Feedforward control is a control strategy that calculates an action from a measured disturbance, a desired reference, or a model of a system, rather than waiting for an error in the controlled output. In control theory, it complements negative feedback: feedforward acts on information about what is expected to happen, while feedback acts on the difference between desired and observed behavior. Feedforward can improve disturbance rejection and reference tracking, but its effectiveness depends on model accuracy, timely measurements, and the physical capabilities of the equipment.
Operating principle
A controlled system, often called the plant, receives a manipulated input and produces an output. Disturbances are additional influences that affect that output but are not directly commanded by the controller.
Consider a heater maintaining the outlet temperature of a flowing liquid. If the incoming liquid becomes colder, an outlet-temperature feedback controller responds only after that change affects its measurement. A feedforward controller instead measures inlet temperature and flow rate, estimates the additional heating requirement, and changes heater power before a substantial outlet-temperature error develops.
Two important forms are:
- Disturbance feedforward: the controller measures or estimates a disturbance and calculates an opposing control action.
- Reference feedforward: the controller calculates the input expected to produce a requested setpoint or trajectory.
Feedforward does not necessarily predict the future. A disturbance measurement can provide useful advance information simply because it is taken upstream of the controlled variable. Reference feedforward can likewise act on a command already available to the controller.
The defining distinction is the information used to generate the action. A feedforward path does not correct its action from the resulting output error. Nevertheless, it can operate inside a larger closed-loop system that also uses feedback.
Mathematical description
Disturbance compensation
For a linear time-invariant system, a simple model is
where is the output, the manipulated input, a disturbance, and the Laplace-transform variable. The transfer functions and describe how the input and disturbance affect the output. The model is commonly expressed in deviations from a steady operating point, with zero initial conditions.
If the feedforward action is
the disturbance contribution becomes
Ideal cancellation therefore requires
This result expresses the basic design principle: the controller must reproduce the disturbance's effect through the manipulated-input path, with the opposite sign and appropriate timing.
Exact cancellation is possible only if the required compensator is physically realizable and the model and disturbance measurement are sufficiently accurate. Merely setting a constant gain may cancel the steady-state effect while leaving a substantial transient error.
Reference tracking
For reference feedforward, an idealized choice is
where is the desired output. If the inverse can be implemented and the model is exact, the output follows the reference.
In practice, designers often specify a desired response and seek a realizable approximation to
Choosing a suitably smooth desired response can avoid demanding instantaneous changes, excessive control effort, or amplification of high-frequency signals.
Combination with feedback
A common arrangement adds feedforward terms to the output of a feedback controller:
where is the feedback controller. Substitution into the plant equation gives
Feedforward shapes the expected response, while feedback reduces errors left by imperfect compensation. A PID controller is frequently used for the feedback component in industrial systems.
In this ideal linear arrangement, adding an external feedforward path does not change the feedback characteristic equation . It does not, however, stabilize an otherwise unstable feedback loop, and an unstable feedforward filter is itself unacceptable. Physical nonlinearities and actuator limits can also invalidate conclusions drawn from the linear equations.
Main implementations
Static feedforward
Static feedforward uses an algebraic relationship, lookup table, or operating map without explicitly modeling transient behavior. Common foundations include a mass balance, an energy balance, or experimentally measured steady-state behavior.
For an ideal flowing-liquid heater, the required thermal power can be approximated by
where is mass flow rate, is specific heat capacity, is desired outlet temperature, and is inlet temperature. Conversion efficiency and heat losses require additional terms when significant.
This calculation establishes the expected heating load but does not by itself describe fluid residence time, heater thermal inertia, or changing heat-transfer conditions.
Dynamic feedforward
Dynamic feedforward accounts for the different response speeds and delays of the disturbance and control paths. Implementations may use lead–lag compensators, delays, or other signal filters.
For example, if
then the ideal compensator is
If , the delay term can be implemented as a nonnegative delay. If the disturbance affects the output sooner than the control action can, so that , exact cancellation requires advance knowledge of the disturbance. Without such preview, a causal controller generally cannot eliminate the initial effect.
Model-based motion feedforward
In robotics and motion-control systems, a desired trajectory supplies position, velocity, and acceleration information. A model can convert these into expected forces or torques.
A common rigid-body model gives the feedforward torque as
where is the desired configuration, the inertia matrix, the velocity-dependent term, and the gravitational term. Friction compensation may also be included. Feedback corrects deviations caused by load changes, modeling errors, and external forces.
Applications and development
Feedforward compensation developed as part of industrial automation and servomechanism engineering, rather than as a single device or invention. Analog implementations used mechanical, pneumatic, hydraulic, or electronic computation. Digital controllers made it easier to incorporate nonlinear calculations, operating maps, multiple measurements, and changing model parameters.
Important applications include:
- Thermal and chemical processes: in process control, measured feed flow, composition, or temperature can adjust heating, cooling, or reagent addition before a product-quality deviation develops. Heat exchangers are a common example.
- Motion systems: electric motor drives use velocity, acceleration, or estimated load information to reduce tracking error during rapid movement.
- Electrical conversion: power electronics controllers can use measured supply or load conditions to adjust switching commands before the regulated output changes substantially.
- Building systems: air-conditioning controls can incorporate measured or forecast thermal loads while retaining room-temperature feedback.
- Predictive process operation: model predictive control can incorporate measured disturbances and future reference values, combining anticipatory action with repeated feedback-based optimization.
A scheduled command or fixed operating map may provide feedforward action, but not every open-loop command is meaningfully compensatory. The term is most useful when the command is explicitly related to a desired response or a known influence on the plant.
Design requirements and limitations
Measurement and model accuracy
Disturbance feedforward requires a usable measurement or estimate. An unmeasured disturbance cannot be directly compensated through that path. Sensor bias, noise, latency, and incorrect calibration can produce an inappropriate control action.
Model parameters may change with operating conditions, equipment aging, fouling, or product composition. A compensator accurate at one operating point may perform poorly elsewhere. Nonlinear models, operating-point-dependent gains, and parameter updates can extend the useful range, but they do not remove uncertainty.
Causality and inverse-model limitations
A plant inverse may require future information, unbounded high-frequency amplification, or unstable dynamics. In particular, direct inversion of a plant with right-half-plane zeros generally produces an unstable inverse. Such nonminimum-phase behavior prevents straightforward exact cancellation by a stable causal controller.
Practical feedforward therefore often uses an approximate inverse, filtering, preview information where available, or a deliberately slower target response. These choices trade ideal tracking or cancellation against realizability and sensitivity.
Actuator constraints
An actuator has limits on magnitude, speed, and sometimes direction of action. A heater cannot supply negative heat unless a cooling mechanism is also available, and a motor cannot deliver unlimited torque.
When feedforward and feedback commands are combined, they may exceed these limits. Saturation can leave a large residual error and cause integral windup in the feedback controller. Implementations commonly coordinate command limits, feedback integral handling, and the transition between operating modes.
Residual errors and robustness
Feedforward alone does not automatically correct an unknown bias, an unexpected disturbance, or a persistent model error. Its performance must therefore be distinguished from the robustness supplied by feedback. Robust control methods address uncertainty explicitly, whereas a nominal feedforward calculation primarily compensates effects represented in its model.
Assessment typically considers peak output deviation, settling time, tracking error, control effort, and sensitivity to parameter changes. Testing under altered loads, measurement noise, and actuator limits reveals limitations that an exact-model simulation can conceal.