A Hybrid Deep Learning Approach for Rate of Penetration Prediction in Deepwater Drilling

Abstract Accurate prediction of the rate of penetration (ROP) is essential for optimizing deepwater drilling operations, enhancing efficiency, and reducing costs. Traditional ROP models often rely on empirical correlations and statistical methods, which do not adequately capture the complex nonlinear interactions between drilling parameters or meet real-time prediction requirements. This study proposes a novel approach that integrates real-time drilling data with feature selection and deep learning techniques to improve ROP prediction accuracy in challenging deepwater environments. The proposed method starts with a Pearson correlation analysis to identify the key drilling parameters, including Weight on Bit (WOB), Torque, Mud Weight (Mw), Revolutions Per Minute (Rpm), and Flow, as well as formation gas parameters (Ethane, Propane, Isobutane) that are closely related to the Rate of Penetration (ROP). These parameters are then input into a deep neural network (DNN) for advanced feature extraction, capturing complex nonlinear relationships. The extracted features are fed into a Transformer model, which utilizes its self-attention mechanism to capture intricate time-series dependencies and pattern variations. This approach allows for a more accurate identification of the impact of parameter changes on ROP, significantly enhancing prediction accuracy. The model was trained and validated using actual drilling data from a deepwater block in the Bohai Bay, encompassing various operational conditions and drilling parameters to ensure its applicability and comprehensiveness. Compared to traditional correlation-based methods and deep learning-driven advanced feature extraction techniques, the principle component analysis (PCA) – DNN parameter sensitivity method effectively captures specific nonlinear patterns and extracts more representative high-level features. Additionally, the Transformer-based prediction method outperformed the conventional Long Short-Term Memory (LSTM) neural network across multiple metrics, particularly in terms of prediction accuracy and model generalization. The self-attention mechanism of the Transformer effectively captures the dynamic and complex interdependencies among drilling parameters. The hybrid deep learning ROP prediction model was applied to three adjacent wells in this study, achieving goodness-of-fit (R²) values of 0.954, 0.924, and 0.946, respectively. These results confirm that the integrated approach not only improves the predictive capability of the ROP model but also provides a more robust framework suitable for the complex conditions encountered in deepwater drilling. This study proposes a novel hybrid approach that combines feature extraction driven by PCA - DNN deep learning with Transformer - based prediction. The proposed method significantly improves the accuracy and robustness of ROP predictions, providing valuable insights and practical benefits for optimizing deepwater drilling operations.