Machine Learning Interpretability Application to Optimize Well Completion in Montney
- Yousef Sheikhi Garjan (Abra Controls ltd) | Mehdi Ghaneezabadi (SAIT)
- Document ID
- Society of Petroleum Engineers
- SPE Canada Unconventional Resources Conference, 29 September - 2 October, Virtual
- Publication Date
- Document Type
- Conference Paper
- 2020. Society of Petroleum Engineers
- 5.8.2 Shale Gas, 2.5.2 Fracturing Materials (Fluids, Proppant), 7.6.7 Neural Networks, 3 Production and Well Operations, 2.4.1 Fracture design and containment, 2 Well completion, 2.4 Hydraulic Fracturing, 5.1.5 Geologic Modeling, 7.6 Information Management and Systems, 5.5 Reservoir Simulation, 7.6.6 Artificial Intelligence, 2.1 Completion Selection and Design, 7 Management and Information, 2.1 Completion Selection and Design, 5 Reservoir Desciption & Dynamics, 7.6.4 Data Mining
- Montney, Optimization, Well completion, Hydraulic fracturing, Machine Learning
- 76 in the last 30 days
- 76 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 5.00|
|SPE Non-Member Price:||USD 28.00|
Recently machine learning has being extensively deployed for oil and gas industry for improving result and expedite process. However, the black box models do not explain their prediction which considered as a barrier to adopt machine learning. This paper is about optimizing hydraulic fracture with machine learning methods and making informative decision with interpreting machine learning model. The solution can show that it could save over million dollars per well and improve well performance significantly. Interestingly, the machine leaning explainability approach was utilized to explain and measure the reason behind of why some wells are performing better than other and vice versa.
Hydraulic fracturing modeling and optimization in tight oil and unconventional reservoir requires substantial geological modeling, fracture design, post-fracture production simulation with excessive sensitivity analysis due to complexity and uncertainty in the nature of data. These types of studies are computationally and monetarily expensive. Furthermore, digital oil technology has facilitated the process of data gathering enabled operators to have access to huge amount of data. Common approaches are no longer suitable to handle this pile of data but machine learning methods could be successfully utilized for this purpose.
In this paper, a variety types of advanced machine learning methods including linear regression, Random forest, Gradient Boost, XGBoost, Bagging, ExtraTrees and neural network were employed to optimize well completion in Montney formation. The objective was to create a robust predictive model capturing all the effective operational well parameters (features) capable of optimizing the first 12 months cumulative of equivalent well production.
Special Individual Conditional Expectation (ICE) plots and Partial Dependency plots(PDP) were used to depict how HF completion features influence the prediction of a machine learning model. Furthermore, a novel approach was employed to explain the model prediction of an existing well by computing the contribution of each feature to the prediction.
Over 1838 hydraulically fractured (HF) wells producing from 2008 till 2019 in Montney formation have been considered for this analysis. The outcome of Explanatory Data Analysis (EDA) revealed that well production performance has not been improved despite of continues enhancement of hydraulic fracture parameters such as proppant injected volume, length of stimulated horizontal wells, and number of stages per well in the course of two years. This finding raises the concern of whether operators are properly optimizing completion design. After comparing all machine learning methods, Random Forest method was chosen as the most appropriate and accurate method to proceed for further analysis. ICE and PDP plots helped to understand the impacts of different fracturing features on production for individual well in addition to define optimum operation features on Montney Formation. Furthermore, quantifying of each feature’s impact on individual well production and linking it to an economic model, we were able to demonstrate potential profit and loss for each well. The model suggests that some wells could have achieved over $1 million extra profit during the first 12-months of production.
In this study, not only a reliable predictive data-driven model has been built for hydraulically-fractured wells in Montney formation, but also a comprehensive workflow of sensitivity and explainatability analysis has been introduced to obtain an optimized fit-to-purpose well completion design.
|File Size||1 MB||Number of Pages||16|
Hirschmiller, J., Biryukov, A, Groulx, B., Emmerson, B., Quinell, S. 2019: The Importance of Integrating Subsurface Disciplines with Machine Learning when Predicting and Optimizing Well Performance —Case Study from the Spirit River Formation. Presented at the SPE Annual Technical Conference and Exhibition held in Calgary, Alberta, Canada, 30 Sep - 2 October. SPE-196089-MS.
Molnar, c., 2019, online book, Interpretable Machine Learning. https://christophm.github.io/interpretable-ml-book/