You are on page 1of 21

SPE-196678-MS

Field Application Study on Automatic History Matching Using Particle


Swarm Optimization

Sanghyun Lee, Korea National Oil Corporation; Karl, Dunbar Stephen, Heriot-Watt University

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Reservoir Characterisation and Simulation Conference and Exhibition held in Abu Dhabi, UAE, 17 - 19 September
2019.

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

Abstract
The traditional trial and error approach of history matching to obtain an accurate model requires engineers
to control each uncertain parameter and can be quite time consuming and inefficient. However, automatic
history matching (AHM), assisted by computers, is an efficient process to control a large number of
parameters simultaneously by an algorithm that integrates a static model with dynamic data to minimize a
misfit for improving reliability. It helps to reduce simulation run time as well.
Particle Swarm Optimization (PSO) is a population based stochastic algorithm that can explore
parameter space combined with the least squares single objective function. The process of AHM can adopt
parameterization and realization methods to reduce inverse problems. In this study, realizations of various
reservoir properties such as porosity, net to gross, relative permeability, horizontal and vertical permeability,
and aquifer size were chosen for controlling throughout the AHM. History matching was conducted to
validate the efficiency of each method. The guidelines for optimized AHM with a stochastic algorithm are
also disccussed.
The realization and parameterization methods improved matching results in a full-field application with
resulting in a reduced misfit and in less. A stochastic algorithm generates multiple models to deduce control
parameters to reduce a misfit. In this study we identified that PSO converged effectively with updated
control parameters. The optimized AHM improved the accuracy of a full-field model although some misfit
remained in the match to bottomhole pressure.
We found that updating with too many parameters makes the problem difficult to solve while using too
few leads to false convergence. In addition, while the simulation run time is critical, a full-field simulation
model with reduced computational overhead is benefitial.
In this study, we observed that the PSO was an efficient algorithm to update control parameters to reduce
a misfit. Using the parameterization and realization as an assisted method helped find better results. Overall
this study can be used as a guideline to optimize the process of history matching.
Keywords: Automatic History Matching, Stochastic Algorithm, Particle Swarm Optimization,
Parameterization, Realization
2 SPE-196678-MS

Introduction
The conventional manual-approach (Manual History Matching, MHM) has been used by engineers for some
time. It helps to obtain a match to dynamic data by adjusting control parameters with a numerical model that
best describes actual characteristics. Automatic History Matching (AHM), sometimes called a computer-
assisted approach, is also becoming well-known for updating models. AHM is more efficient than manual
methods because it uses an optimization algorithm which explores the parameter space to find the best
values such that the match to dynamic data is acceptable. Various optimization algorithms are available in
the oil and gas industry. The term "assisted history matching" may or may not include the automatic element
but it usually uses a quantitative analysis and includes the sensitivity analysis part.
A considerable number of optimization methods have been adopted in the AHM process including
gradient based, hybrid, stochastic, and probabilistic methods and these have been used to search the
parameter space and reduce the misfit [Romero et al. 2000]. Each method has positives and negatives. Fig.
1 summarizes the conceptual AHM process with available options, from which the engineer can choose
suitable methods in each step.

Figure 1—Conceptual AHM Process Assisted by Computer [Sanghyun et al., 2018]

Stochastic methods among optimization algorithms are more explorative and consequently require more
computer resources. They are excellent for solving complex problems and conducting uncertainty analysis,
which provides the guidelines for unpredictable conditions to reduce operational risks by contingency
scenarios. The extra effort spent is often worthwhile. Genetic Algorithm [Radcliffe, 1991], Neighbourhood
Algorithm [Sambridge 1999], Particle Swarm Optimization [Hutahaean et al., 2015; Mohamed et al.,
2010a, 2010b], Differential Evolution Algorithm [Hajizadeh et al., 2009b] and Ant Colony Optimization
[Hajizadeh 2009a] are included in Stochastic methods.
In this work, we use Particle Swarm Optimization (PSO) [Kennedy et al., 1995] with a least squares
single objective function to measure the match quality for full field simulation. Realizations were generated
to be selected for describing reservoir heterogeneity. Investigation of various realizations is one of the
ways to describe heterogeneous rock properties through geostatistics [Hu et al., 2004], and it contributes to
improving the model so that it is more realistic.
SPE-196678-MS 3

Parameterisation is also key to the assisted method to history matching. Chavent and Bissel (1998)
introduced the refined scheme which is attractive for enhancing the control of input variables by focusing on
the dominant properties. The objective is to reduce the number of degrees of freedom to a manageable level
because we can otherwise find it impossible to reach a solution. An appropriate parameterization method
has been used for solving this inverse problem and one of our aims is to invest its suitability. After that, the
best matched model with a minimum misfit accepted will be applied to forecast reservoir performance as
part of the decision making process during field development.

