You are on page 1of 28

Selecting Representative Models for Ensemble-Based

Production Optimization in Carbonate Reservoirs with


Intelligent Wells and WAG Injection
S.M.G. Santos1*, A.A.S. Santos1, D.J. Schiozer1
1
University of Campinas

Summary
Production optimization under uncertainty is complex and computationally demanding, a particularly challenging
process for carbonate reservoirs subject to WAG injection, represented in large ensembles with high simulation
runtimes. Search spaces of optimization are often large, where reservoir models are complex and the number of
decision variables is high. The computational costs of ensemble-based production optimization can be decreased
by reducing the size of the ensemble with representative models (RM). The validity of this method requires that
the RM maintain representativeness throughout the optimization process, where the production strategy changes
at each evaluation. Many techniques of RM selection use production forecasts of the ensemble for an initial
production strategy, which raises questions about the robustness of the RM. This work investigates approaches to
ensure the consistency of RM in ensemble-based long-term optimization. We use a metaheuristic optimization
algorithm that finds sets of RM that represent the ensemble in the probability distribution of uncertain attributes
and the variability of production, injection, and economic indicators (Meira et al., 2020). Our case study is a
benchmark light-oil fractured carbonate with features of Brazilian pre-salt reservoirs and many reservoir and
operational uncertainties. We obtained production, injection and economic indicators using different approaches
to provide valuable insight for RM selection. We inferred about RM fitness for production optimization based on
their adequacy for uncertainty quantification for varying production strategies. Despite the effects of changing
decision variables on RM representativity, our results suggest the possible use of RM for ensemble-based
production optimizations with limitations related to the estimation of the probabilistic objective function due to
mismatches in the probabilities of occurrence. Using production indicators obtained from a base production
strategy decreased RM representativeness when compared to RM selection based on a more robust evaluation of
reservoir performance using a wide-covering well pattern and no restrictions from production facilities. Finally,
our results suggest valid RM selection using production forecasts for intermediate dates of the simulation period,
an important contribution for ensembles with very high simulation runtimes. We also provide a broad theoretical
background on the uncertain reservoir system and on approaches to obtain reduced ensembles and their
applications.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Introduction

The reservoir simulator (also referred to as flow simulator) is known to be one of the best tools to
evaluate the performance of a production strategy over time for a given reservoir scenario. The physical
system (reservoir) is represented in a mathematical model, which is then solved numerically to predict
fluid behaviour over time. This approach can be classified as physics-based as it incorporates the
physical laws governing fluid flow in porous media. Despite computationally expensive, it has the
unique advantage of modelling the complex integration between the production strategy and the
performance of the reservoir system (certain or uncertain).

Despite undeniably important, performing production optimization using reservoir simulation models
is challenging. Defining a production strategy involves defining a large set of variables related to the
production system, including subsurface and surface facilities. In addition, the problem is often
characterized by large search spaces, which may demand thousands of evaluations of the optimizer's
objective function. Given that each evaluation of the objective function equals to one reservoir
simulation, which may take many hours, a production optimization problem may become unfeasible
depending on the available computational infrastructure. This is particularly relevant for production
optimization under uncertainty, such as robust optimization, where optimization is conducted on an
ensemble of different reservoir scenarios. In such cases, each evaluation of the objective function equals
to multiple reservoir simulations (one for each reservoir scenario).

Added challenges exist in Brazilian pre-salt reservoir, light-oil fractured carbonates subject to water-
alternating-gas (WAG) injection, and with intelligent wells. First, representing these reservoirs into
simulation models leads to high simulation runtimes, due to the inherent heterogeneities of carbonate
reservoirs and the physical phenomena related to WAG recovery. Second, due to difficulties related to
reservoir characterization under uncertainty, large ensembles of scenarios are required to ensure that
uncertainty is fully captured. Lastly, the production optimization problem has added variables related
to the WAG injection strategy and intelligent well design and operation.

Different approaches are available in the literature to reduce the computational runtime of each
evaluation during production optimization processes. These can be classified into three mains
categories: (1) simplifications that capture the physics of the process (physics-based); (2)
simplifications that do not capture the physics of the process (data-driven); and (3) hybrid models that
combine physics-based and data-driven approaches.

The most common simplification in the physics-based category is the use of low-fidelity models,
defined as reservoir simulation models that have strong simplifications in rock and/or fluid properties
(e.g., Wilson and Durlofsky, 2013; Aliyev and Durlofsky, 2015). Examples include the use of upscaled
simulation grids and a reduced number of pseudo-components in fluid models. The prime advantage of
this approach is that it easy to implement. However, sub-grid heterogeneity effects and phenomena
related to gas miscibility may be missed

Another class of simplifications within the physics-based category is the use of reduced ensembles,
where a selected number of models is chosen from a large ensemble of uncertain reservoir scenarios
(typically hundreds) (e.g., Schiozer et al., 2004; Scheidt and Caers, 2009; Shirangi and Durlofsky, 2016;
Mahjour et al., 2020a; Meira et al., 2020). Different approaches exist to obtain reduced ensembles,
which can be regarded as a broad class that includes representative models (RM). The prime advantage
of this approach is that it does not require simplification in reservoir representation in the simulation
model, as we have seen for low-fidelity models. However, ensuring that uncertainty is properly
represented in a reduced ensemble is a difficult and complex process.

The most common simplification in the data-driven category is the use of data-fit proxy models, defined
as analytical functions that bypass the reservoir simulator when predicting production forecasts (e.g.,
Scheidt et al., 2007; Feraille and Marrel, 2012; Douarche et al., 2014; Panjalizadeh et al., 2014). The
prime advantage of this approach is that it is extremely fast, allowing thousands of evaluations of the

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
optimizer’s objective function. However, by disregarding the physical laws governing fluid flow in the
reservoir, it is more difficult to validate and ensure reliable forecasts.

Even though often required, using simplifications bears the risk of adding positive or negative biases to
production forecasts and in decision-making. This is because biases are already inherent to the processes
of reservoir representation in simulation models and production strategy definition (Meddaugh et al.,
2017). Thus, the use of simplifications must be chosen carefully and be treated as case dependent.

Why representative models?


Reduced ensembles appear as an interesting approach to reduce the computational intensity of
ensemble-based production optimization in the context of light-oil fractured carbonates subject to
WAG. As the degree of model fidelity is maintained, critical heterogeneities and phenomena related to
WAG injection are adequately represented. In such cases, simpler low-fidelity models and data-fit
proxy models are likely to produce inaccurate forecasts.

However, the validity of solutions obtained for reduced ensembles, such as production optimization,
lays on the premise that the reduced ensemble represents properly the range of uncertainties. Following
the fundamental concept behind RM selection based on percentiles, current methods often base RM
selection on production, injection, and economic indicators obtained for the full lifespan of the reservoir
system, considering an outlined production strategy. However, given that the RM are selected to
perform uncertainty-based analyses (such as ensemble-based production optimization) and not to
evaluate a single production strategy (such as that for the base case), other approaches may be of value.
This includes investigating the need for more robust approaches for evaluating reservoir performance,
such as those related to minimizing potential biases introduced by the use of an outlined production
strategy.

Moreover, during production optimization processes, the production strategy is dynamic, meaning that
different possibilities for the set of variables are sampled at each evaluation. Thus, the RM used for
optimization should maintain the representativity of the uncertain system throughout the process.

Existing RM selection methods must be applied carefully, as not to affect the quality of solutions, i.e.,
the representativeness of the RM may be impaired depending on the input data used for RM selection.
This raises the need for additional studies to investigate the best way to apply existing RM selection
methods aiming to maximize uncertainty representation. Furthermore, despite the increasing number of
works in the literature, further studies are still required looking at current optimization challenges
because most applications consider water flooding and conventional wells.

