PID Tuning For Fast-Ramp Temperature Chambers

Three terms, one ramp

Three terms, one fast ramp.

What PID does

A PID controller is the standard recipe for steering a chamber's temperature. It looks at the difference between the temperature it wants and the temperature it has, and from that error it builds a drive in three parts: proportional to the present error, integral of all past error, and derivative of the error's current change. Together the three steer the heaters and cooling toward the setpoint.

For a fast-ramp chamber the recipe needs care. A long ramp from one extreme to another keeps the error large and persistent in a way no quiet setpoint ever does, and the three terms behave differently under that strain, so the tuning for such a chamber is its own craft. Why the chamber overshoots at all is a matter set out elsewhere; this is how the controller is shaped to ride the ramp without sailing past.

The integral is the villain on a ramp

Of the three terms, the integral is the one that misbehaves on a ramp. Proportional and derivative live in the moment, the P term reading the present error and pushing in proportion, the D term reading how fast the error is changing and damping the response; both see a moving target and adjust to it as it moves, neither building up a debt the controller will later pay. The integral is different. By design it accumulates the error over time, summing it into a stored correction that is meant to drive the steady-state offset away, and on a ramp that summing runs for a long while. While the temperature climbs toward a far setpoint, the error stays large and persistent for second after second, and the integrator dutifully adds each of those errors to its memory, growing a great pile of correction that has nowhere to go. The actuator, meanwhile, is likely saturated for much of that climb, the heaters at full or the cooling at full, so the controller is asking for more than the plant can give and the integrator goes on integrating against a wall. By the time the setpoint is finally reached, the integrator holds a sum the air no longer needs, and it goes on spending that sum into the actuator, driving the heaters past the moment of cutoff or the cooling past the moment of release, and the temperature sails past the mark. That is integral windup, the central problem of tuning PID for a fast ramp, and it explains why a controller that holds a setpoint beautifully will overshoot a ramp by a wide margin under the same gains. The cure is to keep the integrator from winding up in the first place, either by halting its accumulation while the actuator is saturated or by gently undoing what it adds against a saturated output, the family of fixes called anti-windup. Done well, anti-windup lets a brisk I gain serve a fast ramp without paying for it in overshoot when the temperature arrives; done poorly or not at all, the controller fights its own memory at every setpoint.

Anti-windup first

Anti-windup comes first for a ramp.

Anti-windup in practice

Two common forms of anti-windup serve a fast ramp. The first is to clamp the integrator: when the controller's output has saturated, the heater at full or the cooling at full, the integral is held steady and not added to, so its memory grows no further until the output comes off its limit.

The second is back-calculation. The integrator is given a small feedback that subtracts the saturated excess from its own value, gently undoing what the integration was overshooting on, so the integrator drifts down as the actuator runs against its limit rather than piling up unspent.

Each has its tuning, the clamp's choice of when to engage and the back-calculation's gain, but the goal is the same: keep the integrator from carrying a sum the setpoint will never need, so that when the temperature arrives the controller is not still holding yards of correction it must spend on the air past the mark.

Feedforward sees the ramp coming

A coming ramp can be met by feedforward in advance.

Feedforward against the ramp

Feedforward sees what the controller would otherwise have to chase. A ramp is a known disturbance, a setpoint moving at a chosen rate, and a feedforward term can compute the steady drive that rate alone needs and add it to the PID's output before the error ever builds.

With feedforward in place the PID has less to do. The bulk of the drive comes from the feedforward term, set by the ramp's rate and the chamber's known gain, and the PID only corrects the residual, the small departure the model could not predict, which keeps its integrator a great deal smaller.

A chamber with a known load can carry this further, the feedforward shaped to anticipate the load's own draw as well as the ramp, so the controller is told in advance what will be needed and only fine-tunes the rest. The art is in the model the feedforward is built from.

Heat and cold pull through different paths

Heating and cooling pull through different paths, with different gains and different lags.

