You are on page 1of 11

IPTC 11119

Using Production Data to Mitigate Reservoir Connectivity Uncertainty


Hong Tang, SPE, Louisiana State University

Copyright 2007, International Petroleum Technology Conference connectivity. The updated models help capturing the full uncertainty
This paper was prepared for presentation at the International Petroleum Technology ranges for recovery factor.
Conference held in Dubai, U.A.E., 4–6 December 2007. The updated recovery uncertainty as well as reduced modeling
This paper was selected for presentation by an IPTC Programme Committee following review parameters uncertainties help correct business decisions. Furthermore,
of information contained in an abstract submitted by the author(s). Contents of the paper, as the updated response surfaces show confounded higher order
presented, have not been reviewed by the International Petroleum Technology Conference
and are subject to correction by the author(s). The material, as presented, does not nonlinear effects. Therefore, further study based on more detailed
necessarily reflect any position of the International Petroleum Technology Conference, its design is desired. The method in this paper honors available data
officers, or members. Papers presented at IPTC are subject to publication review by Sponsor
Society Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this without over-tuning geological parameters for history matching. It
paper for commercial purposes without the written consent of the International Petroleum also provides guidance for data acquisition to mitigate production
Technology Conference 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 prediction risks.
contain conspicuous acknowledgment of where and by whom the paper was presented. Write
Librarian, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
1. Introduction
The goals of this work are 1) reducing modeling parameters
Abstract uncertainty by integrating dynamic data; 2) changing the reservoir
Even though production data have direct responses of reservoir connectivity to capture correct uncertainty ranges; 3) developing a
heterogeneity and connectivity, they are rarely incorporated into flexible, integrated method to model sedimentary facies connectivity.
reservoir modeling workflow among the geological community. In Stochastic reservoir modeling technology provides geologists and
this paper, a designed simulation method is proposed to mitigate engineers an efficient way to integrate various types of data for
reservoir connectivity uncertainty. This proposed method is more predicting inter-well reservoir properties. At the early stage of
accurate and efficient by integrating production data with reasonable reservoir exploration, well data are usually limited, which pose a big
computational cost. This method, applied on a North Africa shallow challenge of inter-well prediction. Those reservoir properties are then
marine reservoir, includes following four steps. used to compute oil in place (OIP) on which business decisions are
1. A Folded Plackett-Burman design (FPBD) is used to generate usually based. In the study area, existing geological data are limited
typical reservoir facies models using multipoint geostatistics method. with poor quality. One major object is then to integrate all available
Then reservoirs properties (i.e. porosity, permeability and saturation) data to accurately estimate the reservoir properties using accurate
are populated among different facies. The different scenarios of modeling parameters.
reservoir properties represent geological uncertainties. Another objective of this study is to capture the connectivity
2. The recovery factor of each model is computed by using a uncertainty ranges by integrating production data. The original oil in
commercial black oil simulator. A polynomial response surface of place (OIP) is a good proxy for recoverable oil. However, this proxy
recovery factor and modeling parameters is generated as a proxy of might not be true in reservoirs with low net-to-gross ratio or high
flow responses. A Monte Carlo Simulation Bayes Method (MCSBM) heterogeneity such as carbonate reservoirs. Reservoir connectivity
estimates the uncertainty of recovery factors and derives the posterior analysis provides a better alternative to estimate for recoverable oil.
probabilities for modeling parameters. The reservoir connectivity is also important for infill well planning,
3. The facies models are changed by using a probability enhance oil recovery, and current production optimization. However,
perturbation method, which reduces the error between observed the production data such as well transient test data and production
pressure data and simulation prediction. Different from traditional history are less practically used in connectivity prediction. There are
methods, the objective of this step is to mitigate connectivity two major directions for dynamic data integration. Direct analytical
uncertainty by integrating production data instead of forcing history estimation of reservoir properties (such as KH, well distance to faults
match. etc.) from well transient test or PLT logs is informative. Since these
4. Similar to step 2, another polynomial response surface between data can be tested in short time, they are mostly used to estimate
updated recovery factor and modeling parameters is generated by average reservoir properties. However, they are seldom used to
using perturbed reservoir models. MCSBM derived the conditional update reservoir geological models. History match/ assisted history
probability of modeling parameters and probability distribution of match (HM/AHM) are the champion of inversely integrating dynamic
recovery factors. Both derived probabilities are compared with initial data for model updating2,4-6,9,13,15. The major limitation of most
models. current history match methods is that they failed to change the
Sensitivity analysis shows that recovery factor heavy hitters of reservoir model in a way consistent with geological concepts. Thus
updated models are significantly different from initial models. the matched model may not be able to correctly predict the
MCSBM shows that the production updated modeling parameters production. Furthermore, because of non-uniqueness and
have smaller uncertainty ranges than the prior uncertainty ranges, nonlinearality of solutions, the convergence of objective function
which means the production update reduces parameters uncertainty. strongly depends on the initial models; wrong geological models may
The updated recovery factors have wider uncertainty ranges (18 either converge to local minimum or fail to converge9. Some
percent) compared with initial recovery factors (15 percent). This is algorithms (such as metropolitan algorithm, Simulated Annealing, the
understandable because the updated geological models honoring Tunneling Method, and Genetic Algorithms 2,9,13) help find the global
production data may be more heterogeneity and have tortuous minimum of objective function. But these methods usually require
2 IPTC 11119

