Neural Network EOR

Geosystem Engineering, 14(4), 157-164 (December 2011)

Research Paper

157

Selection and Evaluation of Enhanced Oil Recovery Method

Using Artificial Neural Network

Jong-Yong Lee1, Hyo-Jin Shin

2 and Jong-Se Lim

2,*

1Korea National Oil Corporation, Dongan-Gu, Anyang, Gyeonggi-Do, Korea2Korea Maritime University, Dongsam-Dong, Yeongdo-Gu, Busan, Korea

ABSTRACT: Enhanced Oil Recovery (EOR) has gained greatattention as a result of higher oil prices and increasing oil demands. Extensive researches have been conducted to developvarious EOR methods, evaluate their applicability and optimizeoperation conditions. One of the principal areas is to developan effective tool for selection of a suitable EOR method accord-ing to oil field characteristics. The main objective of the studiesis to screen various EOR methods based on field characteristicsand evaluate their technical/economic applicability in an effi-cient way instead of predicting the field performances of all pos-sible competing strategies and comparing their economics. In thispaper, we present an Artificial Neural Network (ANN) ap-proach to enable the petroleum engineer to select an appro-priate EOR method with the given reservoir properties. The ANN developed in this study is a four-layered feed-forward Back Propagation (BP) network consisting of one input and output layer with two hidden layers. The input layer is composedof the key reservoir parameters (reservoir depth, temperature,porosity, permeability, initial oil saturation, oil gravity, and in-situ oil viscosity) while the output layer is composed of thefive EOR methods to be evaluated (steam, CO2 miscible, hydro-carbon miscible, in-situ combustion, polymer flooding). The number of hidden layers and neurons are optimized during thetraining by repeated trial and error. After trained successfully,the ANN is tested and applied to other fields which are not usedfor the training. The noise test is also conducted to evaluate ap-plicability of the model against the error included in the input.A series of the test results show that the ANN developed in thisstudy can be used to select the most appropriate EOR processaccording to reservoir rock and fluid characteristics in a time and cost effective way.

Key words: Enhanced Oil Recovery, Artificial Neural Network,field characteristics, selection, data preparation

Recieved May 15; Revised October 5, 2011; Accepted October 6, 2011

* Corresponding Author: Jong-Se Lim E-mail: [email protected] Address: Korea Maritime University, Dongsam-Dong, Yeongdo- Gu, Busan, 606-791, Korea

INTRODUCTION Higher oil prices and concerns about future oil supply are

leading to increased interest in Enhanced Oil Recovery (EOR) around the world. Because EOR projects are generally more expensive and involve higher front end costs than conventional secondary projects, effective planning takes on added importance (Hite et al., 2004). A large number of studies have been conducted to help the petroleum engineer select efficient EOR methods with limited field information. The main objective of the studies is to select the suitable EOR method in an effective way without predicting the reservoir performance of all possible competing strategies and comparing their economics.

Most of early studies in the EOR selection were to establish the technical screening criteria of each EOR method (Taber and Martin, 1983; Goodlett et al., 1986; Taber et al., 1996a; Taber et al., 1996b). Based on laboratory experiments and field experi-ences, the applicable ranges of the reservoir rock and fluid properties were presented in these studies. The effort has been added in several studies to update the applicable ranges with the current technical and economic conditions (Aladasani and Bai, 2010; Dickson and Wylie, 2010). The problem of selection and implementation of proper EOR techniques was also addressed in some papers as a guide for petroleum engineers (Zerafat et al., 2011).

The improvement of computer technology introduced the artificial intelligence technique into EOR selection (Guerillot, 1988; Elemo and Elmtalab, 1993; Surguchev and Li, 2000; Shokir et al., 2002; Lee and Lim, 2008). Because the values of these models strongly depend on the accuracy of the input data, it should be continuously updated with up-to-date operation data.

In this paper, we developed the Artificial Neural Network (ANN) incorporating the recent database published in the industry. The main goal of the study is to develop the ANN model that can estimate the best EOR method according to the given reservoir rock and fluid properties in a time and cost effec-tive way and evaluate applicability of the model.

ARTIFICIAL NEURAL NETWORK

ANN is an information-processing system that has certain

158 Jong-Yong Lee et al.

Fig. 1. Reservoir depth vs. area of producing EOR projects.

