Machine learning for constrained self-optimizing control

Self-optimizing control (SOC) is a method to select controlled variables (CVs) and keep them constant such that the plant operates optimally. Since the concept of SOC was proposed, some difficult problems in this field have not been solved such as active constraint changes. Previous work either handles the constrained SOC problem with complicated control structure or in the sense of local SOC, and has limitations such as structural complexity and inaccuracy of control. To address the shortcomings of the existing methods, this paper proposed a constrained variable approximation (CVA) method to solve the problem in the global sense using a simple control structure. The constrained variables which may vary between active and inactive are approximated by a nonlinear function of available measurements in the whole operational region using artificial neural network (ANN). Then the CVs are determined as the difference between the nonlinear function and the constrained variables. The system would be near optimal operation when the CVs are controlled at zero. An evaporator process is applied to illustrate the effectiveness of the proposed method.