Application of statistical learning models to predict and optimize rate of penetration of drilling

Modeling the rate of penetration of the drill bit has been essential to optimizing drilling operations. Optimization of drilling – a cost intensive operation in the oil and gas industry– is essential, especially during downturns in the oil and gas industry. This thesis evaluates the use of statistical learning models to predict and optimize ROP in drilling operations. Statistical Learning Models can range from simple models (linear regression) to complex models (random forests). A range of statistical learning models have been evaluated in this thesis in order to determine an optimum method for prediction of rate of penetration (ROP) in drilling. Linear techniques such as regression have been used to predict ROP. Special linear regression models such as lasso and ridge regression have been evaluated. Dimension reduction techniques like principal components regression are evaluated for ROP prediction. Non-linear algorithms like trees have been introduced to address the low accuracy of linear models. Trees suffer from low accuracy and high variance. Trees are bootstrapped and averaged to create the random forests algorithm. Random forests algorithm is a powerful algorithm which predicts ROP with high accuracy. A parametric study was used to determine the ideal training sets for ROP prediction. It was conclude that data within a formation forms the best training set for ROP prediction. Parametric analysis of the length of the training set revealed that 20% of the formation interval depth was enough to train an accurate predictor for ROP.
The ROP model built using statistical learning models were then used as an equation to optimize ROP. An optimization algorithm was used to compute ideal values of input feature to improve ROP in the test set. Surface controllable input features were varied in an effort to improve ROP. ROP was improved to save a predicted total of 22 hours of active drilling time using this method. This thesis introduces statistical learning techniques for predicting and optimizing ROP during drilling. These methods use input data to model ROP. Input features (surface parameters which are controllable on the rig) are then changed to optimize ROP. This methodology can be utilized for reducing nonproductive time (NPT) in drilling, and applied to optimize drilling procedures.