A Methodology To Reduce Uncertainty

Constrained to Observed Data

Célio Maschio and Denis J. Schiozer, UNICAMP; and Marcos A.B. Moura Filho, Petrobras; and
Gustavo G. Becerra, Petrobras Energia S.A.

Summary Guérillot and Pianelo (2000) worked on a method for matching

This paper presents a new method to deal with uncertainty mitiga- production data and seismic data to reduce uncertainty in produc-
tion by using observed data, integrating the uncertainty analysis and tion forecasts. The work proposed simultaneous calibration on
the history-matching processes. The proposed methods are robust both permeability and acoustical impedance of the reservoir for
and easy to use, and offer an alternative to traditional history- an enhancement of the resulting geological model.
matching methods. The main characteristic of the method is the use Queipo at al. (2002) presented a method using neural-network-
of observed data as constraints to reduce the uncertainty of the based efficient global optimization for the problem of estimating
reservoir parameters. The main objective is the integration the distributions of permeability and porosity in heterogeneous
of uncertainty analysis with history matching, providing a natural and multiphase petroleum reservoirs by matching the static and
manner to make predictions under reduced uncertainty. Three meth- dynamic data available. Reis (2006) also used artificial neural
ods are proposed: (1) probability redistribution, (2) elimination of networks (ANNs) to combine risk analysis with history matching.
attribute levels, and (3) redefinition of attribute values. To test the Zabalza-Mezghani et al. (2004) presented several options for
results of the proposed approach, we investigated three reservoir uncertainty management based on experimental-design and proxy
examples. The first one is a synthetic and simple case; the second models. These techniques are applied in the production forecasting
one is a synthetic but realistic case; and the third one is a real with uncertainty quantification (PUNQ) test case, a 3D synthetic
reservoir from the Campos basin of Brazil. The results presented in reservoir model derived from real field data. In this case, a probabi-
the paper show that it is possible to conduct an integrated study of listic distribution of the cumulative oil production was obtained for
uncertainty analysis and history matching. The main contribution several years. Zabalza-Mezghani et al. (2004) also shows the appli-
of this work is to present a practical way to increase the reliability cation of the joint modeling method to a real field with production
of prediction through reservoir-simulation models that incorpo- data. Using an optimization algorithm, this method consisted of
rate uncertainty analysis in the history period and provide reliable obtaining several matches with the same probability, which were
reservoir-simulation models for prediction forecast. extrapolated for production prediction under uncertainty, to obtain
a range of production curves. The history–matching process was
performed five times with different sets of initial deterministic para-
Introduction meters and with different initial geostatistical realizations.
Uncertainty management is important in all steps of petroleum- Kashib and Srinivasan (2006) proposed a method based on
field development. The objective of the uncertainty quantification conditional probability to update permeability distribution taking
and management is to improve the quality and reliability of the into account additional information contained in the dynamic re-
decisions that are based on the predictions generated by reservoir- sponse data. Varela et al. (2006) used seismic-amplitude data to
simulation models. study the influence on the production performance and to reduce
Some methods presented in the literature to estimate uncertain- uncertainty in the forecast of reservoir production. Assessing
ties in fields already in production proposed the generation of multi- uncertainty in production forecasts by constructing a synthetic
ple initial reservoir models, using geostatistical modeling techniques reservoir model, Varela et al. (2006) verified that the seismic-
and history match for each of them or part of the models (Bissel amplitude data did not uniformly improve the variability of the
et al. 1997; Jenni et al. 2004; Bennett and Graf 2002). However, this predictions of water-breakthrough time.
method requires elevated computational effort, and its practical Alvarado et al. (2005) proposed a procedure that includes like-
application within the industry can be difficult or unfeasible. Omre lihood of the forecast period on the basis of the quality of the match
and Lden (2004) used approximate fluid-flow simulators in order along with the objective function (OF). In other words, for a given
to reduce the required effort. Gu and Oliver (2005) applied the number of history-match runs, weighted mean and standard devia-
Kalman-filter method for automatic history matching to estimate tions of the corresponded-forecast runs are calculated and this
uncertainty in future reservoir performance. information is used to guide the history match and, at the same
Lépine et al. (1999) proposed an alternative method to esti- time, to generate confidence intervals of the prediction profiles.
mate uncertainties in the forecasting period. The authors presented Rotondi et al. (2006) presented a method for production fore-
the use of gradient information for uncertainty analysis. It is cast and uncertainty quantification, named the neighborhood algo-
assumed that, if a single simulation model is adjusted, the values rithm (NA), which consists of a stochastic sampling algorithm to
of the matched attributes can be perturbed slightly so that a range find acceptable data-fitting models. Suzuki and Caers (2006) also
of possible future production profiles will be obtained. It is an used the NA in conjunction with multiple-point geostatistics, in
interesting method because it is robust and not computationally which each scenario is depicted quantitatively by a training image
expensive. However, the method assumes that the model response and a geological-model realization is generated stochastically.
varies linearly with the perturbation. Using a similar idea, Landa Most of the methodologies presented in the literature—
and Güyagüler (2003) also used gradient information to assess although presenting good results—are complicated to implement,
uncertainties associated with flow prediction. The disadvantage such as the gradient technique [e.g., Lépine et al. (1999) and
of the gradient method is that the implementation is difficult Landa and Güyagüler (2003)], or require specific optimization
because it requires the assessment of the numerical-simulator methods [e.g., the Kalman filter presented by Gu and Oliver
code to calculate the gradients. Moreover, even with access to the (2005)], or require elevated computational effort. The main objec-
code, the implementation is not trivial. tive of the present paper is to propose a simple, but robust, method
for the incorporation of dynamic data into the uncertainty-analysis
workflow, showing a consistent manner in which to integrate the
Copyright ã 2009 Society of Petroleum Engineers probabilistic analysis to the history-matching process. The goal is
Original SPE manuscript received for review 2 April 2007. Revised manuscript received for
to reduce uncertainties in the history period and, as a conse-
review 17 June 2008. Paper (SPE 111030) peer approved 22 September 2008. quence, to reduce the uncertainties in the prediction period.

