Transforming Oil Well Drilling: Prediction of Real-Time Rate of Penetration With Novel Machine Learning Approach in Varied Lithological Formations

Abstract

The rate of penetration (ROP) is crucial for efficient and cost-effective oil well drilling. This study introduces a novel prediction method for the rate of penetration that pioneers the use of different types of drill bits and lithologies with traditional drilling parameters. Utilizing a comprehensive dataset from 12 diverse wells, it employs advanced machine learning techniques including an adaptive moment estimation-based artificial neural network for developing the algorithm. By integrating various controllable and uncontrollable drilling parameters, the random forest, decision tree, and K-nearest neighbor models demonstrated superior performance. These models achieved a coefficient of determination of approximately 98% and a mean absolute percentage error of only 3.30%, outperforming traditional models such as Maurer and Bingham, as well as other machine learning models. Using 500 testing and 2000 training data points from real-time measurements reduced the risk of overfitting and enhanced model effectiveness in different drilling environments. The predictions of the developed model can modify the input parameters to increase rate of penetration through various formations. This study highlights the importance of lithology and utilizes feature ablation analysis to transition from a black-to-white box model. Additionally, based on the predictions of this work, post-drilling analysis can reduce costs and time by only requiring surface-measured parameters and eliminates the need for extensive study on geological, laboratory, and drilling data prior to drilling activities. This integrated approach sets new standards for machine learning in drilling, representing a robust and adaptive strategy to enhance operational efficiency.