The Route to Model Predictive Control Success

McMillan and Weiner Ask Dennis Cima How He Approaches Making MPC Applications as Successful as Possible

By Greg McMillan, Stan Weiner

Share Print Related RSS
Page 2 of 2 1 | 2 Next » View on one page

Dennis: Sulfur plants have models that last 10 years. Hydrogen plants are basically the dynamic models as developed over 15 years ago. These processes are not subject to degradation or to a continuous progression of technology and changes in feeds and products. In fluid catalytic cracking units (FCCU), there are changes in feed stocks (e.g., virgin gas and oils, atmospheric residues and tar sand oils) and shifts in products (e.g., olefins versus diesel). For small changes, we can simply modify gains online. For major differences, we retest typically using automated step-testing software. If you are committed to making MPC work, you make MPC proficiency part of operator certification. To deal with changes in catalyst capability, we use a model developed near the middle of catalyst life expectancy. The time at the start and the end of a catalyst life is relatively short where the catalyst activity is extremely higher or lower, respectively.

Stan: At Monsanto and Solutia, our models were focused on a particular unit operation with four to 10 models. What is the scope of your models?

Dennis: In a large production unit like a crude distillation unit, you may have 20 to 30 MVs (manipulated variables) and 60 to 80 CVs (controlled or constraint variables). Over the course of three to 10 years, some of the dynamic models change. Fortunately, most stay the same so we can focus on specific areas for retesting.

Greg: We were worried about washout where inputs at the very beginning of the process would be so attenuated by back-mixed liquid volumes that the dynamics were not recognizable near the end of the process. The responses at the end were less than the sensitivity of the measurements or were overwhelmed by competing effects in intermediate volumes. Batch operations for reaction and crystallization often made modeling effects of inputs at the front end on the back end impossible. The time horizon of the production unit models also was a day or more, whereas the time horizon of unit operation models was on the order of hours. I suspect for plug flow gas unit operations, the time horizon and washout concern is a lot less. What are some of the special skills and tools needed to develop and maintain a large MPC? Do you separate the MPC into plant areas such as reaction and purification?

Dennis: Let me answer the second question first. Where it makes sense to combine reaction with separation into a single large MPC application is when separation section constraints limit unit upstream reactor section constraints. For example, if unit throughput is limited by a downstream fractionation constraint, then there will typically be an economic driver to combine reactor section and fractionation section controllers. However, larger controllers are more difficult to maintain and are more difficult to implement to insure model and gain consistency. Singular value decomposition (SVD) and linear program (LP) cost calculation tools are necessary when attempting to build and implement large MPC applications.

Stan: Charlie Cutler in his key note talk at ISA Automation Week 2011 lamented the demise of real-time optimization (RTO). Do you use RTO to maximize the overall plant benefits?

Dennis: We do not use the design quality, open-equation, steady-state models that Charlie was talking about for RTO. The development and maintenance of these models is difficult off-line for plant design. Doing them on-line for RTO adds a whole other level of expertise and commitment. We do multi-unit optimization with simplified and focused models as a practical approach. The less rigorous models enable us to develop and support a larger scope of optimization.

Greg: We finish up here with my favorite topic of PID versus MPC control. The Control Talk blog "When do I use PID, MPC and FLC for Basic Control," totaled the greatest number of views per month. The other blog that created almost as much interest was "The Basics of PID Control Modes."

When do you see you need to move on from PID to the MPC?

Dennis: The first flag PID may not be best is if the control solution is multivariable with complex dynamics. Dynamic decoupling and dynamic feed-forward for PID control requires considerable heuristic expertise and effort. By the time you do it if you have the resources, you could have had an MPC on-line, opening up the door to more extensive opportunities. The second flag is dynamics that go beyond the first- or second-order, plus dead time, self-regulating or integrating process. Particularly difficult for a PID is inverse response, higher order dynamics and secondary effects from heat integration and recycle streams. A more immediate material balance effect may be followed by a slower energy balance effect. For example, changing one side draw flow in a fractionator may shift the temperature profile in the column, requiring a consequential change in reflux flow. Another example is a three-bed catalytic reactor where a change in temperature to the first bed results in an eventual shift in the temperature to the third bed that comes back to affect the first bed from preheat of the feed by the reactor outlet stream. A third flag is multiple constraints and competing optimums. PID overrides and valve position controllers provide a sequential rather than a simultaneous honoring of constraints. Optimization is typically limited to a simple objective such as maximizing feed or minimizing the use of a particular utility (e.g., minimizing boiler or compressor pressure or maximizing a refrigeration unit temperature).

Stan: How does PID tuning affect the ability of the MPC to do its job?

Dennis: MPC dynamic models have embedded in them PID regulatory control dynamics. If you radically change the tuning or performance of PID regulatory controls, this will affect the accuracy of the MPC dynamic models. A poorly tuned regulatory control system is a problem even when not being manipulated by the MPC. We typically want no oscillations from an overly aggressive PID controller. Controller aggressiveness and oscillations can excite nonlinearities in the process.

Greg: The avoidance of oscillations and PV overshoot are general objectives. Some consultants have mistakenly put eliminating PID output (MV) overshoot as a general criterion for tuning PID controllers. Many people don't realize that for control of integrating processes (e.g., gas pressure, level and batch temperature), overshoot is necessary to correct for a disturbance or to achieve a new setpoint. For composition control of large, well-mixed volumes and recovery from upsets, overshoot is needed to get to setpoint in a reasonable time frame.

Greg: Most of my top ten lists were conceived while jogging. Since running is out of the picture with my worn-out knees, I am relegated to finding the moments of inspiration from a mind free to roam. Good thing my mind is not worn out. Here is one I composed returning from New Jersey.

Top Ten Reasons to Write a Top Ten List on an Airplane

(10) Takes your mind off screaming kid.
(9) Airline food heightens your sense of absurdity.
(8) Low humidity enhances your dry sense of humor.
(7) High altitude gives you a perspective.
(6) Engine hum is hummer for creativity.
(5) Money saved by not reading Sky Mall.
(4) Your Kindle's battery is too low to read great literature.
(3) Your mind's battery is too low to think anything serious.
(2) At least your fingers are getting exercise.
(1) Passengers will want a subscription to Control.

Page 2 of 2 1 | 2 Next » View on one page
Share Print Reprints Permissions

What are your comments?

Join the discussion today. Login Here.

Comments

No one has commented on this page yet.

RSS feed for comments on this page | RSS feed for all comments