Real-Time Well Constraint Detection Using an Intelligent Surveillance System

Abstract

Abstract Engineering resources are stretched in a field with large-scale operations. Well performance issues are often overlooked, given the high number of wells generating a large volume of data. We developed an advanced intelligent surveillance system to detect common field problems and constraints faster, thereby reducing the time to analyze and diagnose events contributing to well underperformance and reducing the overall time for field engineers to act. Combining industry expertise with data science techniques and machine learning, a series of workflows were constructed to detect well constraints proactively. The well constraints included hydrate formation, crown valve issues, water-cut increase, flowline blockage, wellhead valve malfunction, scale formation, and nonflowing wells. A detection rate of at least 88% and an accuracy of at least 80% were achieved for all the constraint categories for which the models were built. Most labeled events were detected successfully, and the models also produced several new events in the historical data that were not documented through regular manual analysis, later verified to be genuine well constraints. The underperformance conditions in the study were identified as small changes in well behavior that occur through time and are difficult to detect with a non-digital process, even in trained teams, with human surveillance limitations and errors. The system drastically reduced the response time for taking action in the field, giving operators considerable reduction in well downtime and hydrocarbon production. Several machine learning classification models were evaluated, including regression-based and tree-based techniques to detect real-time operational constraints. A logistic classification model was selected for its strength in interpretability, feature evaluation using statistical confidence, real-time execution efficiency, and robust model implementation. Clues in data patterns assisted in the data labeling process by analyzing the historical time-series data and iteratively verifying with the subject matter experts. Feature engineering techniques were used in both time and frequency domains in a machine learning technique that generates output in terms of event probabilities.