A Hybrid Neural Network Model for Multiphase Choke Pressure Drop Prediction in Managed Pressure Drilling

Abstract

Accurate prediction of choke pressure drop is the core of precise wellbore pressure control in Managed Pressure Drilling (MPD) and a key challenge for the safe development of narrow-margin reservoirs. Traditional choke models often rely on empirical correlations, lacking dynamic adaptability to complex gas-liquid multiphase flows. To address this, this paper proposes a multiphase choke pressure drop model based on a Physics-Informed Hybrid Neural Network (PI-HNN), aiming to improve prediction accuracy and reliability. Based on full-scale gas-liquid multiphase flow experiments, this study innovatively embeds prior hydrodynamic knowledge into the neural network architecture. The constructed model adaptively handles pressure drop variations under different flow patterns. By establishing feature layers based on kinetic energy and normalized flow characteristics, strong physical constraints are introduced. A multi-task learning framework is employed to jointly predict choke pressure drop and quantify uncertainty through a probabilistic output mechanism. Results show that the proposed model significantly outperforms classical empirical correlations and conventional neural networks in terms of accuracy and stability. On a blind test set derived from full-scale experimental data (covering 0–11 MPa), the model achieved a Root Mean Square Error (RMSE) of 0.148 MPa and a Coefficient of Determination (R2) of 0.997. The predicted 95% confidence interval achieved a high coverage rate on the actual data, verifying the model's high reliability from the perspective of uncertainty quantification. Global sensitivity analysis further quantified the non-linear impact of input parameters, revealing that gas flow rate and choke opening are the most significant factors affecting pressure drop. In summary, this research not only provides an efficient choke pressure drop prediction tool but also offers theoretical support and engineering solutions for downhole risk warning and parameter optimization, thereby effectively improving the safety and operational efficiency of MPD.