A Data-based Predictive Model for Distillation Column of Bio-based 2,3-Butanediol

Abstract In bio-based chemical processes, stable and efficient process control is challenging because the feedstock composition made by microbial fermentation is uncontrollable. Many researchers have developed data-driven models that predict the operating condition using a correlation between the process data. Since the performance of a data-driven model depends on the quality of the data, data preprocessing techniques such as noise removal and outlier detection are generally considered to enhance the model performance. Thus, research to establish an appropriate data preprocessing method and range is required to increase the performance of the model. In this study, we developed a data-based prediction model for the bio-based 2,3-butanediol distillation column. We prepared sixty train datasets by applying noise removal and outlier detection with different methods and ranges on the raw data and confirmed the correlation between the data using Pearson's correlation coefficient. We considered two kinds of feature selection depending on data preprocessing and compared the performance (R2 and normalized RMSE (NRMSE)) and reproducibility. The case in which the features were selected by only noise removal showed a similar coefficient value to that of the raw data; therefore, it had a higher performance than the case which applied outlier detection. The best performance provided R2 and NRMSE values of 0.962 and 0.045, respectively, and it could be enhanced by hyperparameter tuning. Based on this result, we plan to develop a robust predictive model that features adaptable real-time prediction and control throughout the entire process time and period.