Prediction of rate of penetration in directional drilling using data mining techniques

Rate of penetration (ROP) represents drilling speed and its productive time during drilling operations in oil and gas wells. A predictive model that links ROP to its influential parameters is essential to optimize ROP for minimizing drilling costs. This study implements a comprehensive data mining approach utilizing Python toolboxes to improve ROP prediction in directional wells, which has not been addressed as much as vertical wells with respect to the downhole weight on the bit (WOB) and cutting transport. To do so, seven functions, including influential parameters, were identified to impact ROP in directional drilling. Drilling data of seven directional wells from an offshore rig in a gas field was compiled to set up the input dataset. The data preprocessing methods, consisting of the modified Z-score and Savitzky-Golay (SG) smoothing filter, were utilized to remove outliers and reduce the noises in the input dataset. Multilayer perceptron (MLP) neural network and random forest regression models were employed comparatively to predict ROP, and their architectures were designed by tuning hyperparameters of the models. The models' accuracy was statistically and graphically assessed by using the K-fold cross-validation and statistical metrics. The random forest model was demonstrated to be superior to the MLP neural network model in terms of accuracy and speed. The results represent that using calculated downhole WOB instead of measured surface WOB in the input dataset reinforces the models' accuracy in the prediction of ROP. Statistical investigations such as partial correlation, mutual information, and permutation feature importance revealed that the cutting transport function can affect ROP in directional drilling as significantly as other influential parameters, which has not been mainly accounted for in the literature when establishing models for ROP prediction.