1
P, I and D each respond to a different part of the error
Proportional acts on the error now, integral on the error accumulated over time, and derivative on how fast the error is changing. Most loops only need the first two.
PID loop tuning is the process of setting the proportional, integral and derivative gains of a controller so a process variable, temperature, flow, level or pressure, holds its setpoint accurately and recovers from disturbances without oscillating. On a food or dairy plant these loops decide product quality and safety directly: the temperature that proves pasteurisation, the flow that holds a hold-tube residence time, the level that keeps a balance tank from running dry. Metromotion Controls is a control systems integrator based in Mount Waverley that tunes and stabilises process loops across Melbourne, Victoria and Australia for dairy, beverage and food plants.
Most loops on a food site run on commissioning numbers or controller defaults, tuned by trial and error under time pressure. The plant runs, but loops overshoot on start-up, hunt around setpoint or respond slowly to disturbances, and the cost shows up as off-spec product at changeover, valves cycling to early failure, and operators running in manual because they do not trust the loop. This article covers how the three terms behave, which loops need which terms, and a structured way to set them.
This post supports our PLC, SCADA and HMI and industrial automation services. Loop tuning is part of reading a process and programming it from how it actually behaves, the same understanding that underpins our work across food and beverage. A loop is only as good as the model behind it, which means measuring the process before adjusting the controller.
What do the P, I and D terms actually do?
A PID controller works on the error, the difference between where the process variable is and where you want it. The proportional term acts on the error right now: the larger the gap, the harder the controller pushes. The integral term acts on the error accumulated over time, so a small offset that proportional alone would leave in place is driven out as it adds up. The derivative term acts on how fast the error is changing, which gives the controller a measure of anticipation, easing off before it overshoots.
Tune proportional too high and the loop oscillates; tune integral too fast and it overshoots and cycles; add derivative on the wrong loop and it chases noise. The skill is matching the terms to the process, and the first decision is usually whether derivative is needed at all.
Why do most flow, pressure and level loops run PI?
Derivative acts on the rate of change of the measured signal, and on a fast, noisy loop that signal is mostly noise. Flow measurement carries fluid turbulence and pump pulsation; pressure carries the same disturbances. The derivative term amplifies high-frequency noise, so feeding it a noisy flow signal produces a jittery output that drives the valve back and forth against noise rather than real process change, wearing the valve and positioner without improving control.
Flow and pressure loops are also fast, responding to a valve change in seconds or less, so they do not need the anticipation derivative provides on a slow loop. Proportional plus integral is enough.
Most level loops run PI for a related reason: they usually do not need to hold a tight setpoint at all. A balance tank exists to absorb variation between an upstream and a downstream flow, so the useful behaviour is averaging control: let the level move within a wide band and change the outflow slowly, so the downstream process sees smooth flow rather than every upstream wobble. That calls for a deliberately loose loop, low proportional gain and slow integral. Derivative on a noisy or sloshing level signal makes it worse for the same reason it hurts a flow loop.
| Loop type | Typical terms | Why |
|---|---|---|
| Flow | PI | Fast and noisy; derivative amplifies turbulence and pump pulsation |
| Pressure | PI | Fast and noisy; same noise problem as flow |
| Level (averaging) | PI, loose | Goal is to smooth flow, not hold a tight setpoint |
| Temperature | PID | Slow with dead time; derivative adds useful anticipation |
Why do slow thermal loops justify full PID?
Thermal loops are the opposite case. Pasteuriser temperature, glycol cooling, steam heating, hot water sets and jacket heating share two features: they are slow, and they carry significant dead time and lag. Heat has to move through a plate pack, a jacket wall or a body of liquid before the sensor registers it, so by the time the reading moves, the process has already moved further than the reading shows. A controller working on the reading alone is always behind.
This is the situation derivative is designed for. Acting on the rate of change, the controller eases off the heat while the temperature is still climbing, which reduces overshoot on a loop that would otherwise sail past setpoint because of its lag. The noise amplification that ruled derivative out on flow loops is far weaker here, because a thermal signal is slow and clean. The guidance that slow temperature loops with significant dead time benefit from derivative action reflects this, and it is why pasteuriser and heating loops are among the few on a food plant where full PID is the right choice.
Anti-windup belongs with every saturating thermal loop
Slow loops have a second problem: integral windup. During start-up a steam or hot water valve can sit fully open and still not reach setpoint for some time, and while it is saturated the integral term keeps accumulating error the valve cannot act on. When the process finally reaches temperature, the controller stays pinned at full output until that oversized integral unwinds, producing a long overshoot well past setpoint. Anti-windup logic prevents this by halting integral accumulation once the output hits its limit, and any thermal loop that can saturate, which is most of them at start-up, should have it enabled.
A structured method beats trial and error
Trial and error can land on gains that work, but it produces no record of why the numbers are what they are, no consistency between loops, and no quick way to retune when the process changes. A structured method fixes all three. The most practical for process loops is lambda tuning, a form of internal model control.
