Abstract
A method for predicting torque and drag has been developed which is based on statistical learning models and which utilizes parameters routinely measured while drilling. Real time prediction of torque and drag during drilling has particular significance in on-the-fly rig floor operational modifications and in overall drilling optimization. These statistical methods are readily incorporated into real time drilling optimization procedures, first through a passive diagnostic tool, and subsequently through an integrated real-time control loop.
This paper presents statistical learning techniques such as bootstrapped regression, random forests, and support vector machines for prediction of downhole torque. An analysis of the aforementioned methods is carried out to evaluate computational efficiency of the processes. The statistical package R has been employed for processing and visualization of data and for displaying results. A comparison is made to real time ROP prediction carried out in previous work, to assess the relationship between predicted torque and ROP. A method is proposed to use classification techniques for predicting whether drilling is proceeding at the optimal level.
Torque values were accurately predicted with all three methods discussed in this paper. However, scalability of the methods must be considered, owing to the increased computational efficiency that is required. Computationally efficient and robust systems are of high priority for field applicability. To that end, ensemble techniques were analyzed along with the aforementioned methods and runtimes were evaluated for all of the methods employed. Results show that this method can be incorporated into real time drilling situations, with attending low runtimes. Development of more accurate methods will require additional computational power and the use of ensemble techniques. Classification proves to be a good technique whenever optimization is required, as indicated by real time predictions. Classification techniques illustrated in this paper explore the viability of various options for correcting undesirable drilling performance.
This paper presents state of the art statistical learning techniques for predicting torque and drag during drilling. Previous torque and drag calculations using finite element models provide a good cross reference with this statistical model for indicating operational anomalies. Unexpected values of torque may forecast imminent drilling problems. These methods can be utilized for prediction of drilling problems and proactive methods for avoiding them, and applied in optimized drilling procedures.
