and the invert 5-spot without considering well
type optimisation results in the lowest NPV. By using both well type and rate for optimising waterflood-
ing, we see that the NPV could increase between 3.55 - 13.10 % from the well rate optimisation. This
implies the benefits of optimising both well type and well rate for waterflooding design.

Figure 9 The Net Present Value (NPV) is plotted versus the iteration of optimisation for four cases: EnOpt with selected starting point
(the triangle-marker black curve), EnOpt with prior knowledge (the dashed curve), EnOpt (the triangle-marker red curve), the inverted
5-spot pattern (the circle-marker curve).

Table 4 Summary of the NPV obtained at the final iteration of case A: homogeneous permeability field
NPV increase (1 x 106 NPV increase in percent
No Scenario NPV (1 x 106 $US) $US) from base case (%) from base case

1 Inverted 5-spot (base case) 181.6 0 0


2 EnOpt 188.0 6.45 3.55
3 EnOpt with prior 189 7.36 4.05
knowledge
4 EnOpt with selected initial 205.4 23.8 13.10
point

The snapshots of oil distribution during 6 months, 2.5 years, and 5 years for the four scenarios are
displayed in Fig. 10, where each column of the plot presents time periods while each row represents the
scenarios. We see from the figure that using the EnOpt with selected initial point can improve oil recovery
as there is less residual oil than the one using the other three approaches. Additionally, the waterflooding
performance from both well type and well rate is much better than using only the well rate optimisation
as the residual oil obtained from only rate allocations (see the first row) is greater than the case with
optimising both well type and well rate. This result is also in line with the NPV shown previously in Fig.
9. In addition, the cumulative oil and water production are displayed in Fig. 11, where we see that the
maximum total oil and water production are from EnOpt with selected initial point. The reason why this
SPE-173829-MS 13

strategy results in the highest water production is that the water disposal cost is much lower than the profit
gained from oil production. Hence, it tends to produce more oil regardless of the amount of water
production. In addition to that, we do not impose a reactive control of water cut in this study.

Figure 10 Each column shows oil saturation for different time periods: 6 months, 2.5 years, and 5 years. The optimisation approaches
are shown in each row, where 1st row: inverted 5 spot, 2nd row: EnOpt, 3rd row: EnOpt with prior knowledge, 4th: EnOpt with selected
initial point(SP). Only the best well configuration from each approach over 50 realisations is displayed here.
14 SPE-173829-MS

Figure 11Cumulative oil (a) and water (b) production are plotted versus time for four scenarios: EnOpt with selected initial point (the
dashed triangle-marker curve), EnOpt with prior knowledge (the circle-marker with the dashed curve), EnOpt (the triangle-marker with
the solid curve), and the inverted 5-spot (the circle-marker with the solid curve).

Since optimal solution from the gradient-based algorithm are generally sensitive to the choice of the
starting point, we re-initialise 20 initial points to compare the performance of our proposed approach to
the well rate optimisation with fixed well type. Figure 12 shows the box plot of optimised NPVs, where
the size of the box denotes the variance of the expected NPVs over 50 realisations from 20 different initial
points. The box indicates a range of NPVs with a lower bound value at 25th percentile and upper bound
at 75th percentile. The horizontal line at the middle of the box indicates the expectation of the NPVs over
20 runs. We see that the expected NPV is significantly higher than the one obtained from the well rate
optimisation.

Figure 12Box plots of the expected NPVs from 50 realisation at the last optimisation iteration are shown. The range of the box
represents the variability of the expected NPVs from different 20 starting points. The box on the left presents the inverted 5-spot pattern
with only well rate optimisation while the right box represents EnOpt for both well type and rate optimisation The optimised NPV are
obtained from 50 realisations.

In order to access the performance of our approach with simply calibrating production rates alone, we
also do a side-by-side comparison of the NPVs at each iteration. This has the benefit of comparing the
solutions of both techniques for similar computational efforts. The box plots with the iteration of
optimisation are shown in Fig. 13. Although our approach began with a relatively lower NPV (1st
iteration), by the 3rd iteration, we obtained a higher expected NPV.
SPE-173829-MS 15

Figure 13c

Case A2: Heterogeneous Permeability Field, in a Non-dipping Reservoir without Localisation This
section illustrates our approach applied to the heterogeneous permeability field shown in Fig. 3, EnOpt
with no prior knowledge (random selected initial point) is compared to the traditional flood patterns,
namely the inverted 5-spot (base case 1) and 5-spot flood pattern (base case 2). The production rates alone
were optimised for the base case studies.
Figure 14 displays the NPVs and the iteration of optimisation for three scenarios: EnOpt, the inverted
5-spot, and the 5-spot case. We see that NPV from the EnOpt is the highest in comparison to the other
approaches. The numerical values for the optimised NPV is shown in Table 5 where the NPV of EnOpt
is 27% higher than one from the base case 5-spot approach, indicating a potentially go-no-go decision
point for project planning.

