NSGA-III Framework for Multiobjective Optimization of Drilling Parameters Based on Cross-Well Generalization: Case Study Using a Stacking Surrogate Model for Rate of Penetration Prediction

Abstract

Rate of Penetration (ROP) is an important indicator for evaluating drilling efficiency, and its accurate prediction is of great significance for shortening the drilling cycle and reducing development costs. However, existing machine learning methods still face limited generalization capability in cross-well prediction scenarios, and the limited transparency of these models also undermines their credibility in engineering applications. To address these issues, this study developed an integrated intelligent framework for ROP prediction and drilling parameter optimization based on while-drilling data from five wells in a block of the Xinjiang Oilfield, China. For model development, a well-based data partitioning strategy was adopted to avoid data leakage. Bayesian optimization was used to jointly optimize candidate model combinations and hyperparameters, and the resulting Stacking ensemble model was constructed using CatBoost and GBDT as base learners and Ridge regression and XGBoost as meta-learners. On the independent test set, the model achieved an R2 of 0.812, an RMSE of 4.998 m/h, and an MAE of 3.484 m/h. Compared with the best-performing single model, Random Forest, the proposed model improved R2 by 9.1% and reduced RMSE and MAE by 14.4% and 23.1%, respectively, demonstrating stronger cross-well generalization capability and robustness. For model interpretability, the SHapley Additive exPlanations (SHAP) method was introduced to analyze the model outputs from both global and local perspectives. The results quantitatively revealed the contribution patterns and effect directions of parameters such as Torque, Depth, and Flow Rate on ROP prediction, providing engineers with an intuitive basis for understanding the model decision logic and enhancing the transparency and engineering interpretability of the model outputs. For drilling parameter optimization, the validated Stacking model was used as the surrogate model, and the field-measured value ranges of each drilling parameter were used to define the feasible operating domain. NSGA-III was then applied to perform multi-objective collaborative optimization of Weight on Bit, Rotary Speed, Flow Rate, Standpipe Pressure, and Drilling Fluid Density, maximizing ROP while maintaining prediction reliability. The average predicted ROP on the test set increased from 15.98 to 17.02 m/h, corresponding to an improvement of 6.54%. This study provides theoretical support for constructing a trustworthy and transferable intelligent ROP prediction system, and holds practical significance for refined drilling parameter control and enhanced field decision-making efficiency under complex operating conditions.