Rolling window for detecting multiple Chan signatures to diagnose excessive water production

Abstract Chan diagnostic plot has proven highly effective in industry as a powerful tool for diagnosing and monitoring excessive water production in oil wells. Chan identified various distinct water problem patterns exhibited by wells during their production lifecycle based on slope features: Constant water-oil ratio (WOR), Normal Displacement, Coning, Channeling, depletion, and Multilayer Channeling. While machine learning models were developed to automate Chan plot interpretation and reduce human bias, they often failed to capture the evolution of water production over time, typically detecting only a single pattern or capturing patterns after it fully materialized, reducing the chance of early diagnosing. This study investigates the effectiveness of incorporating the rolling window function for recognizing patterns dynamically, aiming to distinguish the various water problem patterns for optimal workover candidate selections and quick water treatment. Alaska oil wells’ public production data were analyzed illustrating each pattern’s manifestation. A successful interactive model with the rolling window feature was developed to track slope changes in Chan signatures, resulting in a 7–10% improvement in pattern detection accuracy compared to static features. Throughout, an iterative optimization process, window size was determined as seven points, considering pattern duration. Eight algorithms were evaluated, with Support Vector Machines (SVM) and Random Forest (RF) achieving a remarkable 94% F1 score while the remaining algorithms averaged 93%.