extensive simulations runs, which are prohibitively time consuming An inner loop locally adjusts facies realizations. An outer loop
for complex reservoirs or immature reservoirs such as study area that continuously adds new information till history matched.
has limited well data and short production history. A practical
reservoir modeling workflow should be able to mitigate uncertainties Folded Placket-Burman Design (FPBD) and Response
of reservoir parameters by integrating all available data instead of Surface Method
pushing for history match. Placket-Burman designs (PBD) are very useful economically for
In this paper, a recently developed Probability Perturbation Method detecting large main effects, assuming all interactions are negligible
(PPM) is used in integrating production data by adjusting reservoir when compared with the few important main effects2,3,8,12. However,
models honoring geological concepts 1,5-7. An experimental designed when such assumption is not satisfied, the lack of ability to handle
Monte Carlo Simulation Bayes Method (MCSBM) is used to detect interactions and curvatures may give rise to significant prediction
the factor sensitivity, quantify reservoir parameters’ uncertainties. error12. A FPBD or mirror-image fold-over design is used to increase
Two sets of models are compared in this paper. One is based on a the resolution of PBD. It is obtained by reversing the signs of all the
FPB Design and direct flow responses (initial model); another model columns of the original design matrix. The original design runs are
is based on the same initial model. But each model is updated by combined with the mirror-image fold-over design runs. This
production data using PPM method, which is called updated model. combination can then be used to estimate all main effects clear of any
This paper starts with an introduction of methods and proposed two-factor interaction. A FPBD is a method to build a design with
workflow. It is followed by an application on a North Africa shallow resolution for interactions from a design with resolution for main
marine reservoir. Finally, the application results are discussed. factors only.
Selection of design is a subjective decision. It depends on the purpose
2. Methods of design, number of factors, cost of design runs, and nonlinearity of
Probability Perturbation Method (PPM) responses, For example, for 2 to 4 factors, randomized block design
Probability Perturbation method (PPM) is developed to integrate is used to compare different models; full or fractional factorial
production data without creating any artifacts in geological models1,5- Central composite design or Box-Behnken design is used to
7
. PPM treats production data as a probability cube which is a soft screening models; and Central composite design (CCD) and Box-
constraint for facies modeling and can be manipulated in a similar Behnken design (BBD) is used to construct response surfaces12.
manner as seismic derived probability cube. However, unlike seismic For this study, a design is used to construct response surface to
data, production data are difficult to be converted into 3D probability allow us to estimate interaction effects, and therefore give us an idea
cubes because production data are only available at several well of the (local) shape of the flow response surface. Four design factors
penetrated reservoir cells. Furthermore, the dynamic data are mostly are investigated. Based on above rule of thumb, a CCD or BBD
controlled by nonlinear partial differential equations of fluid flow in should be used. However, considering a large number of simulation
porous media, which needs to be solved globally. Based on Bayes runs may be involved, the more efficient FPBD is selected for this
theorem, Caers1 introduced a simple formulation to convert study.
production data into conditional probability P(A|D): The complexity of recovery factors is expected to require a second
order of polynomial regression model. Due to properties of FPBD,
P(A|D) = (1-rD) i(0)(u)+ rD P(A) (1) the curvature and higher order effects are confounded. The second
Where P(A|D) is the probability of alternative facies providing order model is expressed as
production data D, which is a function of facies prior probability
P(A) and initial realization i(0)(u). i(.)(u) is an indicator function at yˆ = β 0 + ∑ βi xi + ∑∑ β i j xi x j + ε (2)
reservoir location u. The dimensionless perturbation factor rD∈[0,1] i i j
controls how the initial facies are changed. i(0)(u) is adjusted to Where β . is the regression coefficient. It is computed using least
match the production data. To understand how the formulation works,
square method by minimizing the prediction and observation error.
two extremities of rD are discussed here. If rD =1, the P(A|D)=P(A),
which means the alternative facies is not informative and the prior x. is the designed factors. ŷ is predicted value or response. All
facies probability is honored. If rD =0, the P(A|D)= i(0)(u). The initial curvature and higher order effects are confounded into the error
alternative facies is retained in its entirety. By gradually adjusting rD, term ε .
facies realization is adjusted between two end members (initial facies The performance of Regression model is evaluated by R2, which
and prior probabilities) until production data are matched or is a fraction of model interpreted variance vs. total variance. R2
maximum iteration number is exceeded. A gradient based method ranges from 0 to 1. For multivariate regression, the model interpreted
(Brent Algorithm) is used to minimize the error between observation variances may increase by adding variables or factors, although some
and simulation prediction. of this increase in R2 would be simply due to chance variation in that
The PPM algorithm has two loops for facies updating. The intra-loop particular sample. To remove the effects of number of variables
represents small adjustments of facies realization among two end included, the adjusted R2 is used, which is adjusted based on the
members. If satisfying match cannot be achieved by these two end number of variables increased or number of freedom decreased.
members, a more dramatic facies adjustment (another equi-probably
facies model) is added by changing random seed of facies simulation Monte Carlo Simulation and Bayes Theorem
algorithm. Based on previous study, the iteration upper limit for intra- Monte Carlo simulation has been widely used to propagate
loop is 6 and outer-loop is 20. The assumption is that if the uncertainty of input parameters through a performance model (i.e.
production data cannot be approximated within limited runs, the through a polynomial regression responses surface model into
starting geological model is questionable and not likely to be matched. uncertainty of recovery factor)14,17-19. Monte Carlo simulation
The model that has lowest mismatch error is recorded for further requires large number of sampling (104-105) for moderate complex
experimental design analysis. problems. Many other sampling methods such as qausi-Monte Carlo
In summary, the PPM uses a simple formulation to convert dynamic Harmmardy sequence and Latin Hypercubes have been discussed in
data into reservoir properties. To reduce prediction error, two literatures. These methods reduced required sampling size without
iteration loops are used to gradually change the facies realizations. losing Monte Carlo desired properties. However, these alternative
methods need the specific sampling strategy, which is a subjective
IPTC 11119 3

