Over in LinkedIn’s Process Control group, a question was asked:
What is the difference between regulatory and model predictive control. Is it possible that MPC work as stand alone without regulatory control?
Emerson’s Chowyang Neo offered:
Regulatory Control generally refers to good old PID loop control. Gaps between the SP and PV known as errors are being fed to a PID controller so that Process Variables (PV) is then being steered closer towards the setpoint (SP) based on the magnitude of the error. There can be more advanced variants of this simple PID – cascade, ratio, feedforward etc. But all of these are generally considered as Regulatory or Adv Regulatory Control.Chowyang Neo
Sales & Marketing ManagerThe use of model based technique is driven more by the complex, multi-variable nature of some of the processes we see in this industry. The level control of boiler is a classic example with many factors playing a part….Have a read to get more
Emerson’s Lou Heavner added:
There is a hierarchy of process control starting with open loop, where the operator is the controller. He makes manual adjustments to valves and other final control elements to keep the process at operating targets or within operating constraints. This represents the absence of automation. Next up the hierarchy there are feedback controllers that take the place of the operator automatically closing the loop, with PID being the most prevalent by far. This basic level of feedback control is inherently single loop in nature.Lou Heavner
Systems/Project Engineering ConsultantThere is one measured variable that is controlled and one manipulated variable (usually a control valve) that is the process input. A control panel or DCS may have hundreds of single-input-single-output feedback control loops for the operator to manage. The next level up the hierarchy involves a branch. Up one path, there is supervisory control that is used to sequence the process or manage a batch process. That is not the realm of MPC, although MPC may be seen in batch and sequential processes.
The other branch includes multivariable regulatory control. One approach to multivariable control is advanced regulatory control in which single control loops may be cascaded, ratioed, or included as part of complex strategies with feed forward inputs to cancel measured disturbances and override controls to handle constraints. Special features in digital systems also make it easy to handle nonlinearity with characterization, create inferential calculations, and a host of other functions to improve control. Modified PID controllers such as Smith Predictor or “Error Squared” or other controller algorithms can be used to deal with deadtime, nonlinearity, and other challenges.
Occasionally you may see a feedback controller that is not PID per se, but uses something like fuzzy logic. A specific type of advanced control is model predictive control. This type of controller uses dynamic process models and is computationally intensive, compared to PID. But because it has process dynamic models and is projecting where the process will be in the future, it doesn’t have to execute as frequently as PID. In the early days, there were not the tools for process identification that are available today nor the computational power to execute the controls without a host computer. But like many things, the technology has improved and become more accessible to process plants.
Now DCS systems can execute MPC and have their own process identification tools. MPC is perhaps the best means for handling deadtime dominant processes, which tend to be slow due to the deadtime anyway, and which are the most difficult for feedback controllers. It is ideal for interactive processes because it can easily model and accommodate multivariable interactions. It is perhaps the best approach for constraint optimization at the controller level, because it won’t stop seeking a better solution once it hits the first constraint. It will continue to move toward a better position while honoring constraint limits until it uses up all of its degrees of freedom.
However, it is not without its limitations. It can be limited by model accuracy and by process nonlinearity. One common area of process nonlinearity is valve response. So while it is possible to use MPC to manipulate valves directly, it is usually better to have the MPC send a setpoint to a flow loop and let the flow loop manipulate the valve. I would make the same argument for cascade control in many processes, especially level control, where too often, performance is limited by the failure to cascade the level controller to a flow controller. Very fast processes (like flow loops that are easy to control with PID) put quite computational load on controllers executing MPC. While some embedded MPC controllers can execute at a high frequency, like once per second, this generally isn’t the best approach for high speed processes.
MPC has a feedback mechanism, which is usually a model correction factor. As noted in my previous response, there is always going to be some model mismatch and the feedback handles that. But the difference between MPC and PID is that PID controls the present while MPC controls the future. This gives the MPC control the chance to be as robust as PID control. But increasing robustness of either PID or MPC will lead to some degradation of performance. I would recommend in most cases that you don’t limit yourself to just one approach, be that PID or MPC. Use them both where they are the best choice. To do otherwise would be akin to taking your screw driver out of your tool box and limiting yourself to only your hammer.
You can connect and interact with Chowyang, Lou, and other process control specialists in the DeltaV and Ovation tracks of the Emerson Exchange 365 community. If you join us in Orlando for the Emerson Exchange conference, you can find them and many more experts too!
