Pattern Recognition for Fractured Reservoir Characterization using Subsurface Big Data
Open Access
- Author:
- Udegbe, Egbadon Ajibola
- Graduate Program:
- Energy and Mineral Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 16, 2018
- Committee Members:
- Sanjay Srinivasan, Dissertation Advisor/Co-Advisor
Eugene C Morgan, Committee Chair/Co-Chair
Mort D Webster, Committee Member
Hamid Emami-Meybodi, Committee Member
Tieyuan Zhu, Outside Member - Keywords:
- big data
data science
subsurface
fractures
seismic
unconventional reservoirs
shale - Abstract:
- In recent years, there has been a significant increase in the quantity of data generated from monitoring technologies for subsurface operations such as permanent downhole sensors, as well as cross-hole and seismic surveys. Traditional models and techniques have proven inadequate for the purpose of extracting information from Big Data, in support of reservoir management and decision-making. In addition, the last decade has brought about increased exploration of unconventional reservoirs such as shale, due to more favorable economics resulting from advances in directional drilling and hydraulic fracturing. However, existing methods for describing induced and natural fracture characteristics in the subsurface are still evolving, and associated impacts on well performance are not completely understood. To attain optimal development of these resources, we require accurate characterization of fractures and reservoir characteristics from subsurface time-series and spatial data. The above challenges have the potential to be addressed by developing new Big Data analytic tools focused on identifying and characterizing complex subsurface features such as fractures, by exploiting pattern recognition and high-performance computing to uncover masked trends in large volumes of subsurface data. In support of this objective, real-time face detection techniques have been adapted to establish a pattern recognition methodology for feature extraction, statistical learning and probabilistic model evaluation. Under this framework, a set of easy-to-compute features based on Haar wavelets are extracted directly from the data, in order to serve as attributes for training a cascade of probabilistic tree-based ensemble classification models. As a use case for time-series data analytics, production data simulated from hydraulically fractured shale gas wells have been trained to identify candidates for re-stimulation treatment. Results demonstrate the viability of the proposed framework in recognizing favorable re-stimulation candidate wells using solely gas rate profiles, with improved accuracy over conventional tools such as type-curve matches. Secondly, the proposed methodology has been extended to help identify fractures in post-stack seismic data, which has been trained using raw seismic amplitude responses generated using a discontinuous Galerkin finite element seismic wave propagation model. Next, the approach has been validated using 3D post-stack seismic data from the Niobrara Shale interval within the Teapot Dome field in Wyoming. The applicability of the proposed framework has been demonstrated for identifying sub-seismic fractures, by considering the amplitude profile adjacent to interpreted fullbore microimage (FMI) well log data. The up-scaled spatial distribution of the predicted fractures shows agreement with existing geological studies and align with interpreted large-scale faults within the interval of interest.