decision and may cause extra errors in uncertainty estimation. In this This workflow is different from traditional HM/AHM methods in
study, because the computation cost for response surface method is several aspects:
trivial, the uniform sampled MCS is used. Another concern of MCS 1) It iteratively integrates dynamic data for all designed runs instead
is that the input variables should be independent, i.e. orthogonal of limited initial models (i.e. three selected P10-50-90 (OOIP)
assumption. The orthogonal assumption could be safely justified models). Several methods ensure such workflow is computationally
because four input geological modeling parameters represent reasonable. First, an efficient FPBD is selected to represent
different geological properties and there is no obvious correlation geological uncertainty. The FPBD also has good resolution for
between each other. The MCS derived flow responses can be used to interaction effects, which can describe a moderate complex dynamic
compute posterior probability distribution using Bayes theorem. response surface. Second, because each initial model is independent
Bayes theorem is widely used due to its abilities of integrating from each other, they are perfect candidates for parallel computation,
different data, including prior probability and inferring event which dramatically decrease the total simulation time. Finally, the
probabilities. One of the major objectives of this study is to use small intra and outer loop maximum iteration number 7 and 20
production data to infer reservoir factors and subsequently mitigate guarantee the efficiency without forcing history match.
the uncertainty range of reservoir factors. For instance, in this paper, 2) This workflow is able to capture the full uncertainty range of
the Bayes theorem is used to update the distribution of reservoir responses. FPBD regularly samples the geological parameters and
modeling parameters by integrating production data17-19. The Bayes generates diversified geological models. The designed initial
theorem is expressed as: geological models ensure the perturbed final models will not
P(F|R)=P(R|F)*P(F)/P(R) (3) converge to local minima. As a result, the perturbed response surface
The P(F|R) is called posterior probability distribution of modeling could capture the full uncertainty range for correct business decision
factors conditioned to production data. P(F) is the prior probability of making.
modeling factors. P(R|F) is the conditional probability of responses 3) The probability perturbation method adjusts facies connectivity
given factors. P(R) is the boundary probability of responses. The through the perturbation number (rD); by gradually add alternative
prior distribution P(F) can be estimated from the sample data or facies realization, all reservoir models are equally probable as the
existing geological knowledge. Alternatively, if there is no or limited initial models. PPM honors the geological conceptual model and
information of prior distribution, a uniformly distributed prior avoids traditional HM artifacts such as locally changing the reservoir
probability distribution could be used. properties
The P(R|F) is computed from MCS results. The factor values are 4) Because the workflow only changes geological parameters, once
drawn randomly from their prior distributions. The response surface the initial “simulation template deck” is set up, a simple EXCEL
is used to compute recovery factor given factor combination. The VBA script helps geologist quickly QC and update reservoir models.
number of times this joint response is observed is then tabulated into
discrete bins for each factor, which is proportional to the product of 3. Application on a shallow marine reservoir North
P(R|F) and P(R)18,19. After a large number of simulations, the P(R|F) Africa
can be computed. Geological Setting
Using Equation 2, the P(F|R) is computed in simple EXCEL Based on regional geological study and log analysis, the sediments
spreadsheets. include a poorly sorted conglomeratic section caused by movement
along the boundary faults. It was deposited during continued marine
Proposed Workflow transgression and has a pronounced fining upward grain-size profile.
The proposed workflow updates geological models by using The sedimentary system is a combination of both continental and
production data and generates dynamic response surface for the marine systems: (1) swamp-alluvial-fan system with three litho-
updated models. These geological models and response surfaces of facies– swamp facies, paleosoil facies, alluvial-fan facies; (2)
updated models are compared with those of initial models. The transgressive shallow marine facies—with two litho-facies:
proposed workflow includes 4 relating steps. beach/barrier shoreline facies and lagoonal/tidal flat facies.
1. Designed Geological Modeling: Based on geological Petrophysical analysis indicates best reservoirs are beach/barrier
knowledge, uncertainty factors are selected. A Folded facies followed by alluvial fan facies and lagoon tidal facies (Fig. 1)
Plackett-Burman design (FPBD) generates different (Table 1).
combination of modeling factors. Using these factors
combination, a Multiple-Point Statistic (MPS) method Flow Simulation Model Construction
integrate geological and geophysics data into generating Available data include 17 well log data, 2D structure maps and
multiple initial reservoir facies models16. The training reservoir isochors, which are used to build a faulted structure maps.
images are based on neighboring reservoir models. A general property workflow is followed to populate reservoir
2. Flow responses of the initial geological models are properties and manage the geological uncertainties (Fig. 2). To
computed by using a black oil simulator. The Monte Carlo reproduce the geological conceptual model, which is important
Bayes Method (MCBM) uses a polynomial response reservoir with spares or poor quality data, a prior probability cube of
surface to estimate the recovery uncertainties and infer the reservoir quality are created by using well log facies proportion and
conditional probability for factors providing production horizontal geological maps. Multiple Point Statistic Method (MPS) is
data. used to populate reservoir facies models. A simplified binary facies
3. Each initial reservoir model is perturbed to matching the (reservoir and non-reservoir) are modeled. However, the extension to
production data by using the Probability Perturbation multiple facies is straightforward. Moreover, uncertainty
Method (PPM). This step decreases the uncertainty of management plan determines the four major uncertain parameters are
sedimentary connectivity in a way consistent with oil water contact, fault sealing, net to gross ratio, and tar mat. FPBD
geological models. generates multiple facies realizations. Then the petrophysics
4. Similar to step 2, the new response surfaces oil recovery properties of each zone are modeled separately. Empirical
are recomputed based on the perturbed models. Both relationships derived from core data are used for permeability and
responses and conditional probability distribution are saturation transformation. Production bottom-hole pressure (BHP)
compared. data from four wells are available in each fault block. Probability
perturbation method is used to integrate them into geological model.
4 IPTC 11119

