Professional Documents
Culture Documents
2021-SPE-208657-Successful Application of ML To Improve Dynamic Modeling HM For Complex Gas Condensate Reservoirs in Hai Thach Field, Nam Con Son Basin, Offshore Vietnam
2021-SPE-208657-Successful Application of ML To Improve Dynamic Modeling HM For Complex Gas Condensate Reservoirs in Hai Thach Field, Nam Con Son Basin, Offshore Vietnam
This paper was prepared for presentation at the SPE Symposium: Artificial Intelligence - Towards a Resilient and Efficient Energy Industry held virtually on 18 - 19
Oct 2021.
This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.
Abstract
This study aims to apply machine learning (ML) to make history matching (HM) process easier, faster, more
accurate, and more reliable by determining whether Local Grid Refinement (LGR) with transmissibility
multiplier is needed to history match gas-condensate wells producing from geologically complex reservoirs
and determining how LGR should be set up to successfully history match those production wells.
The main challenges for HM gas-condensate production from Hai Thach wells are large effect of
condensate banking (condensate blockage), flow baffles by the sub-seismic fault network, complex reservoir
distribution and connectivity, highly uncertain HIIP, and lack of PVT information for most reservoirs. In
this study, ML was applied to analyze production data using synthetic samples generated by a very large
number of compositional sector models so that the need for LGR could be identified before the HM process
and the required LGR setup could also be determined. The proposed method helped provide better models
in a much shorter time, and improved the efficiency and reliability of the dynamic modeling process.
500+ synthetic samples were generated using compositional sector models and divided into training
and test sets. Supervised classification algorithms including logistic regression, Gaussian, Bernoulli, and
multinomial Naïve Bayes, linear discriminant analysis, support vector machine, K-nearest neighbors, and
Decision Tree as well as ANN were applied to the data sets to determine the need for using LGR in HM. The
best algorithm was found to be the Decision Tree classifier, with 100% and 99% accuracy on the training
and the test sets, respectively. The size of the LGR area could also be determined reasonably well at 89% and
87% accuracy on the training and the test sets, respectively. The range of the transmissibility multiplier could
also be determined reasonably well at 97% and 91% accuracy on the training and the test sets, respectively.
Moreover, the ML model was validated using actual production and HM data.
A new method of applying ML in dynamic modeling and HM of challenging gas-condensate wells in
geologically complex reservoirs has been successfully applied to the high-pressure high-temperature Hai
Thach field offshore Vietnam. The proposed method helped reduce many trial and error simulation runs and
provide better and more reliable dynamic models.
2 SPE-208657-MS
Introduction
Within the oil and gas industry, ML has been applied to a wide range of problems such as predicting and
detecting equipment malfunctions (Sneed 2017, Bangert 2019), production forecasting (Cao et al. 2016;
Mukherjee et al. 2019), automatic depth matching (Zimmermann et al. 2018), shear wave velocity prediction
(Akhundi et al. 2014; Anemangely et al. 2017; Bukar et al. 2019), facies classification (Bestagini et al.
2017, Miller et al. 2019, Tran et al. 2020), well log correlation (Brazell et al. 2019, Maniar et al. 2018), fault
detection (Maniar et al. 2018), seismic facies classification (Chopra and Marfurt 2018), geological feature
prediction (Jobe et al. 2018), sanding prediction (Kanj and Abousleiman 1999), wellbore stability analysis
The well was set to produce at a constant gas rate typical of actual field production for a sufficiently
long period to allow it to reach stable flow. It was then shut in to build up pressure. From the buildup
pressure curve, 1000 pressure points with equal time interval were extracted and normalized for both time
and pressure. In this study, more than 500 synthetic samples were generated.
model is predicted to need LGR and it does need LGR. Between the two types of errors (false positive and
false negative), false negative is worse in this study because it will totally mislead the modeling efforts
and produce very poor HM results. Therefore, close attention will be paid to false negatives during the
evaluation of each ML model.