Scope and Objective


This work investigates approaches to improve RM selection for long-term production optimization
under uncertainty, considering the optimization problems related to the design and operation of
intelligent wells combined with a WAG strategy. We investigate different ways to apply an existing
RM selection method, called RMFinder (Meira et al., 2020), aiming to ensure that full uncertainty is
represented in the RM. We focus on understanding the user decisions that may hamper uncertainty
representation in the set of RM, aiming to ensure an adequate set of RM for future works on ensemble-
based optimization.

The main idea behind RMFinder is the selection of a set of RM that represents both the probability
distribution of uncertain attributes and variability of production, injection, and economic indicators of
the full ensemble. Due to the complex integration between reservoir performance and production
strategy, obtaining these indicators is not a straightforward process. This is because they require the
definition of one production strategy (such as number and placement of wells, and subsurface facilities)
and this choice will affect the variability in production, injection, and economic indicators. In this study,
we obtained indicators of reservoir performance using different approaches to provide decision-makers
with valuable insights into the process of RM selection.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
This work also provides a broad theoretical background on the uncertain reservoir system, its inputs
and outputs, on approaches to obtain reduced ensembles and their applications, as well as some
definitions on production strategy definition considering intelligent wells, WAG recovery, and gas
recycle.

The case study presented in this work is an open-source benchmark in the development phase called
UNISIM-II-D (Correia et al., 2015). Added challenges exist depending on the composition of the
injected gas, such as that in WAG-CO2. Despite the target recovery mechanism of the reservoir
UNISIM-II being WAG-CO2, we did not address it in this work. Instead, we used a low-fidelity model
based on a black-oil fluid model but maintaining geological representation (grid properties) to allow
computational feasibility of different evaluations of the full ensemble (composed of hundreds of
scenarios) for varying decision variables. The simulation model is called UNISIM-II-D-BO, where BO
stands for black-oil.

Note that Meira et al. (2020) demonstrated the use of the RMFinder for the same benchmark case used
in this work, the UNISIM-II-D. However, their application considered the use of conventional wells
and the recovery mechanism was water flooding, while our work considers the adequacy of RMFinder
for a production strategy with WAG injection and intelligent wells. Thus, we do not use the results
presented by Meira et al. (2020).

Theoretical background

Basic concepts and terminology


Let us consider a simulation model of the reservoir and production facilities as a system, composed of
input variables (inputs, for short) and output variables (outputs, for short). Here, system inputs are the
geological, reservoir, fluid, operational, and economic data, in addition to wells and surface facilities.
The outputs are production, injection, and economic indicators (Figure 1).

Np • Time series of field


production indicators
• Time series of field
injection indicators
time
• Grid properties • Time series of field
RF economic indicators
• Rock properties
• Fluid data System Reservoir Economic System • Time series of well
• Rock-fluid data inputs simulator calculator outputs production/injection
• Economic data indicators
time
• … (...) • Time series of well
Gp economic indicators
• …

time

Figure 1 Definition of inputs and outputs of a reservoir simulation system.

In a typical reservoir system, several reservoir uncertainties exist because reservoir and fluid data are
mainly acquired indirectly (e.g., seismic surveys) and, when directly, sparsely (e.g., well coring).
Additionally, because developing a petroleum field is a long-term project, reservoir systems have added
uncertainties related to operational availability of production facilities (e.g., failures in equipment,
maintenance), economic data (e.g., oil price, capital and operational expenditures), among others. As a
result, instead of having one value for each system input, we may have a range of values characterized
by a probability distribution.

Uncertainty in continuous variables, such as the fluid contacts, is described in probability distribution
functions (PDF) while in discrete variables, such as a structural model and geostatistical realizations, it
is described in probability mass functions (PMF). To bring these properties to the simulation system,
we need to create an ensemble of uncertain scenarios, where a scenario is defined here as a particular
combination of one value for all uncertain variables (the system inputs, i.e., geological, reservoir, fluid,

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
operational, and economic data). This definition of uncertain scenario should not be confused with the
definition of geological scenario used in geological modelling, required to generate geostatistical
realizations (i.e., equiprobable spatial distributions of petrophysical properties). These ensembles are
often a simplification of uncertainty representation because combining all possible values of all
uncertain variables would lead to an exponentially high number of scenarios.

Scenarios are generated using statistical sampling techniques such as the Monte Carlo and the
discretized Latin Hypercube combined with geostatistical realizations (DLHG) (Schiozer et al., 2017).
The prime limitation of Monte Carlo sampling, which performs pure random sampling, is that there is
no guarantee that the full range of the distribution is sampled evenly and consistently with a sampling
number lower than many thousands (Mishra, 1998).

Latin Hypercube sampling (LHS) has gain popularity because it incorporates the desirable features of
random sampling and stratified sampling (Helton and Davis, 2003). The DLHG applies the efficient
LHS and integrates all types of uncertainties in the sampling step, i.e., continuous attributes are
discretized and combined with discrete attributes and geostatistical realizations (Figure 2). Thus, the
DLHG appeared as a promising technique to increase the computational efficiency of the sampling
process in reservoir engineering problems, ensuring adequate representation of all uncertainties with
few hundreds of scenarios (Santos et al., 2018a).

• Geostatistical
realizations

• PDF of fluid data


Uncertain DLHG
• PDF of rock-fluid Ensemble
variables sampling
data
• PDF of economic
data (...)
• …

Figure 2 Generation of an ensemble of scenarios using DLHG sampling.

Hence, the inputs of the uncertain system are an ensemble of hundreds or thousands of scenarios, each
one characterized by a combination of the possible values of each uncertain variable (system input).
Forecasting reservoir and economic performances of the ensemble, we obtain a range of values for each
system output (Figure 3). These outputs consist of, for example, production and injection curves over
time, estimated recovery factor, and discounted net cash flows. These ranges of possible outputs can be
used to construct, for a given time, risk curves, which are also referred to as descending or
complementary cumulative distribution functions (CCDF) in the statistics literature.

Generating the outputs of an uncertain system using numerical reservoir simulation may be
computationally expensive or prohibitive, especially for procedures related to production optimization
that demand thousands of evaluations. Over the decades, several approaches have been investigated to
reduce such costs, as we have explored in the Introduction section. A promising line of research for
complex reservoirs consists of reducing the size of the ensemble, which we address in this work.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Np • Time series and
• Ensemble of CCDF of field
scenarios production indicators
• Time series and
time CCDF of field
RF injection indicators
• Time series and
System Reservoir Economic System CCDF of field
inputs simulator calculator outputs economic indicators
time • Time series and
CCDF of well

Cum. prob.
production/ injection
(...) indicators
• Time series and
NPV CCDF of well
economic indicators
• …

Figure 3 Definition of inputs and outputs of an uncertain reservoir simulation system.

Reducing the size of the ensemble


The idea of using few scenarios to represent an uncertain system can be traced back to the classic
approaches based on three scenarios (optimistic, most-likely, and pessimistic) or on percentiles (P10,
P50, and P90) (Figure 4). The prime limitation of using percentiles as the criterion to select the reduced
ensemble is that it only reflects uncertainty in one variable, either input (e.g., pore volume) or output
(e.g., net present value). Although the use of three scenarios is still a common practice in some contexts,
the need for more comprehensive criteria for their selection became evident to ensure proper
representation of uncertainty in procedures related to production optimization and decision analysis.

P90

P50

P10

Figure 4 Risk curve for a system output X, marking the percentiles P10, P50, and P90.

During the past decade, a variety of studies have been published investigating different approaches to
obtain reduced ensembles. The focus shifted from selecting models close to P10, P50, and P90 for one
variable to selecting models that covered the entire variability of one or more variables of the uncertain
system. Today, the literature presents two main approaches to obtain reduced ensembles (Table 1).