The final geological model is converted into a corner point grid for simulation indicates the uncertainty range increase from 15 percent to
numerical simulation. The gridded model has 983,280 cells, of which 18 percent after perturbation.
472,715 are active. The 3D grid preserves the geological details while
still keeps the model small enough to be manageable. Experimental Design and Response Surface Model
In this study, four designed factors are Net-Gross-Value (NTG), Fault
Flow Simulation Sealing (Fault), Oil Water Contact (OWC), and Tar Mat (Tar). The
Flow was simulated with a commercial simulator. Four wells recovery factor is the only response. Three dynamic responses are
naturally deplete the reservoir with reservoir pressure above the initial recovery factor (RFini); updated recovery factor (RFupd) and
bubble point. Four wells penetrate reservoirs with 50 feet above oil predicted recovery factor at 1PV depletion (RpD) (Table 2). A full
water contact. Based on previous simulation report, the PVT, relative three level four factor design will have (81 runs). FPB has reduced
permeability data are assigned and the Kv/Kh is set to 0.2. The PPM total runs from 81 in full factorial to 17 runs. Tar Mat acts as a flow
and an Excel script drive the simulator to automatic approach the barrier right above the oil water contact. Both Tar mat and Fault
production pressure profile. Simulation results are automatically sealing are modeled as transmissibility multiplier. The all factors are
retrieved for response analysis.
Flow simulation fence plots of the fifth run illustrate (Fig. 3) the scaled linearly with xi = ( f i − f i min ) /( f i max − f i min ) where xi is
change of water saturation during the depletion process. The warm the scaled factor, fi is the unscaled factor. The flow response is the
color represents high oil saturation and cold color indicates low oil
saturation. Fig. 3 (a) is the initial saturation distribution; (b) is recovery factor. Recoveries from both initial model and updated
saturation distribution at the water breakthrough time. (c) is for model are included. Based on the updated model, the forecasting
saturation after the 1 year production; (d) is after one pore volume recovery factor after 1 pore volume injection is also recorded. Factor
depletion. The depletion process shows the water finger through the combinations are specified by a 17 Folded Plackett-Burman design
high permeability formation 2 (c). The vertical flow is impeded by (Table 3).
the sedimentary layering which is modeled by vertically proportional Analysis of Variance
gridding and associated Kv/Kh ratio. Response surface model are constructed by using polynomial
regression method11,17-19. As mentioned in equation 3, FPBD cannot
Integrating Production History differentiate the quadratic term (or curvature). A polynomial
The production data integration of well 4 is plotted in Fig. 4. Green regression with major factors and interactions are modeled
curves represent initial models; and red curves represent perturbed (APPENDIX A). The regression factors are illustrated in Table.
facies models. This result may not be true for other reservoirs. Ideally, The integration of production data changed the flow response Table
facie perturbation may visually separate reservoir models into two 4 and Fig. 6. Several results are observed. The fact that adjusted
groups: 1) easy or possible to converged models and 2) difficult to regression R2 adjust (0.96) means 96 percent of total variance are
converged models. Since all these models are equally probable, they interpreted by response surface model. Three out of four linear
are all included in uncertainty analysis in this paper. It is a subjective factors are significant! This indicates the recovery factor are highly
choice to differentiate these two models from each other. However, it sensitive to the change of three modeling factors in a straightforward
will be reasonable to choose the models closer to the production manner (liner relationship). However, after the integration of
history for final history match. Those models could be assumed to dynamic data, the response surface model can only interprets 72
better represent real geology of subsurface reservoirs. percent of total variance with R2 adjust (0.72). The production data
Another limitation in this study is that visually quantification of integration indicates the reservoir dynamic responses are more
reservoir connectivity or continuity is intuitive, further study on complex than our initial guess. The simple response surfaces model is
quantitatively estimate the reservoir connectivity will provide more not adequate to interpret the simulation responses. A higher order
insight on the reservoir static-dynamic relationships. nonlinear response surface is needed to represent the reservoir
The determination of maximum outer loop iteration number is a behavior.
subjective decision, which is a trade off between total computational Sensitivity Analysis
cost and prediction accuracy. Too much iterations waste The regression coefficient can be used to estimate the sensitivity of
computational time without improving the history match. Less the factors. The interpretation of the coefficient is the increment in
iteration runs may miss the correct geological models before reaching response when the factor increase one unit by assuming the rest
the history match. The maximum iteration number of inner loop is factors constant. Usually, a tornado chart is used to visually check the
seven which is based on previous study6. The outer loop iteration uncertainties of factors (Fig. 6). In the initial model, three factors are
number is 20 for this study. So the maximum total iterations for one significant; however, only OWC is significant in the updated model.
model are 140 iterations. This iteration number is acceptable because The interaction between OWC and NTG is significant in initial model;
in this study reservoir is reasonably small. The average simulation but not significant in updated model. Another intuitive way to
time for one run is 5 to 10 minutes. visualize the interaction terms is using 2D response surface plot (Fig
7). The response surface of NTG and OWC interaction illustrate that:
a) The production updated recovery factor is not sensitive to
Change of Facies Realizations and Recovery Uncertainties
the NTG and OWC interaction. However, the initial model
The perturbation changes the reservoir facies realizations (Fig 5). The
has strong NTG and OWC interaction.
reservoir facies is in yellow color; the blue color represents non-
b) In the updated model, recovery factor is only sensitive to
reservoir facies. To match the production data, though the modeling
OWC. Raising the OWC increases the final recovery. This
parameters are the same, the perturbation dramatically changed the
could be explained by rising the OWC will have more high
facies realizations. The prior facies probability cube tends to
permeability formations connect to aquifer that may
distribute reservoir facies around north paleo-high (Fig. 5 (g)). The
provide better pressure support and higher final recovery
production data makes some of the updated facies realizations less
for the reservoir.
continuous, particularly in the southwest block where four producers
Similar response surfaces can be plotted for other factors.
locate. The change of reservoir continuity might increase the
Because other interaction terms are not significant in both updated
reservoir uncertainty ranges, which may finally increase the reservoir
and initial models, it will be not interesting to exam other interaction
recovery factor range. The predicted recovery for all designed
terms. In summary, unlike the initial model, updated model
indicates current response surface model is not adequate to explain
IPTC 11119 5

