Data-Driven Self-Optimizing Control With Parametric Programing Based Constraint Handling

Abstract

Self-optimizing control (SOC) aims to achieve near-optimal operation regardless of variation in disturbances. For constrained operation, the model-based SOC has been solved as an explicit model predictive control (MPC) problem, where the changes in active constraint sets under large variations in disturbance inputs can be determined via parametric programing. For constrained SOC design with historical data only, this paper presents an extended work for a data-driven method that Zhao et al. proposed (2022 “Global Self-Optimizing Control With Data-Driven Optimal Selection of Controlled Variables With Application to Chiller Plant”, ASME J. Dyn. Sys., Meas., Control., 144(2), p. 021008.) for obtaining global SOC solution. While this work remains to adopt the Koopman subspace model that has been used, the SOC-associated constrained optimization problem is solved by applying the multiparameter quadratic programing technique. The solutions consist of piecewise-affine control laws and active constraint sets that are determined via partitions of the disturbance space. The proposed approach is illustrated with a Modelica simulation model of a chilled-water plant, where minimizing the total power consumption is the economic objective. The results show that the proposed method yields the total power very close to that by an offline genetic algorithm-based global optimization procedure, and substantially better than a best-practice rule-based control strategy. The degree of constraint violations under high cooling load is also smaller with the solution of constrained SOC.