Black armbands for APC?
(and other rebellious sentiments regarding advanced process control)
By Allan Kern
People often ask: “Why is advanced process control (APC) utilization low?” Or: “Why are so many manipulated variables (MVs) clamped?” These questions show progress, because many people do not recognize a “defeated” APC, much less understand what brought it about.
APC is generally considered a high level of technological accomplishment and process control sophistication, with people reluctant to criticize it. The APC wave of the past 20 years is settling out, and it is appropriate to take an objective look at APC’s record. Plenty has been written about its successes; this article takes a much-needed look at some of its better known (if not well understood) problem areas.
APC is conventionally considered a universal and near-perfect process control technology with users increasingly coming out of the closet about its shortfalls, including papers suggesting industry has more poor-performing APCs than well-performing ones. A serious problem not to be ignored are clamped MVs. Surprisingly, most current thinking suggests industry should forge ahead with more of the well-established rote APC approach including more performance monitoring, more consultancy-level support, and more operator training. This is a prescription to work harder, not smarter, and ignore past lessons. This prescription is likely throwing good money after bad, if the real nature of APC limitations means changes are needed at the design stage.
Clamped MVs and low MV utilization
Focusing on a single negative aspect of APC has been clamped MVs, wherein MV limits are set so narrowly that the APC controller cannot move them up or down, or worse APC moves them to a less desireable value resulting in operator’s clamping them.
This is one of many takes on how MVs quickly become clamped, often cumulatively rendering an APC “defeated” (on, but effectively off). Further detailed explanation is available, but MV clamping can also be understood on a general level. To someone thinking in terms of the APC control algorithm, rather than in terms of plant operation, any MV can be used to control any controlled variable (CV) for which it has a model, and vice versa. Since processes tend to have lots of interactions, there appears to be lots of opportunity for control and optimization, but the reality is often the opposite in actual plant operation—most MVs have only one CV for which it is practical to effect control, and usually it is not desirable to use that MV to control other CVs, or to control that CV with other MVs. This surely will sound insurrectionary to the APC theorists, but this rule of thumb is generally accurate, as evidenced by the widespread condition of clamped MVs.
To put it another way, while processes are highly multi-variable with many MV/CV interactions, process control in practice tends to be mainly single-variable in that there is often a much preferred pairing of MVs and CVs, and often more reasons not to extend to multi-variable control.
Notice it is desirable not to use some MVs at all. For example, crude column pressure and heater outlet temperature, which APC theorists often see as highly promising MVs, affecting many CVs. In practice, most people learn that it is usually best to keep these setpoints fixed (often an optimized condition), rather than to move them in real time as MVs. APC theory will not tell you this, but crude column operation will usually dictate it.
The problem is not due to operating personnel failing to appreciate the economic or control value of utilizing multiple MVs, as APC personnel historically suppose. Rather, the case more often is the APC personnel do not fully realize the nature of plant operation and process control. Zak Friedman, APC Technical Consultant, writes of seeing an APC with only two of 20 MVs unclamped. In my experience, as well, this is not necessarily a rare observation.
Another way to understand the situation is to notice smart control systems have been around a long time, yet even traditional multiple-input single-output (MISO) control strategies have found relatively limited application (relative to today’s pervasive “big matrix” APCs). If there was no driving force to close a loop or configure an override on a particular MV prior to APC, the availability of APC does not necessarily change that. A common misperception is APC brings many previously unused MVs into play for control and optimization, that is actually not true and does not even make sense—if historically it has not been desirable to cascade even one CV to a particular MV with a traditional regulatory control or advanced regulatory control (ARC) cascade, why would it become desirable to cascade multiple CVs to that MV using APC?
The MVs that survive unclamped in APC are typically the ones that were already under some form of cascade control prior to APC. The main exception to this is inferentials, which can bring new MVs into play, just as a new analyzer would, i.e., not due to anything special about APC. Remember inferentials are easily implemented without APC.