the variation of dynamic responses. A more complex model and more ƒ A case study on a North Africa reservoir indicates the
runs are required to fully explore the recovery responses. production data updated reservoir model has different
heavy hitters for recovery factor from initial models.
Application of Monte Carlo Simulation and Bayes method ƒ MCSBM and posterior study show the modeling
Uncertainty of Recovery Factors parameters conditioning to production data have smaller
The response uncertainty includes measure errors, upscaling errors, uncertainty ranges for both initial and updated models,
oversimplified reservoir models, and biased sampling. Failure in which means the production updating reduces parameters
capture right recovery uncertainty may cause mistakes in business uncertainty. The updated recovery factors have wider
decision. However, the high nonlinearality in flow responses makes uncertainty ranges (18 percent) compared with initial
the uncertainty estimation difficult. Monte Carlos Simulation helps to recovery factors (15 percent). This is understandable
transfer the factor errors into responses uncertainty estimation. The because the updated geological models honoring
uncertain factors are drawn from specific distributions assigned by production data may have wider heterogeneity ranges,
geologist. Besides the MCS computed uncertainties, a measurement which capture the full uncertainty ranges for recovery
error of recovery factors are added. The MCS computed recovery factor.
uncertainty for both models indicates that dynamic updated model ƒ The updated recovery uncertainty as well as reduced
has larger uncertainty ranges than the initial model. A student t test modeling parameters uncertainties help to make correct
assuming unequal variances indicates the two recovery distribution business decision. Furthermore, the updated response
are significantly different (with t value= 65 with confidence interval surfaces shows confounded higher order nonlinear effects.
>95%). Practically the uncertainty range for initial model is also A further study based on more detailed design is needed.
different (Fig. 8). This means the production data update is The above method honors available data without over-
informative and should be considered in decision analysis. tuning geological parameters for history matching. It also
Posterior analysis provides more accurate prediction for risk analysis with
Besides analyzing the uncertainty of recovery factor and, by acceptable computational cost.
combining Monte Carlo Simulation and Bayes theorem (MCSBM),
the MCSBM is used to infer the posterior probability of reservoir
factors condition to responses. For example, in this paper, production Acknowledgement
data is used to derive the posterior probability distribution of The Author would like to extend thanks to Dr. X. -H Wen for
geological factors such as NTG and Fault. The Bayes formula shows providing useful guidance for this work. PPM code is provided by Dr.
that the probability density function (PDF) of factors posterior Hoffman Todd. Dr. N. Liu provides insights and constructive
probability distribution is a function of prior probability P(F) and suggestions. Dr. B. Kurnianwan helps proof read this paper.
condition probability of responses give factors P(R|F).
First, the initial prior probability is assumed uniformly distributed. If Nomenclature
the curved changes to be sharply distributed (less variation), the Roman Symbols
MCSBM is informative. Updated NTG has a skewed distribution fi,fimin,fimaxunscaledfactor
based on RSM from updated model Fig 9. (a). The initial NTG and pprobability density
Fault distributions are more sharply distributed Fig 9. (b). This is
Pprobability
because recovery is more sensitive to these factors in the initial
response surface than the updated response surface. By adding a   Dimensionless perturbation factor
response error, the factor posterior probability distributions have the Nψψ ψrecovery factor at breakthrough
similar trend as the previous models, but the distribution is less x scaled factor
pronouncedly distributed. In other words, the MCSBM becomes less
informative Fig 9 (c) (d). This is not the defect of this method. The Greek Symbols
predictability reflects the data quality. β regressioncoefficients
By adding more confined prior probability (triangular distributions)
uspatial coordinate
to all factors, and using the same response surfaces, the factors
posterior is strongly affected by prior probability. To generation of ε error
triangular prior distribution, a triangular probability density function
is converted to cumulative probability function (CDF) by integration.
Then the CDF converts a randomly drew probability into References
corresponding factor values. Similar to PDFS using uniform priori, 1. Caers, J.and Hoffman, T. 2006, The Probability
MCSBM has more significant effects on the initial models Fig. 10 (b) Perturbation Method: A New Look at Bayesian Inverse
(d) than updated model fig. 10 (a) (c). The deviation from prior Modeling, Mathematical Geology, Vol. 38, No. 1
probability distribution is more pronounced for models without 2. Emanuel, A.S. and Milliken, W.J.: “History Matching
response error added fig. 10 (a) (b) than those with response error (c) Finite Difference Models With 3D streamlines,” paper SPE
(d). 49000 prepared for presentation at the 1998 SPE Annual
In summary, the MCSBM provides uncertainty estimation for Technical Conference and Exhibition, New Orleans, 27–30
responses as well as modeling factors. September.
3. Friedmann, F., Chawathe, A., Larue, D. K.: “Assessing
Summary Uncertainty in Channelized Reservoirs Using Experimental
ƒ A designed simulation method (Folded Plackett-Burman Designs,” SPE Reservoir Evaluation and Engineering,
design) and probability perturbation method (PPM) are August, 2003.
proposed to mitigate reservoir connectivity uncertainty 4. Go´mez, S., Gosselin O., and Barker, J.W.: “Gradient-
with reasonable computational cost. Based History-Matching With a Global Optimization
ƒ A Monte Carlo Simulation and Bayes Method (MCSBM) Method,” paper SPE 56756presented at the 1999 SPE
exam the response surfaces of recovery factors for both Annual Technical Conference and Exhibition, Houston, 3–
initial and updated models. 6 October.
6 IPTC 11119

