Creating Chan water control diagnostic plots is a common well surveillance activity to search for signatures that distinguish and explain mechanisms behind excessive water production in oil wells. The technique involves an engineer who visually classifies patterns or signatures related to a water production mechanism. This study shows how the Chan plot signature identification can be approached as a machine learning (ML) classification problem, where a well can be characterized by the slopes of water-oil ratio (WOR) and WOR time derivative (WOR’) curves. A model tries to find the pattern category to which that well belongs. Having ML models that can predict whether a well belongs to a specific Chan plot signature, or pattern, would be valuable as a well surveillance tool, especially in high-well-count fields. Our previous work focused on using the shape of the Chan plot as features for a radial basis function (RBF) support vector machines (SVM) model. In this study, we examine how features to identify Chan plot signatures can be simplified and how different ML models compare in accuracy. ML models used in this study were: nearest neighbor, SVM, decision tree, random forest, logistic regression, and Naive Bayes. In this study, we use the slopes of WOR and WOR’ as features. As a result, we observed an increase in the accuracy of the ML models that we used. By performing the quality check on the data set after selecting slopes as features, we identified that the dataset contained several incorrectly labeled examples, which we adjusted before we trained the ML models. By comparing the models’ metrics in the context of the test set, we identified that the ML model with the highest f1-score was nearest neighbor at 0.93, whereas the RBF SVM model achieved a value of 0.90. We also compared models’ decision boundaries to find how they differ among all ML models. We obtained an improved accuracy of an ML model by simplifying features as well as raising the quality of data used in the Chan plot signature identification problem. These ML models could be useful in automatic classification whether a well exhibits a specific Chan plot signature, to flag it for a review within a broader petroleum engineering decision framework.
Petroleum engineers widely use Chan water control diagnostic plots to visually examine patterns for mechanisms behind excessive water production in petroleum wells. Distinct signatures reveal constant water-oil ratio (WOR), normal displacement of oil by water, multilayer channeling, and rapid channeling. Visual diagnosis requires extensive practical experience. High well counts amplify the need for timely relevant solutions. This study presents a supervised machine learning (ML) technique, support vector machine (SVM), to automatically identify Chan diagnostic patterns for timely detection and control of excess water production. The project involved publicly available production data. First, we performed manual identification of different Chan plot patterns and split the data set for training, cross-validation, and testing. Next, we normalized each subset, balanced the number of samples per pattern, trained the SVM model, and evaluated its ability on the test data set. The final stage involved model parameter tuning to explore opportunities to improve the model accuracy. When collecting representative samples, we observed uneven sample counts for pattern type subsets. Balancing the training data set was necessary to ensure that the model could generalize well for previously unseen production data. To ensure a relevant comparison among wells within each pattern type, we performed time- and WOR-based normalization. We confirmed that the SVM model was highly accurate across diverse fields and hence did not exhibit a locality bias. Operator companies can harness the prior classification of water control trends by experienced engineers. As more wells start exhibiting specific pattern signatures, the algorithm helps to proactively identify them. The approach maximizes the effectiveness of water production analysis and helps to reinforce technical quality and continuity among engineering personnel. The study demonstrated an application of SVM to the classification task of identifying water control diagnostic plot signatures across different fields. We proposed stricter visual definitions of signature types. Patterns such as normal displacement and multilayer channeling are extremely similar; therefore, we presented feature engineering required to increase classification accuracy in similar edge cases. With various kernels, the described SVM methodology can help production engineers to proactively perform surveillance activities that require identifying signature patterns in Chan plots.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
customersupport@researchsolutions.com
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.