learn · South Africa
PID Tuning Practice: Compare P, PI and PID Responses
Learn PID tuning with measured P, PI and PID simulator results. Compare offset, overshoot, settling and limits, with model assumptions and test notes.
PID tuning becomes easier to understand when you compare controlled, repeatable experiments. Start with a defined simulated process, record the baseline, change one gain and compare the response using more than appearance. A smooth trend can still have a large final error, and a fast response can come with excessive overshoot or output movement.
This tutorial replaces guesswork with a worked P, PI and PID comparison using the educational PID process simulator. The example settings and measured results belong to that model. They are not recommended settings for a real heater, pump, valve or production loop.
What “PID tuning by feel” should mean in a learning exercise
Experienced people recognise patterns in trends, but recognising a shape is the beginning of an explanation, not the end. In training, use intuition to propose a change and measurement to evaluate it. State what you expect before running the model, then check whether the observation supports the expectation.
Avoid a universal instruction such as increasing gain until a real loop oscillates. The consequences of a test depend on the process, operating limits and authorised commissioning plan. An educational simulator gives you a controlled place to explore poor responses without treating those experiments as field procedures.
A useful tuning record has five parts: the process model, initial conditions, controller settings, observed response and conclusion. If one is missing, a later reader cannot tell whether two runs are comparable. Even in a simple simulator, changing the disturbance option between runs can invalidate a conclusion about the effect of a gain.
Define the loop before changing the controller
The process variable, or PV, is the measured value the controller uses. The setpoint, or SP, is the target. In the model used here, error is calculated as setpoint minus process value. The controller output influences a simplified process, and the process response returns through the feedback path.
This chain contains several assumptions. The measurement is represented in the same units as the setpoint. The direction of control action is appropriate for the model. Output has limits. The process does not respond instantly. Before interpreting a trend, identify those assumptions and distinguish them from properties of an actual installation.
If the signal represents engineering units incorrectly, tuning is the wrong first problem to solve. A value of 50 might mean 50 °C, 50 percent or an unscaled raw count. Use the analogue scaling guide and signal-type explanation when the measurement path is unclear.
The model used for the worked comparison
The product’s first-order training challenge has a 60 °C target, an eight-second process time constant and a two-second transport delay. Its numerical evaluator runs with a 0.1-second step, starts the process value at zero and limits output to the range zero to 100. The comparison below uses the challenge without the optional load disturbance.
These are mathematical training conditions, not a description of a physical vessel starting at zero degrees. The model is deliberately bounded and repeatable. It does not include every sensor, actuator, thermal, mechanical or network effect that a real process might exhibit.
The results were calculated on 11 September 2026 with the product repository’s deterministic PID model. We retained the inputs, output traces and source fingerprint for the editorial check. The public product can evolve, so record the challenge and settings you actually use if a later version produces different results.
Understand the three actions and their parameter names
Proportional action responds to the current error. In this example, its contribution is Kp multiplied by error. Increasing Kp changes how strongly the controller responds to a given error, but the outcome also depends on delay, output limits and the process dynamics.
Integral action accumulates error over time. The example model uses an integral gain labelled Ki. Its contribution depends on the accumulated error and Ki, subject to the model’s conditional integration logic when output is limited. Integral action can reduce a persistent offset, but it also changes the transient response.
Derivative action responds to a rate of change. The specific evaluator used for this worked comparison calculates a change in error between time steps and multiplies it by Kd. That is a model implementation detail. Do not assume that every controller takes derivative on error, exposes the same filtering or uses the same parameter conventions.
Ki is not the same field as integral time Ti
Different controller forms express integral and derivative action differently. A field labelled integral time is not interchangeable with an integral-gain field simply because both control integral action. Units, scaling and the surrounding equation matter. A numerical value copied without its controller form can describe a very different controller.
For this lesson, enter values only into the model’s Kp, Ki and Kd fields. If you later study a vendor block that uses Ti or another convention, read its equation and manual first. The Siemens process-control training overview illustrates the distinction between general control learning and training on a specific process-control platform.
Establish a P-only baseline
Begin by setting integral and derivative gains to zero in the educational challenge. This isolates proportional action so you can explain what it does in this particular process. Record the initial state, target and output limits before running.
The first measured run used Kp 0.8. Its final process value was approximately 26.667 against a target of 60, leaving an absolute final error of 33.333. There was no overshoot above the target. That is a clear example of why “no overshoot” does not mean “good control”: the response never approached the requested value closely enough.
A second run increased Kp to 1.8 while keeping Ki and Kd at zero. The final value rose to approximately 38.571 and the final error decreased to 21.429. This improved the offset relative to the first run, but it still did not settle within the model’s acceptance band around the target.
