Handling Disturbances in Modifier Adaptation with Quadratic Approximation*

In model-based real-time optimization plant-model mismatch occurs frequently. Modifier adaptation schemes can be used to find the true plant optimum despite the presence of parametric and structural mismatch in an iterative optimizing procedure. In Gao et al. (2015a) different modifier adaptation schemes are compared in terms of performance in the presence of measurement noise, which is a realistic scenario in process industries. The best performance is attained by iterative gradient modification optimization combined with a surrogate quadratic model of the noisy plant cost function as proposed in Gao et al. (2015b). Key to the success of the construction of the surrogate model based on a minimal amount of plant evaluations is the selection of suitable past setpoints for the regression set. One of the most important criteria in this screening procedure is the allowed “age” of setpoints. This paper examines the Otto Williams Reactor (Williams and Otto, 1960) example that illustrates the trade-off between noise attenuation and rejecting time varying disturbances.