Figure 14 The Net Present Value (NPV) is plotted versus the iteration of optimisation for three cases: EnOpt for both well type and
rate optimisation (the triangle-marker curve), the inverted 5-spot pattern (the circle-marker curve), the 5-spot (the dashed curve).
16 SPE-173829-MS

Table 5Summary of the NPV obtained at final iteration of case A: heterogeneous permeability field
NPV increase (1 x 106 NPV increase in percent
No Scenarios NPV (1 x 106 $US) $US) from base case 1 (%) from base case 1

1 The Inverted 5-spot 181.7 0 0


2 The 5-spot 207.1 25.37 13.40
3 EnOpt for both well type 230.8 49.14 27.05
and well rate

Oil saturation profiles during 5 years of waterflooding from EnOpt, and two base cases, i.e. the 5-spot
and the inverted 5-spot, are shown in Fig. 15. For the base case studies, the arrangement of the wells and
their corresponding configurations vis-a-vis the underlying heterogeneity clearly leads to poor sweep
efficiency. However, by considering dynamic well type and well rate, the sweep efficiency is markedly
improved as shown in Fig. 16(d). During the waterflooding of the 5-spot pattern, the pressure field from
well 2 would force water from well 1 and well 4 to create a channelised flow towards well 5, and
consequently it leads to poor sweep efficiency with residual oil left in the low permeability region
(unswept oil between well 1 and well 2) as shown in Fig. 15(b). However, converting well 5 to an injector
and keeping well 1 and well 4 as producers would be more efficient due to the fact that the pressure field
from well 2 and 5 would sweep oil into the low permeability zone towards the producer (wells 1 and 4),
given better sweep efficiency as shown in Fig. 15(d). For the EnOpt method, we illustrate the distribution
of oil distribution for the first 6 months to 5 years of waterflooding plan in Fig. 16. We see that wells 2
and 5 are the injectors throughout all 5 years while other wells are the producers. The unswept oil zone
after 5 years are restricted to the low permeability region shown in the red zone of Fig. 16(d).

Figure 15The distributions of oil saturation after 5 years are shown for the inverted 5-spot, the 5-spot flood pattern, and EnOpt.
SPE-173829-MS 17

Figure 16 Each figure shows oil saturation obtained from the EnOpt method for different time periods from left to right: 6 months, 1.5
years, 3.5 years, and 5 years.

The cumulative oil production from the EnOpt approach is compared with the 5-spot and the inverted
5-spot flood patterns in Fig. 17. The thin grey curves present the range of NPVs obtained from the EnOpt
method and the red curve denotes the expected NPV. While the triangle- and the circle markers present
the 5-spot and the inverted 5-spot cases. We see that the cumulative oil and water from the EnOpt method
is the highest. Additionally, the worst NPV from the uncertainty of optimal controls from EnOpt is still
better than the base cases, indicating that a robust strategy could be obtained when there is a presence of
uncertainty of the control rates. The reason why EnOpt produces large amount of water production is
given previously from from Fig. 11.
18 SPE-173829-MS

Figure 17Left: The cumulative oil production versus time. Dashed line with a triangle marker indicates the mean of oil production from
the inverted 5-spot while a circle marker indicates a case from the 5-spot pattern. A solid red line shows the mean of oil production from
EnOpt. The thin grey lines are cumulative oil production from 50 realisations. Right: Cumulative produced water production versus
time.

For the study of starting points, we re-initialise 20 initial points to compare the performance of EnOpt
and the other base cases. Figure 18 shows the range of optimised NPVs obtained at the final iteration for
the 20 different starting points. We see that the expected value of the NPVs is significantly higher than
simply optimising production controls, which can be inferred that the EnOpt approach still provides a
better NPV than the base cases for various initial points. Again, we also display a side-by-side comparison
of the NPVs in Fig. 19 for several iterations.