Split gains for heating and cooling

A temperature chamber is not symmetric. Its heating side pours energy in through electrical elements and warms the air with little delay, while its cooling side draws energy out through a refrigeration loop whose compressor, coils, and refrigerant have their own time constants and authorities.

A single PID tuning that pleases one side will misbehave on the other. Gains right for cooling may make the heating overshoot wildly; gains right for heating may leave the cooling sluggish and slow to settle.

Split-range or gain-scheduled PID gives each side its own tuning. The controller carries two sets of P, I, and D, one for the heating range of its output and one for the cooling, and switches between them as the drive crosses zero, so each side is served by gains shaped for it.

On a fast asymmetric chamber this is not a refinement but a basic. A single set of gains forces a compromise that tames the worse side and underuses the better one; split tuning lets each side be driven as hard as it safely can be, which is part of how the ramp rate is won.

Sharp tool, careful hand

The derivative term is sharp, and a sharp tool magnifies any noise it touches.

Filtering the D term

The derivative term, useful for damping a fast response, has a quirk: it amplifies whatever shape the error is currently changing through, including the small jitter of a noisy sensor. Untouched, D translates sensor noise straight into actuator chatter, the heater and cooling buzzing for no real reason.

The fix is a filter. A low-pass on the derivative, or taking the derivative on the measured value rather than on the error, smooths the term enough that genuine slope shows through and noise does not. Without it, a chamber will sound like it is fighting itself even when the temperature is settled.

Cascade is a powerful frame

Beyond the single PID, the chamber's controller often runs in cascade. An outer loop watches the air's temperature against the setpoint and writes its output not to the actuators but to a second, inner loop, which in turn regulates the heater duty or the coil temperature against that command.

The two loops are tuned to different paces. The inner loop runs fast and handles the actuator's own quirks, smoothing the non-linearity of a heater bank or a refrigeration loop into a clean response the outer loop can steer through. The outer loop runs slower and concerns itself only with the air.

Cascade is at its best where a single loop would have to take the actuator's awkwardness into its own gains. Pushing that work down into the inner loop lets the outer loop be tuned for the air alone, and gives the fast ramp a steadier ride than one big loop could manage.

What the tuner must do

Start from a model rather than from a guess. Even a rough first-order picture of the chamber, its gain and its time constant, gives a starting set of gains a useful shape, and refining from there is faster than groping in the dark.

Add anti-windup before the I term is tuned at all. Any integral gain that helps a setpoint will overshoot a ramp without anti-windup beneath it, so the fix goes on first and the tuning then proceeds on the controller that is going to run.

Build feedforward for the ramp rate, and for the load if its draw is known, so the PID is fine-tuning a residual rather than chasing a full disturbance from scratch.

Split the heating and cooling tunings, since one set of gains cannot serve both sides of an asymmetric chamber without leaving one side overshooting or the other dragging.

Filter the derivative and take it on the measured value, so the D term sees slope and not noise, and the actuator is not made to chatter for a sensor's flutter.

Autotune is a start

An autotuner offers a starting point, not the finished tuning yet.

When the autotuner falls short

Many chambers carry an autotuner of some kind, the controller running a short test, a step or a relay oscillation, and computing gains from what it sees. The numbers it returns are a fair first cut for a setpoint-holding job.

On a fast ramp those autotuned gains usually fall short. The tuner has measured the chamber as if it were holding a setpoint, not climbing toward one, so the integrator's behaviour under a long error and the asymmetry between sides go unmeasured. A manual refinement, guided by what the ramp does to the autotuned response, is what finishes the tuning.

The craft that finishes the chamber

A fast-ramp chamber is hardware and control together. The plant supplies the capacity; the controller spends it well; and the tuning is what marries them, deciding how that capacity is applied to the ramp the test asks for.

Done well, the chamber rides its ramp without sailing past, the integrator restrained, the feedforward carrying its share, the two sides each tuned for their own work, and the derivative seeing slope rather than noise. The craft is patient, and it earns the patience a hurried lab does not always wish to give it.