Approach Description Examples from the literature


Selecting • The focus is on geological uncertainty • Scheidt and Caers (2009)
representative • RR are selected from the full set of realizations, disregarding • Yang et al. (2011)
geostatistical representation for all other system inputs and outputs • Armstrong et al. (2013)
realizations • RR are then combined with other uncertain variables (reservoir • Shirangi and Durlofsky
(RR) and fluid) to construct a reduced ensemble of scenarios (2016)
• This is the most common approach in the current literature • Mahjour et al. (2020a)
Selecting • The focus is on more than one system input and/or output • Schiozer et al. (2004)
representative • RM are selected from an ensemble of scenarios • Sarma et al. (2013)
simulation • RM can be selected from the full ensemble or from a reduced • Jiang et al. (2016)
models (RM) ensemble (such as that obtained from representative • Meira et al. (2020)
geostatistical realizations)
Table 1 Main approaches to obtain reduced ensembles.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
This work addresses the selection of representative models (RM). To the best of our knowledge, this
term was first coined by Schiozer et al. (2004). The idea behind the RM is that a few models with
different characteristics can represent an uncertain reservoir system. The work by Schiozer et al. (2004),
pioneer in the area of reservoir engineering, was based on manual selection of RM using cross-plots
(scatterplots) for four system outputs: net present value (NPV), oil recovery factor (RF), cumulative oil
production (Np), and cumulative water production (Wp) (Figure 5). The selection criterion was to
choose models with significant differences in RF that were close to P10, P50, and P90 values for NPV
(Figure 5a). Distributions of Np and Wp were used only to monitor and ensure wide variation in Np
and Wp (Figure 5b and Figure 5c). Based on these criteria, Schiozer et al. (2004) understood that nine
RM would be required to represent pessimistic, most-likely, and optimistic values for two system
outputs simultaneously (in that case, RF and NPV).

Figure 5 Manual selection of representative models based on cross-plots for NPV, Np, RF, and Wp (in
Schiozer et al., 2004).

With the increasing number of criteria for RM selection, the process of manual selection became
extremely time-consuming and difficult to achieve satisfying solutions. This raised the need for
automated procedures for RM selection, which lead to the development of an optimization-based
method implemented in an in-house program called RMFinder (Meira et al., 2016, 2017, 2020).

RMFinder uses a mathematical function that captures the representativeness of a set of models
combined with a metaheuristic optimization algorithm and performs global and local searches. The
main idea behind RMFinder is to select a set of RM that represents both the probability distribution of
the input variables represented by the ensemble (reservoir, operational and economic uncertainties) and
the variability of the main output variables (production, injection and economic indicators) of the
uncertain reservoir system (Figure 6).

Figure 6 Illustrative example of a set of nine RM (in red) selected from an ensemble of 214 scenarios
(in green). Bar charts compare the histograms of two system inputs (uncertain variables) in the set of
RM and the ensemble. Cross-plots (scatterplots) and risk curves (CCDF) make the same comparison
for four system outputs (NPV, RF, Np, and Wp).

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Thus, when selecting the RM, RMFinder takes into account cross-plots, risk curves, and the PMF, each
one accompanied by a user-defined weight that expresses their relative importance in RM selection.
RMFinder also defines the individual probability of each representative model and aims to guarantee
representativeness in dozens of variables.

Main applications of RM
Different processes related to the management of petroleum fields require probabilistic analysis, from
uncertainty quantification to uncertainty management, and from production strategy optimization to
production strategy selection under uncertain. The RM are an important way to make many of these
processes computationally feasible while ensuring adequate representation of uncertainty.

Today, the main applications of RM are those summarized in Table 2. The examples from the literature
listed in Table 2 used RMFinder to select the RM. All those works selected RM using system outputs
obtained for a base production strategy, despite the different applications for which the RM were
selected for.

Before selecting the RM, it is key to define properly the study objective and for which processes the
RM will be used, as it may affect the RM selection process (number of RM, weights of objective
functions, among other user-defined setups of RMFinder). A prime application of the RM (and the focus
of this work) is production optimization under uncertainty, which we address in the following
subsection.

Type of application Examples from the literature


Production optimization • RM nominal optimization (Schiozer et al., 2019; Santos et al., 2020)
under uncertainty • RM robust optimization (Silva et al., 2016)
Uncertainty quantification • Production forecasts under uncertainty to generate risk curves (Santos et al., 2020)
and risk analysis
Decision analysis to manage • Production strategy selection under uncertainty (Schiozer et al., 2015; Santos et al.,
petroleum fields 2020).
Uncertainty management in • Expected value of information (EVoI) analysis (Santos et al., 2020)
petroleum fields • Expected value of flexibility (EVoI) and Chance of Success (CoS) analysis (Ferreira
et al., 2015; Botechia et al., 2018; Gaspar et al., 2020)
Table 2 Summary of types of applications of representative models.

The production optimization problem


Production optimization procedures, such as those described in Table 2, aim to select one final
production strategy that will be implemented in a petroleum field. This includes defining a set of
decision variables related to reservoir development, management and the production facilities, such as:
(1) number, type, and placement of wells; (2) drilling schedules; (3) recovery mechanism; (4) number
and type of production units; (5) fluid processing capacities; and (6) operation and control rules of
equipment over time.

Reservoir properties and size, limitations in logistics of production systems, among other factors, define
the degree of complexity of production optimization processes. Brazilian giant pre-salt reservoirs,
which are light-oil fractured carbonates with high volumes of CO2-rich associated gas, typically involve
many challenges related to gas treatment and recovery mechanism, which increase significantly the
complexity of optimization problems. These cases are often classified as gas-driven production
optimization, where oil production is limited by restrictions in production units, not by reservoir
productivity.

WAG-CO2 injection is an interesting solution for such cases, joining the benefits of increased sweep
efficiency to the ability to safely discarding the CO2-rich associated gas. The decision variables related
to this recovery mechanism include the number of wells operating in the WAG scheme and the
specifications of the injection cycles (per time or injected volume).

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Intelligent wells, where each well is equipped with internal control devices (ICD) or inflow control
valves (ICV), is a technology often implemented in these reservoirs. Each ICV can be operated from
the surface to control fluid flow in production and injection wells, aiming to prevent and delay the
breakthrough of undesirable fluids (water and gas) in the wellbore. Decision variables related to
intelligent wells include ICV design (number and placement of ICV in the wellbore) and control rules
for their operation over time.

The optimization problem under uncertainty using RM can be classified into two main classes (nominal
and robust), as detailed in Table 3. These approaches are not interchangeable, meaning that they are
complementary in the decision-making process under uncertainty and, whenever possible, should be
performed both (Schiozer et al., 2019).

Approach Description Examples from the


literature
RM nominal • Nominal optimization to select a specialized production strategy • Hutahaean et al. (2019)
optimization • Optimization process applied to one RM, individually • Schiozer et al. (2019)
• Deterministic objective function (ex. NPV, Np) • Santos et al. (2020)
• The result is a set of specialized production strategies, i.e., one
optimal solution for each RM
• Provides a set of alternative solutions for field development and
management, giving important insights for uncertainty
management with information acquisition and flexible systems
RM robust • Robust optimization to select a robust production strategy • Yeten et al., (2003)
optimization • Optimization process applied to all RM, simultaneously • Yang et al. (2011)(1)
• Probabilistic objective function (ex. expected monetary value, risk • Silva et al. (2016)
measure)
• The result is one robust production strategy, i.e., the best solution
on average for all RM, but suboptimal for each RM alone
Table 3 Main approaches for production optimization under uncertainty using RM. 1Uses a reduced
ensemble based on RR instead of RM.

Methodology

In this study, we use RMFinder (Meira et al., 2020) to select the RM and investigate two main aspects
related to the generation of system outputs for RM selection: (1) the decision variables used for
obtaining production, injection, and economic indicators, and (2) the simulation time frame of the
reservoir system. These approaches are detailed in the following subsections.

Obtaining system outputs for RM selection: the decision variables of the reservoir system
We considered two approaches to obtain indicators of reservoir performance under uncertainty (Table
4). First, a base production strategy defined as the best solution found for the base case and obtained
from nominal optimization (an approach called here production strategy zero, or PS0). In the second
approach, we evaluate reservoir performance using a well pattern with wide reservoir coverage and with
no restrictions related to the production system, aiming to assess if this approach ensures unbiased
coverage of the full reservoir system (approach called production evaluation zero, PE0).

