A data-quality-aware hybrid deep learning framework for drilling penetration rate prediction from field measurements

Accurate rate of penetration (ROP) prediction is essential for improving drilling efficiency, reducing cost, and supporting intelligent decision-making in subsurface energy engineering. However, field drilling measurements are often contaminated by noise, outliers, missing intervals, and complex nonlinear relationships, which severely limit the reliability of conventional prediction models. To address this challenge, we propose a data-quality-aware hybrid framework for ROP prediction using real-field drilling measurements. The proposed approach integrates multi-stage preprocessing, including principal component analysis, HDBSCAN-based outlier removal, Gaussian process regression-based interpolation, and SG-wavelet denoising, to improve the quality and continuity of the input data. On this basis, CatBoost, multilayer perceptron, and Transformer are combined through stacking, and a residual neural network is further introduced to compensate for systematic errors. Harris hawks optimization is employed to jointly tune the hyperparameters of the full pipeline. Compared with conventional machine learning and deep learning models, the proposed framework improves predictive accuracy and robustness, increasing R 2 from 0.9667 to 0.9791 while reducing RMSE from 2.12 to 1.42 and MAE from 1.39 to 1.17. These results indicate that integrating data-quality enhancement with hybrid learning provides a practical and effective solution for robust ROP prediction under noisy field conditions.