Traditional feedback control is much like a runner running backwards. Without looking at the track ahead, a runner running backwards must rely entirely on their rearward vision to determine their position. When looking backwards, the runner can only stay on the track by adjusting their left and right position as they get closer to the edge line. If the runner goes too fast, it's easy to run out of bounds before the application can correct it (see Figure 1).
Figure 1: Straight ahead. Like a controller with only feedback, a runner facing backward can only see the past position, but this is usually sufficient to stay on the track. Even without looking forward, it is easy to observe the vicinity of the center of the track and compensate for any slow drift to the left or right.
If slow drift is the only disturbance, a pure feedback controller can easily keep the process variable near a constant setpoint. Historical measurements of the process variable typically tell the controller what it needs to know when control problems become easier. Image source: CE
Industrial feedback controllers face the same challenge. To maintain the required temperature, pressure, flow rate, etc., of a controlled process, the feedback controller must wait and monitor the operation, then correct errors and check again. This is typically not a random trial-and-error process. Even if the controller has sufficient knowledge of the process behavior to make educated guesses about necessary corrections, these corrections must always be made retrospectively.
Therefore, like a runner facing backward, a feedback controller must proceed with caution to avoid overcorrecting past mistakes. This is especially true when designing a controller if the knowledge of process behavior is inaccurate or incomplete. If a controller cannot predict the future impact of its current control actions, it has little choice but to act conservatively over longer time intervals rather than actively over shorter ones (see Figure 2).
Figure 2: Running backwards around the curve is more challenging. Runners have already strayed off course when they notice the curve has moved the center of the track away from their path. In the case shown in the figure, the runner, while trying to compensate for the disturbance, went too far to the left and then too far to the right.
If the runner moves too fast, it will cause continuous overcompensation back and forth until the disturbance ends at the end of the curve. If the feedback controller is designed to be too strong or the controlled process is too sensitive to the controller's actions, it will exhibit similar oscillating behavior. In the worst case, even if the disturbance ends, it will not help. The controller may continue to oscillate between fully on and fully off as it continues to overcompensate over and over again.
Look forward, not backward.
The obvious solution to the dilemma faced by runners is to look forward, not backward, during the run. With foresight into the future curve, forward-looking runners will be able to make more informed distance adjustments and run faster.
A sharp-eyed runner can also look down at the track and take preemptive action to stay in the middle of the course when they see an approaching curve. The runner can begin turning left when needed, as shown in the "Run Forward" graphic in Figure 3.
Figure 3: Running straight forward is the easiest. With a view of the track ahead, even at top speed, the runner can almost instantly compensate for any slow drift to the left or right. Similarly, applying a feedforward controller to a process with finite, measurable disturbances makes it easier to keep process variables close to the setpoint.
Running around a curve isn't that difficult. Runners can visually measure any impending disturbance (the curve), predict its impact on the future trajectory, and make corrections as needed rather than afterward. This forward-looking knowledge allows forward-facing runners to run around the curve faster and with less error than backward-facing runners.
Advanced knowledge also makes feedforward controllers more accurate. If a controller can correctly predict how disturbances will affect process variables and how to compensate for them, it can more confidently take on more control work. Doing so can reduce the impact of impending disturbances, much like a runner can stay in the center of the track when predicting an upcoming curve.
Process controllers equipped with sensors capable of measuring precursory sensors for impending disturbances can operate faster and more proactively. Control actions can proceed without waiting for the results of past control efforts to be reflected in measurements. Sensors and controllers work together to observe upcoming disturbances and provide forward information to aid in calculating future control actions.
Advantages of feedforward control
A classic application of feedforward control is in steam distribution systems, such as central boilers that supply steam at a constant pressure to various machines throughout a plant. When idle machines come online and begin drawing steam from the boiler, the pressure controller can preheat and inject additional water into the boiler, provided the system can determine how much steam the machine needs.
If the controller relies heavily on feedback, it must wait until the pressure in the boiler has dropped before attempting to compensate for the additional load. If it can predict the impending disturbance, the pressure controller will be able to proactively prevent the pressure drop, whereas a feedback controller would need to measure the pressure drop before taking action.
The key to effective feedforward control is measuring impending disturbances and accurately predicting their impact on process variables. While a runner moving forward needs little to consider what to do when there's a curve ahead, a feedforward pressure controller must make less obvious decisions. It needs to know not only when a particular machine will be online, but also how much steam it will extract, and the impact of a specific extraction method on boiler pressure over time.
These predictions are typically made using mathematical models that show how a process responds to measurable disturbances. These models can be as simple as a lookup table containing the effects of disturbances in early measurements, or as complex as multivariable differential equations based on first-principles analysis or empirical observations. With technological advancements, future online learning algorithms and other forms of artificial intelligence will be able to help create or improve these mathematical models.
The perfect combination of feedback and feedforward
Since no model can be 100% accurate, and other unmeasurable disturbances can also affect process variables, feedforward controllers are almost always combined with feedback controllers. The feedforward controller makes the best guess at the control effort required to compensate for an impending disturbance, while the feedback controller makes up for its shortcomings. The feedback controller measures the net effect of the disturbance, and then the feedforward control works to compensate for any deviations in the process variables that the feedforward controller cannot avoid.
In some applications, feedforward controllers can be difficult to implement. Designing a feedforward controller becomes challenging when the process behavior is not fully understood, disturbance variables are difficult to measure, or there are too many disturbance variables. A poorly designed feedforward controller can sometimes amplify the effects of disturbances, making the operation of the feedback controller even more difficult.
Reduce energy and wear
If disturbances are frequent or too large for a standalone feedback controller to achieve its purpose, then adding a feedforward controller is highly worthwhile. A successful feedforward controller can reduce the peak values of the main disturbances in the process variables. If this also eliminates the oscillating behavior of the feedback controller, the combined feedforward/feedback controller will reduce movement, thus allowing less energy to be used. Fewer control movements also reduce wear on actuators used in process control.
Disclaimer: This article is a reprint. If it involves copyright issues, please contact us promptly for deletion (QQ: 2737591964). We apologize for any inconvenience.