5. Hoffman, B. T., and Caers, J., 2003, Geostatistical history History Matching Production and Seismic Data”, SPE
matching using the regional probability perturbation Journal, June 2002
method, in Proceedings of the SPE Annual Conference and
Technical Exhibition, Soc. Petrol. Eng., Paper no. 84409, Appendix
Denver, CO, October 5–8, 16 p. Response Surfaces for initial and updated models
6. Hoffman, B. T., “Geologically Consistent history matching RFupdated = 0.0627 + NTG * 0.0029 + Fault * -0.00332 + OWC * -
while perturbing facies “Ph.D. Thesis at Stanford 0.0042 + Tar * -0.00023 + NTG * Fault * 0.002278 + NTG * OWC *
University, 2005. 0.00065 + NTG * Tar * -0.000829 + Fault * OWC * 0.002153 +
7. King, P.R., Harding, T.J., and Radcliffe, N.J.: Fault * Tar * -0.00208 + OWC * Tar * -0.0017
“Optimization of Production Strategies using Stochastic
Search Methods,” Proc., 5th European Conference on the RFinit = 0.0595+ NTG * 0.0021 + Fault * -0.0027 + OWC * -
Mathematics of Oil Recovery, Leoben, Austria,3–6 0.00099 + Tar * 0.00072 + NTG * Fault * -0.0006 + NTG * OWC * -
September 1996. 0.0021 + NTG * Tar * -0.000028 + Fault * OWC * 0.00152 + Fault *
8. Journel, A., 2002, combining knowledge from diverse Tar * -0.000153 + OWC * Tar * 0.000347
sources: An alternative to traditional data independence
hypotheses: Mathematical Geology, v. 34, no. 5, p. 573–
596.
9. Kabir, C. S., Chawathe, A., Jenkins, S. D., Olayomi, A. J.,
Aigbe, C., Faparusi, D. B.: ”Developing New Fields Using
Probabilistic Reservoir Forecasting,” paper SPE 77564,
presented at the SPE Annual Technical Conference and
Exhibition, San Antonio, Texas, USA, 29 September – 2
October, 2002.
10. Milliken, W.J., Emanuel, A.S., and Chakravarty, A.:
“Applications of 3D Streamline Simulation to Assist
History Matching,” paper SPE 63155 presented at the 2000
SPE Annual Technical Conference and Exhibition, Dallas,
1–4 October.
11. Narayanan, K., White, C. D., Lake, L. W., Willis, B. J.:
“Response Surface Methods for Upscaling Heterogeneous
Geologic Models,”
12. NIST/SEMATECH e-Handbook of Statistical Methods,
http://www.itl.nist.gov/div898/handbook/, 2004.
13. Liu, N. and D. S. Oliver, Evaluation of Monte Carlo
methods for assessing uncertainty, SPE 84936, SPE Journal,
8(2), 188–195, 2003b.
14. Liu, N. and D. S. Oliver, Automatic history matching of
geologic facies, SPE 84594, Proceedings of the 2003 SPE
Annual Technical Conference and Exhibition, pp. 1–15,
2003a.
15. Ouenes, A. et al.: “A New Algorithm for Automatic
History Matching: Application of Simulated Annealing
Method to Reservoir Inverse Modeling,” paper SPE 26297
available from SPE, Richardson, Texas (1983).
16. Strebelle, S., and A. Journel, 2001, Reservoir modeling
using multiplepoint statistics: Presented at the Society of
Petroleum Engineers Annual Technical Conference and
Exhibition, SPE Paper 71324, 10 p.
17. Wang, F., White, C. D.: “Designed Simulation for a
Detailed 3D Turbidite Reservoir Model,” paper SPE 75515,
presented at the SPE Gas Technology Symposium held in
Calgary, Alberta, Canada, 30 April-2 May, 2002.
18. White, C.D. et al.: “Identifying Controls on Reservoir
Behavior Using Designed Simulations,” paper SPE 62971
presented at the 2000 SPE Annual Technical Conference
and Exhibition, Dallas, 1–4 October.
19. White, C. D., Willis, B. J., Narayanan, K., Dutton, S. P.:
“Identifying and Estimating Significant Geologic
Parameters with Experimental Design,” SPE Journal
(September, 2001).
20. Willis, B.J. and White, C.D.: “Use of Quantitative Outcrop
Description for Reservoir Modeling,” J. Sedimentary
Research (2000) 70, No. 4, 788.
21. Mantica, S. Cominelli, A. and Mantica, G.: “Combining
Global and Local Optimization Techniques for Automatic
IPTC 11119 7