February 2009 SPE Reservoir Evaluation & Engineering 167

Fig. 1—General aspect of the proposed methodology: (a) re-
distribution of probability levels; (b) uncertainty curves OF
measures the quality of the matching; (c) redefinition of the
distribution values; and (d) the effect of the attribute redistribu-
tion on the production curves. Fig. 2—Scheme of the derivative tree technique (An_l, n = attri-
bute number, l = level number).

Proposed Methodology
also be applied to developed fields. The idea of the method is to
The general aspect of the method proposed in this work is the focus on the more-critical parameters. Sensitivity analysis is used
coupling of the uncertainty analysis in the historical period. The for screening the parameters with less of an impact and to select
reduction of uncertainty in the prediction period is a consequence the more-critical ones.
of the reduction of uncertainties constrained by observed data. In the proposed methodology, the probabilistic analysis is
In order to contextualize the method, one can think about the conditioned to the observed data and, as a result of this integra-
main stages of a petroleum field. In the initial field life, when tion, three methods are presented on the basis of probability
practically no dynamic data are available, traditional uncertainty redistribution
analysis is performed. For a field with a long production period, a • Method 1: using the same attribute values with new
large amount of observed data are normally available, which probabilities
allows the use of the traditional history matching process. Al- • Method 2: elimination of attribute levels
though the proposed method is more appropriate for intermediate • Method 3: redefinition of attribute values
cases, in which some historic data are available but a high degree
of uncertainties is still present in the reservoir description, it can

Fig. 3—Example of probability redistribution and level elimina-

tion: Methods 1 and 2. Fig. 4—Illustration of Method 3—Example 1.

168 February 2009 SPE Reservoir Evaluation & Engineering

 Fig. 6—Permeability region (md) of Model 1 with the true values.

technique. In the derivative tree, the probability-density function

of the uncertain attributes is discretized in a certain number of
levels. The uncertain levels are combined such that each combi-
nation (each branch of the tree) results in a flow simulation
(Fig. 2). The derivative tree is not a purpose of this paper. It is
used only as technique for uncertainty quantification. More details
about the derivative tree can be found in Steagall and Schiozer
(2001), Ligero et al. (2003), and Schiozer et al. (2004).

OF Calculation. The OF is the quantity that represents the mis-

match between the observed production data and the simulated
Fig. 5—Illustration of Method 3—Example 2. production data. The OF can be defined according to the character-
istics of each problem, and, in this work, it is defined as follows:
D
The objective of these methods is to reduce the uncertainty in OF ¼ E; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
the history period and, consequently, to reduce the uncertainty in jDj
the prediction period.
In Fig. 1, a general aspect of the proposed method is presented.
Fig. 1a illustrates an example of a probability-density function and
a discretization in three levels. Fig. 1a also illustrates the redefini-
tion of the probability of the three discrete levels. In Fig. 1b, the
objective function is plotted against the cumulative probability,
which is referred to here as the uncertainty curve. The vertical line
represents the history data. This signifies that if an uncertainty
curve is close to the vertical line, it is well adjusted to the history.
Fig. 1c illustrates the redefinition of the distribution values, and
Fig. 1d shows the effect of this redistribution on the production
curves in the history and prediction periods. The production
curves close to the history represent the reduction of uncertainty
on the production profiles, as a result of the smaller range between
limitsof the attribute. Further, the methods used to redefine the
probabilities and to redefine the attribute values (redefine the
distribution) will be explained better.
In this work, the uncertainty quantification—that is, the sce-
nario’s composition on the basis of the uncertainties—is per-
formed through the derivative tree using flow simulation.
However, other techniques can be used, such as the Monte Carlo

Fig. 7—Permeability map (md) of Model 2: fine-grid model

(upper) and coarse (base) model (lower).