Fig. 2. Reservoir depth vs. temperature of producing EOR projects.

performance characteristics in common with biological neural networks. A typical neural network is a multilayered system consisting of single input layer, single or double hidden layer, and single output layer. Each layer is composed of basic processing elements called neurons. Each neuron is connected to the neurons of the adjacent layer with the connection weights between 0 and 1. The signals between the neurons are multiplied by the associated connection weights and added up together as Eq. (1), and then used as the net input of the neuron.

1n kk kNET I W=∑= (1)

Where NET is the net input of the neuron, I is the input variable, W is the connection weight, k is the index, and n is the number of input variables. Each neuron applies an activation function to its net input to determine its output signal and the signal is transmitted to the next neuron. The sigmoid function in Eq. (2) is a activation function commonly used.

1( )1 NET

f NETe

=+

(2)

The connection weights between the neurons are adjusted

during the training. There are two ways of the training; supervised and unsupervised. For most typical neural network, the connection weights are adjusted by the given input and corresponding output. This process is called as supervised training. One of the widely used supervised networks is the feed-forward Back Propagation (BP) network which adjusts the connection weights during the back propagation process.

In this study, the BP network with the training algorithm of Scaled Conjugate Gradient (SCG) which is a new variation of the conjugate gradient method is used. SCG allows the avoidance of the line search per training iteration of Levenberg-Marquardt approach in order to scale the step size.

DATA SOURCE AND PREPARATION

The data used for training and testing the networks are ex-tracted from the special reports, Worldwide EOR Survey pub-lished by Oil and Gas Journal (Moritis, 2006; 2008; 2010). The reports include the field name, reservoir rock and fluid properties, project maturity, and project evaluation of the field where the EOR was being applied. In this study, the data of the fields where the application of EOR was evaluated to be successful (success and promising) are only used for the training to extract the char-acteristics of the successful EOR projects. The EOR methods which were applied to less than 10 fields are also excluded in the training for training efficiency.

Neurons of the input layer are designed to be the main reser-voir properties. 2-dimensional scatter plots are drawn to screen the input reservoir properties that affect the choice of EOR method. Fig. 1 to Fig. 6 are the scatter plots to analyze the effects

of reservoir rock and fluid properties on the applied EOR methods. It is found that the applied EOR methods tend to be categorized by most of the reservoir rock and fluid properties analyzed except the reservoir area (Fig. 1). Accordingly, the seven reservoir properties which are reservoir depth, temper-ature, porosity, log permeability, initial oil saturation, oil grav-ity, log oil viscosity are selected as the input variables of the ANN model.

A new variable is generated by grouping the input reservoir pa-rameters to sample the data to be used for the training and two-thirds of the total data are selected based on this group variable as summarized in Table 1. For training efficiency, the sampling ratio increases to three-fourths if the number of sampling data is less than ten. The remaining data which are not included in the

2

2

Selection and Evaluation of Enhanced Oil Recovery Method Using Artificial Neural Network 159

Fig. 3. Oil saturation vs. porosity of producing EOR projects.

Fig. 4. Reservoir permeability vs. porosity of producing EOR projects.

Fig. 5. Reservoir depth vs. oil gravity of producing EOR projects.

Fig. 6. Oil viscosity vs. oil gravity of producing EOR projects.

training are used for testing the developed ANN model. Table 2 shows the number of data for the training and the applicability test.

The ranges of the input reservoir parameters are summarized in Table 3. Each input variable is normalized between 0 and 1 before the training for numerical stability as defined in Eq. (3). The normalized input variables are then entered into the input neurons to train the network.

min

max min

actualnorm

X XXX X

−=

−(3)

Where Xnorm is the normalized input variable, Xactual is the

original value of the variable, Xmin is the minimum value of the

variable and Xmax is the maximum value of the variable. The ranges of the input reservoir parameters are summarized in Table 3.

The neurons of the output layer are composed of the EOR methods to be selected. The five EOR methods (steam, carbon dioxide miscible, hydrocarbon miscible, in-situ combustion, polymer flooding) which are being applied in more than ten fields consist the output layer. The target value of the output neu-rons are designed to be +1 in the neurons indicating the success-fully applied EOR methods and -1 in other neurons indicating other EOR methods.

