Evaluation and Interpretation on Data Driven ROP Models From Engineering Perspectives

Abstract

Abstract Recently machine learning (ML) technology has been adopted in the Oil and Gas industry as solutions to the most demanding problems like drilling parameters estimation and prediction, drilling incidents detection, optimal well planning etc. Rate of penetration (ROP) modeling is complicated since the ROP is affected by many interacted parameters/factors, and there is no obvious correlation between a single drilling parameter and the target ROP value. Different variables, e.g., weight on bit (WOB), rotary speed (RPM), standpipe pressure (SPP), formation/bit properties, interact such that it is difficult to formulate an accurate mathematical model to describe their correlations. Thanks to a big amount of available field data, the ML becomes a powerful tool to develop data-driven ROP models. It is a positive sign that many good results from data-driven ROP models have been presented in recent years. However, the limitations of ML approaches like data generality, quality and selection, model evaluation and interpretation, model generality, robustness and stability, and model validation might be the barriers to implementing data-driven techniques to drilling operations in real life. Therefore, our study aims to evaluate and analyze data-driven ROP models with different engineering perspectives, including: 1) Data: how does data affect the data-driven ROP model behavior? 2) Performance metric: which one is suitable to evaluate the error of the estimation/prediction, e.g., coefficient of determination (R2), mean absolute error (MAE), mean absolute percentage error (MAPE) or others? 3) Dynamics: How to evaluate the ROP model dynamics? 4) ML models: how well do the data-driven ROP models perform in the ROP optimization? From the case studies in this research, the comprehensive results are shown to evaluate the developed ROP models with respect to the above-mentioned aspects. The work focusing on model evaluation and engineering interpretation facilitates a better understanding of the data-driven models and their applications. It also presents a new evaluation method using trend analysis.