The method has three steps. First, a small open-loop step test: with the loop in manual, make a single step change to the valve and record the response. From it you read the process gain (how far the variable moves per unit of valve change), the time constant (how quickly it responds) and the dead time (how long before it responds at all). Second, choose lambda, the closed-loop response time you want: larger is slower and more robust, smaller is faster and tighter. Third, calculate the gains from the model and the chosen lambda. The approach is documented in process-control references on lambda and IMC tuning, and its value is that the engineer chooses the behaviour deliberately rather than discovering it by adjustment.
Structured tuning is auditable and repeatable: the same step test and lambda choice give the same gains, so a retune after a plate-pack change or pump replacement is a measurement rather than a guess. It also makes the choices explicit: a level loop gets a large lambda because you want it slow, while a pasteuriser temperature loop gets a smaller one because the diversion valve has to act decisively.
Why auto-tune often misbehaves on water and flow loops
Many controllers offer an auto-tune. On slow, well behaved thermal loops it often works, because the process response is large relative to the noise and the routine identifies the model cleanly. On fast water and flow loops it frequently does not: the noise is large relative to the small response the routine is trying to read, so it identifies the wrong model and returns aggressive gains that leave the loop oscillating. On those loops a short manual step test and a lambda calculation is more reliable than trusting the routine.
How are CIP loops tuned differently from production?
A circuit tuned for production will not behave the same way under clean-in-place. Flow paths and fill levels change, and the thermal load of heating hot caustic through an empty vessel is nothing like holding product temperature, so a flow loop tuned for production rate may be sluggish or unstable at CIP flow and a temperature loop may overshoot badly heating a wash solution.
The control goal changes too: under CIP the loops hold chemical concentration, temperature and flow long enough to meet a contact-time target, not a production setpoint. The practical answer is to give the CIP state its own tuning parameters, switched automatically by recipe or sequence state. This is the loop-level counterpart to the recipe and phase thinking in our guide to CIP automation: the same valve and sensor, controlled to a different goal with different gains depending on what the plant is doing.
Common mistakes that undo good tuning
Three mistakes account for most loops that will not settle, and none of them are fixed by adjusting gains.
- Leaving integral windup unhandled. The long overshoot after start-up is read as the loop being tuned too aggressively, so the gains get detuned, which makes the loop sluggish in normal running and does nothing about the windup. The fix is anti-windup, not weaker gains.
- Tuning around a sticky valve. A control valve with stiction or excessive deadband does not move smoothly with the controller output; it sticks, then jumps, and the loop oscillates no matter what the gains are. Detuning the controller hides the symptom and makes the loop slow. The fix is the valve: service the positioner, address the stiction, then tune. A loop is never better than the final control element it drives.
- Ignoring valve characterisation. A loop that is stable at one operating point and unstable at another usually has a valve whose installed flow characteristic does not match the process, so the effective process gain changes across the valve's travel and no single set of gains is right everywhere. Selecting or configuring the characteristic so flow changes roughly linearly with output lets one set of gains hold across the operating envelope.
The thread through all three is that the controller is only one element in the loop. Tuning it to compensate for a fault in the sensor, valve or process produces a loop that is poor everywhere to hide a problem that is real somewhere.
A worked example, illustrative only
The following is a generic illustration, not a measurement from any Metromotion Controls installation. Consider a hot water set heating a jacketed vessel. An open-loop step opening the steam valve 10 percent raises the water temperature by about 8 degrees Celsius after settling: the temperature first moves roughly 30 seconds after the step and completes most of its change over about three minutes, giving a process gain of about 0.8 degrees per percent, a dead time of 30 seconds and a time constant of 180 seconds. The dead time is significant relative to the time constant, which confirms the loop benefits from derivative and warns that aggressive proportional gain will make it oscillate. A lambda roughly equal to the time constant gives a robust PID setting that approaches setpoint firmly without overshoot, with anti-windup enabled so a cold start does not produce a long overswing. Push lambda below the dead time and the loop turns oscillatory, because no controller can react faster than the process can report. The numbers are illustrative; the method, measure first, then choose the response, is what carries across.
Where tuning makes the difference on a food plant
The loops worth tuning properly are the ones that decide product and safety: pasteuriser hold temperature and diversion, glycol and chilled water that hold cold chain, steam and hot water that drive cook and CIP, the flows that set residence time, and the levels that keep balance tanks stable. On those loops the difference between adequate and good is consistent product, valves that last, and operators who trust the loop instead of running in manual.
Good tuning is not a one-time setting. Plates foul, pumps wear, valves drift and recipes change, and a loop that was right at commissioning drifts with the plant. A structured method matters most here, because retuning from a fresh step test is quick and repeatable, where trial and error is an afternoon of guessing every time. For the broader control layer this work sits within, see our overview of PLC, SCADA and HMI work.
References
Sources for the tuning methods and term behaviour described above:
