Deep learning approach to prediction of drill-bit torque in directional drilling sliding mode: Energy saving

Abstract

Directional drilling, a sophisticated well-drilling technique, enables precise wellbore navigation toward inaccessible reservoirs via vertical wells while optimizing hydrocarbon recovery and minimizing environmental impact. This method encounters significant challenges in managing torque and drag, particularly during sliding mode, where the drill string remains stationary while the bit rotates, causing unpredictable torque fluctuations. Traditional torque measurement methods, relying on downhole sensors, are often costly and complex. This research examines advanced machine learning (ML) models that leverage commonly available drilling data to predict drill-bit torque during the sliding mode, thus removing the reliance on expensive sensors. The study presents an innovative approach using Deep Auto-Regressive Network (DARN) and Deep Neural Network (DNN) models, specifically designed to predict torque based on directional drilling parameters like Weight on Bit (WOB), Revolutions Per Minute (RPM), and Standpipe Pressure. Using a dataset of 2,746 data rows from four directionally drilled wells in a Middle Eastern oil field, encompassing scenarios such as casing milling, opening the sidetrack drilling window, and navigating various trajectory sections with different build rates and hold intervals to predict drill-bit torque (TQ), these models were trained and evaluated against Support Vector Machine (SVM) and Decision Tree (DT) benchmarks. Results indicate that DARN achieved superior accuracy with an RMSE of 49.6 and an R² of 0.9986, outperforming other models due to its ability to capture complex temporal dependencies. This predictive model facilitates real-time, cost-effective torque management, significantly enhancing operational efficiency in sliding mode directional drilling.