Enhancing Well Intervention Strategies: A Data-Driven Framework for Water-Shutoff Candidate Selection and Post-Treatment Assessment

Abstract

Summary Although the recent research explored the role of data science in reducing well intervention costs through the automated ranking of water shutoff (WSO) candidates, those studies did not consider the feasibility assessment of post-treatment well performance - during the design phase. This oversight could result in unexpected re-intervention cost or reserve loss, after job execution. This paper introduces an innovative, physics-informed, data-driven approach that addresses this research gap and proposes optimum recommendations for well intervention strategies. A dataset of 30 oil producers from different fields was used to evaluate workflow. Each well was characterized by (1) varying water production behaviors and (2) extensive well intervention activities. Methodology employed a hybrid approach, which integrated computer vision, unsupervised ensemble learning, and piecewise linear approximation (PWLA). Computer vision was utilized to pinpoint potential WSO candidates, while unsupervised learning provided insights into well dynamics and reservoir performance based on analysis of previous operations. Backward PWLA was implemented to uphold petroleum engineering principles and to anticipate the optimum WSO technique that could maximize recovery and minimize the need of post-job re-intervention. The proposed workflow is vital for large oilfields, enabling an instantaneous recommender system for WSO operations. Physics-based machine learning model was found effective in overcoming the preceding limitations and assumptions, permitting the auto- segmentation of Arps water-cut forecasting model, decline curve analysis (DCA), and water control diagnostic plots. This approach simplified the complex patterns and behaviors observed in mature reservoirs. Overall, the key alterations in water production driving mechanisms were detected with a mean average precision of 0.84, while the hidden anomalies in reservoir and well data, which negatively impact reserve estimation, were identified with an F1-score of 0.76. The water-cut forecast, post-WSO, was predicted at the design phase with a confidence score of 0.86. Altogether, the primary advantage of this hybrid approach is its ability to effectively segment and distinguish changes in well and reservoir performance from those changes caused by surface operational variations. This enabled better understanding of reservoir response to WSO - whether mechanical or chemical - as well as optimizing the selection of treatment techniques that maximizes ultimate recovery gain. This paper investigated a research gap in the existing WSO frameworks. It presents a novel machine learning approach, which not only ranks WSO candidates but also evaluates the feasibility of diverse WSO treatment techniques for each well individually. This workflow reduces the potential costs of well re-intervention and helps asset teams optimize field development.