Feedforward Control
Feedforward acts on a measured reference or disturbance before it affects the output, complementing feedback's after-the-fact correction.
Acting Before the Error
Feedback control corrects errors only after they appear in the output. Feedforward control acts earlier: it uses a model of the plant to compute, from a measured reference or disturbance, the input needed to produce the desired output directly, without waiting for an error to develop. Feedforward anticipates; feedback corrects.
Two uses of feedforward
- Reference feedforward: given a target trajectory, compute the input that would produce it from an inverse plant model, so tracking is fast and does not rely on error buildup.
- Disturbance feedforward: measure a disturbance directly and inject a compensating input before it perturbs the output.
- Both require a plant model; their accuracy is limited by how good that model is.
The combination with feedback
Feedforward and feedback are complementary and almost always used together. Feedforward handles the large, predictable component of the command or disturbance quickly and without lag, while feedback cleans up the residual caused by model error and unmeasured disturbances. This division lets each do what it does best: feedforward for speed, feedback for robustness.
Why feedforward alone is not enough
Pure feedforward is open loop: it has no way to correct for model inaccuracy, unmeasured disturbances, or drift, because it never looks at the actual output. A small modeling error accumulates uncorrected. This is exactly why feedback is added, to provide the closed-loop correction that feedforward structurally cannot.
Practical requirements
Disturbance feedforward requires a sensor on the disturbance itself, which is not always available. Reference feedforward requires a stable, causal approximation of the inverse plant, which is problematic for non-minimum-phase plants whose exact inverse is unstable. In such cases an approximate or filtered inverse is used.
Large fusion machines use feedforward heavily: precomputed coil-current and heating trajectories drive the discharge along a planned path, while feedback trims the residual position and shape errors that the feedforward plan cannot foresee.