Transfer learning for batch process optimal control using LV-PTM and adaptive control strategy

Abstract

In this study, we investigate a data-driven optimal control for a new batch process. Existing data-driven optimal control methods often ignore an important problem, namely, because of the short operation time of the new batch process, the modeling data in the initial stage can be insufficient. To address this issue, we introduce the idea of transfer learning, i.e., a latent variable process transfer model (LV-PTM) is adopted to transfer sufficient data and process information from similar processes to a new one to assist its modeling and quality optimization control. However, due to fluctuations in raw materials, equipment, etc., differences between similar batch process are always inevitable, which lead to the serious and complicated mismatch of the necessary condition of optimality (NCO) between the new batch process and the LV-PTM-based optimization problem. In this work, we propose an LV-PTM-based batch-to-batch adaptive optimal control strategy, which consists of three stages, to ensure the best optimization performance during the whole operation lifetime of the new batch process. This adaptive control strategy includes model updating, data removal, and modifier-adaptation methodology using final quality measurements in response. Finally, the feasibility of the proposed method is demonstrated by simulations.