Advancing Predictive Analytics for Gas Sweetening Plants Through Machine Learning and Feature Selection

Abstract

Predictive models employing random forest regression and support vector machines (SVMs) were developed to predict output parameters in an industrial natural gas sweetening plant. Extensive data comprising 550 input/output variables from a gas processing facility in western Iran was leveraged to construct and evaluate the models. The key output forecast was rich amine loading (mole of acid gas per mole of amine). The dataset was partitioned into training (80%), optimization (10%), and testing (10%) subsets after normalization. An R-squared value of 0.97 and a Mean Absolute Error (MAE) of 0.008 were achieved by the random forest regression, outperforming SVM's R-squared score of 0.91 with an associated MAE of 0.012. Furthermore, the random forest model was optimized using particle swarm optimization (PSO), a metaheuristic technique. The pivotal innovation entails exploiting comprehensive empirical data with hundreds of variables to build data-driven models capable of exceptional predictive fidelity exceeding 0.9 R-squared. This research establishes random forest regression, especially after optimization with PSO, as a highly efficacious and robust methodology for the simulation and optimization of natural gas treating plants