Stability of nn-based MPC in the presence of unbounded model uncertainty

摘要

Uncertainty associated with neural network models used in control is unstructured, can be unbounded and is also difficult to be characterized in an a priori sense using robustness measures. Feedback limitation is an "emergency" technique that alleviates in an on-line way the stability problems inflicated by unbounded model uncertainty. Like relay autotuning, feedback limitation forces the closed-loop through an oscillation, thus producing additional process information suitable for hierarchical, localized model adaptation.