(37c) Durable Model Predictive Control

This paper discusses methods for developing robust and durable linear models for large scale Model Predictive Control (MPC) applications that encompass a whole process unit. MPC models are usually identified from process data using statistical techniques. The MPC model is an approximation of the physical process. The "raw" statistical model should be refined using process engineering knowledge and mathematical techniques so that the optimizer driving the MPC application does not attempt to utilize degrees of freedom that do not exist in the real physical process. This is especially true when ill-conditioned sub-models exist in the larger overall MPC model. This paper will discuss methods for ensuring that the model used by the MPC application more closely resembles the physical structure of the process. The methods will be illustrated with examples from typical refinery process units.