Objective Function
To analyse the accuracy of selected models when applying AHM, we need to quantify the accuracy of the
forecast. A misfit is set up for this purpose and hence we can select the best matched models. The misfit is
calculated from the magnitude of the residual vector which estimates the difference between observed and
modelled values. For multiple time step, wells and various properties, the total misfit is then

(1)

where Ni, Nj wells and Nk are the number of time steps, wells and variables respectively. represents the
parameter vector. and are the observed and modelled values, resepectively, of the kth variable in
the jth well at the ith time step. describes the standard deviation of the errors in observations assumed to
be time dependent in general [O’Sullivan et al., 2005; Rotondi et al., 2006]. The variables include oil and
water production rate, Gas Oil Ratio, Water Cut, and BHP. Meanwhile, the optimized misfit cannot be zero
(0) because there may be errors derived as [Shah et al., 1978]:
(2)

(3)

where is the true value, , , and are the errors of the observed data, the model precision
error, and model error from choosing the wrong parameters. It is pointed out that these errors can arise
through data acquisition, geological modelling, and discretization (model errors are not covered here). The
parameter errors should be reduced along with the misfit. However, the data and model precision errors
cannot be reduced.

Particle Swarm Optimization (PSO)


The PSO was initially introduced to describe the social behavior of bird flocks, and then applied to AHM
as a population-based algorithm recently [Mohamed et al., 2010a]. It is noted that the PSO uses velocity
and position information of individual models, called particles, to explore the multidimensional parameter
space within the range of uncertainties. The behavior of each particle undergoes an optimization process of
minimizing the misfit through an objective function. Meanwhile, it memorizes the best result from previous
calculations and uses these to select parameters at the next iteration. The PSO, which remembers best the
previous solution, is excellent for finding the optimal solution quickly [Mohamed et al., 2010c]. However,
it does not always provide a better outcome because it tends to be more exploitative by memorized velocity
and position information.
The particles of the swarm, in the PSO algorithm, are positioned at initial parameter values assigned by
random generation. After that it updates its velocity and position by attempting to move towards the best
models encountered so far while the particle swarm moves to the next step. If a particle has a better misfit
than at its previous positions or overall, that position is memorised for the next iteration.
4 SPE-196678-MS

The velocity vector element, , is used to claculate the particle’s position, , as shown [Eberhart
et al., 1995; Onwunalu et al., 2009]:
(4)
where p and q are the iteration and parameter counters respectively. The velocity, which is moving each
particle, is updated using following equation:
(5)

where ω, c1, and c2 are coefficients which weight previous velocity (inertia), best known position ,
and position of neighborhood particle (gbest ), r1 and r2 are generated by random numbers between 0 and
p

1 the weight and gbestp respectively for acceleration towards the new best positions. In this study,
ω = 0.73, and c1 and c2 were set to 1.496 from the best practice. The particles keep moving to next best
positions until they converge over defined iterations.

Parameterization
Parameterization is the method used to update reservoir properties so that we can improve the accuracy
of the prediction to match with observations. It should ideally retain the general geological concept and is
constrained by the observed petrophysical data of the reservoir. In this study, we define the spatial domains
(i.e. the linked grid cells) as zones which have similar characteristics. This approach allows properties to be
updated by a single multiplier function. Thus, the parameterization of a zone easily leads to a simultaneous
modification of the initial properties with a smaller number of controlling parameters. The parameterization
technique links to the input data of the simulation model by a geostatistical method [Roggero, 1997].
However it is not simple to get meaningful parameters that will reduce a misfit. In order to establish a good
representation of highly heterogeneous reservoirs, local parameterization may be required while in simpler
cases zonation may be used. This is a traditional approach to reduce the number of control parameters which
update reservoir properties [Shah et al., 1978].

