Enhanced Rate of Penetration Prediction with Rock Drillability Constraints: A Machine Learning Approach

ABSTRACT: The rate of penetration (ROP) is directly related to the drilling cost. Accurate prediction of the ROP is of great significance to reduce the drilling cost. At present, there are two main types of methods for predicting ROP. One is based on physical mechanism analysis. However, due to many kinds of parameters affecting ROP, the prediction results are often very different from the actual ROP. The other is the data-driven prediction model of ROP. For the sample data with no obvious characteristics or "small sample" data, compared with the actual results, the prediction results of data-driven model have large errors. To solve the above problems, a new machine learning ROP prediction model based on physical mechanism constraints is proposed in this paper. Based on the data of 4 wells in X block of T oil field, the random forest model is selected from the three machine models including support vector regression, random forest, and regression tree, and the rock drillability constraint is added to predict the ROP. The prediction results that do not meet the rock drillability constraints are optimized to reduce the prediction error. The results show that the random forest model with constraints has the best performance, and the prediction error is reduced from 14.52% to 11.27%. 1. INTRODUCTION ROP prediction methods can be divided into two categories: one is the physical model based on parameter relationships, and the other is the machine learning model based on big data (Gan, 2019). Maurer (1962) proposed a model for predicting the ROP of rolling cutting tooth bit based on WOB, formation drillability, rock compressive strength, and bit size. Bingham (1965) introduced the weight index based on Maurer’s model to predict the ROP. Bourgoyne and Young (1974) considered the formation strength, formation depth, compactness, bottom hole differential pressure, bit size, bit weight, bit wear, string ROP, and bit hydraulics, and established a multiple regression ROP prediction model. Warren (1987) gave the prediction model of cone bit penetration rate based on rock compressive strength, string rate of penetration, torque, and bit size. Hareland et al. (2010) established the prediction model of ROP by considering the physical number of insertions in contact with rock, the physical number of insertions through rock, the cutting angle of formation, and the ultimate strength of rock. The ROP prediction model based on the mechanism model takes into account the limited parameters affecting the ROP, and the mechanism relationship between each parameter and the ROP is mostly given through experimental assumptions. Therefore, the accuracy of predicting the ROP based on the physical model is low.