Data-driven framework for predicting rate of penetration in deepwater granitic formations: A marine engineering geology perspective with comprehensive model interpretability

Abstract

Deepwater oil and gas resources are vital for meeting global energy demand, supporting economic growth, and ensuring energy security. The marine engineering geology of deepwater environment presents significant challenges for drilling operations, with rock behavior of deep granitic formations increasing the risk of well control incidents. Rate of Penetration (ROP) is a crucial parameter for evaluating efficiency, ensuring operational safety and controlling economic costs of deep-water drilling. In recent years, data-driven methods have provided new ways for predicting ROP in deepwater drilling. In this work, the database is derived from actual deepwater drilling operations at depths ranging from 2203 to 2938 m in China and three different data-driven methods are used to predict ROP based on field measured deepwater drilling data. After preliminary screening, the results show that the method of LightGBM has the best prediction performance. Subsequently, hyperparameter optimization has been conducted based on Bayesian principle. A comprehensive model interpretability approach based upon SHAP and PDP is adopted to conduct explanatory analysis on the improved LightGBM from global and local perspectives. The contribution degree of different feature variables to ROP is obtained as follows: TORQUE, BitTime, Pump Time, WHO (weight on hook), SPP (standpipe pressure), WOB (weight on bit), FLWPump (flow pump), and RPM (revolution per minute). Furthermore, the impact of important feature variables on ROP is analyzed with consideration of actual operating conditions and drilling hydraulics.