ID Approach Description Hypothesis being


evaluated
PS0 Evaluation of reservoir • A nominal optimization process is applied to the base case The base production
performance using a to obtain a base production strategy strategy is valid for
base production • Specialized solution for the base case RM selection
strategy • This is the most common approach in the current literature
PE0 Evaluation of reservoir • A production strategy is not implemented RM selected from
performance using a • A 5-spot well pattern is defined covering the entire these system outputs
well pattern with wide simulation grid are representative for
reservoir coverage • No production unit is installed or any physical restrictions all production
related to production systems for potential unbiased strategies
coverage of the full reservoir system
Table 4 Approaches for obtaining system outputs for RM selection: the reservoir decision variables.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
For the nominal optimization of the base production strategy (PS0), we used an assisted hierarchical
procedure, as follows: (1) number and type of wells and size of the production unit (fluid processing
capacities), (2) well placement and fine-tuning the production unit constraints, (3) drilling schedule and
simplified scheme of WAG cycles (initial estimate), (4) ICV design (number and placement of ICV),
(5) reservoir and well-operating control rules over time, (6) ICV control rules for producers over time,
(7) WAG injection strategy and ICV control rules for injectors over time, and (8) fine-tuning of ICV
control rules for producers and size of the production unit. The optimizer used in this work is
implemented in an in-house software and is based on the iterative Latin hypercube sampling (von
Hohendorff Filho et al., 2016).

Obtaining system outputs for RM selection: the reservoir simulation time frame
We selected RM using system outputs obtained for different dates throughout the lifecycle of the
petroleum reservoir, ranging from the first years of the production phase until field abandonment (Table
5). Here, the focus is on understanding if there is value in using an intermediate date of the simulation
time frame in RM selection. In addition to the evident potential to reduce computational costs, reservoir
performance at field abandonment is strongly dependent on the production strategy, meaning that few
years of production may suffice to predict the productivity of a reservoir scenario.

Approach Description Hypothesis being evaluated


Production, injection, and • Reservoir performance is forecasted for the entire RM selection must follow the
economic indicators at lifecycle of the field, from reservoir development to same fundamentals of the
field abandonment (tf) abandonment classic approach based on
• This is the most common approach in the current percentiles
literature
Production, injection, and • Reservoir performance is forecasted for a partial part Few years of production
economic indicators at an of the field lifetime, from reservoir development until suffice to predict reservoir
intermediate date of the some date after the completion of the development performance for a given
simulation time frame (ti) phase and before field abandonment reservoir scenario
Table 5 Approaches for obtaining system outputs for RM selection: the reservoir simulation time frame.

Quantification and validation


To assess the validity of each approach, we perform qualitative and quantitative analysis based on risk
curves for representative models when compared to those for the full ensemble. Despite evaluating
different intermediate dates for RM selection, our objective is always to ensure representativeness for
the time frame of production optimization (long-term optimization, in our case). Thus, we always assess
the adequacy of the RM in representing risk curves at field abandonment.

Due to the reduced number of data points, risk curves for RM are drawn using a stair-step graph. To
quantify the approaches used for RM selection, we determine the area between the RM and the ensemble
risk curve, called absolute error. To allow comparison across different indicators and sets of RM, we
normalize the absolute error by the standard deviation of the ensemble and call it normalized error. An
illustrative example is provided in Figure 7.

At the end of the study, we test the validity of each approach by applying the different sets of RM to a
control production strategy used only for validation (called here production validation, or PV), i.e., not
used for RM selection.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 7 Example of risk curve for the ensemble (continuous curve in blue) and the RM (stair-steps in
red). The coloured area in grey between curves is used as the error measure in quantitative analysis.

Case study

Our case study is the UNISIM-II-D-BO, an open-access benchmark created by the UNSIM research
group and made available to the scientific community. Using public data, the available literature and
analogues, Correia et al. (2015) constructed a synthetic carbonate fractured reservoir with features found
in Brazilian pre-salt fields, such as high-permeability thin layers, commonly known as super-k.

The case study is in the initial field development phase under uncertainties and has 1.5 years of
production data for the Wildcat. The reservoir is represented in dual-permeability simulation models
with a corner-point grid with 46 x 69 x 30 cells measuring, on average, 100 x 100 x 8 m (total 64,110
active cells). Reservoir depth varies between 5,000 m and 5,500 m from sea level, initial pressure is 450
kgf/cm2, the temperature is 59°C, and oil viscosity is 1.14 cP (28° API).

The case study has a set of geological and reservoir uncertainties that were combined using the DLHG
sampling to generate scenarios that, after ensemble-based data assimilation procedures, resulted in an
ensemble of 199 scenarios (Table 6). These reservoir scenarios were then combined with operational
uncertainties (Table 7) using the DLHG to generate 199 scenarios that include both reservoir and
operational uncertainties. A base case (base0) was selected from the ensemble as the model closest to
P50 for field indicators.

Table 8 shows the economic data used for the NPV calculations. Platform investment (Invplat), in US$
millions, is given by Equation (1) (Santos et al., 2018b) where Cpo is the oil processing capacity (103
m3/day), Cpl is the liquid processing capacity (103 m3/day), Cpw is the water processing capacity (103
m3/day), Ciw is the water injection capacity (103 m3/day), Cpg is the gas processing capacity (106
m3/day), and nw is the number of well slots.

𝐼𝑛𝑣$%&' = 417 + 12.2 × 𝐶𝑝2 + 3.15 × 𝐶𝑝% + 3.15 × 𝐶𝑝5 + 3.15 × 𝐶𝑖5 + 9.61 × 𝐶𝑝9 + 0.1 × 𝑛5 (1)

Uncertain Type (PMF1) Description (probability)


system input
199 geostatistical realizations of porosity, matrix permeability, fracture
Geostatistical realizations2
Static properties permeability, fracture spacing, net-to-gross ratio, and rock type, and
(uniform)
include uncertainty in reservoir top and bottom (0.005)
Relative Table2 kr[-1] kr[0] kr[+1]
permeability (triangular) (0.32) (0.41) (0.27)
Rock Discretized scalar 19.1 x 106 56.0 x 106 79.8 x 106
compressibility (triangular) (0.22) (0.58) (0.20)
Table 6 Geological and reservoir uncertainties and their probabilities, represented in an ensemble of
199 scenarios after data assimilation procedures (modified from Santos et al., 2018b). 1Probability
mass function. 2Files available at https://www.unisim.cepetro.unicamp.br/benchmarks/en/unisim-
ii/unisim-ii-d

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Uncertain system input Type (PMF1) Value (probability)
-1 (0.33) 0 (0.33) +1 (0.34)
Group availability Discretized scalar (uniform) 0.91 0.96 1.00
Platform availability Discretized scalar (uniform) 0.90 0.95 1.00
Production well availability Discretized scalar (uniform) 0.91 0.96 1.00
Injection well availability Discretized scalar (uniform) 0.92 0.98 1.00
Well index multiplier Discretized scalar (uniform) 0.70 1.00 1.40
Table 7 Operational uncertainties of the UNISIM-II-D-BO case study in Santos et al., 2018b).
1
Probability mass function.

Type Attribute (unit) Value


Market Oil price (US$/m3) 257.9
variables Annual discount rate (%) 9.00
Royalties (%) 10.00
Taxes Social taxes (%) 9.25
Corporate taxes (%) 34.00
Oil production cost (US$/m3) 48.57
Water production cost (US$/m3) 4.85
Water injection cost (US$/m3) 4.85
OPEX
Gas production cost (US$/m3) 0.013
Gas injection cost (recycled gas) (US$/m3) 0.014
Abandonment (% of well investment) 8.20
Drilling of vertical well (US$ millions) 23.40
Completion of vertical well (US$ millions) 26.94
Well-platform connection (US$ millions) 13.30
CAPEX
1st Inflow Control Valve (ICV) (US$ millions) 1.00
2nd and 3rd ICV (US$ millions) 0.30
Production unit (US$ millions) Eq. (1)
Table 8 Economic model of the UNISIM-II-D-BO case study (modified from Santos et al., 2018b).

