Enhancing Hydrocarbon Recovery Factor Prediction Using Ensemble Machine Learning Workflow

This paper seeks to develop a machine learning workflow for accurately predicting hydrocarbon recovery factor (RF), a crucial property for exploration and production strategies.The study employed agnostic method and Bayesian Optimized Decision Tree (DT) model to assess causal relationships between input and target variables.The decision tree model yielded the lowest error metrics (0.1 MAE and 0.132 RMSE).To determine how much each input reservoir parameter contributes to the target (RF), the permutation feature importance and Shapley Value metrics were used to analyze the input features.The decision tree model was optimized and trained using the chosen parameters, yielding better results of 0.02 MAE and 0.03 RMSE.The features selected by Shapley values greatly improved the predictive model outcome because it demonstrated a causal relationship with recovery factor estimation.The proposed method outperformed several reported models, demonstrating the value of machine learning in petroleum engineering for reservoir characterization.Improved data pre-processing, analysis, and visualization techniques were also utilized in this study.