February 2009 SPE Reservoir Evaluation & Engineering 169

where Method 1—Probability Redistribution. This method uses the
distance between the simulation models and the observed data to
X
N
redistribute the occurrence probability of the attribute levels. The
D¼ ðdiobs  disim Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
new probabilities (PL) of each level of a given attribute are calcu-
i¼1
lated according to Eq. 4:
and
ð1=jSjÞðEL Þ1
X
N PL ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
P
k
E¼ ðdiobs  disim Þ2 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3) ð1=jSjÞðEL Þ1
i¼1 L¼1

where N is the number of observed data and diobs and disim are where
observed and simulated data, respectively. The square error (E) !
presented in Eq. 3 is a common manner to mathematically repre- X
ML
EL ¼ Ej ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5)
sent the difference between observed and simulated data (mis-
j¼1
match). The term D/|D| in the Eq. 1 is used to define the sign of
the mismatch. It indicates the position of simulated-data curves and
with respect to the observed-data curves. For instance, if a
simulated production curve (e.g., water rate) is predominantly D
S¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (6)
under the observed water rate, the sign of the mismatch is posi- jDL j
tive. On the other hand, the sign is positive. This is important to
better understand the distribution of the models with respect to the In Eqs. 4 through 6, k is the number of level; ML is the number of
observed data in the uncertainty curve. models related to the level L; and the term EL is the sum of square
error (EL ) of the ML models related to the level (L); and S is the
symmetry of the level. The term EL is fundamental in the proposed
methodology. To better understand it, consider the three models
represented by A2B0, A2B1, and A2B2 in Fig. 3. The term EL
(Eq. 5) for Level A2 is the sum of the square error E (calculated
through Eq. 3) corresponding to these three models. For Level A0,
the term EL is the sum of E correspondent to the models repre-
sented by A0B0, A0B1, and A0B2, similarly for Level A1.
For the attribute B2, the curves are grouped according to Levels
B0, B1, and B2. Therefore, EL for the level B0 is the sum of
E correspondent to the models B0A0, B0A1, and B0A2 and
similar for B1 and B2.
The use of symmetry (S) in Eq. 4 was set in order to give a
larger probability for uncertainty levels with curves more uni-
formly distributed around the observed data. In this work, a
maximum value of 5 for the term 1/S (if 1=jSj  5, then
1=jSj ¼ 5) was set to avoid weights being too high and conse-
quently influencing of this factor excessively (in practice, |S| was
varied from 0.2 to 1). This value can be calibrated according to
the case studied.
The key idea of Method 1 is illustrated in Fig. 3. The example
shows theoretical curves for nine scenarios composed with
two attributes with three uncertainty levels for each one. The
three groups of curves represent the combination of the
three levels of Attribute A (A0, A1, and A2) with each level
of Attribute B. According to Method 1, Level A2 has the lar-
gest probability because all correspondent models are closest to
the observed data. On the other hand, Level A1 has the smallest
probability.

Method 2—Elimination of Attribute Levels. The elimination of

a given uncertainty level follows two criteria: (1) the probability
of the level is smaller than a given cut-off value and (2) the
symmetry equals 1 or -1. Considering the example shown in
Fig. 8—Cross section (Depth) and 3D model of NTG (Case 3). Fig. 3, Level A1 is eliminated. After the elimination of a given

170 February 2009 SPE Reservoir Evaluation & Engineering

level, the probability of the remaining levels is recalculated such The new upper limit is given by
that the sum of the probabilities equals 1.
LNS ¼ A1 þ ðA0  A1Þ  P½A1; . . . . . . . . . . . . . . . . . . . . . (11)
Method 3—Redefinition of Attributes Values. This method
and the new lower limit is given by
consists in the redefinition of the uncertainty levels according to
the distribution of the models related to each level of the attri- LNi ¼ L þ ðA1  Li Þ  ð1  P½A1Þ: . . . . . . . . . . . . . . . . . . (12)
butes. Given the example in Fig. 4, the new limits are given by
This means that the new upper limit of Attribute A is between A1
LNS ¼ Ls ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (7) and A0 and the new lower limit is between A1 and the original lower
limit. With the new limits and new probable level, according to
and
the triangular distribution, the new pessimistic and optimistic levels
LNS ¼ A2: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (8) (A1N and A2N, respectively) are calculated. This method implies
new simulations once the values of the attribute are changed.
This means that the new limit of Attribute A is the original upper
limit, and the new inferior limit is the delimited by Level A2. Applications
The new probable level is calculated as Case 1. The application of the method was first illustrated with a
simple reservoir model (Fig. 6), used by Moura Fillho (2005). It is
A0N ¼ ðLNs þ LNi Þ=2: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (9) an idealized homogeneous reservoir with four flux regions in a
five-spot pattern (four producers and one injector). Each region
Consider now the example in Fig. 5; in this case, the new proba- has different permeability values between 1,109 and 7,084 md
ble level is calculated as (the true permeability values of the four regions are shown in
    Fig. 6). The simulation of this model generated a synthetic pro-
A0  P½A0  1=jS0j þ A1  P½A1  1=jS1j duction history for 10 years. The base case was generated assign-
A0N ¼     :(10) ing the same permeability value (4,500 md) to all flux regions.
P½A0  1=jS0j þ P½A1  1=jS1j The problem consists of four uncertain attributes, given by kx1,

Fig. 9—Water-rate production for PROD1 grouped by uncertain- Fig. 10—Water-rate production for PROD3 grouped by uncer-
ty level of kx1. tainty level of kx3.

