Data-Driven Models for Predicting Rate of Penetration Based on Machine Learning Algorithms

ABSTRACT Data-driven models are used extensively for predicting rate of penetration (ROP). However, what data-driven algorithm is best suited to ROP prediction is currently undecided. In this paper, the data-driven model based on back propagation neural network (BP-ANN) and random forest (RF) algorithms are proposed respectively to predict ROP. The features that include both the engineering and formation parameters are selected as the model inputs by combining physical drilling laws and correlation analysis. The optimal hyperparameter combinations of the model are found by cross-validation. The MAE, MSE and R2 are adopted as the indicators to evaluate the model performance. The case study illustrates how accurately and rapidly the data-driven model can be used to predict the ROP. The results indicate that the RF model can track the data-based ROPs more accurate even if hyperparameters optimization is ignored. The MAE, MSE and R2 of the optimized model are 0.243m/h, 1.599m2/h2, 0.989, respectively. While for the BP-ANN model, the predicted ROPs can achieve a more desired result after hyperparameters optimization, but it still cannot come close to the result of RF model. The present model provides some methodological bases for the further study on optimizing drilling parameters. INTRODUCTION Rate of penetration (ROP) is one of the key factors affecting drilling period and operating cost. Over the past few decades, researchers in drilling engineering field have been devoted to accurately predict and obtain an optimal value of ROP. Several physics-based models, e.g. the Maurer model (Maurer, 1962), the Warren model (Warren, 1987), the B-Y model (Bourgoyne and Young, 1974), the Hareland model (Hareland and Rampersad, 1994), the Detournay model (Detournay et al., 2008), the Motahhari model (Motahhari et al., 2010), are therefore proposed to predict ROP. All of these models take into account the engineering and geological factors to various extent, however, it still exists a great many limits in itself, especially in the extremely complex downhole environments. The primary reason is that the strong nonlinear relations between the ROP and the governing factors are not thoroughly understood (Zhang et al., 2022).