Machine learning-driven multiphase flow prediction for wet gas: a temporal data perspective incorporating fluid property analysis and explainable AI

• Phase flowrate prediction of wet gas flow using various machine learning models. • Advanced feature engineering for ML modelling against pure sensor data modelling to investigate the effectiveness of adding physically meaningful features. • Implementing PVT equation of states (EoS) modelling for density prediction to boost and consider the uncertainty of density values as model input. • Using Explainable AI methodologies such SHAP and PFI to enhance the interpretability of developed models for better comparison and insights. • Providing insights on the implication of pure data-driven modelling for gas and liquid flowrate prediction and their application in the field conditions. The complexity and nonlinear behaviour inherent in multiphase flow systems make wet gas flow measurement a challenging aspect of monitoring and optimizing numerous production processes. Accurately determining phase flow rates with minimal uncertainty remains a critical challenge for operators. This study investigated the use of rapidly advancing technologies in industrial applications, focusing on eight machine learning (ML) and neural network (NN) algorithms to predict the gas and liquid flow rates. The prediction models were built by integrating the Venturi tube sensor temporal data and phase densities as primary features, in addition to adding a feature transform perspective. Furthermore, SHAP (SHapley Additive Explanations) and PFI (Permutation Feature Importance) analysis were used to enhance the model interpretability and transparency that leads to provide insights for the key feature’s importance and aligning the results on model’s performance. In terms of model performance, XGBoost, RF, and GRU consistently produced more accurate and stable output than the other models, with RF and XGBoost achieving a Mean Absolute Error (MAE) as low as 0.0003 for liquid flowrate and 0.0012 for gas flowrate. While still the best-performing neural network, GRU’s errors were slightly higher, with an MAE of 0.0145 for liquid and 0.0265 for gas. This study also provides insights into the application of data-driven techniques for the real-time monitoring of multiphase systems in field conditions and how the changing dynamics of flow would be challenging for AI-based solution methodologies to ensure accurate and robust measurements.