Figure 18 Box plots of the expected NPVs from 50 realisation at the last optimisation iteration are shown. The range of the box
represents the variability of the expected NPVs from different 20 starting points. The box on the left presents the inverted 5-spot pattern
with only well rate optimisation while the middle box represents EnOpt for both well type and rate optimisation, and the right box
presents the 5-spot pattern.
SPE-173829-MS 19

Figure 19 Box plots of the expected NPVs from 50 realisation for several optimisation iterations are shown. The range of the box
represents the variability of the expected NPVs from different 20 starting points. The box on the left presents the inverted 5-spot pattern
with only well rate optimisation while the middle box represents EnOpt for both well type and rate optimisation, and the right box
presents the 5-spot pattern.

Impact of localisation
Well control rates can be correlated in time and space. This study however considers only the impact of
a correlation length in time. We first make assumption that the rates of any two wells are uncorrelated,
which means that the information of the control rates from one well does not have an influential effect on
the control rates of the other wells. We mainly consider the effect of the smoothness imposed on the
control rates within the same well at different time. Therefore, the localisation matrix Cx is a block
diagonal matrix with off-diagonal values set to zero, as shown below:
(14)

(15)

where the covariance matrix Cvpr is a function of the reservoir volume flow rate for each well at each
time step i and j and Nt stands for total number of time steps; the covariance matrix Cinj is dependent on
injection rate for each well at each time step and the matrix Ct is the corresponding covariance matrix of
the conversion time.
Case B: Localisation Studies of Homogeneous Permeability Field, in an Inclined Reservoir This
section presents the impact of the localisation length for the same starting points where the initial point
is randomly drawn from a uniform distribution. We used the EnOpt method with the reservoir shown in
20 SPE-173829-MS

Fig. 2 with a homogeneous permeability field. Figure 20 display the NPV for four correlation length
imposed on the control rates. We see that the shorter correlation length provide the higher NPV.

Figure 20 The expected NPV over 50 realisation is plotted versus the iteration of optimisation. Each curve presents the NPV obtained
from different correlation lengths.

For a comparison of each optimisation iteration, Fig. 21 shows the variability and the expected values
of the NPV for 50 realisations at initial, first, third, and final iteration. The initial ensemble of 50
realisations are the same for each localisation length. Therefore, all the boxes in Fig. 21(a) are identical.
Even though we begin with the same set of 50 realisations, the optimisation with the shorter correlation
length quickly outperforms the other lengths. A smaller correlation length tends to have a better gradient
approximation by keeping highly correlated variables and removing weak correlations. By choosing the
correct length, we are able to screen out spurious correlations from the use of small ensemble size.

Figure 21Box plots of the NPVs from 50 realisation for several iterations of optimisation are shown. Each box in the sub-figure
represents the NPVs obtained from different correlation lengths.
SPE-173829-MS 21

The similar experiment for 20 different starting points is repeated here for different correlation lengths
in Fig. 22 and with the side-by- side correlation length with the optimisation iteration in Fig. 23. We see
that the shorter correlation length tends to outperform the longer one for both Figs. 22 and 23. Since we
use a relatively large time step of 6 months for the optimisation, it is reasonable that correlation of the
control rates between time steps are likely to be a yearly basis (2 time steps) or less. For this reason, we
expect to observe the different results if we select the shorter time steps for each control, but we have not
studied this effect in this work.

Figure 22Box plots of the NPVs from 50 realisation for several optimisation iterations are shown. The boxes in the sub-figure presents
NPVs obtained from the EnOpt method using different correlation lengths.

Figure 23Expected NPVs over 50 realisation from different 20 starting points for several iterations with different degree of correlation
length, i.e., 1, 4, 6, 10.
22 SPE-173829-MS

Conclusions
Our results indicate that optimising well types and well control rates simultaneously is critical to obtaining
significantly improved project economics and sweep efficiencies in comparison to simply either optimis-
ing only well controls or only well configurations. Our result are summarised as follows:
1. Well type optimisation could improve the NPV in comparison to the fixed well type waterflooding
strategy such as the inverted 5-spot and 5-spot pattern. We highlighted the significance of well
type optimisation to the NPV.
2. The superior performance of simultaneous well type and well rate optimisation is indicated by the
increase in NPV, cumulative oil production, and sweep efficiency within a predefined waterflood-
ing period. This is significant because we have a positive impact on project economics and
recovery of precious resources without the environmental impact associated with drilling new
wells.
3. We presented that the performance of EnOpt with simultaneous well type and control optimisation
with the known initial point gave the highest NPV. The selection of the starting point is critical
for gradient-based algorithms and this is often governed by prior knowledge. However, even with
random initial configurations, the performance of our algorithm is clearly superior to the tradi-
tional arrangement of wells using the 5-spot and inverted 5-spot pattern.
4. We showed that the correlation length of the control rates could impact the optimised NPV.
However, this is also dependent on the chosen control step. We found that using the shorter time
step within 6 months to 1 years provided a higher NPV than one using the longer correlation
lengths.