February 2009 SPE Reservoir Evaluation & Engineering 171

 Fig. 12—Uncertainty curves for PROD3 (Case 1).
Fig. 11—Uncertainty curves for PROD1 (Case 1).

Fig. 13—Comparison of the initial scenarios and after the appli- Fig. 14—Comparison of the initial scenarios and after the appli-
cation of Method 3 (water rate of the PROD1—Case 1). cation of Method 3 (water rate of the PROD3—Case 1).

The conditional sequential-indicator-simulation (SIS) method

kx2, kx3, and kx4. In Table 1, we show the probability distribution has been used to model the geological facies. Within the reservoir
of each uncertainty level for each of the four attributes, and the rocks, porosity and permeability were assigned with noncondi-
upper and lower limits. The OF analyzed in this case was asso- tional sequential Gaussian simulation (SGS). A constant porosity
ciated with the water rate of the four producer wells. The objec-
tive of this simple model was to check the consistency of the
methods through a controlled case with known response.

Case 2. Case 2 is a realistic reservoir model built using geostatis-

tical techniques. The reservoir characterization was based on the
log data of six wells. The interpretation from the logs allowed the
identification of the intervals with reservoir rocks. Two major
facies—sandstones and shale—have been distinguished on the
basis of well-log interpretation.

Fig. 15—Sensitivity analysis for the field.

172 February 2009 SPE Reservoir Evaluation & Engineering

 Fig. 17—Uncertainty curves for field (Case 2).
Fig. 16—Sensitivity analysis for Well PROD1.

Fig. 19—Curves of average reservoir pressure for the initial

Fig. 18—Curves of field water rate for the initial scenarios and scenarios and after the application of Method 3, Steps 1 and 2.
after the application of Method 3, Steps 1 and 2.
The reference model was discretized with a grid of 504634
of 5% and permeability of 0.5 md were assigned to the shale rock blocks, with block dimensions of 80602 m. The base model,
type. Two different geological models were generated. One of after the upscaling, was discretized with a grid of 252334 blocks.
them was used as a reference model to generate a production and The drainage strategy was composed of 12 vertical wells, six
pressure history of 10 years. The other model was scaled up and producers and six water injectors.
was used to generate the base model.

Fig. 21—Water-rate curves for the initial scenarios and after the
Fig. 20—Uncertainty curves for Well PROD1. application of Method 3, Steps 1 and 2 (PROD1)

February 2009 SPE Reservoir Evaluation & Engineering 173

 3D maps of the models are shown in Fig. 7, which also shows value of kx1 is 1,109 md. Therefore, the closest value is Level 1
the position of the wells. The upper one is the fine-grid model, and of the kx1, which has an initial value of 2,096 md, according to
the lower one is the coarse (base) model. The uncertain attributes Table 1. As expected, it must receive the higher probability. For
for Case 2 are described in Table 2. These parameters were chosen kx3 (Fig. 10), the new probabilities are 25.9% for Level 0, 68.3%
on the basis of the knowledge of the premises of the reference for Level 1 and 5.8% Level 2. Level 2 of kx1 and kx3 is eliminated
model and also on a previous analysis of the base model. according to Method 2.
The uncertainty curves for Wells PROD1 and PROD3 are
Case 3 (Real Reservoir). The third case application (Fig. 8) is a presented in Figs. 11 and 12, respectively. The proximity of these
field case from the Campos basin of Brazil. It is a developed field curves with respect to the vertical line denoting the zero value
with more than 20 years of production, of which approximately represent the quality of the matching of each model represented
one-half was used in the study. The reservoir is drained by 33 by a point in the curve. In the two cases, a significant reduction of
producer wells and 13 water-injector wells. The injection began uncertainty can be noted in the curves obtained with Methods 1
after 5 years of primary production. There are two main oil zones, and 2. However, Method 3 yields a better match.
with an aquifer. Although a reasonable history period, the field A comparison of the initial scenarios and the ones obtained
presents some difficulties to be matched. The parameterization is after the application of Method 3 is shown in Figs. 13 and 14. The
based on the remaining uncertainties in the petrophysical proper- reduction in the range of water-rate curves for the final scenarios
ties (mainly horizontal, vertical, and relative permeability) and the is expressive, and they are well distributed around the history.
influence of the aquifer (Becerra 2007). Table 4 shows the new limits and uncertain levels after the appli-
The reservoir model is represented by a corner-point grid cation of Method 3. In order to check the consistency of the
with 65358 cells. The uncertain attributes taken into account method, one can observe—as expected—that the new most-
in the analysis are described in Table 3. Each attribute is char- probable values of the four attributes are closer to the true values.
acterized by a triangular probability distribution, discretized This simple reservoir model was used as a test case to check
in three uncertain levels: Level 0 (referred to as the probable the consistency of the proposed method. In the next sections are
level), with 60% of occurrence probability, and Levels 1 and presented the results for a realistic reservoir model (Case 2) and a
2 (called the pessimistic and optimistic levels, respectively), real reservoir (Case 3). The number of simulations run for this
both with 20% of occurrence probability. The OF analyzed was case was 81 (= 34 that correspond to four attributes with three
the field water rate. levels) for Method 1 and Method 2, and 81 for Method 3.