Indeed, the most common approach to APC support is to check active constraints in the morning and work with operations to correct those that are not desired. But this is exactly how processes were managed prior to APC, except the unwanted conditions were flagged with alarms only, not control overrides. The difference is not that big, and one could conjecture that where overrides were preferred, they were configured long ago, since bumpless override algorithms have been in distributed control systems (DCSs) since the 1980s. Looked at this way, APC, in many cases, may not be a step forward at all, yet APC users frequently beam at having overrides on every interaction as a stroke of technology par excellence, even though the end result is often clamped MVs, not better control.
Taking opposite approach to APC design
This perspective suggests approaching APC matrix design from the opposite direction. In the traditional “big matrix” approach, an envelope is drawn around the entire process, and all interactive variables are included in the APC matrix by default. On a crude unit, this typically results in a matrix size of 20x50 or larger. But ultimately, the effective matrix can be as small as two MVs. And all the defeated variables contribute mainly to operator (and engineer) confusion, in addition to several direct cost implications.
Approaching the problem from the opposite direction means looking at existing regulatory and ARC controls to identify groups of related controllers known to be important (to stability or economics) and may already be under a cascade control strategy, and may involve feedforwards or overrides. Such groupings reflect real APC opportunities, probably with matrix size in the single digits.
A good example from the power industry is “coordinated control” of power generation (megawatts) and turbine steam pressure (although the jury is still out whether this out-performs traditional ARC). A likely example from oil refining is a crude column where, instead of a big matrix, a small matrix is implemented around product draw inferential controls, which is the source of most crude column APC benefits.
For years, APC theorists ignored the persistent clamped MV syndrome, and now we are beginning to understand why. Similarly, the APC community traditionally ignores the paradox surrounding APC benefits. On the one hand, fantastic ROIs are claimed, amounting to millions of dollars per year; while on the other hand, sophisticated “performance monitoring” and other data acrobatics are usually unable to show these earnings convincingly, and operations do not usually feel them palpably. Connecting these dots (between defeated APCs and missing-in-action benefits) is only too obvious to anyone not still in denial about them. Most tangible “APC benefits” stem from inferential control or base layer improvements, not APC.
Perhaps even more sobering than the size of benefit claims is the realization that the average APC may actually have something closer to a 0% return and the large financial sums and expensive talent used represent a huge opportunity cost to business for 20 years with an installed base of APC that is possibly more of a support burden than a profit-making asset. Hindsight is 20/20; it is easy to imagine a more lasting and fundamental improvement had more of these resources been invested in smart regulatory and advanced regulatory controls, as well as other important process control competencies.
Time for black armbands?
As a gimmick, industrialized personal computer vendors sported black armbands at trade conferences in the 1980s, implying the imminent demise of the programmable logic controller (PLC). Today, neither the PLC nor the PC dominates, and both have found their place in process control. A similar trajectory can be expected for APC, so I would not get my black armband back out just yet.
But a couple changes do seem likely including a more discriminating approach to APC application selection and matrix design, as described above. Simply putting a big matrix on every process is decidedly not best practice. (Or: “When you find yourself in a hole, stop digging!”)
Another positive change will be greater emphasis on applying the intelligent control capabilities at the DCS and ARC level, which are actually much more adaptable to the wide variety of control and automation challenges facing most facilities—APC is a good algorithm, but it is only one algorithm; DCSs have hundreds.
There are many process control competencies of equal or greater importance than APC. Going forward, the process control community needs to balance these and, when it comes to promising new technologies, and remember to keep their eyes on the cattle, not the hats.
ABOUT THE AUTHOR
Allan Kern (email@example.com) has 30 years of process control experience and has authored several papers on multivariable control, expert systems, and decision support systems, with emphasis on operation and practical process control effectiveness. He is a licensed professional engineer (inactive) and a senior member of ISA.