Machine Learning Workflows for Detection of High Water Cut Unconventional Wells Using Petrophysical Logs
Abstract
For many years, high water-cuts in the Delaware basin have been the source of much frustration for oil and gas operators producing in the region. The goal of this thesis was to construct automated workflows which are capable of predicting very early on in a horizontal well’s lifetime whether or not it will produce a substantially higher amount of water compared to hydrocarbon. With the intent to accomplish this goal, two different data-driven workflows have been developed. Each workflow focused on the differentiation of high water-producing wells (HWPs) and low water producing wells (LWPs) using machine learning (ML) algorithms. Both data-driven workflows use well log data, which provide information about the rock properties surrounding a given wellbore. The first data-driven workflow extracted out summary features from the well logs with respect to depth intervals below the kick-off point of a given wellbore, which is the point which a wellbore begins to transition from vertical to lateral. Using features extracted from well log data from 20 horizontal wells from the Delaware basin, supervised ML algorithms were trained to differentiate and predict which wells would be HWPs and LWPs. Logistic regression proved to be the most accurate supervised ML algorithm for the first proposed workflow. This workflow produced promising median F1 and Mathew’s correlation coefficient (MCC) scores of 0.96 and 0.92, respectively, for 100 cross-validation training iterations. The second data-driven workflow used unsupervised ML algorithms to assign a predicted lithology to every sample for 500 ft of well log data for 17 wells from the Delaware basin. This resulted in 5 unique lithologies which were found when all 17 wells were combined together. Using these predicted lithologies as a guide, features were extracted for all 17 wells and then used to train supervised machine learning algorithms to differentiate the two well classes: HWP and LWP. Using 100 cross-validation training iterations, three supervised algorithms
proved very comparable: K-Nearest neighbors, logistic regression, and support vector machine. Each of these supervised algorithms produced a median MCC score of 0.90. The geologic meaning of the most informative features from both workflows were also interpreted.
Citation
Foster, Jonathan Dominic (2021). Machine Learning Workflows for Detection of High Water Cut Unconventional Wells Using Petrophysical Logs. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /195372.