Realization
Geostatistical techniques are used to generate a geological model to represent the reservoir heterogeneity in
3D reservoir based on the petrophysical data at the wells [Didier et al., 2010] combined with the conceptual
model and also spatial variables such as the range of a semi-variogram. Reservoir properties are distributed
using various methods which may include Sequential Indicator Simulation (SIS), Sequential Gaussian
Simulation (SGS), Fractal Simulation, and Kriging. With each of these methods, different realization may
be created stochastically, controlled by the seed number. Such realisations have the same global statistics.
Multiple realizations are recommended to help explore uncertainty in the geological properties and thus
reduce risks [Pallister et al., 2000; Odai et al., 2011]. Nevertheless, it is almost impossible to generate a
perfectly correct model that represents reality.

Project Outline
To illustrate the optimized AHP process, we define objectives of the study as

• Obtain the best matching models with minimum misfit for a number of cases using the workflow
in Fig. 2
SPE-196678-MS 5

Figure 2—Schematic Outline of Full Field Application

• Define the effects of control parameters on the outcome. These control parameters are the
multipliers to zones discussed above.
• Quantify the effects of parameterization which includes defining zones in different ways.

• Identify the feasibility and efficiency of PSO compared to MHM

Throughout the AHM process, the optimum control parameters, realization, and parameterization method
is chosen based on the minimum misfit. Further applications of AHM were carried out to optimize the
control parameters including local parameterization. The number of control parameters was optimized for
the PSO algorithm to reduce the number of models needed even though we considered that we might lose
control on the heterogeneity updates.

Application to Field Study


The reservoir used in this work had a deep marine depositional environment with mud and sand deposited
as turbidites mainly by gravity flows which established an upward coarsening grain size distribution. The
quality of the reservoir is lower at the base therefore. The field contains light oil, 38 API, has an initial
pressure at 3650 PSI which is close to the bubble point pressure at 3431 PSI. Initially the field was produced
by depletion and has been producing hydrocarbons since August 2005 through 7 horizontal producers
supported by a strong aquifer. To maintain the reservoir pressure, water was injected into the bottom of the
reservoir in the North-East area to enhance sweep efficiency from 2 injectors. Fig. 3 shows the simulation
model which consisted of 43×52×129 grid cells, (a total of 288,444), that means each cell was 60×60
m×3.048 m with a numerical aquifer supported at the eastern edge of the reservoir, and locations of the wells.
6 SPE-196678-MS

Figure 3—Reservoir Model with 7 Producers and 2 Injectors, showing distribution of initial oil saturation.

The reservoir model consisted of three geological facies including high quality (blocky) sand,
heterogeneous (heterolithic) sand, and low quality (heterolithic) mud after scale up based on the field data.
The Net to Gross (NTG) is modeled by SGS of geostatistical techniques using a variogram constrained by
the well data. The porosity (PHI) in the facies is coKriged with NTG and then permeability was assigned
based on the function of porosity as described in Eq. (6) taken from correlation in logs
(6)
where kh is an absolute permeability linked to porosity, kx and ky are x and y direction permeability
respectively. The vertical permeability is distributed by 0.01 × kh according to the core analysis. The
constructed model is highly heterogeneous and the properties are illustrated in Fig. 4.

Figure 4—Reservoir Porosity Model

Relative permeability curves (Fig. 5) for the three main facies in the field were obtained by an upscaling
exercise (not discussed here).
SPE-196678-MS 7

Figure 5—Relative permeability before and after manual history matching


for (a) Heterolithic mud, (b) Heterolithic sand and (c) Block sand.

The history data comprised oil and water production and GOR for the 7 producers, P1-7 and injection rate
for the 2 injectors, I1 and I2. BHP for all wells was also used. Fig. 6 illustrates that field history for the initial
model has a large misfit for water production although oil is relatively well matched. The total initial misfit
is 67038 from the given initial model. At the individual well level, producers P2 and P3 match oil production
rates very well and were equally matched after production. For the other wells water breakthrough is too
early as abserved for the whole field (Fig. 6). This was indicated to be due to flow through thief zones in
the 20th and 56th layers of the model along with capillary pressure effects identified by pre-review. The
early breakthrough reduces the oil production by 20% to 50% of the historical values. Secondly, the model
predicts high BHP over production at all producers which causes a high value of misfit.

