Abstract

Oil and gas (O&G) well integrity issues, including sustained casing pressure and casing vent flow, could cause significant environmental and climatic risks by releasing methane-containing fluids into the environment. The large number of O&G wells makes routine inspection or spatiotemporally adequate screening of well integrity issues expensive and sometimes intractable. Machine learning (ML) algorithms with the ability to predict O&G well integrity issues could help prioritize such inspections. We manually extracted well characteristics from 1223 O&G well completion reports in Bradford County, PA, appended geospatial metrics, and used them to predict well integrity issues reported by field tests. Different ML models (e.g., Random Forrest, XGBoost, and Logistic Regression) were compared, and an accuracy of 68% was achieved by comparing the predictions with unseen field test records. Important predictive features were identified by the Random Forrest model, such as the length of casings, amount of cement, and well operator. Moreover, the wells with integrity issues were geospatially clustered in our study region, which could be explained by the clusters of important physical features. Overall, the predictions would help prioritize well integrity inspection and have the potential to guide well design and well location selection for mitigating integrity issues that lead to methane emissions.