I wanted to bring additional visibility to a great thread in the LinkedIn Process Control group on advanced regulatory control versus advanced process control. The post, So, ARC Can “Replace” APC? is based on a blog post, Advanced regulatory versus model-predictive control. The LinkedIn post’s author draws two points from the blog post:
- ARC is implemented at the Process Control System level while APC (or MPC [model predictive control]) more precisely needs usually a server and communications to the PCS
- ARC has a limited number of variables compared to MPC.
Any other differences?
Lou wrote:If the choice is between PID with override and feed forward vs MPC, there are simply some things that MPC can do that PID cannot and some things that based on training and skill are arguably easier with MPC. Deadtime dominant processes are a challenge for feedback control and the deadtime limits the ability of feedback control. A Smith Predictor may help, but may not be as robust as MPC. MPC can more easily and robustly handle deadtime dominant processes and I think there is little dispute on that point.
On the issue of constraint limits, with override control, the override controller takes over when a constraint is about to be violated in both MPC and ARC. But in a multivariable controller (typical MPC) the process can ride on one constraint and still drive toward another constraint until the degrees of freedom are saturated. This just doesn’t really happen in ARC. Finally, in interactive processes, feed forward and/or tuning can be used to decouple the interactions. Feed forward can be difficult for many to tune. It can be on the level of difficulty as MPC with modern process dynamic identification tools or worse. MPC will not only decouple the interaction, but can actually coordinate the controller moves in a way that optimizes the response and which feed-forward is unlikely to do.
Greg wrote:The PID can do dead time compensation by the simple insertion of a dead time block in the external reset path of the PID. Only a dead time parameter needs to be set whereas the Smith Predictor also required the identification and setting of the process gain and time constant. This dead time can be written to. The online calculation and setting of the dead time is critical particularly when the source is a transportation delay. The PID reset time can be greatly reduced if the dead time setting is accurate. The benefit of dead time compensation is not seen until the reset time is reduced to much lower values than found from tuning rules.
The PID can provide directional move suppression that is also computable and settable online by the use of setpoint rate limits in the secondary loop or analog output block and external reset feedback (e.g., dynamic reset limit). The PID can also inherently prevent oscillations from violation of the cascade rule by the use of external reset feedback of the secondary loop PV and fast digital valve controller (DVC) readback of actual valve position. The PID can stop limit cycles from backlash and stick-slip by the integral deadband or an enhancement of the PID where integral action is suspended if there is no appreciable change in the process variable. The enhanced PID can also suppress oscillations and eliminate the need for detuning when an analyzer cycle time is larger than the process dead time and process time constant. The PID can do feedforward control and decoupling with dynamic compensation so the preemptive correction signal arrives at the same place at the same time in the process as the disturbance.
Sounds great, but the expertise required is largely undocumented and not automated. My latest book Tuning and Control Loop Performance – 4th Edition published by Momentum Press attempts to remedy this situation but even after 556 pages, there are still gaps.
Meanwhile, the MPC implementation is very much automated and the tuning often simplifies to setting a move suppression parameter (e.g., penalty on move). Another parameter may be set to provide more or less emphasis on a controlled variable or constraint (e.g., penalty on error). The dynamics of decoupling, feedforward, optimization for multiple variables are inherently addressed and signal characterization can be used for gain nonlinearities. If MPC move suppression and model dead time and remaining nonlinearities can be adapted online and something akin to external reset feedback is available, the lone remaining advantages of the PID are largely gone and come down to the applications where a special PID structure (e.g., no integral action for unidirectional response), high PID gain and rate time for open loop unstable processes, or where a fast PID execution rate due to process dead times and times constants less than 1 second are needed. For more details see my March 29 Control Talk Blog on the Control Global site.
The Process Control LinkedIn group requires membership to view the group, so if topics like these are of interest, consider joining. You can also connect and interact with Lou and Greg in the Emerson Exchange 365 community in the DeltaV and Train & Develop groups.