Data Integration Enabling Advanced Machine Learning ROP Predictions and its Applications

Abstract

Abstract Predicting and optimizing the rate of penetration (ROP) is a crucial part of drilling optimization. While numerous machine learningbased ROP prediction methods exist, their practical utilization remains underexplored. This paper introduces an innovative approach, employing data integration from different sources with multiple machine learning algorithms for accurate ROP prediction. The proposed model is validated using field data and several potential applications are discussed and implemented. The data from multiple vertical wells (real-time drilling data, insights into formation drillability obtained through coring, information related to the Bottom Hole Assembly (BHA) and PDC drill bit) are meticulously collected, screened, pre-processed, and seamlessly integrated through a data integration process. This enriched dataset then becomes the cornerstone of advanced analytics. Leveraging the strength of six machine learning-based regression models, namely the Artificial Neural Network (ANN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine Regression (SVR), Polynomial Regression (PR), and Multiple Linear Regression (MLR), a comprehensive strategy for the Rate of Penetration (ROP) prediction is meticulously implemented. Field data from the Tarim basin in Xinjiang, China, was utilized to validate the proposed approach. Among all the methods, decision tree-based algorithms exhibit the best performance, with XGBoost standing out with a coefficient of determination (R-square) as high as 0.98, and a remarkably low mean absolute percentage error (MAPE) of 4.8%. Random Forest (RF) shows very similar results. Furthermore, both the Artificial Neural Network (ANN) and polynomial Regression demonstrate good performance, boasting R-square values around 0.93, and MAPE figures of approximately 10%. However, the Support Vector Machine Regressor (SVR) and Multiple Linear Regression (MLR) display comparatively lower scores with R-square values of 0.8 and 0.87, respectively. Regrettably, their MAPE values are considerably high at 22.39% and 19.98%, rendering them unsuitable for recommendation. Utilizing the fine-tuned ROP prediction model, practical applications such as on-the-fly ROP optimization, real-time drilling advisory systems, drill bit recommendation systems, etc. are explored and implemented. The majority of the analysis in this paper is the result of an automated data analysis pipeline, enabling a seamless deployment in future autonomous drilling operations. This paper presents a novel machine learning approach for ROP prediction, which incorporates integrated data encompassing formation drillability information as well as information about the drilling Bottom Hole Assembly (BHA) and drill bit. This approach not only revolves around ROP prediction but also emphasizes the practical application potential of the predictive model, an aspect that received limited attention in prior research.