Figure 6—MHM Field Production compared to Observed and Initial Model: Qo (left) and Qw (right).

Manual History Matching


Traditional manual history matching was initiated to improve the global match of field liquid rate data rather
than individual well production and BHP data. The static reservoir parameters and flow properties in Table
1 were modified to minimize the misfit. These included relative permeability curves of the three lithofacies
for solving inverse problems casusing the over prediction of water production, illustrated in Fig. 5. Fig. 6
presents production data after MHM indicating significant improvement was obtained. The model is almost
obscured by the production data such is the quality of the match. However, the water production indicates
that some work is required. The matched model also had a large misfit in BHP. We address local updating
afterwards.
8 SPE-196678-MS

Table 1—Manual HM Control Parameters

Parameter Unit Initial setting Updated Value

NTG ft/ft Assigned by SGS 1.1 × original

Vertical permeability Multiplier 0.01 × Kh 0.1 × Kh

PHI in of localized thief zone Multiplier coKriged with NTG 1.1× original

Permeability in of localized vertical Multiplier Relationship with PHI by Eq. (6) 10 × original
baffle

Relative permeability mD/mD Figure 5 Figure 5

Automatic History Matching Application


Particle Swarm Optimization (PSO) algorithm was used for AHM as a stochastic method. In order to
generate multiple models, the PSO used 30 models per iteration. This population factor is optimized using
a simple HM practice which is not included in this study.

Objective Function
The objective function for optimization of history matching included 27 variables initially: field oil and
water production rates, oil and water rates from 7 producers, water injection rate for 2 injectors, and BHP
from 9 wells. The history production rates among objective values were weighted by 1.7 times compared
to the pressure data. This was based on the experience that BHP data generally has large errors caused by
depth conversion from well to reservoir, as well as damage of wellbore and resolution. Throughout history
matching, the early breakthrough of water, low oil production, and relatively high BHP at wells were the
target errors in the models and the remaining misfit will be addressed.

Initial Model
Control Parameters. The correlation between porosity and permeability from well data was employed
to reduce the number of control parameters. Six control parameters that imply high uncertainties as input
data described by the operator in the initial model were used (Table 2) as multipliers or set directly, and
each of the parameters were combined to identify improved reservoir properties based on how they govern
flow behavior. The control parameters were modified based on the ranges of the uncertainties and errors
indicated by the operator.

Table 2—Defined Control Parameters and Ranges

Control Parameter Method Parameter Range Assumption

1 Porosity, PHI Multiplier 0.8 to 1.2 Heterogeneity

2 NTG Multiplier 0.8 to 1.2 Heterogeneity

3 Permeability, K Multiplier 0.1 to 10.0 Heterogeneity

4 Vertical permeability, Kz Update kv/kh ratio 0.01 to 0.2 Heterogeneity

5 Relative Permeability, Kr Update Swc 0.2 to 0.4 Error range

6 Aquifer Size Multiplier 0.01 to 5.0 Uncertainty

Realization. Six model realizations considering facies variation were generated to help solve inverse
problems based on data provided by the operator (Fig. 7). One realization was conditioned to the distribution
of facies. The reservoir was divided into 19 intervals based on the log analysis and three facies were
distributed throughout using SIS method. Assuming that the petrophysical properties of the facies can vary
between each interval, separately, this gives 57 zones in total. Further to this study, the operator's estimate
SPE-196678-MS 9

of facies distribution was used to create independent zones for multiplication of permeability, mainly. SIS
was used to generate five realizations with 40 zones. In other words, models were generated as if there are
40 rock types. So that these rock types were controlled to change the properties.

Figure 7—Realizations; each color illustrates a separate zone as a contiguous set of cells
assumed to control the same property multiplier. Various properties were altered to get
a history match. Items 1-5 in Table 2 were changed by zone-wide multipliers or values.

Parameterization. The porosity was considered to be the key variable. The zonation method, which
controlled property data independently for each zone, was used for parameterization that follows the
geological conceptual model to simplify the inversion process. Zones could be randomly merged so that
the same multiplier is applied to a group of zones. From a maximum of 57 zones, the total was gradually
reduced to a minimum of 5.
Meanwhile, to solve the local issues where early breakthrough was caused along with low oil production,
multi-scale parameterization was adopted to subdivide control parameters in the selective zones (Table 3).