Case 1. Sensitivity Analysis. The first step in the study of Case

Results
2 was the sensitivity analysis. The main objective of sensitivity
Case 1. In order to illustrate the probability redistribution of the analysis is to study the influence of the attributes over the
uncertain levels, the water-rate production curves for PROD1 and OF related to the six producer wells and across the field. The
PROD3 are presented in Figs. 9 and 10, respectively, grouped by influence of the attributes for the field and PROD1 is shown in
levels. The new probabilities calculated according to Method 1 for Figs. 15 and 16, respectively. It can be seen that for the field the
kx1 is 12.1% for Level 0, 82.6% for Level 1, and 5.3% for Level 2. most important attribute is porosity. For the well PROD1, five
It can be observed that the models that contain Level 1 are closer attributes exert significant influence (more than 10%). For the
to the history, and, consequently, a higher probability is assigned other wells, the same five attributes are important. Therefore, on
to this level. Additionally, one can observe in Fig. 6 that the true the basis of the sensitivity analysis, five attributes were selected
for the further uncertainty analysis.
Quantification and Reduction of Uncertainties. The quantifi-
cation and reduction of uncertainties process for Case 2 was
carried out in two steps. In the first one, the OF analyzed was
the field water rate.
The uncertainty curves for the field are shown in Fig. 17. The
expressive reduction of uncertainty obtained with Methods 1 and
2 can be seen in Fig. 16. In Method 2, Level 1 of porosity (PorSD)
was eliminated. It can also be seen that in the region near Percen-
tile P10, there is a very good match in the uncertainty curves
corresponding to these two methods. It is clear that Method 3
yields the best results. The uncertainty curve obtained with this
method is very close to the vertical line crossing the OF axis
zero. The curve related to the Method 3 (Step 2) shows a slight
improvement with respect to Method 3 (Step 1).
In Figs. 18 and 19, the curves of the water rate and average
reservoir pressure, respectively, are presented for all scenarios in
the history period before (initial scenarios) and after the

174 February 2009 SPE Reservoir Evaluation & Engineering

application of Method 3. A significant reduction in the range of simulations run related to the sensitivity analysis were used again
curves is observed. It is important that the history data are sym- in the derivative tree. For Case 2, five attributes with three levels
metrically positioned in the envelope after the reduction of uncer- each were selected from the sensitivity analysis, yielding 243
tainty, mainly in the water-rate curves. A good match in the models composed according to the derivative tree.
average reservoir pressure and a great reduction in the curve Prediction. All initial scenarios and those resulting from the
dispersion is observed. quantification and reductions of uncertainties (Method 3, Step 2)
After the first step, well analysis was performed in a second step. in the history period were extrapolated for production forecast. In
It was noted that two attributes, water/oil contact (WOC) and kxA, practice, generally a few models are chosen for prediction pur-
still exerted a significant influence, mainly on three wells (PROD1, pose, for example some percentiles such as P10, P50, and P90;
PROD4, and PROD5). Therefore, a second iteration of Method 3 however, the objective of extrapolating all scenarios is to show
was applied for these two attributes. The uncertainty range related to that all models after the application of Method 3 compose a small
the other three attributes was maintained as obtained after the first uncertainty range. For the prediction step, 243 simulations were
application of Method 3 (Step 1). Therefore, in Method 3, Step 2, run (Table 7).
five attributes composed the analysis, such that three of them have The results for the cumulative oil production, cumulative water
the uncertainty range obtained in the previous step. Additionally, the production, and for the average reservoir pressure are presented in
reason to choose a well analysis in the second step is that globally, Figs. 22, 23, and 24, respectively. The extrapolation corresponds
the water rate of field was well adjusted. Therefore, the objective to a period of 10 years. The vertical lines in the plots indicate the
was to improve the matching in some wells. end of the history period and the beginning of the prediction. Note
In Fig. 20, we show the uncertainty curves for PROD1. In this
case, the initial scenarios correspond to the scenarios obtained
after the application of Method 3 in Step 1. Method 2 was not
applied in this case. In Fig. 21, a comparison of the range of
the water rate for the initial scenarios is shown [in this case the
original scenarios (e.g, before the application of Method 3)], the
range of the water rate after the application of Method 3 in Steps 1
and 2. The evolution of in the improvement can be observed.
In Table 5, the new limits and uncertainty levels after the
application of Method 3, Steps 1 and 2, is shown. The reduction
in the range of the attributes kxSD and WOC, from Step 1 to Step
2, can be observed.
Table 6 shows the overall reduction (%) of mismatch before
and after the application of Method 3 (Step 1) and Method 3 (Step
2) for the field and for the six wells. In general, significant reduc-
tions of the mismatch can be observed after the first application of
Method 3. Only in Well PROD6 is this not observed. Another
observation is that for the field, there was a smaller reduction in
the second step, and that was expected because of the large reduc-
tion in the first step.
The number of simulations run in the study of Case 2 is shown Fig. 23—Uncertainty reduction in prediction of cumulative
in Table 7. It should be observed that for Methods 1 and 2, the water production.