Table1. Sedimentary Facies Description by Zones


Formation Litho-Facies Desciption Porosity (%) Sw (%) Permeability (md)
Avg. Stdev. Avg. Stdev. Avg. Stdev.
Swamp alluvial fan system with swap
F1 facies, paleosoil facies, and alluvial 14.0 3.0 48.0 5.0 124.0 41.0
facies
Transgressive shallow marine system
F2 24.0 4.0 33.0 5.0 159.0 6.0
including beach/barrier facies
Shallow marine system including
F3 16.0 5.0 16.0 3.3 35.8 2.8
lagoonal/tidal flat facies

Note: the F1 is the youngest zone, follows by F2 and F3. F2 has highest permeability and may cause uneven oil depletion.

Table2. Experimental Designed Factors


F ac tor R anges and S c aling
F ac tor M appings
Index F ac tor N am e S y m bols U nits -1 0 1
1 N et to G ros s R atio N TG % 0.2 0.5 0.8
2 F ault S ealing * F ault 0 0.5 1
3 O il W ater C ontac t OW C ft -10000 -10250 -10500
4 Tar m at * Tar 0 0.5 1
* F ault s ealing and Tar m at are defined as different trans m is s ibility m ultiplier
0 m eans c om pletely s eal, 1 m eans no barrier, 0.5 m eans partially s eal

Table3. Folded Placket-Burman Design


Factors Response
NTG Fault OW C Tar RF_ini RF_upd RF_1pv
1 -1 -1 1 0.070 0.074 0.176
1 1 -1 -1 0.060 0.069 0.149
Upper Fold

1 1 1 1 0.055 0.051 0.153


-1 1 1 1 0.059 0.052 0.155
1 -1 1 1 0.062 0.063 0.288
-1 1 -1 1 0.053 0.055 0.105
-1 -1 1 -1 0.058 0.057 0.256
Folded PB

-1 -1 -1 -1 0.059 0.067 0.180


CPR 0 0 0 0 0.060 0.074 0.188
1 1 1 1 0.060 0.061 0.206
1 1 -1 1 0.060 0.065 0.144
Lower Fold

1 -1 1 -1 0.059 0.061 0.279


-1 1 -1 -1 0.052 0.054 0.123
1 -1 -1 -1 0.069 0.066 0.164
-1 -1 -1 -1 0.060 0.071 0.191
-1 -1 1 1 0.060 0.055 0.231
-1 1 1 -1 0.057 0.053 0.167
Note: RF_ini=Recovery Factor of initial model; RF_upd= RF of updated model;
RF_1pv=RF after 1 pore volumn injection for updated model; CPR=Center Point
Run

Table 4. Sensitivity Analysis of Response Surface


Recovery Factor
Simulation Model Initial Updated
R2 Adjusted 0.96 0.72
Number of Regressors 10 10
Parameter Estimated
Constant 0.0595 0.0627
Interaction Terms Linear Terms

NTG 0.0021 0.0029


Fault -0.0027 -0.0033
OWC -0.0010 -0.0042
Tar 0.0007 -0.0002
NTG*Fault -0.0006 0.0023
NTG*OWC -0.0021 0.0007
NTG*Tar 0.0000 -0.0008
Fault*OWC 0.0015 0.0022
Fault*Tar -0.0002 -0.0021
OWC*Tar 0.0003 -0.0017
Note: Significant coefficients are shown
in Bold type, and insignificant
coefficients are italicized.
8 IPTC 11119

Sedimentary System Co nceptual M o del

MARINE SEDIMENT
ACTIVE
FAULT PALAEOZOIC
MARGIN
TRIASSIC
PASSIVE MARGIN
BASEMENT

(From Sghair and A lami)

Figure 1.The sedimentary system includes both continental and marine systems.
The petrophysical study shows the best reservoirs are shore face /barrier facies
followed by alluvial fan facies and lagoon tidal facies.

Probability Cube+MPS
Vertical (from well) and Area (geological prior)

Experimental Designed Modeling


to determine the range of reservoir values

Porosity Permeability Transformation


Using empirical correlation

Probability Perturbation Method to


Integrate Production data

Figure2. Reservoir Property Modeling Workflow

(a) (b)

Vertical Exaggeration is 2

(c) (d)

Figure 3. Reservoir Depletion Process Fence Plot

Warm color represents high oil saturation; cold color represents low oil saturation.
(a) initial saturation (b) saturation distribution at water breakthrough (c) Saturation
distribution at 1 year depletion (d) saturation depletion at 1 pore volume depletion.
IPTC 11119 9

5000

4500
Initial Presure
History Data
4000

3500

WBHP pressure (psi)


3000

2500

2000

1500 Updated Presure

1000

500

0
0 200 400 600 800 1000 1200 1400 1600
time (days)

Figure 4. PPM integration of Production history from well 4.

