Use of data analytics has become pervasive to optimize results and increase the efficiency of operations. Similarly data obtained during drilling can be used to predict rate of penetration (ROP) using statistical and machine learning techniques. However, one of the most basic forms of statistical analysis i.e. multivariate regression can be used for inference to improve the computational efficiency of more complicated models.
Multivariate regression (MVR) has been applied to a field data set to predict the rate of penetration (ROP) based on features of the data. While MVR works as a predictor, its use as an inferential technique shows more promise. Regression techniques have been combined with bootstrapping techniques to analyze residuals when data set features are plotted against ROP. Analysis of these graphs provides insight into the non-linearity in the relationships among the data. Bootstrapping the residuals also yields a good predictor. The regression model has also been subjected to the bootstrap technique separately to observe its behavior on predictions.
Using the aforementioned techniques, features with both linear and non-linear relationships with ROP can be identified. This saves computational power when in investigating real time applications of the data. Features having a linear relationship with ROP can be modeled with linear techniques and the rest using non-linear techniques. These concepts have been implemented in this paper using general additive models to yield excellent accuracy. More complicated models employ ensemble techniques which can be computationally robust and provide excellent accuracy. When ensemble techniques are used, identifying the data associated with each method can be crucial to efficiency of the technique. The data used in an ensemble technique for prediction illustrates the important role that this technique can play in reducing error as well as increasing computational efficiency. The model developed using linear techniques with bootstrapped residuals also gives a low error rate in prediction results.
Using linear methods for determining the nature of relationships within the data is important in developing computationally robust models. Using this as an inferential technique can serve as a key component of data exploration in any operation. Moreover, these methods can be a stepping stone to improving the use of ensemble techniques in real time drilling operations.