Fig. 22—Uncertainty reduction in prediction of cumulative oil Fig. 24—Uncertainty reduction in prediction of average reser-
production. voir pressure.

February 2009 SPE Reservoir Evaluation & Engineering 175

 Fig. 25—Uncertainty curves for the water rate of the field (Case 3).

that in the initial scenarios, there are models that do not honor Method 3 are shown in Fig. 27 and Fig. 28, respectively. A
observed oil-production data (Fig. 22) between approximately significant reduction in the range of production-water rate in com-
2,000 and 3,600 days. These models are composed by the pessi- parison to the initial curves (before reduction of uncertainty) is
mistic level of the porosity combined with the other attributes that shown.
result in insufficient oil volume.
It is possible to observe that the initial uncertainty in the oil Impact of the Uncertainty Reduction on the Field Prediction.
production was almost 5.3 million m3. After the uncertainty The reduced ranges of the critical attributes consequently lead
reduction, this range was reduced approximately 95%. A very to a reduction of production-prediction spread, which varies
expressive reduction of uncertainty in water production and according to the method used. In the models extrapolated for
reservoir pressure was also observed. prediction, the producer wells were controlled by minimum
bottomhole pressure and maximum liquid rate and the injectors
Case 3. A sensitivity analysis indicated four critical attributes were controlled by maximum bottomhole pressure and maxi-
(krw, vm, aq, and kv), which had more impact in the OF analyzed. mum water rate. In this case, only the cases representing the
These attributes were then used in the next stages. 10% (P10) and 90% (P90) percentiles were chosen, according
Quantification and Reduction of Uncertainty. The uncertain- to the uncertainty curves. The field oil and water rate, in the
ty curves for the water rate of the field are shown in Fig. 25, history and prediction period, before and after the application
which presents a comparison of the three methods. In this case, of the presented methodology, are shown in Figs. 29 and 30,
the OF is normalized with respect to the largest value. In Fig. 26, respectively. Similarly, Figs. 31 and 32, respectively, show the
we show the uncertainty curves after two iterations of Method 3. field oil and water rates after the reduction of uncertainty.
In the second iteration, the uncertainty is slightly reduced com- These figures show an important reduction in the range be-
pared with the uncertainty related to the first iteration. The field- tween P10 and P90. The reduction of uncertainty observed after
water-rate curves, grouped according the uncertainty levels of krw, the application of the method can decrease the risk associated
before reduction of uncertainty and after the application of with the decisions taken with the support of the predictions.

Fig. 26—Uncertainty curves for the water rate of the field after two iterations of Method 3 (Case 3).

176 February 2009 SPE Reservoir Evaluation & Engineering

 Fig. 27—Water-rate curves grouped according to the uncertainty levels of krw before reduction of uncertainty.

General Discussion but it has a higher probability of obtaining better answers and
The methods presented in this work are appropriate to cases with a predictions.
high level of uncertainty, in which matching is difficult to achieve For complex and realistic cases, additional techniques or a
because multiple answers are possible. The advantage is that it division of the process into steps may be necessary. In this work,
naturally yields prediction under uncertainty. Another advantage sensitivity analysis was used in order to select the most influential
is that the implementation is simpler than the gradient method attributes. Other techniques, such as proxy models, based on
(Lépine et al. 1999; Landa and Güyagüler 2003) used to estimate experimental design or an ANN, for example, can also be used.
uncertainties in the prediction period. Besides the implementation, The application of these techniques in conjunction with the pro-
another important point is that the gradient method is applied only posed method can be addressed in future works.
for linear problems, which makes the technique very restrictive. The methods presented in this work are more appropriate to
The method presented here can be used in complex nonlinear cases with a high level of uncertainties. However, they also can be
problems, which form the bulk of cases involving reservoir applied to other cases where matching is difficult to achieve,
simulation. yielding multiple answers.
The methods are also flexible, allowing adaptation to each
problem (i.e., OF, number of levels, different criteria to Conclusions
eliminate levels). Methods 1 and 2 are faster because no A consistent and flexible method to integrate the history-
additional simulation is required after the uncertainty quantifica- matching process with the uncertainty analysis was presented
tion. Method 3 yields a higher number of simulation runs, in this work. The method yields a natural way of prediction

Fig. 28—Water-rate curves grouped according the uncertainty levels of krw after application of Method 3.