DEVELOPMENT OF ANN MODEL Design and training of the ANN is conducted in Matlab

2

2

2

2


Table 1. Data sampling method by group variable

STEP 1Divide the input variables into two groups by their effects on the selection ∙Group 1 : porosity, log permeability, oil saturation, log oil viscosity∙Group 2 : depth, temperature, oil gravity

STEP 2Multiply the variables for each group∙V1 = porosity × log permeability × oil saturation × log oil viscosity∙V2 = depth × temperature × oil gravity

STEP 3 Generate the group variable by dividing V1 by V2STEP 4 Rank each data by group variable and group each three dataSTEP 5 Sample two data for each group

Table 3. Ranges of the input reservoir parameters

Reservoir parameters Minimum Average MaximumReservoir depth, ft 200.0 4,079.0 13,750.0Reservoir temperature, ℉ 45.0 126.9 290.0Porosity, % 3.0 23.3 65.0Permeability, md 0.1 1,283.6 11,500.0Initial oil saturation, % of OOIP 26.5 62.8 98.0Oil gravity, °API 8.0 24.9 57.0In-situ oil viscosity, cp 0.1 26,594.4 200,000.0

Table 2. The number of data used for the training and the applicabilitytest

EOR type Total Training TestingSteam 103 70 33Carbon dioxide miscible 65 45 20Hydrocarbon miscible 32 22 10In-situ combustion 15 11 4Polymer flooding 15 11 4Total 230 159 71

Fig. 7. Sensitivity on the number of neurons of the hidden layers.7.12.0 (Beale et al., 2010). The object function in the training is the mean square error as defined in Eq. (4) and the convergence tolerance is initially designed to be 0.001.

( )2

1

1 ( )pNi i i

p

Error y f xN =∑= − (4)

Where Np, yi, and f(xi) indicate the number of data, the meas-ured output, and estimated output by the model respectively. As an activation function, the tangent sigmoid function is used for the first hidden layer and the logistic sigmoid function is used for the second hidden layer. For the output layer, the linear func-tion is used (Lee and Lim, 2008).

The structure of the ANN model, that is the number of neurons of the hidden layers, is optimized during the training by repeated trial and error. The number of iterations per each training case is compared in Fig. 7. Maximum number of iteration is set to 10,000. As shown, the network can reach the given convergence tolerance when the number of neurons in both of the hidden lay-

ers are greater than 8. Among the networks evaluated, the one consisting of 10 neurons in the first hidden layer and 8 neurons in the second hidden layer is selected as the work basis. The fi-nally selected structure of the ANN model is shown in Fig. 8.

Applicability of the ANN model is tested against the new data which are not used for the training. A number of convergence tolerances are tested with the model during the training to avoid the over- or under-training. The test results indicate that the ANN model show good performance with the all given con-vergence tolerances (Table 4). The best performance is shown in the model trained by 0.0001 which is eventually selected as the final model. The prediction performance of this model is summarized in detail in Table 5.

APPLICATIONS

After the ANN model was trained and tested successfully, the noise test is conducted to evaluate the model applicability


Table 4. Prediction performance of the ANN model according to the convergence tolerance

Convergence tolerance

0.001 0.0001 0.00001 0.000001No. of data set % No. of data set % No. of data set % No. of data set %

Total 71 100 71 100 71 100 71 100Succeed 66 93 69 97.2 68 95.8 66 93Failed 5 7 2 2.8 3 4.2 5 7

Table 5. Detailed prediction performance of the developed ANN model

EOR typeSucceed Failed Total

No. of data set % No. of data set % No. of data setSteam 33 100 0 0 33Carbon dioxide miscible 19 95 1 5 20Hydrocarbon miscible 10 100 0 0 10In-situ combustion 3 75 1 25 4Polymer flooding 4 100 0 0 4Total 69 97.2 2 2.8 71

Table 6. Oil viscosity vs. oil gravity of producing EOR projects

Field Cold Lake[1] WasonDenver[1] Twofreds[2] Cocorná[3]

Location Canada TX, USA TX, USA ColombiaOperator Imperial Oil Occidental HNG Fossil Fuel Company Ecopetrol S.A.

Formation type Unconsolidated Sandstone Dolomite Sandstone SandstoneDepth, ft 1,509 5,200 4,820 2,500

Temperature, °F 55 105 104 109Porosity, % Unknown Unknown 20.3 20-30

Permeability, md 1,500 8 33.4 1,080Oil saturation, % 70 51 Unknown 64

API gravity 10.2 33 36.4 13.1Oil viscosity, cp 10,000 1.2 1.467 Unknown

Method Steam (CSS) CO2 miscible CO2 miscible Steam* Data source[1] Lee and Lim, 2008; [2] Chung and Carroll, 1995; [3] Trujillo et al., 2010

against the error that may be included in the input data. A certain amount of noise is given to input data and accuracy of the estima-tion is measured with increasing error up to 30%. As shown in Fig. 9, the result shows that the ANN model can estimate the successfully applied EOR method with accuracy of 80% even though the input reservoir properties contain 30% of noise.