In addition, the accuracy score was calculated as follows:
In this study, the three algorithms with the highest accuracy scores in decreasing order were Decision
Tree, K-nearest neighbors, and Random Forest, as shown in Table 1.
Training 1.00
Decision Tree Classifier
Test set 0.99
Training 0.81
Bernoulli Classifier
Test set 0.85
Training 0.86
Multinomial Classifier
Test set 0.86
Training 0.99
K-Nearest Neighbors
Test set 0.99
Training 0.79
Gaussian Naïve Bayes
Test set 0.78
Training 0.94
Support Vector Machine
Test set 0.96
Training 0.97
Random Forest
Test set 0.97
The confusion matrix report system was also used in this study to evaluate the accuracy of each model.
The confusion matrix is as follows:
The confusion matrices for all ML models for 153 datasets in the test set are summarized in Table 2. It
was observed that the Decision Tree classifier had an outstanding performance with only 1 false positive
and no false negatives. KNN was the second best with no false positives but two false negatives. Random
Forest came in third with only 1 false positive but three false negatives. Other methods did not perform as
well as the aforementioned top three methods. Therefore, only Decision Tree, KNN, and Random Forest
algorithms were selected for further study.
SPE-208657-MS 5
Bernoulli Classifier
Multinomial Classifier
Random Forest
For the ANN model, 1000 runs were conducted. The accuracy was relatively high, as shown in Fig. 2.
There was also a good fit with a minimal gap between the two final loss values, as shown in Fig. 3. However,
the results were not as good as those using Decision Tree, KNN, and Random Forest methods. Therefore,
ANN was not pursued further in this study.
For the size of the LGR area, the results are summarized in Table 3. The Decision Tree classifier was
again the best ML method, with 89% accuracy on the training set and 87% accuracy on the test set. Neither
KNN nor Random Forest method could perform as well as Decision Tree for either data set. The confusion
matrices, as summarized in Table 4, also showed that the Decision Tree classifier was the best algorithm
among the three methods.
Training 0.89
Decision Tree Classifier
Test set 0.87
Training 0.83
K-Nearest Neighbors
Test set 0.79
Training 0.82
Random Forest
Test set 0.81
K-Nearest Neighbors
Random Forest
SPE-208657-MS 7
For the range of transmissibility multiplier, the results are summarized in Table 5. The Decision Tree
classifier was again the best ML method, with 97% accuracy on the training set and 91% accuracy on the
test set. Neither KNN nor Random Forest method could be as good as Decision Tree for either data set. The
confusion matrices, as summarized in Table 6, also showed that the Decision Tree classifier was the best
algorithm among the three methods.
Training 0.93
K-Nearest Neighbors
Test set 0.89
Training 0.91
Random Forest
Test set 0.89
K-Nearest Neighbors
Random Forest
Figure 4—Typical poor history matching results of well HT-C using conventional models.
Comprehensive and in-depth review of the reservoir revealed that it was severely faulted with a network
of sub-seismic faults. Furthermore, there was core-based evidence that these faults could be filled and
have low permeability, therefore acting as flow baffles within the reservoir. As a consequence, LGR with
transmissibility multiplier was applied around well HT-C to the previously best dynamic model without
LGR. Finally, well HT-C was successfully history matched for the whole history, including all drawdown
and buildup periods, as shown in Fig. 5.
SPE-208657-MS 9
Figure 5—Good history matching results of well HT-C with LGR and transmissibility multiplier.
ML prediction using HT-C buildup data showed that LGR would be needed to history match this well,
as shown in Table 7, consistent with actual dynamic modeling results.
10 SPE-208657-MS
Table 7—HT-C prediction results for the need to use LGR in history matching.
K-Nearest
Buildup No. Decision Tree Random Forest
Neighbors
1 1 1 1
2 1 1 1
3 1 1 1
4 1 1 1
5 1 1 1
Moreover, prediction using HT-C buildup data also showed that an area larger than 3×3 cells would
need LGR and the transmissibility multiplier of less than or equal to 0.02 would be needed. Those results
were consistent with the actual LGR setup used to successfully history match HT-C (LGR of 9×9 cells
with transmissibility multiplier of 0.02). In summary, the ML model was successfully validated using actual
production data as well as actual dynamic modeling results of well HT-C.