February 2009 SPE Reservoir Evaluation & Engineering 177

 Fig. 29—Field oil-rate prediction: Models P10 and P90 before reduction of uncertainty.

under uncertainty after a reduction of uncertainty in the history dsim = simulated data
period. Among the three proposed methods, based on redistri- D = error between observed and simulated data
bution of probabilities, the use of the same vales and values E = square error between observed and simulated data
(Method 1) and level elimination (Method 2) provided a fast EL = sum of misfit (square error) of all models of Level L
way to redefine the uncertainty trends. The advantage of these k = number of levels
methods is that no additional simulation is needed after the
kh = horizontal permeability
uncertainty analysis and they yield a good reduction in uncer-
tainty curves. Method 3 is more efficient in the reduction of krSD = relative permeability of the sandstone
uncertainty; however, it requires additional simulations. The krw = water relative permeability
necessity of applying Method 3 depends on the problem kv = vertical permeability
and on the precision required, similarly for its application in kxSD = horizontal permeability of the sandstone
several steps. kxSH = horizontal permeability of the shale
kxz = ratio between vertical and horizontal permeability
Nomenclature LiN = new inferior limit
aq = aquifer LsN = new superior limit
A0 = probable level ML = number of models of Level L
A0N = new probable level N = number of observed data
A1 = pessimistic level Np = cumulative oil production
A2 = optimistic level porSD = porosity of sandstones
Avg P = average reservoir pressure vm = pore volume multiplier
dobs = observed data (history) Wp = cumulative water production

Fig. 30—Field water-rate prediction: Models P10 and P90 before reduction of uncertainty.

178 February 2009 SPE Reservoir Evaluation & Engineering

 Fig. 31—Field oil-rate prediction: Models P10 and P90 after reduction of uncertainty (Method 3).

SL = symmetry of Level L Becerra, G.G. 2007. Uncertainty Mitigation Through the Integration with
S0 = symmetry of Level A0 Production History Matching. MS Thesis, Petroleum Engineering
S1 = symmetry of Level A1 Department, State University of Campinas, Brazil.
S2 = symmetry of Level A2 Bennett, F. and Graf, T. 2002. Use of Geostatistical Modeling and Auto-
matic History Matching to Estimate Production Forecast Uncertainty—
P = probability
A Case Study. Paper SPE 74389 presented at the SPE International
Pn = probability of Level n
Petroleum Conference and Exhibition in Mexico, Villahermosa,
P[A0] = probability of Level A0 Mexico, 10–12 February. DOI: 10.2118/74389-MS.
P[A1] = probability of Level A1 Bissel, R.C., Dubrule, O., Lamy, P., Swaby, P., and Lépine, O. 1997.
P[A2] = probability of Level A2 Combining Geostatistical Modeling With Gradient Information for
History Matching: The Pilot Point Method. Paper SPE 38730 pre-
sented at the SPE Annual Technical Conference and Exhibition, San
Acknowledgments Antonio, Texas, USA, 5–8 October. DOI: 10.2118/38730-MS.
The authors would like to thank the Brazilian Council for Research Gu, Y. and Oliver, D.S. 2005. History Matching of the PUNQ-S3 Reser-
and Development—Conselho Nacional de Desenvolvimento Cien- voir Model Using the Ensemble Kalman Filter. SPEJ 10 (2): 217–224.
tı́fico e Tecnológico (CNPq/PROSET), Finep/CTPETRO, Petro- SPE-89942-PA. DOI: 10.2118/89942-PA.
bras, Engeria S.A., CEPETRO, and Agência Nacional de Petróleo Guérillot, D. and Pianelo, L. 2000. Simultaneous Matching of Production
(ANP) for supporting this research and development project. Data and Seismic Data for Reducing Uncertainty in Production Fore-
casts. Paper SPE 65131 presented at the SPE European Petroleum
Conference, Paris, 24–25 October. DOI: 10.2118/65131-MS.
References Jenni, S., Hu, L.Y., Basquet, R., de Marsily, G., and Bourbiaux, B. 2004.
Alvarado, M.G., McVay, D.A., and Lee, W.J. 2005. Quantification of History Matching of Stochastic Models of Field-Scale Fractures:
uncertainty by combining forecasting with history matching. Issues in Methodology and Case Study. Paper SPE 90020 presented at the SPE
Mental Health Nursing 23 (3): 445–462. DOI:10.1081/LFT- Annual Technical Conference and Exhibition, Houston, 26–29 Sep-
200031057. tember. DOI: 10.2118/90020-MS.

Fig. 32—Field water-rate prediction: Models P10 and P90 after reduction of uncertainty (Method 3).