We calculated the NPV using a simplified net cash flow formulation based on the Brazilian concession
fiscal regime (Equation (2)) where NCF is the net cash flow, Rev is the gross revenue, Roy is the amount
paid in royalties, ST is the amount paid in social taxes, OPEX are operational expenditures, T is the
corporate tax rate, CAPEX are investments on equipment and facilities, and AC is the abandonment
cost.

𝑁𝐶𝐹 = [(𝑅𝑒𝑣 − 𝑅𝑜𝑦 − 𝑆𝑇 − 𝑂𝑃𝐸𝑋) ∗ (1 − 𝑇)] − 𝐶𝐴𝑃𝐸𝑋 − 𝐴𝐶 (2)

Results and discussion

Generating the system outputs for RM selection


We performed nominal optimization to obtain one specialized solution for the base case (base0). This
production strategy is called PS0 and has 10 producers and 8 WAG injectors. All wells are vertical and
equipped with ICV, except the Wildcat. Producers have three multi-position ICV and injectors have
two (the top ICV is multi-position, while the bottom is on/off). After optimizing the well placement, we
defined ICV placement using permeability logs for pseudo-wells obtained from base0.

The production unit has the following capacities: Cpo and Cpl 180,000 bbl/day, Cpg 8,000,000 m3/day,
Cpw 50,000 bbl /day, Ciw 150,000 bbl/day, and 18 well slots. All gas produced is reinjected without
separation of gas streams. Water injection rates were controlled to maintain the hydrocarbon volume
weighted average at the initial reservoir pressure.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
We defined reactive control rules for ICV installed in producers using triggering conditions evaluated
every 6 months and based on the gas-oil ratio (GOR). ICV rules in injectors were optimized together
with the WAG injection strategy and were defined based on streamlines to control gas production at
producers. During the first 10 years of production, all the eight injection wells operate within the WAG
scheme. From that date onwards, six wells operate within the WAG scheme, one operates as a gas
injector and another as a water injector. WAG cycles are of 6 months.

For the approach called PE0, we did not perform optimization. Instead, we placed 56 wells in a 5-spot
well pattern that covered the entire reservoir grid (modified from Mahjour et al., 2020b). To capture the
phenomena related to WAG recovery and the restrictions related to gas recycle, we defined the 28
injection wells as WAG injectors operating in 6-month cycles. Due to the inherent vertical
heterogeneities of carbonate reservoirs and the uncertainty in reservoir top and bottom, we did not place
ICV in the wellbore to permit unconstrained assessments of reservoir performance and reduce potential
biases introduced during ICV placement and operation.

Figure 8 shows the well placement for PS0 and PE0. We used a commercial black-oil reservoir
simulator and an in-house economic calculator to obtain production, injection, and economic forecasts
until field abandonment for the ensemble of 199 scenarios using approaches PS0 and PE0. Then, we
truncated these outputs in different time series based on the timeline of the case study (Figure 9), as
illustrated in Figure 10 for Np. We stored these results in a database and used them for RM selection,
which did not require any additional reservoir simulation runs.

UNISIM-II-D - PS0 UNISIM-II-D - PE0

p20 p21 p22 p23


PRK085 IRK063
IRK056 i25 i24 i26 i27 i28
PRK083 PRK084 p27
p15 p16 p17 p18
34 34
i20 i19 i21
PRK052 i22 i23
31 31
PRK060
p11 p12 p13 p14 p26
IRK050 IRK049 28 28
i18 i14 i13 i15 i16 i17
25 25
IRK028 p19 p08
p07Wildcat p09 p10
IRK036 Wildcat PRK061 21 21

18 i09 i08 i10 i11 i12 18


IRK029
PRK045 15 p03 p04 p05 p06 15

12 i06 i07 i03 i02 i05 12


PRK028 8 p01 p02 8
0.00 0.25 0.50 0.75 1.00 mile 0.00 0.25 0.50 0.75 1.00 mile
5 i01 i04 5
0.00 0.50 1.00 km PRK014
IRK004 0.00 0.50 1.00 km
2 p25 p24 2

Figure 8 Map of total oil per unit area for the base case, showing the well placement used for obtaining
system outputs in approaches PS0 (left) and PE0 (right).

Initial date of End of


reservoir production Begin of field End of field 31/12/2023 31/12/2025 31/12/2035
(tm+14) Field
simulation history development development (tm+2) (tm+4)
abandonment

30/09/2016 01/03/2018 31/12/2021 31/12/2022 31/12/2024 31/12/2030 31/12/2040 30/09/2046


(td) (tm) (tm+1) (tm+3) (tm+9) (tm+19) (tf)
28/02/2018

Figure 9 Timeline of UNISIM-II-D-BO, marking in yellow the intermediate and final dates used for RM
selected. The timeline is not-to-scale.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 10 Example of four out of the eight Np time series used for RM selection.

Selecting the RM
We selected nine RM using the system outputs stored in a database: six field indicators (Np, Wp, Gp,
Wi, RF, NPV) and 18 well indicators for PS0 (one indicator per well), while 56 well indicators for PE0
(again, one per well), totalling 24 system outputs for PS0 and 62 for PE0. The indicator used here for
the well performance is the well economic indicator (WEI) (Ravagnani et al., 2011), which quantifies
the individual contribution of each well to the field NPV (i.e., CAPEX, OPEX, and revenues, if
applicable). It is desired to maximize WEI for producers (reflecting maximum hydrocarbon production)
and to minimize WEI for injectors (reflecting maximum fluid injection).

Given that RMFinder uses a probabilistic optimization method with a random component, we executed
the method five times to selected five different solutions of RM (S1 to S5) to ensure the robustness of
results obtained for each approach. As our analysis focused on risk curve representation, we attributed
a higher weight to this parameter when selecting the RM. We used sets of nine RM, as recommended
by Schiozer et al. (2004).

Despite evaluating different intermediate dates for RM selection, our objective is always to ensure
representativeness for the time frame of production optimization (long-term optimization until field
abandonment, in our case). Thus, we always assessed the representativeness of the RM when compared
to the risk curves of the ensemble at field abandonment (not at the intermediate date).

Evaluating the approaches for RM selection


Figure 11 gives an overall summary of the analyses conducted in this work: (a) to the left, results for
RM selection using system outputs of PS0 at different dates; and (b) to the right, results for RM selection
using system outputs of PE0 for the same dates. Boxplots characterize the range of errors (i.e., the
normalized area between the RM risk curve and the ensemble risk curve for production strategy PS0 at
field abandonment, or 2046) for all system outputs (field and well indicators) for a given set of RM.

Figure 11 shows the results for solution S1. We chose this solution blindly without observing the quality
of each solution in representing the targeted risk curves (i.e., risk curves for PS0 at field abandonment).
The blue curve in Figure 11 marks the mean normalized error obtained for the five solutions (S1 to S5).

Risk curves using RM selected from PE0 recorded higher errors than those obtained directly for PS0.
This was expected to a certain degree, given that minimum errors were anticipated for the RM of PS0
in itself. However, understanding the underlying limitations in the RM selected from PE0 is essential
to validate or discard this approach. To understand better these results, we detail analyses in the
following subsections.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Results obtained for both PS0 and PE0 revealed that risk curve error obtained for the RM selected using
few years of production data led to comparable errors to those obtained for the RM selected using
system outputs for field abandonment. To better understand these results, in the following subsections
we detail analyses for RM selected at time 2023 (2 years after the end of the development phase) and
2046 (the date of field abandonment).

To enhance the understanding of the error used to quantify the adequacy of each approach, we show
four examples of RM risk curves for varying degrees of the normalized error (Figure 12). Quantitative
error analysis supported by visual inspection of risk curves reveals that significant mismatches start to
be recorded in normalized errors of around 0.25.

Figure 11 Boxplots characterizing the range of errors in representativity of RM risk curves when
compared to the risk curve of the ensemble of PS0 at field abandonment (2046). To the left, results for
RM selection using system outputs of PS0 at different dates. To the right, results for RM selection using
system outputs of PE0 for the same dates. Boxplots illustrate one of the five solutions (here S1), while
the blue curve marks the mean error obtained for the five solutions (S1 to S5). Crosses mark outliers
and circles the mean of each boxplot.

Figure 12 Risk curves for four indicators (field indicators Np and Wp and well economic indicators for
producers PRK061 and the Wildcat) of PS0 at the time of field abandonment (2046). Green circles: risk
curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Grey fill: normalized error between risk curves. Here, we show the RM solution S1.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Evaluating the RM selected using different decision variables
Figure 13 to Figure 15 show risk curves at field abandonment for PS0 using the ensemble of 199
scenarios and two sets of RM, one selected using the outputs of PS0 at field abandonment, while the
other the outputs of PE0 at the same date (risk curves show results for solution S1, as an example).

Overall visual inspection of risk curves (qualitative analysis) reveals that the PE0 approach has
drawbacks in capturing uncertainty for some system outputs, namely for injectors IRK049, IRK050,
and IRK056. These imprecisions were expected given that normalized errors (quantitative analysis)
were higher than 0.30, averaging 0.35 (Figure 11b). In most cases that recorded mismatches, the nine
RM covered the entire spread of the risk curve (evenly and consistently, in some cases), meaning that
they failed to match the correct cumulative probability. The accuracy of the risk curve relies also on the
proper estimate of the probability of occurrence of the RM, which was not the case for these wells as
revealed by the risk curves. This suggests that the RM can be representative of the uncertain system
provided that their probabilities of occurrence be adjusted.

We selected RM using well economic indicators to reflect, in a single value, a combination of Np, Wp
and Gp for producers, and Gi and Wi for WAG injectors, which are important to support production
strategy definition. As our target production optimization is a gas-driven problem, where oil production
is limited by restrictions in gas processing capacity, WAG injection and ICV operations are often aim
to control gas production. As a result, in addition to evaluating RM uncertainty representation for well
economic indicators, we assessed it for indicators obtained per ICV in producers and injectors. Figure
16 shows the gas-oil ratio (GOR) per ICV over time for three illustrative producers and Figure 17 the
cumulative gas injected (Gi) per ICV for three WAG injectors. Results show that RM obtained from
both approaches covered the range of the ensemble, with approach PE0 recording better uncertainty
representation for some cases, namely ICV1 and 2 of producer PRK028 and ICV2 of injector IRK028.

Figure 13 Risk curves for field indicators of PS0 at the time of field abandonment (2046). Green circles:
risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 14 Risk curves for injection wells of PS0 at the time of field abandonment (2046). Green circles:
risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

Figure 15 Risk curves for production wells of PS0 at the time of field abandonment (2046). Green
circles: risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs
of PS0 at 2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 16 GOR time series per ICV of three producers of PS0. Grey: ensemble. Red: RM selected using
system outputs of PS0 at 2046. Blue: RM selected using system outputs of PE0 at 2046.

Figure 17 Gi time series per ICV of three injectors of PS0. Grey: ensemble. Red: RM selected using
system outputs of PS0 at 2046. Blue: RM selected using system outputs of PE0 at 2046.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Given that the RM cover the range of the risk curves for most system outputs as well as the time series
for ICV indicators, the RM may be valid for ensemble-based production optimization with limitations
related to the estimation of the probabilistic objective function. In cases with weighted functions based
on the probability of each scenario, biases may be introduced to the optimization process. Further
investigations are needed to quantify the potential impacts of this mismatch.

Results obtained for RM selected based on approach PE0 evidence the inherent relation between the
RM and production strategy. On another note, the use of a high number of wells may have been
unfavourable to a fair evaluation of reservoir performance. Moreover, the use of a high number of
system outputs for RM selection may have hampered the procedure. This is because the more variables
we use, the more difficult it is to represent them evenly in one small set of RM. Thus, further
investigations are recommended to improve the application of approach PE0.

Evaluating the RM selected using different simulation time frames


Figure 18 to Figure 20 show risk curves at field abandonment for PS0 using the ensemble of 199
scenarios and two sets of RM, one selected at 2023 and the other at 2046, both using system outputs for
PS0 (risk curves show results for solution S1, as an example). Visual inspection reveals a strong
similarity between risk curves for the two sets of RM. Figure 21 quantifies the range of normalized
errors for the five solutions of RM selected using outputs at 2023 (to the left) and 2046 (to the right),
confirming the qualitative analysis. The consistency of results obtained for S1 to S5 shows the
robustness of the nine RM selected by RMFinder.

As expected, the normalized error is slightly higher for the RM selected for 2023 than for those obtained
directly for 2046 but are highly comparable and averaging less than 0.25. The prime advantage of using
a reduced time frame is the strong reduction of the computational costs required to simulate the
ensemble for generating system outputs for RM selection. While simulating one reservoir scenario of
the ensemble over a lifespan of 30 years takes, on average, around 1 hour and 25 minutes using 12
processors in high-performance computing, the simulation runtime for the reduced time frame takes, on
average, less than 10 minutes using the same computational infrastructure (Figure 22).

Figure 18 Risk curves for field indicators of PS0 at field abandonment (2046). Green circles: risk
curves of the ensemble. Continuous red stair-steps: risk curves for RM selected using system outputs of
PS0 at 2046. Dashed red stair-steps: risk curves for RM selected using outputs of PS0 at 2023.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 19 Risk curves for injection wells of PS0 at field abandonment (2046). Green circles: risk curves
of the ensemble. Continuous red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Dashed red stair-steps: risk curves for RM selected using system outputs of PS0 at 2023.

Figure 20 Risk curves for production wells of PS0 at field abandonment (2046). Green circles: risk
curves of the ensemble. Continuous red stair-steps: risk curves for RM selected using system outputs of
PS0 at 2046. Dashed red stair-steps: risk curves for RM selected using system outputs of PS0 at 2023.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 21 Boxplots characterizing the range of errors in representativity of risk curves of PS0 at field
abandonment (2046) plotted using the five solutions of RM (S1 to S5). To the left, RM selected using
system outputs for PS0 at 2023 and, to the right, at 2046. Crosses mark outliers and circles the mean
of each boxplot.

Figure 22 Boxplots characterizing the simulation runtime of the ensemble of 199 scenarios considering
the full lifespan of the field (September 30, 2046) and until the intermediate time (December 31, 2023).

Validation
Results obtained from previous steps were applied to a control production strategy for validation (called
PV), not used for RM selection. This strategy is not an optimized solution, representing an initial guess
of well number and placement in a production optimization procedure. To test fully our approaches, we
ensured that well placement differed from that used in PS0 and PE0.

We constructed risk curves for PV at field abandonment using four sets of RM, comparing them to
those obtained for the ensemble. These sets are the RM selected from system outputs of PS0 at 2023
and 2046, and the RM selected from system outputs of PE0 at 2023 and 2046. Similarly to the previous
sections, we show here the best RM solution (S1 to S5) for PS0 and PE0, that is, the one that recorded
the smallest mean error based on PS0 and PE0 alone (i.e., blindly for PV).

Figure 23 reveals comparable mean and range of errors for the RM selected for PS0 and PE0 at the
intermediate time (2023), with the latter recording slightly higher quartiles that PS0. Conversely, RM
selected using outputs for field abandonment (2046) lead to significantly lower errors for PE0. Not only
did the RM for PE0 recorded a smaller mean error, but the error variability was also reduced, revealing
robustness in the degree of risk curve representativeness for the main indicators. Except for two outliers
(field Gp and WEI of producer P01), the RM for PE0 recorded normalized errors lower than 0.35, while
normalized errors for PS0 went up to 0.50.

