Active directional modifier adaptation for real-time optimization

Abstract

Modifier adaptation is a real-time optimization (RTO) methodology that uses plant gradient estimates to correct model gradients, thereby driving the plant to optimality. However, obtaining accurate gradient estimates requires costly plant experiments at each RTO iteration. In directional modifier adaptation (DMA), the model gradients are corrected only in a small subspace of the input space, thus requiring fewer plant experiments. DMA selects the input subspace offline based on the local sensitivity of the Lagrangian gradient with respect to the uncertain model parameters. Here, we propose an extension, whereby the input subspace is selected at each RTO iteration via global sensitivity analysis, thus making the approach more reactive to changes and robust to large parametric uncertainties. Simulation results performed on the run-to-run optimization of two different semi-batch reactors show that the proposed approach finds a nice balance between experimental cost and optimality.