Machine learning assisted detection of excess water-producing wells in unconventional shale plays

Abstract

Excess water production from oil/gas wells drilled in unconventional shale formations of Permian Basin leads to environmental concerns related to the disposal of produced water and also leads to significant expenses in water management programs. In the Delaware Basin, water-to-oil ratios are as high as 10:1 for an oil-producing well. This paper analyzes subsurface data from 20 wells targeting the Bone Springs sands and the Wolfcamp shale formation, which are the primary unconventional-resource targets in the Delaware Basin. The dataset was expanded by adding 53 wells from both the Fort Worth Basin and Gulf Coast Region. For purposes of this study, a data-driven workflow was developed to use 5 well logs from the 300-feet depth interval above or below the kick-off point for each well to detect high water-producing well. The 300 ft of logged data above or below the KOP are divided into six 50-ft intervals. From these banded intervals, 7 statistical parameters were computed for each of the 5 logs. In doing so, 210 features were generated for each well, which were then reduced to most informative 21 features by first thresholding based on p-value and F value, followed by the use of Pearson Correlation. The dimensionally reduced 21 features were used to train classification methods to predict whether a well is a high-water producer (HWPs) and low-water producer (LWPs). Logistic regression performs the best. In Delaware Basin, the prediction performance has a median F1 score of 0.96 and a median Matthew's Correlation Coefficient of 0.92. Median scores for both the Fort Worth and Gulf Coast datasets are above 0.80. For the Delaware Basin, the statistical parameters computed using deep resistivity and porosity logs from the nearest 50-ft band just below the KOP are the most informative features for detecting excess water-producing well.