I am currently working on a problem where I need to evaluate the performance of a model predictive controller whose control objective is predictable disturbance rejection. The optimization is based on a linear state-space model and I have successfully shown the improvement of the closed loop with experiments, but am curious to the limit of the MPC. Does anyone know a good way to evaluate this? thanks!
Classic measures for goodness of control are related to the controlled variable (CV) closeness to the setpoint (SP). The actuating error (e) is the difference, and many use integral of the error, integral of error squared, etc. as a performance error.
However, to me, excessive changes in the manipulated variable (MV) are an undesirable outcome of minimizing CV performance measures. So, I like including "travel" the sum of absolute deviations in the MV or "energy" the sum of squared changes.
However, there are many other important issues to consider in MPC: How easy is it to develop (obtain the models, and tune for multivariable responses, and assign weights related to constraints)? For instance can an operator do it, or does it take a PhD to understand the procedure and guide it away from traps?
What is the computational burden? Can it always find a solution within the scan time desired?
What is the robustness to faults, noise and the many sorts of events that Nature will contrive?
When handling constraints, does it absolutely prevent a constraint violation? If not the magnitude and duration of a constraint violation will be a consideration for goodness. The constraint categories are on the MV (rate of change and magnitude), and on both the CV and other (auxiliary) process variables (AuxV). Typically it is easy to place hard constraints on the MV, but we use soft constraints on the CVs and AuxVs.
There should be no windup when on a constraint. As soon as the constraint condition is removed, the controller should take action, not remain wound-up.
Does it provide bumpless transfer form MAN to AUTO? Or, because it was not adequately initialized in MAN mode, does it change the MV upon the switch and cause a deviation in the CV?
Does the controller explain itself to the operators, or is its action a mystery? If it takes MV action now to avoid a constraint that will happen in the future, that is good. But if the operators are confused about why it is making the move, that is bad.
I think you should refer Process Control: Modeling, Design, and Simulation by B. Wayne Bequette. it provides you a better idea.
- Badges
- Science topic
- Similar topics
- Engineering
- Control Theory
- Control Systems
- Control Systems Engineering