Application
Well HT-B was selected for the application in this case study. For each buildup, pressure data was loaded
with a minimum time step of 30 minutes to obtain one thousand data points. The top three algorithms
(Decision Tree, K-nearest neighbors, and Random Forest) all predicted that LGR would be required to
achieve good HM for well HT-B, as shown in Table 8.
Table 8—HT-B prediction results for the need to use LGR in history matching.
K-Nearest
Buildup No. Decision Tree Random Forest
Neighbors
1 1 1 1
2 1 1 1
3 1 1 1
Regarding the LGR setup, the ML models predicted that an area larger than 3×3 cells would need LGR
and a transmissibility multiplier less than or equal to 0.02 would be required to history match HT-B.
Based on the ML prediction results, LGR with transmissibility multiplier was applied to HT-B dynamic
model. Well HT-B had only one year of downhole gauge data, therefore it was history matched using tubing
head pressure control. HM for HT-B was previously done without LGR and a good match could not be
obtained, as shown in Fig. 6. Gas production rate was overestimated while BHP was underestimated. Based
on the ML prediction, LGR was then applied around HT-B to the previously best model without LGR. With
a 10×11 LGR area and transmissibility multiplier of 0.005, a good history match was obtained, as shown
in Fig. 7.
Downloaded from http://onepetro.org/SPEIOGS/proceedings-pdf/19IOGC/2-19IOGC/D021S007R001/2506334/spe-208657-ms.pdf by Son Hoang on 19 October 2021
11
Figure 7—Good history match of HT-B model with LGR and transmissibility multiplier.
12
SPE-208657-MS 13
Conclusions
This paper reports a successful case study of applying ML to improve the HM process for gas-condensate
wells producing from geologically complex reservoirs, making it easier, less time-consuming, and more
reliable, by determining with almost perfect accuracy whether LGR with transmissibility multiplier would
be required. The required LGR area and range of the required transmissibility multiplier could also be
determined with high accuracy before the history matching process begins. As a result, the ML prediction
results have significantly reduced the number of trial-and-error simulation runs, helped provide better
models in a much shorter time, and greatly improved the efficiency and reliability of the dynamic modeling
References
Akhundi, H., Ghafoori, M., and Lashkaripour, G.R. 2014. Prediction of Shear Wave Velocity Using Artificial Neural
Network Technique, Multiple Regression and Petrophysical Data: A Case Study in Asmari Reservoir (SW Iran). Open
Journal of Geology 4:303–313. http://dx.doi.org/10.4236/ojg.2014.47023.
Alakbari, F.S., Elkatatny, S., and Baarimah, S.O. 2016. Prediction of Bubble Point Pressure Using Artificial Intelligence AI
Techniques. Paper SPE 184208 presented at the SPE Middle East Artificial Lift Conference and Exhibition, Manama,
Kingdom of Bahrain, 30 November-1 December. https://doi.org/10.2118/184208-MS.
Anemangely, M., Ramezanzadeh, A., and Tokhmechi, B. 2017. Shear Wave Travel Time Estimation from Petrophysical
Logs Using ANFIS-PSO Algorithm: A Case Study from Ab-Teymour Oilfield. Journal of Natural Gas Science and
Engineering 38:373–387. https://doi.org/10.1016/j.jngse.2017.01.003.
Bangert, P. 2019. Predicting and Detecting Equipment Malfunctions Using Machine Learning. Paper SPE 195149
presented at the SPE Middle East Oil and Gas Show and Conference, Manama, Bahrain, 18-21 March. https://
doi.org/10.2118/195149-MS.
Bestagini, P., Lipari, V., and Tubaro, S. 2017. A Machine Learning Approach to Facies Classification Using Well Logs.
SEG International Exposition and 87th Annual Meeting, Houston, TX, USA, 24-29 September. https://doi.org/10.1190/
segam2017-17729805.1.
