Optimizing Models for Predicting Torque on Bit Using Data From the Volve Field in Norway

Abstract

Torsional oscillations can cause severe damage to downhole tools and may result in expensive fishing and sidetracking operations. The drilling industry is aware of this problem and still looking for suitable solutions to determine the drivers of the oscillations, and to quantify their effects. The mitigation of this problem requires a detailed knowledge of the parameters controlling the drilling process. Nowadays, modeling is a useful tool for describing the downhole processes using the stream of data acquired from the sensors installed in the drilling equipment. This paper focuses on torque on bit which is directly connected with torsional oscillations. The model generation is performed, either by fitting empirical models with measured data or by creating new machine learning models. Five empirical literature models are parametrized using the optimization module of the Python library SciPy. Machine learning models are generated using Scikit-learn with measurement data from the Volve field in Norway. For the current testing dataset Random Forest showed the highest accuracy with a R2-score of 0.767. Other machine learning algorithms showed a comparable accuracy. However, empirical models failed to achieve reliable results. In future, the generated models can be used to optimize drilling parameters to prevent technical drilling problems.