What has industry learned about model-based multivariable control?
And where do the lessons lead going forward?
By Allan Kern, PE
When model-based multivariable control made its debut in the 1980s, it was expected that process models, once acquired through a plant step test, would be durable and long lived. However, this assumption proved to be mistaken, revealed in the form of "clamped" manipulated variables and "degraded" performance. Multivariable control technology has struggled with this problem ever since.
Experience has shown that most models are short lived, and many are essentially a moving target. They depend on feedstocks, feed rates, product grades, equipment health, catalyst condition, and many other factors. The model-based control community has been slow to recognize this, but in retrospect, the single-loop community has known it all along, and industry attempts at auto-tuning have also recently refreshed this lesson.
In single-loop control, feedback is universally the first choice, due to its low cost, ease of implementation, and high rate of success and long-term reliability. Feedforward (which is the single-loop equivalent of model-based control) is only used sparingly, because of its greater difficulty of implementation and risk of unstable performance due to model error, sooner or later. (Boiler drum level control is almost the only known example where feedforward is routinely used in initial control design, made possible by the durable and instantaneous relationship between steam out and feedwater in. Few other models are so accommodating.)
The term "model-based multivariable control" has been so ubiquitous that many people do not realize multivariable control, no less than single-loop control, can be implemented based on feedback, with or without the addition of feedforward models. Feedback multivariable control has the potential to eliminate more than 90 percent of models, if the historical incidence of feedforward in industry's installed base of single-loop control (less than 10 percent) is any indication.
Industry's recent experience with auto-tuning also reminds us of the wisdom of relying on feedback and avoiding the use of feedforward models to the extent possible. Like model-based control, auto-tuning initially gave industry hope that the recurring task of loop tuning might be solved once and for all. But auto-tuning has largely gone dormant now, having run into the same conundrum: If models change, then updating them, even in real-time, is not the right answer (because the next disturbance may not be the same as the last disturbance). This is especially relevant as the model-based control community weighs adaptive modeling, which is essentially auto-tuning on a multivariable scale.
Optimization belongs in the business layer
After model maintenance, the embedded optimization programs that are integral to conventional multivariable control technology form a large part of application cost, maintenance, and complexity. Fortunately, experience and evolving technology make it possible to leverage optimization results from the business layer and eliminate optimization from the control layer, where it becomes largely redundant and is probably inappropriate.
The business layer has a much more complete and global optimization solution, because it has access to much more information. With today's technology, business-side optimization solutions (or parts thereof) can be updated at higher frequency if necessary. Any results that affect constraint limits or optimization targets in the control layer (which in practice are actually quite few) can easily be pushed down to the control layer, either via connectivity or simply via the operating chain of command. Both of these methods are common practice in industry today.
Moreover, optimization at the control network layer is probably inappropriate from a process control and automation principles point of view. End users expected multivariable control maintenance and support to decrease as the technology matured, but instead they have steadily increased with no sign of abating. As this has unfolded, many have lost sight of the fact that industrial automation applications should be robust and deterministic, and carry minimal support and maintenance requirements, in order to minimize unnecessary activity on the control network, for reliability and security purposes. Optimization by its nature does not meet these application criteria.
If it is hard to imagine (at first) that the (often esteemed) task of multivariable control and optimization can be accomplished without large burdens in modeling, optimization, and maintenance, then it helps to realize that this, too-the essential role of multivariable control in industrial process operation-has also come into a more practical and realistic focus with the hindsight of experience.
Multivariable control, traditionally viewed as a complex, monolithic piece of automation with often difficult-to-discern objectives and benefits, can now be seen as a fundamental aspect of nearly every industrial process operation, and (therefore) as a fundamental part of process control and automation going forward. The functional specifications of multivariable control applications can be succinctly captured in matrix diagrams, with the objectives and benefits as plain as those of closing any loop-multivariable control closes the loop on the matrix. (See sidebar.)
In this emerging multivariable control paradigm:
- Multivariable control, like single-loop control, is predominantly feedback, with selective (not wholesale) use of feedforward models.
- Optimization is removed from the control layer and leveraged as necessary from the business layer.
- Applications can be succinctly defined using a matrix format, with scope, objectives, and benefits as plain as closing any loop.
With the elimination of 90 percent of modeling and optimization, multivariable control becomes much simpler, more agile and affordable, easier to use, and more operation friendly overall. Smaller, more numerous applications can be expected to proliferate, since optimization will no longer drive monolithic matrix designs. Table 1 lists additional potential aspects of an emerging multivariable control paradigm.
Multivariable control: Closing the loop on the matrix
When console operators adjust controller set points and outputs in the course of a shift, they are doing manual multivariable control. When a piece of process automation is deployed that manipulates those set points and outputs automatically, that is automated (or closed-loop) multivariable control (also known as advanced process control [APC]).
Multivariable control simply means adjusting the available single-loop controllers to keep related process variables within limits, and to move toward more optimal operation to the extent possible. This is an inherent and natural aspect of nearly every industrial process operation, from a feed drum to a crude oil distillation column (figure 1).
Perhaps the most important contribution to industry from the model-based multivariable control era will ultimately prove to be-not model-based control-but popularizing the concept of the matrix within the process control and operation communities. The matrix provides a concise method to diagram the multivariable nature of any process. A matrix consists of the available set points and outputs along one axis (also known as "handles" or manipulated variables [MVs]), the controlled variables (CVs) along the other axis, and models (or at least model directions) at various locations within the matrix, which indicate which MVs can be used to control which CVs.
A knowledgeable team, which comprises process engineers, control engineers, and experienced operations personnel, can develop a matrix diagram, including constraint limits, optimization targets, and other key parameters, for most processes in a single meeting (with no plant test). The matrix then becomes the heart of multivariable control, whether manual or automated. (Matrix diagrams have not migrated out of the multivariable control space to become common operation and training aids, but a good case can be made for availing matrix diagrams in this role.)
Automated multivariable control has obvious benefits over manual multivariable control, just like closing any loop. Benefits come in timeliness and consistency of action, reliability of constraint control, and maximizing optimized operation to the extent possible. Automated multivariable control also offloads this time-consuming task from operators and process engineers, who then have more time for other priorities.
Figure 1. When an operator adjusts controller set points in the course of a shift, that is manual multivariable control. When this is automated, that is closedloop multivariable control. Multivariable control is an inherent aspect of nearly every industrial process operation, from a feed drum to a crude oil distillation column. Automated multivariable control brings the same important benefits as closing any loop.
Figure 2. Model-less multivariable control (XMC) automates the way operating teams have always carried out multivariable constraint control and optimization manually. Notably, this method does not require detailed models or embedded optimizers. XMC uses rate-predictive control as its internal feedback control mechanism.