Machine Learning-Based Adaptive Regression to Identify Nonlinear Dynamics of Biochemical Systems: A Case Study on Bio 2,3-Butanediol Distillation Process

Developing an accurate process model is essential to efficiently operate a process and maximize its economics. While offline data-driven models utilizing historical data generally exhibit satisfactory performance, their effectiveness diminishes in accurately predicting real processes characterized by constant changes and uncertainties over time. Hence, there is a need for an adaptive model that is capable of effectively handling dynamic behavior. In this study, we propose an adaptive data-driven regression model that leverages subset selection techniques and decision thresholds. In addition, a comprehensive analysis was performed to determine the best adaptive regression model, considering case studies with different model parameters and training window sizes, taking into account statistical indicators of model accuracy as well as nonstatistical indicators such as the number of updates, update period, and computation time. The developed adaptive regression model has been successfully demonstrated on a bio 2,3-Butanediol distillation column at GS Caltex, Republic of Korea, suggesting its potential applicability to similar process systems and providing opportunities for future research in process optimization and control.