Feedforward Control: Correcting Disturbances Before They Arrive

Feedforward Control: Correcting Disturbances Before They Arrive is a control strategy that measures a known, measurable disturbance and applies a calculated corrective action to the manipulated variable before that disturbance can move the process variable. Unlike feedback, which waits for an error to appear at the output and then reacts, feedforward acts at the source. When the disturbance is well understood, feedforward can hold a process variable nearly flat through upsets that would otherwise cause large, slow excursions.

Feedforward Versus Feedback

A feedback controller closes the loop on the controlled variable. It compares setpoint to measurement, computes an error, and drives the final control element to remove that error. This is robust and self-correcting, but it is fundamentally reactive: the process variable must first deviate before the controller responds, and with long dead time or large lags that deviation can persist for minutes.

Feedforward inverts the logic. Instead of watching the output, it watches an incoming disturbance. When the disturbance changes, the controller predicts how it will affect the process and adjusts the manipulated variable in the opposite direction so the two effects cancel at the process variable. A well modelled disturbance is rejected before it is ever seen at the output. But feedforward is blind to anything it does not measure and depends entirely on its internal model.

The Process Model: Static Gain and Dynamic Lead-Lag

Feedforward requires an explicit model relating the disturbance to the correction needed. That model has two parts:

• Static gain. The steady-state ratio between a change in the disturbance and the change in manipulated variable required to offset it. For example, a drop in feedwater temperature may require a proportional increase in heating.
• Dynamic lead-lag. The disturbance and the correction usually reach the process variable through different paths with different time constants and dead times. A lead-lag block shapes the timing of the correction so it arrives neither too early nor too late. Getting the dynamics wrong produces a transient overshoot in the wrong direction even when the static gain is exact.

A first-order-plus-dead-time approximation of each path is often enough, but the model must be identified from real plant data, not assumed.

Why Feedforward Is Almost Always Combined With Feedback Trim

No feedforward model is perfect. Gains drift, sensors read slightly off, and unmeasured disturbances still enter the process. Pure feedforward has no way to detect or correct its own residual error, so it will slowly wander off target. So feedforward is nearly always a supplement to feedback, not a replacement. The feedforward path handles the bulk of a large, fast disturbance, while feedback provides trim: it removes the small residual error the model leaves behind and absorbs what the feedforward path never measured. The principles in PID controller tuning apply directly to that trim controller.

Ratio Control: A Common Feedforward Form

Ratio control is the most widely used form of feedforward. Here one flow, the wild or uncontrolled stream, is measured while a second flow is manipulated to hold a fixed ratio between them. A combustion air-to-fuel loop is the classic example: fuel flow is the measured disturbance, and combustion air is driven to hold the target ratio before flame composition drifts. Ratio control is often layered underneath a slower outer loop in a cascade control arrangement, where an analyser trims the ratio setpoint.

Common Feedforward Applications

Where Feedforward Shines, and Where It Struggles

Feedforward earns its place under specific conditions:

• Large process or measurement lags. When feedback alone would respond too slowly, acting on the disturbance directly avoids a long, costly excursion.
• A measurable, dominant disturbance. Feedforward only helps if the disturbance can be instrumented reliably and accounts for a meaningful share of the upsets.
• A stable, identifiable relationship. The disturbance-to-output path must be consistent enough to model.

It struggles when the dominant disturbances are unmeasured, when the relationship is strongly nonlinear, or when the disturbance measurement is noisy. Then the correction can inject as much error as it removes.

Model Accuracy and Periodic Review

Because feedforward performance decays as the plant ages, valves wear, and surfaces foul, the model gain and lead-lag settings should be reviewed on a schedule, not set once and forgotten. A drifting feedforward loop degrades quietly: the feedback trim masks the growing residual until it can no longer keep up. Logging the identification data and last-verified settings in a system such as Fabrico makes it obvious when re-identification is due. To see how loop-health checks fit into a preventive maintenance schedule, Book a Fabrico demo.

Frequently Asked Questions

Can feedforward control be used on its own?

It can, but it rarely should be. Pure feedforward has no mechanism to correct its own residual error or to reject unmeasured disturbances, so the process variable slowly drifts. Feedback trim adds both fast disturbance rejection and long-term accuracy.

What is the difference between feedforward and cascade control?

Feedforward acts on a measured disturbance entering the process. Cascade control nests two feedback loops so an inner, faster loop rejects disturbances before they reach the outer variable. They are complementary and often used together.

Why does feedforward need a lead-lag block?

The disturbance and the correction reach the process variable through paths with different dynamics. The lead-lag block times the correction so it arrives in step with the disturbance effect. Without it, a correct static gain can still cause a transient overshoot.

How often should a feedforward model be re-checked?

There is no universal interval, but any change in equipment condition, fouling, or operating range is a trigger. Reviewing whenever the feedback trim starts working harder than usual prevents a silently degrading model.