Optimizing Shale Assets: A Hybrid Approach Combining Classification and Prescriptive Analytics to Boost Well Performance in the Eagle Ford Basin

Abstract This Project seeks to explain the characteristics and variables that differentiate good wells from bad wells in the Eagle Ford Basin using Net Present Value (NPV) as a classifier. The first part of the project comprises of building a robust classifier that can differentiate good wells from bad wells. The second part of the project is the development of different classification models, variable selection and feature engineering, and the selection of the best performing model. The third part of the project entails identifying features that impact NPV. Lastly, the model is used to characterized what engineering variables best improve the NPV of deficient wells. The dataset was cleaned and processed for missing values and outliers. The variables in the dataset were divided into 3 main classes. The first class is the target or dependent variable which is the NPV. Geological independent variables which cannot be controlled or engineered is the second class of variables, and engineering independent variables which can be controlled is the last. A classifier was created using a simple rule and tested to be balanced. Different classification algorithms, Generative, Discriminative and Ensemble was applied to the dataset and the model with the best classification score was selected. Prescriptive analytics was carried out with the selected model on the wells with negative NPV to determine the optimum engineering variables that can make these wells positive. Random Forest, an Ensemble model, outperformed all the other models. For most of the models, the location of the well and TVD (geological variables), the proppant volume & lateral length (engineering variables) strongly influences well NPV classification. This methodology serves as a quick way for oil and gas operators to strongly identify variables that affect NPV and also to select the optimum engineering values which can complement other established methods such as Linear Regression Analysis, Rate Transient Analysis or Reservoir Simulation on a fairly large dataset.