Assessment of Machine Learning Techniques for Real-Time Prediction of Equivalent Circulating Density

The equivalent circulating density (ECD) is crucial in avoiding fluid losses or kicks while drilling. It's more critical in wells where the pore pressure gradient is close to the fracture pressure gradient. The conservation of mass and momentum determine the ECD, but this method does not account for other factors like torque, rotating speed, weight on bit, etc. These may affect the ECD directly or indirectly. The aim of this study is a practicality to predict the ECD using various machine learning techniques and to determine their effectiveness. The complete drilling dataset of an oil well from Texas was acquired. Over 16000 data points were obtained after the removal of the null values. The data was prepared by scaling it and conducting principal component analysis (PCA). PCA reduced the dimensionality of the dataset while retaining the information. Following this, five different machine learning regression techniques were used to predict the equivalent circulation density, namely, XGBoost, Random Forest, Support Vector Machine, Decision Tree, and Elastic net regression. The performance of these techniques was judged by comparing their R2 scores, mean squared errors (MSE), and root mean squared errors (RMSE). The results showed that ECD prediction through all the above machine learning techniques is a vital reality. Random forest regression emerged superior from the different methods used, illustrating the highest R2 score and the lowest MSE and RMSE. Its R2 for our model was 0.992, which is an excellent fit. It was followed by SVM, which had the second-lowest RMSE and an R2 of 0.987, close to the random forest technique. Elastic Net, Decision tree, and XG Boost in the respective order were at the bottom of the pool. Machine learning is a powerful tool at our disposal to effectively predict quantities in real-time that directly or indirectly depend on several parameters. It can even be effective when no direct correlation between the quantities is known. Thus, machine learning can significantly enhance our ability to optimize drilling operations by having quicker and more accurate predictions. The work shown in this study, if implemented, can provide the crew more time to respond to situations such as the occurrence of kicks and thus will lead to safer operations.