Data Analytics Techniques for Stimulation Treatment Performance Prediction/ Classification

Abstract While academic, the paper provides an overview of several data analytics techniques for the prediction or classification of the success, or failure, of any types of stimulation treatments performed for productivity improvement. The premises of the research consist of comparing the performance of five machine learning approaches to predict the production response, or classify the wells in poor, maybe, good, and very good candidates, to enable the ranking and execution of the best stimulation candidates. The study is flexible and can be applied to any type of stimulation treatments, such as cyclic steaming, chemical well cleaning, small frac-packs, and/or acid jobs. A database of stimulations was aggregated using public information available on the "California Department of Conservation" from a heavy oil reservoir located in Santa Maria County. The five technologies tested in this research, namely Logistic Regression, Decision Tree, Random Forest, Neural Networks, and XGBoost, which are relatively common in the machine learning domain, provided relatively different performance capabilities. The models used the same training data set, and the validation was performed using a blind dataset for comparison.