Brazell, S., Bayeh, A., Ashby, M., Burton, D. 2019. A Machine-Learning-Based Approach to Assistive Well-Log
Correlation. Petrophysics 60(4):469–479. https://doi.org/10.30632/PJV60N4-2019a1.
Bukar, I., Adamu, M.B., and Hassan, U. 2019. A Machine Learning Approach to Shear Sonic Log Prediction. Paper SPE
198764 presented at the Nigeria Annual International Conference and Exhibition, Lagos, Nigeria, 5-7 August. https://
doi.org/10.2118/198764-MS.
Cao, Q., Banerjee, R., Gupta, S., Li, J., Zhou, W., and Jeyachandra, B. 2016. Data Driven Production Forecasting Using
Machine Learning. Paper SPE 180984 presented at the SPE Argentina Exploration and Production of Unconventional
Resources Symposium, Buenos Aires, Argentina, 1-3 June. https://doi.org/10.2118/180984-MS.
Chopra, S. and Marfurt, K.J. 2018. Seismic Facies Characterization Using Some Unsupervised Machine Learning
Methods. Proc., SEG International Exposition and 88th Annual Meeting, 15 October, 2056–2060. https://
doi.org/10.1190/segam2018-2997356.1.
Honorio, J., Chen, C., Gao, G., Du, K., and Jaakkola, T. 2015. Integration of PCA with a Novel Machine Learning
Method for Reparameterization and Assisted History Matching Geologically Complex Reservoirs. Paper SPE 175038
presented at the SPE Annual Technical Conference and Exhibition, Houston, Texas, USA, 28-30 September. https://
doi.org/10.2118/175038-MS.
Illarionov, E., Temirchev, P., Voloskov, D., Gubanova, A., Koroteev, D., Simonov, M., Akhametov, A., and Margarit, A.
2020. 3D Reservoir Model History Matching Based on Machine Learning Technology. Paper SPE 201924 presented at
the SPE Russian Petroleum Technology Conference, Moscow, Russia, 12-14 October. https://doi.org/10.2118/201924-
MS.
Jobe, T.D., Vital-Brazil, E., and Khait, M. 2018. Geological Feature Prediction Using Image-Based Machine Learning.
Petrophysics 59(6):750–760. https://doi.org/10.30632/PJV59N6-2018a1.
Kanj, M.Y. and Abousleiman, Y. 1999. Realistic Sanding Predictions: A Neural Approach. Paper SPE 56631 presented at
the SPE Annual Technical Conference and Exhibition, Houston, TX, USA, 3-6 October. https://doi.org/10.2118/56631-
MS.
Maniar, H., Ryali, S., Kulkarni, M.S., and Abubakar, A. 2018. Machine Learning Methods in Geoscience.
Proc., SEG International Exposition and 88th Annual Meeting, 15 October, 4638–4642. https://doi.org/10.1190/
segam2018-2997218.1.
14 SPE-208657-MS
Miller, R.S., Rhodes, S., Khosla, D., and Nino, F. 2019. Application of Artificial Intelligence for Depositional Facies
Recognition – Permian Basin. Paper URTEC-2019-193-MS presented at the Unconventional Resources Technology
Conference, Denver, Colorado, USA, 22-24 July. https://doi.org/10.15530/urtec-2019-193.
Mukherjee, T., Burgett, T., Ghanchi, T., Donegan, C., and Ward, T. 2019. Predicting Gas Production Using Machine
Learning: A Case Study. SEG International Exposition and 89th Annual Meeting, San Antonio, TX, USA, 15-20
September. https://doi.org/10.1190/segam2019-3215692.1.
Pham, D.H., Hoang, S.K., Trinh, V.X., and Tran, T.V. 2020. Condensate Banking Characterization and Quantification of
Improvement from Different Mitigations Using Pressure Transient Analysis: A Case Study in Hai Thach Field Offshore
Vietnam. Paper OTC 30142 presented at the Offshore Technology Conference Asia, Kuala Lumpur, Malaysia, 2-6
November. https://doi.org/10.4043/30142-MS.