Do not generalise those exact values to every P-only loop. They are results for the stated model, starting conditions and output behaviour. Their teaching value is the controlled comparison: one gain changed, the error decreased, and a substantial offset remained.
What to write in your notebook
State the prediction, such as “increasing Kp should reduce the offset in this model.” Then record the observations supporting it and the limitation that remains. Avoid writing only “better” or “stable.” A bounded response can still be unsuitable for the learning objective.
Include the output trend where available. An output that spends much of the run at a limit tells a different story from one that remains comfortably inside its range. Looking only at PV can hide the reason a change has little effect during part of the experiment.
Add integral action and compare the same measures
Keep Kp at 1.8 and introduce Ki 0.15 with Kd zero. In the recorded model run, final error fell to 0.459 and the reported settling measure was 45.8 seconds. The run had no overshoot above the target. Integral action addressed the large persistent offset seen in the P-only examples.
Next, increase Ki to 0.25 while leaving the other conditions unchanged. The recorded final error was 0.005, and the reported settling measure reduced to 23.2 seconds. Integrated absolute error also decreased. This is evidence of improvement for the selected model and comparison, not a rule that larger Ki is always better.
The correct next question is what happens outside the single successful case. Would a different delay, a different starting condition or an output limit alter the result? Does the controller recover well from a load change? A tuning that looks attractive in one experiment is not proven across all operating conditions.
A measured comparison table
The following values come from the deterministic training-model runs described above. “Not reached” means the run did not establish settling within the model’s recorded time window. It does not imply that a physical process has been tested or certified stable.
| Run | Kp | Ki | Kd | Final absolute error | Overshoot | Integrated absolute error | Reported settling measure |
|---|---|---|---|---|---|---|---|
| P baseline | 0.8 | 0 | 0 | 33.333 | 0% | 2151.481 | Not reached |
| Higher P | 1.8 | 0 | 0 | 21.429 | 0% | 1431.991 | Not reached |
| PI comparison | 1.8 | 0.15 | 0 | 0.459 | 0% | 536.485 | 45.8 s |
| Stronger I in this model | 1.8 | 0.25 | 0 | 0.005 | 0% | 388.732 | 23.2 s |
| Small D addition | 1.8 | 0.25 | 0.05 | 0.004 | 0% | 383.227 | 22.6 s |
Read across the table rather than ranking on one column. Both P-only runs show zero overshoot but large errors. The final two runs are close on several measures. The small derivative addition does not transform the result in this example, so the data does not justify a claim that derivative is always essential.
Interpret overshoot, settling and accumulated error carefully
Overshoot describes how far the response exceeds a reference under a stated definition. The model used here expresses the peak excess as a percentage of its fixed 60-unit target. Other analyses may normalise overshoot to the size of a setpoint change. State which definition you use before comparing percentages from different tools.
Settling also requires a defined band and observation window. This evaluator tracks the last sampled time outside a band of plus or minus 1.2 units around the target during its approximately 60-second run. Its displayed measure therefore has a specific sampling and window convention. It should not be presented as an infinitely precise continuous-time settling calculation.
Integrated absolute error adds the magnitude of error over time. A response that remains far from target for a long period accumulates more error than one that approaches it promptly, all else equal. It is useful for comparing repeated runs with the same target and duration, but it is not a universal score that makes unrelated processes directly comparable.
Final error reports the gap at the end of the observation window. It can be small even if the earlier response was poor. Conversely, a slow response may still be moving towards the target when the window ends. Keep the trend, duration and metrics together so the conclusion reflects the whole experiment.
Add derivative only to answer a defined question
In the measured example, adding Kd 0.05 to the PI settings produced a small reduction in integrated error and the reported settling measure. The difference was modest. A responsible interpretation is that the added term slightly changed this model’s response, not that these settings are a universally optimal PID tune.
Derivative behaviour depends on the implementation and signal quality. A learning model with a clean deterministic signal does not establish how a real controller will react to noisy measurements. If a vendor controller provides derivative filtering or a choice of signal on which derivative acts, those settings need their own explanation.
A useful advanced exercise compares the controller forms explicitly. Record whether derivative uses error or PV, what filtering is present and how the sample interval enters the calculation. Without those details, two sets of identically named Kp, Ki and Kd values may not represent equivalent behaviour.
Output limits and integral accumulation
A controller can request more output than the model permits. The limited output then differs from the unconstrained calculation. This matters because the process can only respond to the applied output, not to a value that exists inside the equation but cannot be delivered.
Integral accumulation during a prolonged mismatch can influence what happens when the response catches up. The example evaluator uses conditional integration around its output limits. Other controllers use different anti-windup mechanisms, and their behaviour should be read from the relevant documentation rather than inferred from a generic description.