Table 3—Selected Local Zonation which occurred early water breakthrough

Zone 1 Zone 2 Zone 3

Zonation 20th horizontal layer 56th horizontal layer 20th vertical layer

Control Parameter PHI PHI Kh, Kz

Data QC. Fig. 8 shows the first few history matching runs where it is indicated that while the models
converged, the misfit, 64971, was unsatisfactory. If the precision and parameter errors are zero in Equation
3 such that the model is the "truth" then the misfit is the total number of measurements for each variable
and time step and for all wells, and corresponds to the data error only. This might be the minimum misfit
we can obtain under perfect conditions. However, there were 114 time steps recorded and 27 variables from
10 SPE-196678-MS

all the wells. The number of misfit for noise is then 114 × 27 = 3078. The misfits in Fig. 8 are therefore
too high. A QC of the data was then performed to identify obvious data problems. Fig. 9 shows that wells
P1, P2, P4, P6, I1, and I2 recorded abnormally constant BHP which is impossible to explain depleting
reservoir performance with production rate, and unreasonably high injection rates in the water injectors.
The abnormal and incorrect data set prevents convergence of the automatic algorithms. It can also decrease
efficiency and result in an unrealistic model with high misfit and therefore incorrect properties. Rather than
trying to find ways to match the data (for example by assuming a dynamic variation to the Productivity
Index, PI, of the wells), the unreliable data were screened out by setting a zero weighting factor for each
data point in the objective function. The data removed (Fig. 9) includes BHP for P1, P2, P4, P6, I1, I2 and
several of the injection rates for I1, I2.

Figure 8—Initiation of AHM to Check Data Quality; the final minimum misfit over 200th model by PSO indicates
unacceptable misfit, 64971. Blue and orange symbols indicate the PSO runs with 57 and 40 zones resepectively.

Figure 9—BHP and Injection Rate History for Data QC; (a) abnormally high or constant BHP data points on Well
P1, P2, P4, P6, I1 and I2 (circled), (b) example of two measurements of unreasonablly high observed data (circled).

Results
Effect of Control Parameters. We investigated how various properties might affect the history matching
by successively adding them to the set of control parameters. So that we started with porosity (PHI) and
then added relative permeability (Kr) and vertical permeabilith (Kz) followed by net to gross (NTG) and
then the aquifer properties (Aquifer). Fig. 10 shows the result of this process. The rate of convergence was
SPE-196678-MS 11

consistent but the misfits ended up being smaller overall. Note that the initial variation in the models is
caused by random variations between models.

Figure 10—Effect of changing control parameters with using 10 zones

Nevetheless, adding more variables did not guarantee a better solution as the best result. In fact, with
more degrees of freedom, the solution actually becomes harder to find and we may get trapped in a local
minimum. PSO seems to have less ability to control a large number of parameters (Fig. 11). We discuss this
in the Feasibility and Efficiency of PSO Algorithm Section. The number of zones used is important as well
as the number of variables. As shown in Fig. 11, the case with 46 parameters ("46 Para") became trapped
around a poorer fitting model overall while the best models were found with 28 parameters ("28 Para"). It
is emphasized that choosing the number of parameters is itself an optimization issue as the smallest number
of parameters also resulted in a less accurate set of models.

Figure 11—PSO Performance with 30 models per iteration depending on the Number
of Control Parameters (PHI in global and local zonation, Kz, Kr) on 1st SIS Realization

Effect of Realization. The six realizations which represent various facies distributions in Fig. 7 were also
examined and each was used as a set of zones. Fig. 12 demonstrates the misfit evolution for these cases; the
5th SIS realization produced the lowest misfit and was chosen for further AHM. It is observed that, as shown
12 SPE-196678-MS

in Fig. 12, multiple realizations should be investigated to improve the changes of reaching a satisfactory
solution.

Figure 12—Reliability of Realizations with PSO with 30 models per iteration; controlling PHI, Kz, and Kr

Effect of Parameterization. The zonation was coarsened by grouping the multipliers in the 5th SIS
realizations in a random fashion but also to avoid large zones being formed. The efficiency of PSO and
the overall level of the final misfits were improved so that 5 appears to be optimal (Fig. 13). On the other
hand, 3 zones were too coarse and lead to larger final misfits. In addition, it was found that the impact of
choosing different parameterization settings was affected by the control parameters themselves. When a
control parameter included the size of the aquifer, updating based on 10 zones resulted in models with the
best fit as shown in Fig. 13. The aquifer that is based on the field information is connected to the eastern edge
of the reservoir which was modeled by numerical aquifer with 8.5% porosity and 48.4 mD permeability.

Figure 13—Optimization of Coarsening Zonation on 5th SIS Realization


by PSO. Zones are used to control changes to PHI, NTG, Kz, and Kr.

In a heterogeneous reservoir, updating the model to achieve a global match often misses many of the
local problems associated with heterogeneity. Thus, in this work, multi-scale parameterization is more
attractive to target those areas that govern a large influence on improving the history match. Pre-evaluation
of the initial simulation model to define target areas for updating and other issues such as water coning
SPE-196678-MS 13

or thief zones should be carried out to identify local targets for updating. After going through this process
we obtained a history match by reducing a number of control parameters, as shown in Fig. 14. The case
controlling PHI, NTG, Kr, Kz, multi-scale, and aquifer has minimum misfit, 17498. The converged control
parameters were chosen for both global and local properties within the uncertainty range shown in Table 3.

Figure 14—Effect of Control Parameters with 5th SIS realization by using the PSO with 30 models per iteration.

Feasibility and Efficiency of PSO Algorithm. Using the full field model, AHM took 2.3 hours to complete
running 200 models generated by the PSO algorithm using 8 parallel CPU processes with 3.5 GHz. It
became obvious that the time to run the simulation model should be multiplied by the total number of
models required for convergence. This run time will increase for complex simulations. Other parameter
combinations scaled the total run time. The PSO results were better as fewer zones were used (Fig. 15). The
key observation is that one might consider fewer zones to be a better choice in the lack of the knowledge
for saving cost with less convergence time (Fig. 15). On the other hand, considering our experience that the
PSO has limited capacity for handling a large number of control parameters, optimized control parameters
should be investigated.

Figure 15—Efficiency of PSO Convergence depending on the number of control parameters; efficiency improved.

Optimized AHM Result. By optimizing the AHM process, the scenarios (Table 4) that identified the
minimum misfit at each step were chosen for a run of the PSO. The minimum misfit of the best overall
14 SPE-196678-MS

match was achieved at 17498. Fig. 16 shows misfit calculated by PSO and again, most of the remaining
misfit was due to BHP.

Figure 16—AHM Misfit with PSO with 30 models per iteration; final misfit of 17498

Table 4—Optimized AHM Scenario. PSO used 30 models per iteration

AHM Process Case

Realization 5th SIS realization

Control parameters 10 PHI, 10 NTG, 1 Kz, 3 Kr, 4 Multi-scale, 1Aquifer

Parameterization 10 zones with multi-scale parameterization

No. of control parameter 29

AHM Algorithm PSO

Validation of the Matched Model. Fig. 17 illustrates the improvement in misfit by variable and as a total.
The original oil in place of the base case model was changed from around 500 MMstb by 9.5% after history
matching. This was a result of changing porosity and connate water saturation, shown in Table 5. The PSO
algorithm improved the history matching result not only for production but also BHP compared to the
initial model and MHM, which are demonstrated in Fig. 18 through to Fig. 21. In Fig. 16, many parameters
converged quickly and the convergence of the misfit is relatively straightforward after the 60th model.
SPE-196678-MS 15

Figure 17—Misfit change for various variables (a) each objective function and (b) total misfits

Figure 18—AHM Result on (a) Field Oil and (b) Field Water Production Rate

Table 5—Comparison of History vs. Matched Model. All variables are in MMstb

Variable (MMstb) History Matched Model Difference

STOIIP 495.2 542.2 9.5 %

Cum. Oil Production 52.2 51.1 −2.1 %

Cum. Water Production 1.2 1.6 36.1 %

Misfit 67038 17498 −73.9 %

The most of the remaining misfit are due to BHP errors. Nevertheless, AHM with PSO improved the
quality of history matching and the reliability in reservoir model forecasting.
Figures 19-21 show the oil, water and pressure matches for individual wells. The oil rates are matched
very well, with better match after history matching as in Fig. 18a. The water match is also improved in
most of the wells. P2 and P4 consistently have no water production. The original model wrongly had
water production in P1 and P3 which is now corrected. Well P6 has been improved with a small room for
improvement. P5 still has some water production which should not be there. P7 has been over corrected. It
is these last two wells which require further attention possibly by more local updates. The pressure matches
16 SPE-196678-MS

are more variable though much of this can be corrected by changing the PI multiplier. Some wells show
dynamic variation in PI which may be due to collapse and subsequent clean up.

