Comparing Drilling Anomaly Prediction by Purely Data-Driven and Hybrid Analysis Methods - Case Study of Utah FORGE Geothermal Wells

Abstract

Abstract Automatic prediction of drilling incidents can be conducted through either a purely data-driven approach or a hybrid approach. In the first approach, the variable space is typically limited to surface measurements and downhole sensor data, while in the second approach, the variable space is expanded to include information from physics-based models. This paper analyzes the additional value of incorporating physics-based information to predict drilling incidents such as stuck pipe, illustrated using data from the Utah FORGE geothermal wells. In our study, we trained three anomaly detection models with two distinct variables spaces. In the first one, we considered the real-time signals only, while in the second one, we included physics-based information derived from cuttings-transport, tortuosity, and torque-and-drag models. We selected three models that showed promising results in recent studies and represent the taxonomy of machine-learning-based anomaly detection algorithms. Specifically, we utilized recurrent neural networks, autoencoders, and clustering. Finally, a comparison between the two approaches was performed in terms of the fidelity of the warnings they generated. We observed that the inclusion of physics-based information is key to improving the performance of models for predicting drilling incidents. Specifically, we noted a reduction in the number of false alarms, which, in turn, increases the reliability of the models. In addition, we found that physics information can guide the selection of prediction time windows when drilling anomalies develop, thereby eliminating bias in the models' construction. Finally, we observed that some drilling anomalies, which were previously believed to occur suddenly with little warning, can, in fact, be predicted in a timely manner with hybrid models. These observations demonstrate that the use of hybrid models can significantly increase the performance of drilling anomaly predictions, providing sufficient forewarning time for their prevention and associated NPT avoidance. State-of-the-art methods that implement purely data-driven and hybrid approaches have individually demonstrated high accuracy in predicting incidents on specific datasets. However, no previous comparative study has been conducted to analyze the value of incorporating physics-based information. This paper is the first to perform such an analysis for models aiming at the early detection of drilling anomalies. The results from this study provide valuable guidance for future NPT avoidance in drilling operations.