Another lesson learned from Figure 23 is that the normalized errors for RM selected using outputs of
PS0 at the intermediate date (2023) recorded lower variability than that obtained for the RM selected
using outputs at field abandonment (2046). This suggests the increasing effects of production strategy

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
on the system outputs over time, which may be mitigated using a shorter time frame for simulation than
the field's entire lifespan. Conversely, system outputs from PE0, which are not based on a predefined
production strategy, benefited from a longer time frame of simulation to better capture reservoir
performance. Visual inspection of Figure 24 to Figure 26 confirm the quantitative analysis of Figure
23, showing the improved representation of risk curves for RM selected using indicators for PE0 at field
abandonment (2046) than those obtained for PS0.

Figure 23 Boxplots characterizing the range of errors in representativity of risk curves of PV at field
abandonment (2046) plotted using the RM selected from system outputs of PS0 and PE0. “PS0 2023”
and “PS0 2046”: RM selected from system outputs of PS0 at 2023 and 2046, respectively. “PE0 2023”
and “PE0 2046”: RM selected from system outputs of PE0 at 2023 and 2046, respectively. Crosses
mark outliers and circles the mean of each boxplot.

Figure 24 Risk curves for field indicators of PV at the time of field abandonment (2046). Yellow circles:
risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 25 Risk curves for injection wells of PV at the time of field abandonment (2046). Yellow circles:
risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs of PS0 at
2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

Figure 26 Risk curves for production wells of PV at the time of field abandonment (2046). Yellow
circles: risk curves of the ensemble. Red stair-steps: risk curves for RM selected using system outputs
of PS0 at 2046. Blue stair-steps: risk curves for RM selected using system outputs of PE0 at 2046.

However, risk curves in Figure 24 to Figure 26 show global indicators either at field and scope level,
raising the need to verify the representativity of the RM for production and injection indicators obtained
per ICV. Visual inspection of Figure 27 and Figure 28 reveals that RM representativity is limited for
both sets in some indicators (e.g. ICV1 for I01), but has good coverage for the remaining, with slight
better representation for PE0.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Figure 27 GOR time series per ICV of producers of PS0. Grey: ensemble. Red: RM selected using
system outputs of PS0 at 2046. Blue: RM selected using system outputs of PE0 at 2046.

Figure 28 Gi time series per ICV of WAG injectors of PS0. Grey: ensemble. Red: RM selected using
system outputs of PS0 at 2046. Blue: RM selected using system outputs of PE0 at 2046.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
We used GOR to assess RM representativity of reservoir performance per ICV zone because we defined
GOR as the monitoring variable on the triggering conditions to close ICV in producers (in PV, ICV is
closed when GOR exceeds 1,500 m3/m3). Due to the effects of WAG injection in fluid rates on
producing wells, we observe a high-frequency oscillation at a frequency of 6 months on the GOR. These
effects make interpreting Figure 16 and Figure 27 difficult, but ensuring adequate RM representation
for this monitoring variable is essential for reliable ensemble-based optimization.

Given that PV has a different number and placement of wells, different ICV design and operation over
time, different WAG injection strategy, and different production unit than PE0, our results suggest that
despite having normalized errors higher than the accepted limit of 0.25, the use of PE0 ensured, to a
certain degree, improved representativity of the RM for varying production strategies, when compared
to the alternative approach that uses the base production strategy.

Conclusions

This work investigated approaches to improve the selection of representative models (RM), aiming to
ensure adequate representation of uncertainty during long-term ensemble-based production
optimization processes. We started by providing a broad theoretical background on the uncertain
reservoir system and on approaches to obtain reduced ensembles and their applications.

We investigated two main aspects related to the generation of system outputs (production, injection,
and economic indicators) for RM selection: (1) the decision variables (production strategy) used for
obtaining these indicators, and (2) the simulation time frame of the reservoir system.

Our results showed that changing decision variables affects the representativity of the RM. However,
in many cases, the RM covered the range of risk curves but failed to attain the correct probabilities of
occurrence. This suggests a possible use of the RM for ensemble-based production optimizations with
limitations related to the estimation of the probabilistic objective function. In cases with weighted
functions based on the probability of each scenario, biases may be introduced to the optimization
process. This suggests an open area for research to quantify the impacts and reduce the effects of this
mismatch on production optimization.

Using production, injection and economic indicators obtained from a base production strategy revealed
some drawbacks in RM representativeness when compared to a more robust approach based on the
evaluation of reservoir performance using a well pattern with wide reservoir and no restrictions from
production facilities. Despite similar errors across indicators, they were higher than the accepted limit,
suggesting the need for further improving this approach. The use of a high number of wells may have
been unfavourable to a fair evaluation of reservoir performance. This also led to a high number of
system outputs for RM selection, which may have hampered the procedure.

Our results also suggest that RM can be selected using production and injection indicators obtained for
an intermediate date of the simulation time frame, without requiring simulating the full lifespan of the
project. This approach benefited the generation of system outputs using an outlined production strategy
by reducing of the effects of this strategy on the forecasts of reservoir performance. Given that running
reservoir simulation of a full ensemble is computationally demanding, these findings show the potential
to accelerate analyses. To improve RM selection, this approach could be combined with data-driven
methods to estimate production and injection indicators for the remaining time frame of the project.

We used a stair-step plot to construct RM risk curves and defined the area between the RM risk curve
and the ensemble risk curve as the error measure for quantitative analyses. The low number of data
points of the RM risk curve has the inherent drawback of reduced precision when compared to the
ensemble risk curve, composed of hundreds or thousands of data points. Moreover, we calculated the
area between percentiles ~P0 and ~P100 (in our case study, cumulative probability of 0.25% and
99.75%, respectively) to capture the precision of the RM risk curve in the less-likely but high-
consequent events, represented by the two ends of the risk curve. However, increased imprecision is

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
expected in these regions for the RM risk curve. Research is currently ongoing on ways to improve the
construction of the risk curves for reduced ensembles, with added precision for error estimates. In
addition, we recommend investigating ways to calculate the area to prevent adding potential errors into
the quality estimates
of solutions.

To conclude, this work inferred about the adequacy of RM for ensemble-based production optimization
based on insights of the RM fitness for uncertainty quantification for varying production strategies. Our
RM showed fair uncertainty representation for different placement of wells, ICV design and operations,
and WAG injection strategies, suggesting adequacy for ensemble-based production optimization.

We focused on problems related to reservoir development, where the focus of production optimization
is on design variables and infrastructure. Further studies are suggested focusing on RM selection for
reservoirs in the production phase, where the focus of production optimization is on the management
of equipment over time.

Acknowledgements

We gratefully acknowledge the support of EPIC – Energy Production Innovation Center, hosted by the
University of Campinas (UNICAMP) and sponsored by Equinor Brazil and by FAPESP – São Paulo
Research Foundation (grant number 2017/15736-3). We acknowledge the support of ANP (Brazil’s
National Agency of Oil, Natural Gas and Biofuels) through the R&D levy regulation.
Acknowledgments are extended to the Center for Petroleum Studies (CEPETRO) and School of
Mechanical Engineering (FEM) at UNICAMP. We also acknowledge the funding from Energi
Simulation and thank the Computer Modelling Group Ltd. (CMG) and Mathworks for software licenses
and technical support.

