Model Adaptation for Real-Time Optimization in Energy Systems

Abstract

Real-time optimization (RTO) is widely used in industry to operate processes close to their maximum performance. The models used for RTO need to be adapted using real-time data to ensure feasibility of the model optimal inputs and convergence to the real plant optimal point. Heat and power systems are suitable for being optimized in real-time because of their fast dynamics and the benefits achievable by reacting to changes in power prices and steam demand. This work proposes a modifier adaptation strategy that exploits the structure of certain problems to make the adaptation faster and more reliable, which is proven to be particularly useful for heat and power systems. The adaptation is performed in the equations that predict efficiencies or performance of unit operations. By identifying the variables that modify each performance factor, the number of data sets needed for gradient correction is reduced. This makes the proposed strategy suitable for real-time optimization of processes with a large number of inputs. Two alternatives are proposed to implement the approach: gradient calculation by finite differences and quadratic regression using current and past data. The features and behavior of this approach are shown through two case studies: (i) a simple model with three processes, and (ii) a heat and power system of a sugar and ethanol plant. A comparison with other existent approaches shows a better performance in terms of operating cost and sensitivity to noise.