Prior model identification for stochastic optimal control of continuous aqueous two-phase flotation

Abstract

In chemical process control, where an accurate model of the system dynamics is often not available, advanced control strategies such as stochastic optimal control promise superior control performance as opposed to nominal approaches neglecting the – often significant – uncertainty associated with the model predictions. A crucial prerequisite for stochastic optimal control is a suitable description of the uncertainty associated with the available model as well as a computational description of how this uncertainty evolves as more measurements become available. In this work, we exemplify how a stochastic model might be identified from experimental data and illustrate how non-stochastic models fail to describe the available data in the presence of high inter-experimental variation within the dataset. To this end, model identification from experimental data of the continuous aqueous two-phase flotation serves as a case study. In a second step, we showcase the performance of an optimization-based control strategy which is based on the identified stochastic model in closed-loop experiments.