Red curves are production profile from initial geological models. Green curves are PPM updated
geological models. Production history is plotted in a dotted solid line The Updated model
separate into two groups; one converges or possible to converge to history data. These models
represent the correct connectivity and are possible for further History match. Another group drifts
away from production history after perturbation, which may not correctly represent the geological
connectivity and is less possible be matched later.

Initial Facies Distribution Updated Facies Distribution

80% Run5 – high

(a) (b)

20% Run6 – low

(c) (d)

50% Run9-Base

(e) (f)

Prior probability

(g)

Figure 5. Comparison of facies realizations before and after production update and facies prior probability

Yellow represents reservoir, blue represents non-reservoir. Left column is for initial facies distribution with different NTG;
right column represents updated facies distribution with different NTG. Bottom row is the initial facies prior probability cube.
(a) (b) are high initial/updated facies with 80 % NTG. (c) (d) are low initial/updated facies with 80 % NTG. (e) (f) are middle
initial/updated facies with 80 % NTG. The production update changes the original reservoir facies distribution and
connectivity, which may subsequently change the final recovery.
10 IPTC 11119

Fig. 6 Plato-Chart for sensitivity analysis


The sensitivity analysis indicates updated models even using the same modeling parameters, the major
hitter is OWC (a). Rest of factors is less significant. However the major hitters for initial response surfaces
are Fault, NTG, and interaction of NTG and OWC (b). The significant factors are marked with a *.

0.068 0.068

0.066 0.066

0.064 0.064

0.062 0.062
0.06 0.06
0.058 0.058
0.056 0.5 0.056 0.47
0.44 0.054
0.054 0.38
0.38
0.052 0.052
0.32 NTG NTG
0.05 0.05 0.29
0.26
-10500
-10500

-10450
-10450

-10400
0.2
-10400

0.2

-10350
-10350

-10300
-10300

-10250
-10250

OWC OWC

Fig. 7 Response Surfaces

The interaction term in the updated model response surface model is not significant (a); however,
it is significant in the origin model.

0.7
Original Recovery
Probability Density Function

0.6 Uncertainty
Updated Recovery Ucertainty
0.5

0.4

0.3

0.2

0.1

0
0.05 0.055 0.06 0.065 0.07 0.075
Recovery Factor (%)

Fig. 8 Comparison of recovery factors uncertainty

Updated uncertainty of recovery factor is larger than initial recovery uncertainties. This may be
because of the change of connectivity increase the uncertainty range of connectivity.
IPTC 11119 11

0.16
0.2

0.18 0.14

0.16 (a) 0.12 NTG (b)

Probability Density
0.14

Probability Density
NTG
Prior
0.1
0.12 Fault
0.08 OW C
0.1
Fault
0.08 OW C
0.06
0.06
0.04
0.04
Tar Tar
0.02 0.02

0 0
-1 -0.5 0 0.5 -1 -0.5 0 0.5
Factor Value Factor Value
Updated factor pdf conditional to production data Initial factor pdf conditional to produc tion data
0.2
0.12
0.18

0.16
(c) 0.1

0.14 NTG
Probability Density

Probability Density
NTG 0.08 Prior
0.12 Fault

0.1 OW C
0.06
Fault
OW C
0.08

0.06 0.04

0.04

0.02
Tar 0.02
Tar (d)
0 0
-1 -0.5 0 0.5 -1 -0.5 0 0.5
Factor Value Factor Value

Updated fac tor pdf including response error Initial factor pdf inc luding response error

Figure 9.The probability density function of factors posterior probability distribution.

If the curved is sharply distributed, the MCSBM is informative. (a) Updated NTG has a skewed distribution. b)
initial NTG and Fault distributions are more sharply distributed. This is because recovery is more sensitive to these
factors in the initial response surface. (c) (d) by adding the response error, the factor posterior probability
distributions have the similar trend as the previous models, but the distributions less pronouncedly distributed.

0.3
0.3

0.25
(a) Tar 0.25 (b)
Probability Density

Probability Density

0.2 OWC
0.2
NTG
Fault
0.15 Fault 0.15

prior 0.1
Tar
0.1
NTG
0.05
0.05
OWC

0
0 -1 -0.5 0 0.5
-1 -0.5 0 0.5 Factor Value
Factor Value

Updatedl factor pdf without response error Initial factor pdf without response error

0.25
0.3
Fault
Fault
0.2 OWC 0.25

Tar
(c) (d)
Probability Density

Probability Density

0.2
0.15 NTG

0.15
0.1 Tar OWC

NTG 0.1

0.05
0.05

0
0
-1 -0.5 0 0.5 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Factor Value Factor Value

Updatedl factor pdf including response error Initial factor pdf including response error

Figure 10 the probability density function of factors posterior probability distribution.

By adding more confined prior probability (triangular distributions), the factors posterior is strongly affected by
prior probability. MCSBM skewed the sensitive factor poster from symmetric triangular distribution (the prior
distributions). Updated PDFs are skewed but less uncertain (less sharply distributed) (a). Initial PDF of Fault and
NTG are skewed with sharp distribution;

You might also like