Stuck pipe events continue to be a major cause of non-productive time (NPT) in well construction operations. Considerable efforts have been made in the past to construct prediction models and early warning systems to prevent them—a trend that has intensified in recent years with the increased accessibility of artificial intelligence tools. This paper presents a comprehensive review of existing models and early-warning systems and proposes guidelines for future improvements.
In this paper, we review existing prediction approaches for their merits and shortcomings, investigating five key aspects: (1) the frequency and spatial bias of the data with which the models are constructed, (2) the selection of the variable space, (3) the modeling approach, (4) the assessment of the model's performance, and (5) the model's facility to provide intuitive and interpretable outputs. The analysis of these aspects is combined with advancements in anomaly detection across other relevant domains, such as the internet of things, to construct guidelines for improvement of real-time stuck pipe prediction.
Existing solutions for stuck pipe prediction face numerous challenges, allowing this problem area to remain a missing component in the broad scope of progressive drilling automation. In our analysis, we looked at notable approaches, including decentralized sticking prediction, sophisticated data-driven models coupled with explanation tools, and data-driven models coupled with physics-based simulations (hybrid sticking predictors). However, even these sophisticated approaches face challenges associated with: general, non-specific applicability; robustness; and interpretability. While the best approaches tackle some of these challenges, they often fail to address all of them simultaneously. Furthermore, we found that there is no standardized method for assessing these models’ performance or for conducting comparative studies. This lack of standardization leads to an unclear ranking of (the merits and shortcomings of) existing prediction models. Lastly, we encountered cases where unavailable information, i.e., information that would not be available when the model is deployed in the field for actual stuck pipe prediction, was employed in the models’ construction phase (we will refer to this as "data leakage"). These findings, along with good practices in anomaly detection, are compiled in terms of guidelines for the construction of improved stuck pipe prediction approaches.
This paper is the first to comprehensively analyze existing methods for stuck pipe prediction and provide guidelines for future improvement to arrive at more universally applicable, real-time, robust and interpretable stuck pipe prediction. Moreover, the application of these guidelines is not limited to stuck pipe, and can be used for predictive modeling of other types of drilling abnormalities, such as lost circulation, drilling dysfunctions, etc. Additionally, these guidelines can be leveraged in any drilling application, whether it is for oil and gas recovery, geothermal energy, carbon storage, etc.