Figure 19—AHM Result for Oil Production Rate; the model is well matched.
SPE-196678-MS 17

Figure 20—AHM Result for Water Production Rate; final matched model was improved by AHM.
18 SPE-196678-MS

Figure 21—AMH Result for BHP; although PSO improves matching result.
SPE-196678-MS 19

Discussion
In this field application, the process of AHM required a wide range of approaches to generate a reliable
model. Choosing a single parameterization scheme or realization could result in a poor match, ultimately.
Various combinations with assistant methods can be made more efficient with AHM and good decision
still needs to be made by the engineer. Testing of different options was needed. These include the selection
of objective functions, the initial realization, efficient parameterization, the choice of uncertainty range,
and also the algorithm to use. Optimizing the number of parameters can also be important to improve the
efficiency of convergence time and avoid getting stuck in a poorly matching model (i.e. a local minimum).
By considering the geological concepts used to build the model, the up-scaled reservoir properties in the
initial grid cells created various problems. It is noted that localized parameterisation can help to solve these
problems. Global control variables did not provide sufficient flexibility while too many local options make
the solution very hard to find. Limited parameterization was applied to check its feasibility for updating but
the misfit remained high. Further work is required to address the issue of high misfit in the BHP data.
In order to reduce the misfit, productivity index (PI) could be considered as a property to be altered during
the production period. Usually this is done after matching produced volumes. However, it may be that this
would lead to a bias in the search. However, there are many properties associated with the productivity
index such that the number of parameters might be updated too much. It is also noted that the data and
model precision errors through modeling and simulation should be considered. The errors could affect the
quality of history matching if not handled effectively [Stephen, 2007].
The PSO algorithm clearly seems to give an efficient solution. In some cases, the convergence is much
faster in Figure 15 where the best models were found. It should be noted however that we did not investigate
the various parameters of the PSO problem. It is possible that tuning these can give a better response. The
number of simulations required for the PSO was 150 as a best practice to get convergence. At this point we
do not know if it takes an hour to run a single model it would be efficient to use AHM to find the better
solutions.

Conclusions
1. PSO can help explore the control parameter space as an AHM method leading to fast convergence.
In unknown uncertainties of reservoir properties assigned in the grid cells, PSO reduced effort to find
out optimized value to match simulation with observed data.
2. Optimal combination of the realization and parameterization led to better solutions. The geological
models generated by the stochastic SIS method allowed flexibility in the parameterization.
3. The efficiency of PSO depends on the combination of control parameters. If too many are used, a
trapping in a local minimum may occur. On the other hand, too few parameters means that there is
not enough flexibility.
4. The data should be checked for quality and unrealistic data removed or considered later.

Acknowledgements
We thank Schlumberger Geoquest for allowing us to use their software. I, Sanghyun, acknowledge financial
support from Korea National Oil Corporation.

References
1. Chavent. G., and Bissell, R., 1998, Indicator for the refinement of parameterization, 2nd
International Conference in Inverse Problems in Engineering, ISIP 98, pp. 309–314, 12 March
1998, Nagano, Japan
20 SPE-196678-MS

