Real process models provide predictive behavior, but also very important process information that sometimes surprises operators and process engineers. "Models allow us to more easily target process physical limitations when MPC continually hits the same constraints," he says. "This can help a lot in economic justifications for process upgrades by quantifying percent of time against a constraint. Examples can be fan/heat exchanger cooling capacity or limited pumps and hydraulics."
MPC removes the operator "safety cushion" that some control systems put in to avoid excursions because of their variability. "It minimizes perfectly normal human limitations such as a long shift, lack of experience or too many units to control at the same time," he explains. "The safety cushion varies depending on operator knowledge/experience, and it keeps the process away from its real constraints."
"Skill degradation for operators is another issue," he adds. "They easily and quickly lose the feeling of how much they should move manipulated variables once MPC is out of service. Model-based, high-fidelity operator training simulators are the best solution, but they are still too expensive for most users and hard to maintain. MPC is not designed at all for abnormal situations, meaning process knowledge is still mandatory for operators."
McCormick likes real-time dynamic optimization sitting on top of MPC applications. "There are differences between rigorous steady-state optimization versus dynamic optimization," he says. "The first one requires developing chemical engineering equations for the whole process to be optimized, but it can only run under steady-state conditions. With dynamic optimization, such as Honeywell's Profit Optimizer that sits on top of Profit Controllers, you can attach off-line process simulators for nonlinear processes and make it run at certain frequencies to automatically update optimizer parameters on the fly."
