Data-driven self-optimizing control: constrained optimization problem

Abstract

Self-optimizing control (SOC) is a technique used in selecting controlled variables (CVs) for a process plant control structure with a view to operating the plant optimally in the presence of uncertainties and disturbances. Existing SOC approaches are either local which result to large losses or too cumbersome to be applicable to real systems. In this work, a novel method of CV selection based on data was developed. In the method, a compressed reduced gradient of a constrained optimization problem was proposed to be estimated using finite difference scheme. The CV function was then used to approximate the necessary condition of optimality (NCO) using data only in a single regression step. The new approach was applied to a simplified case study and its performance was compared to an existing SOC methodology. An excellent goodness of fit was obtained during the regression with a R²-value of 1.0 associated with one of the designed CVs. The formulated CVs were found to be very robust with performance similar to that of NCO approximation method. A zero loss was incurred with one of the CVs.

Keywords: compressed reduced gradient, constrained problems, controlled variable, data-driven, necessary condition of optimality, self-optimizing control