In a learning record, identify periods when output reaches a limit. Ask whether the observed delay is due to the process dynamics, the output constraint or both. If increasing a gain produces the same applied output for the first part of the run, a similar early PV response is not surprising.
This also explains why repeated uncontrolled edits are unhelpful. If you change gains, limits and starting state together, it becomes difficult to identify which mechanism changed the result. Use the same evidence discipline as PLC troubleshooting: one hypothesis, a relevant observation and a repeatable check.
Setpoint tracking is different from disturbance recovery
The worked table compares the model’s response to its fixed target from the stated initial condition. It does not establish how well the settings reject a later disturbance. Treat disturbance recovery as a separate experiment with its own baseline and acceptance criteria.
If the product’s disturbance challenge is available under your current access, record when the load change occurs and compare the response after that event. Keep the controller settings and model conditions identical between the baseline and comparison. Read the current PID lab access and challenge details before assuming every challenge is included in a free session.
For the analysis, distinguish the initial approach to the target from recovery after the disturbance. A controller may perform differently in those two phases. Explain which part of the trend supports your conclusion, and avoid averaging away a poor recovery simply because the initial response was good.
Use the lesson in a South African training programme
For an instrumentation learner, connect PID practice to measurement quality and engineering units. For a PLC learner, connect it to the controller block’s execution and data flow. For an operator-focused course, emphasise interpreting trends and recognising when the observed behaviour differs from the expected process response.
Ask a training provider which process model or hardware rig is used and whether each learner performs a measured comparison. A demonstration of a smoothly controlled tank can be instructive, but it does not show that every learner can explain the gains, limits or test conditions.
The instrumentation course guide helps compare broader study routes. ISA’s process automation and control course catalogue is a primary-source example of how measurement, control fundamentals and specialist subjects can be separated into defined learning areas. Confirm current regional availability and prerequisites with the relevant provider.
If you are studying remotely around shifts, keep one saved comparison sheet per session. A short session that changes one condition and records a conclusion can be more useful than repeatedly chasing a better-looking trace without notes. Bring the unresolved question to your instructor or supervised practical session.
Assess your understanding with a reproducible handover
Give another learner the model name, challenge, initial conditions and one set of gains. Ask them to predict whether the response will have a large final offset, then run the same case and compare the result. If they cannot reproduce it, investigate differences in model version, fields, disturbance selection or starting state.
Next, ask them to explain why zero overshoot does not prove that the P-only baseline is satisfactory. A good answer refers to the large final error and accumulated error. This checks whether they understand the measures rather than remembering which row has the smallest number.
Finally, ask for one limitation of the experiment. Suitable answers include the simplified process, clean signal, fixed target, output limits, observation window and controller form. Recognising a limitation is part of competent interpretation; it does not make the exercise useless.
Your portfolio should include the table, a trace or screenshot where appropriate and a short paragraph explaining the selected next experiment. Label the work as simulation. Do not claim that these results demonstrate tuning competence on an unidentified real process.
Frequently asked questions about PID tuning practice
Can I learn PID tuning without advanced mathematics?
You can begin by defining the signals, changing one setting and comparing measured responses. Mathematical models become increasingly useful as the problems become more complex. Start with a clear experiment and build the theory alongside the observations rather than treating intuition and mathematics as competing approaches.
Why does P-only control show no overshoot but still perform poorly here?
In the recorded model runs, the response remained well below the target. That creates zero overshoot while leaving a large final error. You need several measures and the response trace to judge the result accurately.
Should every PID loop use derivative?
No universal rule follows from this exercise. The value of derivative depends on the process, implementation, signal quality and objectives. In the worked model, a small derivative term produced only a modest change compared with the stronger PI case.
Can I copy these Kp, Ki and Kd values into TIA Portal or Studio 5000?
No. The values are tied to the educational model, its units and controller equation. A vendor block can use a different form or parameter convention, and a real process has its own identification and commissioning requirements.
Is integral gain the same as integral time?
They are related ways of parameterising integral action in particular controller forms, but their fields and numerical values are not directly interchangeable. Read the controller equation and units before interpreting or converting them.
What should I do when two runs give different results?
Compare the model version, challenge, gain values, initial state, disturbance option and duration. If those match, record the discrepancy rather than changing more settings. A reproducible experiment depends on a reproducible starting point.
What is the next useful exercise after this table?
Choose one unanswered question: the effect of delay, output limitation, a disturbance or measurement noise in a model that supports it. State the prediction and compare like-for-like runs. For the current product, begin with the available PID simulator challenges and their documented scope.