Integrating Machine Learning in Identifying Sweet Spots in Unconventional Formations
- Saurabh Tandon (The University of Texas at Austin)
- Document ID
- Society of Petroleum Engineers
- SPE Western Regional Meeting, 23-26 April, San Jose, California, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. Society of Petroleum Engineers
- Regression Analysis, Machine Learning, Sweet Spot Analysis, Data analytics
- 43 in the last 30 days
- 325 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 9.50|
|SPE Non-Member Price:||USD 28.00|
Productive zones or "sweet spots" in unconventional reservoirs depend on their geomechanical and petro-physical rock properties. Machine learning algorithms can significantly improve workflows used for evaluating sweet-spots in such complex reservoirs. The objectives of this paper are to: (i) quantity the effects of rock mechanical properties on fracturing treatments using data analytics and (ii) use regression-based machine learning algorithms and improve sweet-spot assessment in complex mudrock reservoirs.
We used a hydraulic fracturing simulator that couples fluid-flow with fracture deformation in discrete fracture networks to model field-scale hydraulic fracturing treatments. First, we selected several geomechanical properties related to rock fracability. We obtained wide variation in aforementioned properties using a quasi-random design approach. Then, we performed 200 slick-water fracturing simulations with quasi-random distribution of design parameters using the hydraulic fracturing simulator. We quantified the performance of fracture treatments by calculating the effective short- and long-term Stimulated Reservoir Volume of the reservoir (SRV). We finally analyzed the results of numerical simulations by applying regression analysis to improve the assessment of sweet-spots in complex reservoirs.
The regression analysis involved the following simulation variables: shear modulus, poisson's ratio, fracture friction coefficient, principal horizontal stress anisotropy, fracture toughness, fracture closure stress, shear dilation angle, and initial fracture aperture. The SRV results were analyzed using: linear regression, linear regression with beta coefficients, ridge and lasso regression, and principal component regression algorithms. The regression analysis revealed that linear models can explain 73.1% and 59.2% variance in short- and long-term SRV values, respectively. The ridge and lasso regression and beta linear regression analysis revealed that stress anisotropy, fracture dilation angle, and fracture friction coefficient show the highest effect on the aforementioned SRV values. In all the regression models, shear modulus and critical fracture toughness did not have a significant effect on SRV but these parameters are important as they are correlated to other parameters that directly impact fluid flow.
The results of using data analytic approaches demonstrated that factors related to unpropped fracture conductivity play a critical role in success of hydraulic fracturing treatments. We have also introduced and compared the performances of different machine learning algorithms that might be used to assess the impact of geomechanical properties on fracturing treatments. Such supervised and unsupervised machine learning algorithms can help in integrating legacy field data in the analysis of productive zones in complex reservoirs. Such analysis can also be used to develop data-based models that might improve the study of sweet-spot and fracturing treatment performance assessment in complex reservoirs.
|File Size||1 MB||Number of Pages||17|
Banerjee, A. 2004. Validating clusters using the Hopkins statistic. IEEE International Conference on Fuzzy Systems: 149–153. doi:10.1109/FUZZY.2004.1375706.