References

Aliyev, E. and Durlofsky, L.J. [2015] Multilevel Field Development Optimization under uncertainty
using a sequence of upscaled models. Mathematical Geosciences, 49(3), 1-33.
Armstrong, M., Ndiaye, A., Razanatsimba, R. and Galli, A. [2013] Scenario reduction applied to
geostatistical simulations. Mathematical Geosciences, 45(2), 165-182.
Botechia, V.E., dos Santos, D.R., Barreto, C.E.A., da Silva Gaspar, A.T.F., da Graça Santos, S.M. and
Schiozer, D.J. [2018] Estimating the chance of success of information acquisition for the Norne
benchmark case. Oil & Gas Science and Technology–Revue d’IFP Energies nouvelles, 73, 54.
Correia, M., Hohendorff, J., Gaspar, A.T.F.S and Schiozer, D. [2015] UNISIM-II-D: Benchmark case
proposal based on a carbonate reservoir. SPE Latin American and Caribbean Petroleum
Engineering Conference, SPE-177140-MS.
Douarche, F., Da Veiga, S., Feraille, M., Enchéry, G., Touzani, S. and Barsalou, R. [2014] Sensitivity
analysis and optimization of surfactant-polymer flooding under uncertainties. Oil and Gas Science
and Technology – Rev. IFP, 69(4), 603-617.
Feraille, M. and Marrel, A. [2012] Prediction under uncertainty on a mature field. Oil and Gas Science
and Technology – Rev. IFP, 67(2), 193-206.
Ferreira, C.J., Davolio, A. and Schiozer, D.J. [2015] Improving the estimation of the chance of success
of a 4D seismic project based on representative models. EUROPEC 2015, SPE-174386-MS.
Gaspar, A.T.F.S., Santos, S.M.G., Ferreira, C.J. and Schiozer, D.J. [2020] Estimation of the chance of
success of a four-dimensional seismic project for a developed oil field. ECMOR XVII.
Helton, J.C., Davis, F.J. [2003] Latin hypercube sampling and the propagation of uncertainty in analyses
of complex systems. Reliability Engineering & System Safety, 81(1), 23-69.
Hutahaean, J., Demyanov, V. and Christie, M. [2019] Reservoir development optimization under
uncertainty for infill well placement in brownfield redevelopment. Journal of Petroleum Science
and Engineering, 175, 444-464.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Jiang, R., Stern, D., Halsey, T.C. and Manzocchi, T. [2016] Scenario discovery workflow for robust
petroleum reservoir development under uncertainty. International Journal for Uncertainty
Quantification, 6(6), 533-559.
Mahjour, S.K., Santos, A.A.S., Correia, M.G. and Schiozer, D.J. [2020a] Developing a workflow to
select representative reservoir models combining distance-based clustering and data assimilation
for decision making process. Journal of Petroleum Science and Engineering, 190, 107078.
Mahjour, S.K., Santos, A.A.D.S., Correia, M.G. and Schiozer, D.J. [2020b] Two-stage scenario
reduction process for an efficient robust optimization. ECMOR XVII, WeA15.
Meddaugh, W.S., Meddaugh, W. and McCray, B. [2017] Quantitative assessment of the impact of
sparse data and decision bias on reservoir recovery forecasts. SPE Annual Technical Conference
and Exhibition, SPE-187402-MS.
Meira, L.A., Coelho, G.P., Santos, A.A.S. and Schiozer, D.J. [2016] Selection of representative models
for decision analysis under uncertainty. Computers and Geosciences, 88, 67-82.
Meira, L.A., Coelho, G.P., Silva, C.G., Abreu, J.L.A., Santos, A.S. and Schiozer, D.J. [2020] Improving
representativeness in a scenario reduction process to aid decision making in petroleum fields.
Journal of Petroleum Science and Engineering, 184, 106398.
Meira, L.A., Coelho, G.P., Silva, C.G., Schiozer, D.J. and Santos, A.S. [2017] RMFinder 2.0: an
improved interactive multi-criteria scenario reduction methodology. SPE Latin America and
Caribbean Petroleum Engineering Conference, SPE-185502-MS.
Mishra S. [1998] Alternatives to Monte-Carlo simulation for probabilistic reserves estimation and
production forecasting. SPE Annual Technical Conference and Exhibition. SPE-49313-MS.
Panjalizadeh, H., Alizadeh, N. and Mashhadi, H. [2014] A workflow for risk analysis and optimization
of steam flooding scenario using static and dynamic proxy models. Journal of Petroleum Science
and Engineering, 121, 78-86.
Ravagnani, A.T.F.S., Mazo, E.O.M. and Schiozer, D.J. [2011] A case study of the structure of the
process for production strategy selection. International Journal of Modeling and Simulation for
the Petroleum Industry, 4(1), 9-15.
Santos, A.A.S., Gaspar, A.T.F.S., von Hohendorff Filho, J.C., Correia, M.G., Santos, S.M.G. and
Schiozer, D.J. [2018b] Case study for field development and management - selection of production
strategy based on UNISIM-II. University of Campinas, June 20, 2018. Available at:
https://www.unisim.cepetro.unicamp.br/benchmarks/files/UNISIM-II-D.pdf
Santos, L.S., Santos, S.M.G., Santos, A.A.S., Schiozer, D.J. and Silva, L.O.M. [2020] Influence of
representative models in the expected value of information applied to petroleum fields
development. SPE Europec featured at 82nd EAGE Conference and Exhibition, SPE-200525-MS.
Santos, S.M.G., Gaspar, A.T.F.S. and Schiozer, D.J. [2018a] Comparison of risk analysis
methodologies in a geostatistical context: Monte Carlo with joint proxy models and discretized
Latin hypercube. International Journal for Uncertainty Quantification, 8(1), 23-41.
Sarma, P., Chen, W.H. and Xie, J. [2013] Selecting representative models from a large set of models.
SPE Reservoir Simulation Symposium, SPE-163671-MS.
Scheidt, C. and Caers, J. [2009] Representing spatial uncertainty using distances and kernels.
Mathematical Geosciences, 41(4), 397-419.
Scheidt, C., Zabalza-Mezghani, I., Feraille, M. and Collombier, D. [2007] Toward a reliable
quantification of uncertainty on production forecasts: adaptive experimental designs. Oil and Gas
Science and Technology – Rev. IFP, 62(2), 207-224.
Schiozer, D.J., Avansi, G. D. and Santos, A.A.S. [2017] Risk quantification combining geostatistical
realizations and discretized Latin hypercube. Journal of the Brazilian Society of Mechanical
Sciences and Engineering, 39(2), 575-587.
Schiozer, D.J., Ligero, E.L., Suslick, S.B., Costa, A.P.A. and Santos, J.A.M. [2004] Use of
representative models in the integration of risk analysis and production strategy definition. Journal
of Petroleum Science and Engineering, 44(1), 131-141.
Schiozer, D.J., Santos, A.A.S. and Drumond, P.S. [2015] Integrated model based decision analysis in
twelve steps applied to petroleum fields development and management. EUROPEC 2015, SPE-
174370-MS.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event
Schiozer, D.J., Santos, A.A.S., Santos, S.M.G. and Hohendorff Filho, J.C. [2019] Model-based decision
analysis applied to petroleum field development and management. Oil and Gas Science and
Technology – Rev. IFP, 74, 46.
Shirangi, M.G. and Durlofsky, L.J. [2016] A general method to select representative models for decision
making and optimization under uncertainty. Computers and Geosciences, 96, 109-123.
Silva, L.O.M., Santos, A.A.S. and Schiozer, D.J. [2016] Otimização da estratégia de produção sob
incertezas geológicas e econômicas. Rio Oil & Gas Expo and Conference 2016, IBP1513_16.
von Hohendorff Filho, J.C., Maschio, C. and Schiozer, D.J. [2016] Production strategy optimization
based on iterative discrete Latin hypercube. Journal of the Brazilian Society of Mechanical
Sciences and Engineering, 38(8), 2473-2480.
Wilson, K.C. and Durlofsky, L.J. [2013] Optimization of shale gas field development using direct
search techniques and reduced-physics models. Journal of Petroleum Science and Engineering,
108, 304-975.
Yang, C., Card, C., Nghiem, L. and Fedutenko, E. [2011] Robust optimization of SAGD operations
under geological uncertainties. SPE Reservoir Simulation Symposium, SPE-141676-MS.
Yeten, B, Durlofsky, L.J. and Aziz, K. [2003] Optimization of nonconventional well type, location, and
trajectory. SPE Journal, 8(3), 200-210, SPE-86880-PA.

ECMOR XVII – 17th European Conference on the Mathematics of Oil Recovery


14-17 September 2020, Online Event

You might also like