MPC, Fuzzy or PID?

This guest post is authored by Greg McMillan.

In the ISA Automation Week Mentor Program I am providing guidance for extremely talented individuals from Argentina, Brazil, Malaysia, Mexico, Saudi Arabia, and the USA. We will be sharing a question and the answers each week. If you would like to provide additional answers, please send them to Susan Colwell at ISA. This question is from Hector Torres of Mexico:

MPC, Fuzzy, or PID – how do you pick the correct solution now that all are available at a lower cost? How do you prevent people from abandoning APC/MPC systems when the expert leaves?

  1. If you want a comprehensive optimization of a continuous process the solution is MPC.  Linear Programs (LP) and Real Time Optimization (RTO) need the prediction to violation of constraints inherent in MPC. If you are satisfied with a local piecemeal optimization of fed batch as well as continuous unit operations, PID valve position control is a quick solution. See my article “Don’t Over Look PID in APC”  and my blog “A Smorgasbord of Batch Cycle Time Minimization Opportunities – Part 1” . Your skill base will also sway the argument. When I did some pH control consulting at a refinery, the manager said they would startup on PID control but convert to MPC within 2 weeks because they don’t have any Shinskey types. In refineries and large mature continuous processes MPC is the norm. MPC was originally developed by Charlie Cutler at Shell.
  2. The PID with the deadtime compensation option has been proven by McAvoy and Shinskey to provide the best load rejection. There are many PID options under- utilized, such as dynamic reset limiting, setpoint velocity limiting, the enhanced PID for wireless (PIDPlus), adjustable alpha and beta factors, and deadtime compensation by the simple addition of a deadtime block. The combination of AO block velocity limits and dynamic reset limiting provides the move suppression found so useful in MPC for limiting the transfer of variability from the process variable to the controlled variable to reduce interaction. If decoupling is needed, a simple feedforward of one PID output to another PID output can help. Chapter 14 McMillan14Rev1.pdf explores many PID opportunities in the book PID Control in the third Millennium: Lessons Learned and New Approaches.
  3. A PID can be set to do better than a Fuzzy Logic Controller via use of the extensive PID options and a “Full Throttle Batch and Startup Response
  4. You can interactively explore the capability of the PID for various situations and dynamics by the use of process control lab website. The website was created by me, Jack Ahlers (Monsanto), and Charlie Schliesser and is available for general use. The simulation and control modules were configured as a virtual plant. The Process Control website is accessible at no cost via LogMeIn, a free service. Jim Cahill recorded 12 live demo-seminars called “Deminars” showing the use of the website for PID control solutions.
  5. In my ISA pocket guide Models Unleashed, I compared the performance of the MPC and PID and found the MPC to be much more sensitive to changes in process dynamics. An overestimate of the loop deadtime makes an MPC unstable where it just makes the PID sluggish from slower tuning based on the larger deadtime. I also found that deadtime compensators were adversely affected by changes in dynamics similar to the MPC.
  6. Shinskey and I found that the improvement and robustness of deadtime compensation greater for a process dominated by a lag (t) rather than a deadtime (q) (deadtime compensation works best for t >> q), contrary to public opinion and what is often stated in the literature. Note that the term lag is commonly used for time constant.
  7. An APC solution using PID is much more likely to stay in-service once the expert leaves the building because the PID is familiar to the operator and configuration engineer-technician. In a plant with mostly PID, you need to have more training for an MPC.
  8. In general the most powerful way to keep a control solution in-service is to provide online process metrics showing the dollar benefits with a capture of “before” and “after” cases. Normally, the best way to show this is a cost per lb or kg of intermediates or product (e.g. raw material or energy $ cost per lb). Get with a process engineer to develop the calculations. Production rate should also be computed to show benefits from capacity related benefits (e.g. $ profit per lb product) and to monitor capacity effects on costs. Synchronize input flows (e.g. utilities and raw materials) to output flows (e.g. intermediates or products) by passing the input flows through a deadtime and filter block. Filter the final calculation to eliminate noise and inverse response. Modes and setpoints should be trended with metrics and suspend an update of metrics when the unit operation is shutdown. If you can’t get agreement on costs and prices, simply compute a ratio of a key output to an input. For more details, take a look at my June 23 post “Top Ten Limitations – Value Analysis


Print Friendly

, , , , , , , , , , , ,