I have been working through some forehearth control questions recently. They invariably come back to the same fundamental objective: establishing and maintaining repeatably high thermal stability in a moving mass of glass, despite process and ambient disturbances.
That leads quickly into the question of how best to achieve it—and into the increasingly imprecise language surrounding “advanced process control”. Advanced Process Control (APC), Model Predictive Control (MPC), adaptive control, self-tuning control, fuzzy logic, expert systems and, increasingly, AI are often discussed as though they were competing versions of the same thing. They are not.
Some describe broad categories of control; others describe particular mathematical methods or capabilities. Several can coexist within the same control system. Before deciding which is appropriate for a particular process, it helps to establish what these terms actually mean.
Start with PID
PID remains the foundation of industrial process control for good reason. It is inexpensive, familiar and available in essentially every PLC and DCS. For fast, reasonably linear and weakly interacting processes, a properly implemented PID loop can provide excellent control. Flow, pressure and many local temperature applications require nothing more.
Even PID, however, is not entirely uniform. Manufacturers differ in algorithm form, gain conventions, integral and derivative definitions, filtering, anti-windup provisions and digital implementation. Two controllers can therefore require different numerical tuning to produce essentially the same process response.
The name of the method tells us less than we sometimes assume.
APC is a category, not an algorithm
Advanced Process Control (APC) is best understood as an umbrella term for methods that extend or supplement conventional regulatory control. It can include feedforward and dead-time compensation, adaptive and rule-based control, constraint handling, multivariable control and model predictive control.
Thus:
MPC⊂APC⊂Process Control
MPC is an APC technology. APC is not necessarily MPC. This distinction is often blurred in commercial literature, where APC may simply refer to a supplier’s particular control package.
Model-based control and MPC
Model-based control is another broad category. It means that the controller explicitly incorporates some mathematical representation of process behavior—gain, time constant, dead time, interaction or more complex dynamics.
A Smith Predictor, for example, uses a process model to compensate for dead time. It is model based, but it is not normally considered MPC.
Model Predictive Control (MPC) goes further. It uses a dynamic process model to predict how the process will respond to potential control actions. It calculates control moves over a future horizon, implements the appropriate current action, measures the resulting process condition and calculates again.
Conceptually:
Measure→Predict→Calculate→Act→Measure again
That predictive capability becomes particularly valuable when the consequence of a control action does not appear immediately—as in a forehearth.
Adaptive, self-tuning and autotuning
Adaptive control means that some aspect of the controller changes as observed process behavior changes. It might modify PID gains, update model coefficients or select between models.
MPC can be adaptive but need not be. PID can also be adaptive without becoming MPC.
Self-tuning is a form of adaptation in which the system identifies process behavior and adjusts its controller parameters accordingly. A self-tuning PID remains a PID.
Autotuning is different again. It usually describes an episodic procedure used to determine controller settings. Once tuned, the controller may continue indefinitely with fixed parameters and therefore is not adaptive at all.
Fuzzy logic and expert systems
Fuzzy logic control represents process relationships through degrees of membership and rules rather than requiring the same explicit dynamic model used by conventional model-based methods. It can be effective where experienced operators understand process behavior qualitatively even when the relationships are difficult to model analytically.
Expert or rule-based systems similarly encode operating knowledge into logical decisions. They may change setpoints, select operating modes or intervene under defined process conditions.
Neither fuzzy nor rule-based control is inherently predictive or adaptive, although either can be combined with other methods.
RTO: optimizing where to operate
Real-Time Optimization (RTO) operates at a different level. Its principal purpose is not to hold a process variable at its target, but to determine where the process should operate to best satisfy economic or production objectives.
An RTO system can use current plant conditions, process constraints and economic information to establish preferred operating targets. Those targets are then implemented by MPC, supervisory control or conventional regulatory loops.
In simplified terms:
RTO: Where should we operate?
↓
MPC/APC: How do we get there and remain there?
↓
Regulatory control: Execute the required actions
MPC itself involves optimization, but generally over a shorter prediction horizon and primarily to achieve control objectives. RTO normally addresses the higher-level economic operating point.
And then there is “AI”
Artificial intelligence may now be the least precise term of all.
Neural networks, machine learning, fuzzy systems, adaptive identification, reinforcement learning and combinations of conventional control methods can all appear in systems described as AI. The label is also increasingly attached to control mathematics that predates modern AI by decades.
“AI-based control” therefore tells us little unless the underlying method is identified.
For an industrial control system, the more useful questions are:
What does it measure? What does it model? What does it predict? What does it manipulate? How does it handle disturbances, dead time and interaction? And what happens when actual process behavior departs from the model?
Those answers tell us considerably more than the marketing category.
These categories overlap. A system could legitimately be described as adaptive, model-based, multivariable predictive control operating above conventional PID regulatory loops.
That tells us something. “Advanced AI control” does not.
The more important question for glass
Glassmaking contains plenty of variables for which conventional PID is entirely adequate. There is little economic justification for replacing a well-performing flow, pressure or straightforward regulatory loop simply because more sophisticated mathematics is available.
But some of the most valuable variables in the glasshouse present a different problem.
Glass temperature, glass level and thermal conditioning can involve substantial dead time, long process memory, interaction and disturbances whose consequences continue developing long after the initiating event. What we measure now may substantially reflect actions and disturbances that occurred many minutes—or hours—earlier.
That changes the objective.
It is no longer simply:
How well can we tune the feedback loop?
It becomes:
How much better can we control the process if the controller understands its dynamics and uses that knowledge to anticipate where the process is going? The short answer is often by a lot.
It is also the question that leads us toward model predictive control.