A Data-driven Approach for Selecting Controlled Variables

Abstract

This paper presents a novel data-driven approach for selecting controlled variables (CVs) following the self-optimizing control (SOC) strategy. Existing SOC methods rely on known process models and are developed based on linearization of original nonlinear process models, which requires rigorous knowledge of the plant and may result in large losses due to the locality inherited from linearization. In this work, we propose a data-driven approach firstly to derive a regression model for the economic cost as a function of independent variables using operation data, then to determine CVs by incorporating the concept of necessary conditions of optimality (NCO). Compared to existing SOC methods, the proposed approach doesn't need a process model to be a priori. Moreover, it is able to achieve a better self-optimizing performance under uncertainties, as demonstrated by a case study of the exothermic reactor example.