Sequence-to-sequence digital twin model in chemical plants with internal rolling training algorithm

Abstract

A data-driven digital twin based on the sequence-to-sequence (StS) model is introduced in this study. The rolling training algorithm is presented as a plug-in in StS training to achieve long-term rolling predictions. The model performs long-term horizontal predictions over 8-days based on initial observed data and manipulation sequence conditions without any new historical inputs. The digital twin model is validated using the dynamic simulation of the vapor-recompression C3 process. When the rolling training algorithm is used, the StS model demonstrates the best predictions than the traditional artificial neural network model and simple StS model for 1- and 8-days tests. Moreover, noise added cannot affect the prediction performance of StS with rolling training, and the step-change test validated the StS with rolling training containing physical meaning. Hence, the proposed StS model is a data-driven digital twin having physical interpretability. The StS model can implement long-term rolling predictions for days.