Finally, the developed model is applied to several world-wide successful EOR projects - Cold Lake, Wason Denver, Twofredsand Cocorná fields. The reservoir characteristics of each field are summarized in Table 6. In this paper, one of the reser-voir properties per each field is assumed to be unknown property to consider the uncertainty that we may be included at the initial project stage.

Cold Lake

Cold Lake is one of the largest oil sand fields located in Alberta, Canada. To produce these huge amounts of bitumen, the CSS (Cyclic Steam Stimulation) is applied. Porosity is assumed to be

unknown for this field. Within a typical porosity range, 10 to 50%, the model selected the steam as the best EOR method for this field.

Wason DenverWason Denver is located in Texas, U.S.A. The oil is being

produced by the carbon dioxide flooding for this field due to relatively deep reservoir depth and light oil compositions. Porosity is assumed to be unknown parameter as same as Cold Lake field. Within the given range of porosity, the model se-lected the carbon dioxide miscible flooding as the best EOR method.

TwofredsTwofreds is located in west Texas, U.S.A. The carbon dioxide

operations for the east side of the field were initiated in February 1974 after the field was nearing its waterflood economic limit. As a result of CO2 operations, oil production increased from 170 bopd to over a 1,000 bopd (Flanders and DePauw, 1993). For


Fig. 8. Structure of the ANN model developed in this study.

0

20

40

60

80

100

0 5 10 15 20 25 30

% of noise

% o

f acc

ura

cy

Fig. 9. Accuracy performance of the noise test.

this field oil saturation is assumed as unknown parameter while the actual oil saturation in the field is 57%. Within a tested range of oil saturation, 30 to 80%, the ANN predicted that carbon diox-ide flooding is the best EOR method.

CocornáCocorná field, operated by Ecopetro S.A., is located in the

Middle of Magdalena Valley basin in Puerto Perales town, Columbia. Trukillo et al.(2010) studied and concluded that the field is a good candidate to implement continuous steam

injection through the staged EOR screening methodology con-sisting of binary technical screening, analogies, benchmarking, and analytical predictions. The oil viscosity is 722 cp, but it is assumed as unknown property in this study. The ANN predicted that steam flooding is the most favorable in the field when the oil viscosity is greater than 100 cp.

CONCLUSIONS 1. A four-layered ANN model is developed to select the most

suitable EOR method based on the field characteristics. The in-put layer consists of the seven reservoir parameters and the out-put layer consists of the five EOR methods to be selected. The number of neurons in the hidden layers is optimized during the training; ten for the first hidden layer and eight for the second hidden layer.

2. After trained successfully with the successful EOR field data, the ANN model is tested against the data excluded in the training. The model correctly selected the best EOR method with the accuracy greater than 95%.

3. The noise test is performed to examine whether the ANN model overcomes the error included in the input data. The result showed that the model predicted the successful EOR method with the accuracy greater than 80% with up to 30% error in-cluded in the input data.

4. Finally, the ANN is applied to the worldwide successful EOR fields. The model selected the applied EOR methods as the best method even one of reservoir parameters was assumed as unknown.

5. A series of the tests results showed that the ANN model


developed in this study can be used to select the most appro-priate EOR process with the limited data in a very short time and cost effective way.

ACKNOWLEDGEMENT

This work was supported by the Energy Efficiency & Resources of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Ministry of Knowledge Economy, Republic of Korea (No. 2010201030001A).

REFERENCES Aladasani, A., and Bai, B, 2010, Recent Developments and Updated

Screening Criteria of Enhanced Oil Recovery Techniques: paper SPE 130726 presented at the CPS/SPE International Oil & Gas Conference and Exhibition, Beijing, China, June 8-10.

Beale, M.H., Hagan, M.T., and Demuth, H.B., 2010, Neural Network ToolboxTM 7, MathWorks, Massachusetts, U.S.A.

Chung, T.-H., and Carroll H.B., 1995, Application of Fuzzy Expert Systems for EOR Project Risk Analysis: paper SPE 30741 pre-sented at the SPE Annual Technical Conference and Exhibition, Dallas, Texas, U.S.A., October 22-25.

Dickson, J.L. and Wylie, P.L., 2010, Development of Improved Hydrocarbon Recovery Screening Methodologies: paper SPE 129768 presented at the SPE Improved Oil Recovery Symposium, Tulsa, Oklahoma, U.S.A., April 24-28.

