If it were possible to get on with the whole business of multivariable control without the whole business of modeling, the implications for process automation would be far-reaching. But little is yet known about what a model-less multivariable controller might actually look like (it looks a lot like a model-based multivariable controller) or how it might work (it's less complicated). This column discusses the circumstances that give rise to the idea of model-less multivariable control, while a later article (Part 2) will provide an engineering look at an actual model-less multivariable controller design.
Manual multivariable control
Multivariable control is usually considered a product of the computer age, but nearly all processes are multivariable, and multivariable control has always been with us. In the pre-computer age, multivariable control was carried out by the operating team, who adjusted controllers and valves manually to keep related process variables within constraint limits and at economic targets. This basic approach to managing the multivariable nature of most processes remains a prominent aspect of plant operations today, whether in lieu of, or in conjunction with, modern automated multivariable controllers.
That's a step on the path to model-less multivariable control – to reflect that it's always been with us, albeit in manual mode. Ergo, automated model-less multivariable control should also be possible, and should bring all the benefits of timeliness and consistency that normally accrue with any automation, even if it doesn't bring 100% of the benefits deriving from the use of models (although that is by no means being conceded at this point).
Model-based multivariable control (MPC)
With the advent of computers in process control, it became possible to automate multivariable control, with (as mentioned) obvious potential to improve the quality of constraint control and optimization. Industry, armed with promising model-based theory and new-found computer power, opted early on for the model-based path. But several associated assumptions have played out unexpectedly (1):
- Nobody foresaw that achieving "model fidelity" (a combination of accuracy and durability) would be so elusive. In retrospect, this can now be seen as having the same root cause that has always plagued single-loop tuning – the process disturbances we seek to control often alter the very models (or tuning parameters) we employ to control them. (In Part 2, we will see a novel control algorithm that is inherently adaptive to dynamic changes in process gain.)
- Nobody had yet realized that "error minimization," industry's de facto control performance criteria, would prove inappropriate for many multivariable control purposes. For high-level constraint control and optimization, less aggressive and more careful control action, with an eye to managing risk and preserving process stability, is usually preferred from an operations standpoint (Figure 1).
- And nobody realized that the "big matrix" design approach would net many inappropriate and ultimately unwanted models that serve to compound these problems and compromise the control solution much more than they strengthen it. MPC engineers have instinctively trimmed matrices over the years, but few have yet realized that the sensible bottom to this trend is to use primarily the variables and models that have historically been employed and are already proven in prior manual operation, i.e. the "small matrix" design approach (2).
The general consequence of this—of industry's commitment to a technology that has not matured as expected—has been stasis. Instead of the cost and complexity of MPC coming down, they remain high. Instead of performance improving, it remains low. And instead of evolving a more affordable and agile tool, suitable to the everyday needs of a plant operating environment, MPC remains an inflexible and taxing technology to own. Meanwhile, countless medium-range multivariable control applications throughout industry remain unfulfilled for lack of such a tool.