A New Rate of Penetration Prediction Model Based on the XGBoost and Explanation Method

Abstract

ABSTRACT: Machine Learning (ML) studies are carried on in the hydrocarbon exploration and production. The Rate of penetration (ROP) is one of the investigations related to ML. Previous studies could not fit the ultra-deep well and most of machine learning ROP prediction are black box models lacking of enough explanation. ML ROP prediction need more accurate ROP models. In this paper a new and reliable calculation method of ROP prediction for well drilling is proposed by Extreme Gradient Boosting (XGBoost) algorithm. It is compared with Random Forest Regression algorithm on ROP prediction models. According to the importance ranking, rotating torque is the most impacted factor in this dataset. The conclusion is that ROP prediction model based on XGBoost has smaller prediction mean square error and shows higher efficiency than Random Forest, showing the superiority of the XGBoost. Considering the SHAP value ranking, torque and RPM are the most important features and they are positively impacted ROP in well drilling ROP model established by XGBoost machine learning method can make reasonable use of drilling parameters and provide reference for well drilling optimization. ROP model optimization derived from XGBoost can extremely reduce the expenses by reducing the drilling time. 1. INTRODUCTION Hydrocarbons, including oil and gas, are often stored in sedimentary rocks in deep formations (Yang et al. 2017). wells are drilled both onshore and offshore with more drilling cost (Ma, Ping, and Jian 2016). One of the most important factors is the Rate of Penetration (ROP) which affect the drilling efficiency and cost (Eskandarian, Bahrami, and Kazemi 2017). It is difficult to enhance ROP for ultra-deep well drilling as the well depth is increased. Increasing ROP is an approach of drilling optimization and cutting costs by reducing the drilling time. Existing ML models considering less parameters and data points cannot meet current ultra-deep well developing demands. In order to cut costs and increase the amount of drilling operations, finding out the relationship between the ROP and parameters is very important. Make it clear that variables and their relationships with ROP are the main factors which is helpful to promote ROP. Traditionally instantaneous ROP refers to the drilling footage in time unit, like the data used in this work. And some researchers want to estabilish some models of ROP using drilling parameters. These models can be used to make real time drilling optimization by equipment design guidance and drilling parameters selection.