Machine Learning Application for Wellbore Damage Removal in the Wilmington Field

Abstract

Abstract The use of acid is an important well maintenance tool in removing near wellbore damage to restore a reservoir’s natural permeability and represents one of the most economic options in managing base decline. The selection of acid maintenance candidates however can be a complex process, particularly in wells completed across multiple sands, involving many factors both on the surface and subsurface. As a consequence, individual acid maintenance jobs have had a mixed success rate historically, with certain jobs resulting in a lack of response, or worse, higher water production rates and equipment failures. This paper uses the Wilmington Oil Field, located in southern California, as a case study to examine the typical characteristics associated with low volume acid maintenance success and provides a novel approach using machine learning (ML) algorithms to aid in the screening and selection of future candidates. The developed algorithms, which make use of the open-source statistical software R, is trained based on results from over 500 producer and 3900 injector acid maintenance jobs that were executed at the field and incorporates predictors from the following groupings determined from literature and subject matter experts (SMEs): Production/injection history, Reservoir properties, Acid type and volume, Delivery mechanism, Formation damage, Well completion design, and Surface facility properties. Over 100+ predictor variables were compiled and screened using supervised feature selection to identify those variables providing the greatest explanatory power. A series of machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Machines (GBM)) were then used in a classification task to successfully predict whether a producer acid maintenance job would be economic. The logistic regression model ultimately yielded the best classification results (71% prediction accuracy for the producer jobs and 77% for the injector jobs) and proved to be the ML algorithm with the best balance of accuracy, interpretability, and ease of implementation in the field. The model robustness is examined by applying the algorithm outside of the training and test datasets, to acid maintenance jobs executed in 2016-17 and shows similar predictive accuracy. As a result the model is being actively used to automatically screen for treatment candidates among all 700+ producers and 400+ injectors in the Wilmington Field, which are then validated by SMEs before being executed. The overall process has resulted in significant cost savings by both improving the performance of the acid maintenance program and greatly reducing the amount of time spent by technical staff in selecting candidates. These results indicate that ML algorithms can be effective analytical tools not only for ‘big data’ problems (i.e. large-n, time series datasets) which are featured heavily in industry literature, but also for smaller datasets thus opening up a variety of potential applications that can be deployed by surveillance teams alongside traditional approaches.