Time-series clustering approach for training data selection of a data-driven predictive model: Application to an industrial bio 2,3-butanediol distillation process

In this study, we propose a time-series clustering approach that selects optimal training data for the development of predictive models. The optimal number of clusters was set based on the variation of within-cluster sums of squares. A predictive model was developed with the selection ratio of training data from each of those clusters. Based on the results, a regression model was developed to predict the performance of the model. The search space was applied to the regression model, and the optimal training data ratio were selected satisfying the objective function and constraints. The effectiveness of the method is demonstrated by addressing a commercial bio 2,3-butanediol distillation process. As a result, the number of data for model training was reduced by 49.20% compared to the base case without clustering. The coefficient of determination (R²) showed the same level of performance, and the root-mean-square error was improved up to 14.07%.