A Pipe Sticking Prediction Method Based on Random Forest

Abstract

This study aims to address the challenge of predicting pipe sticking accidents in petroleum drilling engineering. In response to the high blindness of tradi-tional empirical qualitative operations, the difficulty of mathematical models in accurately reflecting drilling laws, and their poor operability, machine learning methods such as neural networks (NNs), support vector machines (SVM), and random forests are introduced for pipe sticking prediction re-search. Adhering to the principles of sample representativeness, category balance, and diversity, 9 drilling parameters including depth, drilling speed, weight on bit (WOB), torque, friction coefficient, outlet temperature, rota-tional speed, inlet flow rate, and outlet flow rate were selected. Drilling data from 30 wells in the Yan'an area were collected. Measurement data within a certain period before pipe sticking were used as pipe sticking samples, which were cross-arranged with normal samples to construct the training dataset. By comparing the performance of SVM (linear kernel, radial basis kernel) and random forest models, the results show that the random forest performs the best. In its confusion matrix, the number of correctly predicted samples for true labels 0 and 1 is 190 and 207 respectively, with only 9 and 5 misclassifi-cations. The accuracy rate reaches 0.966. In summary, this study confirms that the random forest has significant advantages in balancing "accuracy and efficiency" in pipe sticking prediction tasks, providing a reliable early warning model and technical support for safe drilling operations.