Hybrid Approach Using Physical Insights and Data Science for Stuck-Pipe Prediction

Early detection of stuck-pipe incidents is crucial because of the enormous costs of recovering from such incidents. Previous studies have leaned significantly toward a physics-based or data science approach. However, both approaches have challenges, such as the uncertainty of the physics-based model and the lack of data in the data science approach. In this study, we propose a hybrid approach using physical insights and data science and discuss the possibility of stuck-pipe prediction. The proposed method comprises two steps. In the first step, a data-driven model with physical insights is trained using the historical data of the in-situ well to estimate some of the drilling variables. In the second step, the risk of stuck-pipe occurrence (hereafter referred to as sticking risk) is calculated based on the historical and current measured data and the estimation of the trained model. This approach is expected to overcome the limitations of the previous methods as it allows the construction of a detection model tuned to the in-situ well. In the case studies, models for estimating the topdrive torque and standpipe pressure were constructed. The performance of the models is discussed using actual drilling data from drilling fields, including 21 stuck-pipe incidents during drilling operations. The proposed method was first examined using short-term output. The output confirmed that the sticking risk increased shortly (up to 20 hours) before the stuck-pipe incident occurred in 15 cases. This increase in sticking risk was consistent with physical considerations. Subsequently, this study examined the long-term output over several months; this was rarely done in previous studies. Even within this long-term output, some cases had good performance with only a few false alarms, while others had problems with many false alarms. For cases of low performance, several model improvements, such as adding mud information or making the learning and threshold-setting methods more robust to outliers, were found to have the potential to improve performance. The novelty of our research lies in creating a broad framework for the stuck-pipe prediction by using both physical insights and data science methods. The proposed hybrid approach demonstrated the potential to reduce false alarms and improve interpretability compared with previous methods. The framework is highly extensible, and further performance improvements can be expected in the future.