2. Didier, D., Frédéric, R., 2010, History Matching Geostatistical Model Realizations Using a
Geometrical Domain Based Parameterization Technique, Mathematical Geosciences, Vol.42(4),
pp.413–432
3. Eberhart, R., and Kennedy, J., 1995, A New Optimizer using Particle Swarm Theory, Sixth
International Symposium on Micro Machine and Human Science, MHS’95, pp.39–43
4. Hajizadeh, Y., Christie, M. A., and Demyanov, V., 2009a, Ant Colony Optimization Algorithm
for History Matching, Society of Petroleum Engineers, doi:10.2118/121193-MS
5. Hajizadeh, Y., Christie, M. A., and Demyanov, V., 2009b, Application of Differential
Evolution as a New Method for Automatic History Matching, Society of Petroleum Engineers,
doi:10.2118/127251-MS
6. Hu, L., Ravalec-Dupin, L. M., and Hu, L., 2004, Elements for an Integrated Geostatistical
Modeling of Heterogeneous Reservoirs, Oil and Gas Science and Technology, Mar.-Apr. 2004,
Vol.59(2), pp.141–155
7. Hutahaean, J. J., Demyanow, V., and Christie, M. A., 2015, Impact of Model Parameterisation
and Objective Choices on Assisted History Matching and Reservoir Forecasting, Society of
Petroleum Engineers, doi:10.2118/176389-MS
8. Kennedy, J., and Eberhart, R., 1995, Particle Swarm Optimization, The Proceedings of the
Inter’l Conf. on Neural Networks, pages 1942-1948, [available: http://dx.doi.org/10.1109/
icnn.1995.488968]
9. Mohamed, L., Christie, M. A., Demyanov, V., Robert, E., and Kachuma, D., 2010a, Application
of Particle Swarms for History Matching in the Brugge Reservoir, Society of Petroleum
Engineers, doi:10.2118/135264-MS
10. Mohamed, L., Christie, M. A., and Demyanov, V., 2010b, Reservoir Model History Matching
with Particle Swarms: Variants Study, Society of Petroleum Engineers, doi:10.2118/129152-MS
11. Mohamed, L., Christie, M. A., and Demyanov, V., 2010c, Comparison of Stochastic Sampling
Algorithms for Uncertainty Quantification, Society of Petroleum Engineers, doi:10.2118/119139-
PA
12. Odai, L., and Ogbe, D.O., 2011, An Approach for Ranking Realizations to Characterize
Reservoirs for Fluid Flow Simulation, Society of Petroleum Engineers, doi:10.2118/150738-MS
13. Onwunalu, J. E., and Durlofsky, L. J., 2009, Development and Application of a New Well Pattern
Optimization Algorithm for Optimizing Large Scale Field Development, Society of Petroleum
Engineers, doi:10.2118/124364-MS
14. O’Sullivan, A., and Christie, M., 2005, Error models for reducing history match bias,
Computational Geosciences, Vol.10, pp.450, doi:10.1007/s10596-006-9027-5
15. Pallister, I. C., and Ponting, D. K, 2000, Asset Optimization Using Multiple Realizations and
Streamline Simulation, Society of Petroleum Engineers, doi:10.2118/59460-MS
16. Radcliffe, N. J., 1991, Formal Analysis and Random Respectful Recombination, Proc.
of the Fourth Inter’l Conf. on Genetic Algorithms, pages 222-229, [available: http://
citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.34.5083&rep=repl&type=pdf]
17. Roggero, F., 1997, Direct Selection of Stochastic Model Realizations Constrained to Historical
Data, Society of Petroleum Engineers, doi :10.2118/38731-MS
18. Romero, C. E., Carter, J. N., Gringarten, A. C., and Zimmerman, R. W., 2000, A Modified
Genetic Algorithm for Reservoir Characterisation, Society of Petroleum Engineers,
doi:10.2118/64765-MS
19. Rotondi, M., Nicotra, G., Godi, A., Contento, F. M., Blunt, M. J., and Christie, M., 2006,
Hydrocarbon Production Forecast and Uncertainty Quantification: A Field Application, Society of
Petroleum Engineers, doi:10.2118/102135-MS
SPE-196678-MS 21

20. Sanghyun, L., and Stephen, K. D., 2018, Optimizing Automatic History Matching using
Genetic Algorithm and Particle Swarm Optimization, Society of Pertroleum Engineers,
doi:10.4043/28401-MS
21. Sambridge, M.S., 1999, Geophysical inversion with a neighbourhood algorithm-I. Searching a
parameter space, Geophysical Journal International, Vol.138(2), pp.479–494
22. Shah, S., Gavalas, G. R., and Seinfeld, J. H., 1978, Error Analysis in History Matching: The
Optimum Level of Parameterization, Society of Petroleum Engineers, doi:10.2118/6508-PA
23. Stephen, K. D., 2007, Scale and Process Dependent Model Errors in Seismic History Matching,
Oil and Gas Science and Technology, Vol. 62, No. 2, pp. 123–135, doi:0.2516/ogst:2007011

You might also like