In the domain of process engineering, LP-MPC cascade control systems are commonly used for controlling high-dimensional MIMO systems subject to constraints. In general, LP-MPC methods are made up of two layers. The LP (linear program) layer computes optimal set-points for the inputs and outputs of the steady-state system, and is continuously fed disturbance information by the MPC. The MPC (model predictive control) layer then ensures the system is closely tracking those set-points in real time. Plant operators who are tasked with supervising an LP-MPC system often run into a problem where the set-points returned by the LP abruptly shift in a way that hinders performance. It is then the responsibility of the operators to diagnose the root cause of the set-point change, and to figure out how to adjust the operating limits in order to rectify the behaviour of the controller. Both of these tasks are difficult, especially when there are dozens or even hundreds of input and output variables. This work uses tools from inverse optimization in order to help practitioners easily trace the origins of unexpected set-point changes, and to ensure that the LP returns reasonable set-points.
Thiago Vasco (Thu,) studied this question.