dc.description.abstract | This research investigates oil production in the modestly studied Niobrara shale play,
using data containing information about well-design and production volumes from wells
drilled in the period 2011 - 2018. Firstly, machine learning techniques were employed to
conduct a descriptive analysis, with the motive of identifying drivers of well-productivity.
Models of increasing spatial resolution were applied to isolate the effect of high-grading
of geological conditions from well-design choices. The statistical models employed were
different Random Forest (RF) configurations, Geographical Random Forest (GRF), and
the well-established technique of Regression Kriging (RK). It was found that spatial effects
were attributed slightly short of 40% of the relative importance in explaining variations in
the first-year production volumes of oil. Further, it was found that models attributed too
much importance to well-design variables if spatial effects were not adequately accounted
for. It was also found that the more data-driven and less restrictive RF and GRF
performed slightly better than the widely recognized RK. Secondly, a predictive analysis
was conducted in an attempt at identifying undrilled locations with favorable geology for
future drilling. For this part of the research, data containing information about geological
conditions were utilized, alongside the production data. It was found that applying RF
and GRF yielded quite disappointing results when the task was formulated as a regression
problem. However, lowering the precision by rephrasing the task as a classification problem
resulted in a RF classifier that outperformed random guessing to some extent. A visual
assessment of the classifier’s generated heatmap of predictions suggested that the model
was highly capable of identifying geological settings associated with the most extreme
wells, in terms of productivity. | en_US |