Worst-case and stochastic optimization schemes are used to safely operate chemical processes, with operating conditions that are guaranteed to be feasible in the presence of a plant-model mismatch, often at the expense of a less optimal operating point. Modifier adaptation (MA) is a methodology of real-time optimization (RTO) which uses measurements to iteratively modify the operating conditions until convergence, which is guaranteed to satisfy the Karush–Kuhn–Tucker conditions of the plant. However, MA is not guaranteed to remain feasible for every iteration, thus undermining the eventual more optimal operating conditions. This paper develops three multi-model RTO approaches using robust optimization techniques and a new filter which guarantees feasible iterates under a structural plant-model mismatch if the model uncertainty can upper-bound the second derivative of the plant. These approaches are compared to MA in the simulation of a chemical reactor, where all three converge to the plant optimum without violating the constraints.