Interpretable Machine learning model for predicting Ethane-Ethylene composition in binary distillation process

Abstract

Accurate prediction of ethane and ethylene compositions in binary distillation columns is critical for optimizing industrial separation processes. Traditional modeling approaches often struggle with process nonlinearities and variabilities, resulting in inefficiencies. This study evaluates non-ensemble models (SVR, DTR, KNN) and ensemble models (RF, GBR, XGB) to significantly enhance predictive accuracy and address these challenges. While the DTR model outperformed other non-ensemble models, the ensemble models exhibited superior performance. Among these, the XGBoost model demonstrated the highest accuracy, achieving R² values of 0.968 (test) and 0.993 (train) for ethane composition, with a mean absolute error (MAE) of 0.21 and root mean square error (RMSE) of 0.27. For ethylene composition, XGBoost achieved R² values of 0.971 (test) and 0.987 (train), with MAE and RMSE values of 0.32 and 0.43, respectively. To enhance model interpretability, Shapley Additive exPlanations (SHAP) analysis identified the reboiler duty to feed flow ratio as the most significant feature, contributing + 1.79 SHAP units for the bottom product and + 0.24 for the top product. Aditionally, a real-time application platform was developed to deploy the XGBoost model, enabling non-expert users to make informed decisions in industrial operations. This study addresses key challenges in ethane-ethylene separation, providing solutions to improve thermodynamic efficiency and sustainability in industrial distillation processes.