The reliability of public water production data for tight oil wells

Objectives/Scope: The accuracy of publicly available water production estimates for shale wells has been controversial. As a result, the robustness of models based on production data is uncertain. The produced water estimates are a key to evaluating the environmental impact and economic viability of hydraulic fracturing. The uncertainty in the data models and how these impact the estimates being used to impact decision making have not been systematically examined. In this study, publicly available estimates of produced water per well were compared to proprietary data for ten thousand tight-oil wells in New Mexico (1200 wells), North Dakota (1300 wells), Texas (7700 wells), and Utah (100 wells). Methods/Procedures/Process: In order to understand these uncertainties several proprietary data sets were compared to matched public reports. They were compared in raw form as well as after being transformed and normalized and as both water production and water cut. The comparisons included Pearson’s correlation and Spearman’s rank-correlation. One thousand bootstrapped samples were taken to determine uncertainty in the correlation. Then, for wells in the Permian basin, well engineering, reservoir, and administrative factors (such as the size of the lease and date the well was drilled) were compared through building a gradient boosting machine learning models to find what could be causing large discrepancies in reported production volumes and whether the discrepancies could be corrected. Results/Observations/Conclusions: Estimates derived from public data varied greatly state-by-state. In Utah, the public results always accurately reflected proprietary data, with an R-squared above 0.9. In New Mexico and North Dakota, early production showed an R-squared of about 0.6, and later production values had R-squared values above 0.9. Lastly, in Texas, R-squared was never above 0.6, falling to 0.1 for early water production. In all states, correlation metrics improved as time-on-production grew. In Utah, North Dakota, and New Mexico, water production from the first year on is accurate, but in Texas, public reporting poorly reflects real water production. Applications/Significance/Novelty: There are several implications of inaccurate public water production data for Texas. These include difficulty in assessing the hydraulic fracture flow-back volumes, disposal costs, and environmental impact for water produced from tight-oil wells. Incorrectly predicting water production can impact when wells become uneconomical from water lifting and disposal costs and are shut in, affecting reserves estimates. Possible remedies for this problem include using salt-water disposal volumes to estimate the water produced, machine learning modeling work such as ours, Texas instituting regulatory changes, and researchers partnering with operators and data brokers to increase the fidelity of their models.