Real-Time Torque and Drag Prediction in Oilwell Drilling: A Comparative Study of Machine Learning Models

This study focuses on evaluating some of the most common machine learning models for real-time Hookload prediction in oilwell drilling operations, namely, Linear regression, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). The dataset, derived from a deviated geothermal well drilled in Utah, USA, served as the basis for comparative analyses. Among the models tested, Linear regression emerged as the most effective for real-time Hookload prediction, showcasing superior accuracy, durability, and swift training times, making it an optimal choice for practical implementation. The study underscores the importance of the number of future steps in anomaly detection, revealing that predictions farther into the future enhance the model’s ability to identify drilling anomalies. The findings strongly advocate for the utilization of the Linear regression model for real-time Hookload prediction in oilwell drilling, provided that it is fed a sufficient amount of training data. The study recommends predicting at least 25 future steps to optimize anomaly detection in this specific scenario, offering valuable insights for enhancing drilling operation efficiency and safety. Considering that actual data points are used as inputs for making future predictions, using fewer prediction steps than 25 could lead to masking drilling anomalies.