Acknowledgement
The authors thank Dr. Deepak Devegowda at the University of Oklahoma and Prof. Dean Oliver at the
University of Bergen for the guidance and valuable discussion for this project. Funding for Petvipusit,
provided by OU Consortium on Ensemble Methods and Mewbourne School of Petroleum and Geological
Engineering (MPGE) at the University of Oklahoma, is greatly appreciated. Computer resources were
provided by the OU Supercomputing Centre for Education and Research. Academic ECLIPSE licenses
were donated by Schlumberger.

Nomenclature
x Control variable
S(x) Objective function
f NPV function
Cx Covariance matrix of control variables
Cf, x Cross-covariance matrix of control variables and the NPV function
weight factor for the objective function

Abbreviations
BHP Bottom-Hole Pressure
EnOpt Ensemble-based Optimisation

References
1. Alhuthali, A., Optimal waterflood management under geologic uncertainty using rate control :
Theory and field applications, SPE Annual Technical Conference and Exhibition held in New
Orleans, Louisiana, 4-7 October 2009, 2009.
SPE-173829-MS 23

2. Alhuthali, A., O. A., and A. Datta-Gupta, Optimal waterflood management using rate control,
SPE Reservoir Evaluation & Engineering, 10(5), 539 551, 2007.
3. Asadollahi, M. and G. Naevdal, Waterflooding optimization using gradient based methods,
SPE-125331, Proceedings of the SPE/EAGE Reservoir Characterization and Simulation Confer-
ence, 19-21 October 2009, Abu Dhabi, UAE, 2009.
4. Asheim, H., Maximization of water sweep efficiency by controlling production and injection
rates, SPE European Petroleum Conference. London, UK, SPE-18365, 1988.
5. Brouwer, D. R. and J. D. Jansen, Dynamic optimization of waterflooding with smart wells using
optimal control theory, SPE Journal, 9(4), 391402, 2004.
6. Chen, Y, Ensemble-based closed-loop optimization optimization, Ph.D. thesis, University of
Oklahoma, Norman, Oklahoma, 2008.
7. Chen, Y., D. S. Oliver, and D. Zhang, Efficient ensemble-based closed-loop production optimi-
zation, SPE Journal, 14(4), 634 645, 2009.
8. Leeuwenburgh, O., P. J. P. Egberts, and O. A. Abbink, Ensemble methods for reservoir life-cycle
optimization and well placement (SPE 136916), Proceedings of SPE/DGS Annual Technical and
Symposium and Exhibition, 04-07 April 2010, Al-Khobar, Saudi Arabia, 2010.
9. Li, L. and B. Jafarpour, A variable-control well placement optimization for improved reservoir
development, Computational Geosciences, 16(4), 871889, 2012.
10. Li, L., B. Jafarpour, and M. R. Mohammad-Khaninezhad, A simultaneous perturbation stochastic
approximation algorithm for coupled well placement and control optimization under geologic
uncertainty, Computational Geosciences, 17(1), 167188, 2013.
11. Lien, M., D. Brouwer, T. Mannseth, and J. Jansen, Multiple regularization of flooding optimi-
zation for smart field management, SPE-99728, SPE inteligent Energy Conference and Exhibition
held in Amsterdam, The Netherlands, 11-13 April 2006, 2006.
12. Nwaozo, J., Dynamic optimization of a water flood reservoir, Masters thesis, University of
Oklahoma, 2006.
13. Petvipusit, R., Dynamic Well Scheduling and Well Type Optimization Using Ensemble based
Method (EnOpt), Master thesis, University of Oklahoma, 2011.
14. Sarma, P, K. Aziz, and L. J. Durlofsky, Implementation of adjoint solution for optimal control of
smart wells, SPE-92864, 2005.
15. Su, H. and D. S. Oliver, Smart well production optimization using an ensemble-based method
(SPE 126072), SPE Reservoir Evaluation & Engineering, 2010.
16. Tarantola, A., Inverse problem theory and methods for model parameter estimation, Society for
Industrial Mathematics, 2005.
17. Wang, C., G. Li, and A. C. Reynolds, Production optimization in closed-loop reservoir manage-
ment, SPE-109805, SPE Journal, 14(3), 506 523, 2009.
18. Yeten, B., D. Brouwer, and K. Aziz, Decision analysis under uncertainty for smart well
deployment, Journal of Science and Engineering, 44, 175191, 2004.

Вам также может понравиться