Enhancing Distillation Column Impurity Prediction: A Novel Machine Learning and Deep Learning Approach

Abstract

In this study, we address the Dow data challenge by predicting distillation column impurity levels using advanced machine learning. Our goal is to surpass the accuracy of existing Dow process models (Qin et al., 2021). Distinctive data preprocessing strategies were incorporated during model training, including feature selection, noise reduction, and online bias learning (OBL), by implementing 14 regression models and a deep learning model. The integration of preprocessing methodologies and various models collectively enhances predictive accuracy. Our study effectively narrowed the gap between R2 and Q2 values, indicating successful mitigation of overfitting in divergent training and validation data sets. The long short-term memory (LSTM) model, with convolution smoother integration, is the most proficient model, boasting an outstanding R2 of 0.9998 and an impressive Q2 of 0.9996, indicating exceptional predictive performance. This study makes a substantial contribution to the prediction of distillation column impurities, which aligns seamlessly with industrial objectives.