Abstract

Objectives/Scope: Understanding drainage dynamics of hydraulically stimulated horizontal wells through time is among the most critical parameters for optimizing field development strategies in unconventional reservoirs. Several years ago, we established a novel geochemical-based methodology using production allocation of produced oils collected in time-series and calibrated to core extracted oils to cost effectively ascertain vertical drainage. While the core-calibrated methodology has proven successful, it requires time-consuming project-specific model building, uncontaminated proximal core, depth shifting and triangulation, and parameter restriction to the extracted fraction. To overcome these challenges, an innovative processing technology was developed that exploits data analytics and machine-learning of the existing Time-Lapse Geochemistry (TLG) database. Methods/Procedures/Process: We created a 3-D full field compositional model across ConocoPhillips' Eagle Ford acreage using high confidence produced oils. The best field-wide compositional parameters distinguishing drilling targets were determined by rigorous statistical filtering with principal components and cluster analyses on uniformly aligned and indexed gas chromatography data. Oil compositions per landing zone were back-calculated and matched to contributions by layer in each oil mixture that were known a priori from previously core-calibrated allocations. Compositions vertically in-between drilling targets were interpolated from the known landing zones. Multi-layer compositional depth profiles were generated at each sampled project and extrapolated across the field using correlation gridding with geological parameters derived from a calibrated 3-D basin model. Results/Observations/Conclusions: Mapped geochemical parameter grids were stacked and amalgamated to construct the 3-D compositional grid system, which forms a field wide 17-layer vertical profile of geochemical endmembers. An integrated platform was built to visualize the compositional model in combination with indexed TLG oils. Python was used to digitize the compositional model and explore multiple machine learning algorithms as a back check on grid system quality. Random TLG samples served as blind validation data. We identified a random forest ensemble regression algorithm from the Scikit-Learn package as the optimal model to forward predict the 3-D location of every produced oil. Production allocations were further assisted by a random forest classifier that calculated prediction probabilities for drainage contributions from each layer. The machine learning toolkit aids in selecting appropriate endmember depth profiles and provides optionality for running production allocations through either the numerical mixing optimizer or machine learning algorithm. Applications/Significance/Novelty: This is the first example of a 3-D geochemical model of produced oil compositions generated for an unconventional reservoir. This next-generation workflow couples a consistent and robust grid system across the field with machine learning assistance made possible by data analytics of a high density, high resolution TLG dataset. Access to a wider molecular range, expanded spatial reach away from core control, and dramatically reduced allocation run times have greatly increased production allocation versatility. Tremendous business value is created by having the ability to quickly ascertain vertical drainage dynamics across multiple well configurations, completions, enhanced recovery projects, and infill drilling as feedback for long-term strategy decisions in addition to incorporating geochemistry with any multivariate investigation of subsurface challenges.

This content is only available via PDF.
You can access this article if you purchase or spend a download.