Elemo, R.O., and Elmtalab, J., 1993, A Practical Artificial Intelligence Applicationin EOR Projects: SPE Computer Applications, V. 4, No. 2, p. 17-21.

Flanders, W.A., and DePauw, R.M., 1993, Update Case History: Performance of the TwofredsTertiary CO2 Project: paper SPE 26614 presented at the SPE Annual Technical Conference and Exhibition, Houston, Texas, U.S.A., October 3-6.

Goodlett, G.O., Honarpour, F.T., Chung, F.T., and Sarathi, P.S., 1986, The Role of Screening and Laboratory Flow Studies in EOR Process Evaluation: paper SPE 15172 presented at the Rocky Mountain Regional Meeting, Billings, Montana, U.S.A., May 19-21.

Guerillot, D.R., 1988, EOR Screening With an Expert System: paper SPE 17791 presented at the Symposium on Petroleum Industry Applications of Microcomputers, San Jose, California, U.S.A., June 27-29.

Hite, J.R., Avasthi, S.M., and Bondor, P.L., 2004, Planning EOR Projects: paper SPE 92006 presented at the SPE International Petroleum Conference, Puebla, Mexico, November 8-9.

Lee, J.-Y., and Lim, J.-S., 2008, Artificial Neural Network Approach to Selection of Enhanced Oil Recovery Method: Journal of the Korean Society for Geosystem Engineering, V. 45, No. 6, p.719- 726.

Moritis, G., 2006, Worldwide EOR Survey: Oil&Gas Journal, V. 104, Issue 15, p. 45-57.



Shokir, E.M., Goda, H.M., Sayyouh, M.H., and Fattah, Kh.A., 2002, Selection and Evaluation EOR Method Using Artificial Intelligence: paper SPE 79163 presented at the 26rd Annual International Technical Conference and Exhibition, Abuja, Nigeria, August 5-7.

Surguchev, L., and Li, L., 2000, IOR Evaluation and Applicability Screening Using Artificial Neural Networks: paper SPE 59308 presented at the SPE/DOE Improved Oil Recovery Symposium, Tulsa, Oklahoma, U.S.A., April 3-5.

Taber, J.J., and Martin, F.D., 1983, Technical Screening Guides for the Enhanced Recovery of Oil: paper SPE 12069 presented at the 58th Annual Technical Conference and Exhibition, San Francisco, California, U.S.A., October 5-8.

Taber, J.J., Martin, F.D., and Seright, R.S., 1996a, EOR Screening Criteria Revisited - Part1: Introduction to Screening Criteria and Enhanced Recovery Field Projects: paper SPE 35385 presented at the SPE/DOE Improved Oil Recovery Symposium, Tulsa, Oklahoma, U.S.A., April 21-24.

Taber, J.J., Martin, F.D., and Seright, R.S., 1996b, EOR Screening Criteria Revisited - Part2: Application and Impact of Oil Prices: pa-per SPE 39234 presented at the SPE/DOE Improved Oil Recovery Symposium, Tulsa, Oklahoma, U.S.A., April 21-24.

Trujillo, M., Mercado, D., Maya, G., Castro, R., Soto, C., Pérez, H., Gómez, V. and Sandoval, J., 2010, Selection Methodology for Screening Evaluation of Enhanced-Oil-Recovery Method: paper SPE 139222 presented at the SPE Latin Americal& Caribbean Petroleum Engineering Conference , Lima, Peru, December 1-3.

Zerafat, M.M., Ayatollahi, Sh., Mehranbod, N., and Barzegari, D., 2011, Bayesian Network Analysis as a Tool for Efficient EOR Screening: paper SPE 143282 presented at the SPE Enhanced Oil Recovery Conference, Kuala Lumpur, Malaysia, July 19-21.


Jong-Se Lim

Jong-Yong Lee is a reservoir engineer of Global Technology & Research Center, Korea National Oil Corporation. He holds BS andMS degrees in Energy & Resources Engineering from Korea MaritimeUniversity. (E-mail: [email protected])

Hyo-Jin Shin is a graduate student of the Department of Energy & Resources Engineering at Korea Maritime University. She holds a BSdegree in Energy & Resources Engineering from Korea Maritime University. (E-mail: [email protected])

Jong-Se Lim is a professor of the Department of Energy& ResourcesEngineering at Korea Maritime University. He holds BS, MS and Ph. Ddegrees in Petroleum Engineering from Seoul National University. (E-mail: [email protected])

Jong-Yong Lee Hyo-Jin Shin

Documents

Neural Network EOR