February 2009 SPE Reservoir Evaluation & Engineering 179

Kashib, T. and Srinivasan, S. 2006. A probabilistic approach to integrating Simulation Symposium, Houston, 11–14 February. DOI: 10.2118/
dynamic data in reservoir models. J. of Petroleum Science and Engi- 66399-MS.
neering 50 (3–4): 241–257. DOI:10.1016/j.petrol.2005.11.002. Suzuki, S. and Caers, J. 2006. History Matching With an Uncertain
Landa, J.L. and Güyagüler, B. 2003. A Methodology for History Match- Geological Scenario. Paper SPE 102154 presented at the SPE Annual
ing and the Assessment of Uncertainties Associated With Flow Pre- Technical Conference and Exhibition, San Antonio, Texas, USA,
diction. Paper SPE 84465 presented at the SPE Annual Technical 24–27 September. DOI: 10.2118/102154-MS.
Conference and Exhibition, Denver, 5–8 October. DOI: 10.2118/ Varela, O.J., Torres-Verdı́n, C., and Lake, L.W. 2006. On the value of 3D
84465-MS. seismic amplitude data to reduce uncertainty in the forecast of reser-
Lépine, O.J., Bissel, R.C., Aanonsen, S.I., Pallister, I.C., and Barker, J.W. voir production. J. of Petroleum Science and Engineering 50 (3–4):
1999. Uncertainty Analysis in Predictive Reservoir Simulation Using 269–284. DOI:10.1016/j.petrol.2005.11.004.
Gradient Information. SPEJ 4 (3): 251–259. SPE-57594-PA. DOI: Zabalza-Mezghani, I., Manceau, E., Feraille, M., and Jourdan, A. 2004.
10.2118/57594-PA. Uncertainty management: From geological scenarios to production
Ligero, E.L., Maschio, C., and Schiozer, D.J. 2003. Quantifying the Im- scheme optimization. J. of Petroleum Science and Engineering 44
pact of Grid Size, Upscaling, and Streamline Simulation in the Risk (1–2): 11–25. DOI:10.1016/j.petrol.2004.02.002.
Analysis Applied to Petroleum Field. Paper SPE 79677 presented at
the SPE Reservoir Simulation Symposium, Houston, 3–5 February. University of Campinas (UNICAMP), where he develops
DOI: 10.2118/79677-MS. research in the areas of reservoir engineering and simulation.
Moura Filho, M.A.B. 2005. Integration of Uncertainty Analysis and E-mail: [email protected]. He holds a mechanical
History Matching Process. MS Thesis, Petroleum Engineering Depart- engineering degree from UNESP (Universidade Estadual
ment, State University of Campinas, Brazil. Paulista), and an MSc and PhD degrees in mechanical engi-
Omre, H. and Lden, O.P. 2004. Improved Production Forecasts and neering from UNICAMP. Denis José Schiozer is a professor in
the petroleum engineering department at UNICAMP, where
History Matching Using Approximate Fluid-Flow Simulators. SPEJ 9
he conducts several research projects in the areas of reservoir
(3): 339–351. SPE-74691-PA. DOI: 10.2118/74691-PA. engineering and simulation. E-mail: [email protected].
Queipo, N.V., Pintos, S., Rincón, N., Contreras, N., and Colmenares, J. br. He is coordinator of the UNISIM-UNICAMP Research Group
2002. Surrogate modeling-based optimization for the integration of in reservoir simulation. Schiozer holds an aeronautical engi-
static and dynamic data into a reservoir description. J. of Petroleum neering degree from the Instituto Tecnológico da Aeronáu-
Science and Engineering 35 (3–4): 167–181. DOI:10.1016/S0920- tica (ITA), an MSc degree in petroleum engineering from
4105(02)00238-3. UNICAMP, a PhD degree in petroleum engineering from
Reis, L.C. 2006. Risk Analysis With History Matching Using Experimen- Stanford University, and an MBA degree from Fundação
tal Design or Artificial Neural Networks. Paper SPE 100255 presented Getúlio Vargas (FGV). Marcos Moura Filho is a reservoir engi-
neer at Petrobras in Brazil. E-mail: mmourafilho@petrobras.
at the SPE Europec/EAGE Annual Conference and Exhibition, com.br. He worked at Norse Energy in Rio de Janeiro on reser-
Vienna, Austria, 12–15 June. DOI: 10.2118/100255-MS. voir simulation and field operations. Moura Filho holds an MSc
Rotondi, M., Nicotra, G., Godi, A., Contento, F.M., Blunt, M.J., and degree in petroleum science and engineering from UNICAMP
Christie, M.A. 2006. Hydrocarbon Production Forecast and Uncertain- and a BSc degree in mechanical engineering from Federal
ty Quantification: A Field Application. Paper SPE 102135 presented at University of Ceará (UFC), Brazil. Gustavo G. Becerra is now a
the SPE Annual Technical Conference and Exhibition, San Antonio, senior reservoir engineer with Petrobras Energia S.A. working
Texas, USA, 24–27 September. DOI: 10.2118/102135-MS. currently at the research and development center in Rio de
Schiozer, D.J., Ligero, E.L., Suslick, S.B., Costa, A.P.A., and Santos, Janeiro, Brazil, on history matching and predictive and uncer-
tainty analysis with reservoir simulation. E-mail: gustavo.becer-
J.A.M. 2004. Use of representative models in the integration of risk
[email protected]. Previously, he worked covering several
analysis and production strategy definition. J. of Petroleum Science and positions in reservoir engineering at headquarters of Petrobras
Engineering 44 (1–2): 131–141. DOI:10.1016/j.petrol.2004.02.010. in Venezuela and Research Center in Neuquén, Argentina. He
Steagall, D.E. and Schiozer, D.J. 2001. Uncertainty Analysis in holds an MSc degree in petroleum science and engineering
Reservoir Production Forecast during the Appraisal and Pilot Pro- degree from UNICAMP and a BSc in hydraulic engineering
duction Phases. Paper SPE 66399 presented at the SPE Reservoir from National University of La Plata, Argentina (UNLP).

180 February 2009 SPE Reservoir Evaluation & Engineering