A guide for prompt rate of penetration prediction using machine learning models with comparative analysis

Abstract

The rate of penetration (ROP) is a critical parameter in drilling which it reflects drilling time, efficiency, and costs. Several approaches have been investigated to handle complex and non-linear drilling parameters for precise prediction for this parameter. Machine Learning remains a promising tool for data analysis and for predicting drilling parameters with acceptable accuracy. However, most of ML algorithms are mainly coded using Python, MATLAB, or R, which require proficiency in programing skills to perform this task. In order to reach an easy, prompt, and precise results, this study conducted Dataiku Data Science Studio (DSS) ML software for predicting ROP through in-memory ML algorithms only, without using advanced or hybrid technologies. The case study was being conducted in an offshore directional gas field in the Nile delta, Egypt. The ML models used in this study are Extreme Gradient Boosted Machine (XGBoost), Gradient Boosted Trees (GBT), K-Nearest Neighbors (K-NN), and Support Vector Machine (SVM). The evaluation metrics used in this study are R2 score, Correlation Coefficient (CC), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). All models’ performance has been validated through learning curves and errors distribution analysis. The output results showed very strong overall performance, and demonstrated that XGBoost model performed best testing scores with R2 = 0.965, RMSE = 1.339, followed by GBT with R2 = 0.936, RMSE = 1.809, then K-NN with R2 = 0.849, RMSE = 2.78, then SVM achieved least performance with R2 = 0.801 and RMSE = 3.586. The study further validates feature importance analysis and SHAP values interpretation to identify dominant features governing ROP prediction.