Machine Learning-Based Predictive Modelling of Monoethylene Glycol Injection Rate for Hydrate Prevention in Gas Dehydration and Dehydrocarbonization Processes a Case Study of Songosongo Gas Field

Abstract Accurate forecasting of monoethylene glycol (MEG) injection rates is fundamental to the reliable gas dehydration and effective hydrocarbon dew point control in natural gas processing facilities. In this work, we develop a data-driven machine learning framework to predict MEG injection rates at the Songosongo Gas Field in Tanzania This study employed the Random Forest Regression (RFR) model to be trained using operational process data that included seven key independent variables: rich and lean MEG concentrations, hydrocarbon dew point, water dew point, heat exchanger level, suction pressure, and suction temperature. Model performance was evaluated using standard regression metrics on an out of sample test dataset. For the Random Forest Regressor, R2 value obtained for training set was 97.19% with MAE and RMSE of 2.24 m3/h and 3.40 m3/h, respectively, while it maintained a strong predictive capability on unseen data (R2:89.89%), where the corresponding values were MAE: 5.25 m3/h and RMSE:7.14 m3/h; in comparison evaluation, this model outperforms all benchmarks models of Linear Regression, Decision Tree, Gradient Boosting and Support Vector Regression combined as well, feature importance analysis shows that rich MEG concentration and hydrocarbon dew point have the strongest impact on model predictions, contributing jointly to more than 46% of total explanatory value. This is in line with process expectations, considering both of the variables relate directly to hydrate inhibition and phase behavior modulations. Overall, the results indicate that Random Forest Regression can predict MEG injection quantities with excellent fidelity. From a practical standpoint, this enables several features, including more targeted chemical dosing, minimizing over-injection, and overall cost efficiency. The general framework can be applied to similar gas processing environments beyond Songosongo. It provides a pragmatic, field-tested framework for predictive chemical management in Sub-Saharan Africa, where such data-driven interventions are rarely implemented in academic and industrial settings.