Enhanced Liquid Detection in Wet Gas Metering Via Microwave Sensing and Random Forest Regression

Abstract This study explored the integration of machine learning regression models with a microwave transmission line sensor for estimating liquid volume fraction and liquid flowrate in wet gas flows. Under low liquid loading conditions (gas volume fraction 95–99.9%), four models: Bruggeman, support vector regression, Gaussian process regression, and random forest regression were evaluated. Random forest regression delivered the best tradeoff between accuracy, robustness, and computational efficiency, achieving a relative absolute error of 2.23% for liquid volume fraction and approximately 5% for liquid flowrate, with a Durbin–Watson statistic of 2.02 indicating minimal residual autocorrelation. Feature importance analysis identified the mixture dielectric constant as the dominant predictor (approximately 97% contribution), while other dimensionless parameters had a limited impact. Support vector regression failed to generalize, and although Gaussian process regression showed slightly higher accuracy, its computational cost limited real-time applicability. Overall, random forest regression combined with microwave sensing offers a scalable, nonintrusive solution for wet gas metering, with future validation needed under industrial hydrocarbon–water conditions and liquid loading flow regimes.