Real-Time Modeling of Drill String Torque

Abstract

Onshore and offshore rotary drilling suffers from stick-slip vibrations, a severe mechanical issue that damages equipment and slows operations. Current downhole sensors transmit data too slowly to predict and proactively avoid those events. This paper presents a deep learning approach to predict drill string torque before damage occurs. Using data from the OpenLab drilling simulator, a Long Short-Term Memory (LSTM) neural network was trained to forecast future torque at the drill bit. Relying on only three available sensor inputs, the model predicts torque for 1, 5, and 10 seconds into the future. Model accuracy was evaluated using the coefficient of determination (R^2). Results demonstrate high accuracy for 1-second (R^2 = 0.974) and 5-second (R^2 = 0.868) forecasts, offering warning window for automated surface controllers to adjust and avoid those instances. While 10-second predictions show some error (R^2 = 0.696), this method proves that proactive real-time vibration control is highly effective even with limited sensor data.