How should PID control parameters be adjusted for different types of processes?

Hello dear colleagues,
Should an integrating level loop, a dead-time-dominant thermal process and an oscillatory motion system be tuned with the same PID strategy?
No. One of the most common mistakes in PID control is to treat Kp, Ki and Kd as three independent numbers that can be increased or decreased until the trend somehow looks acceptable. The same controller settings that produce a smooth response on a storage-tank level loop may cause repeated cycling in a delayed furnace, while those same values may be far too slow for a pressure or motion-control application.
A better question is therefore not "What should Kp be?" but "How does this process respond when the manipulated variable changes?" PID tuning becomes meaningful only after the dynamic character of the process is understood.
First Identify the Process, Then Tune the Controller
When operating conditions and safety rules permit, a small controlled change in controller output can reveal much more than repeatedly editing PID parameters. Watching the process-variable trend after a small output change helps identify whether the process is self-regulating, integrating, dominated by dead time or naturally underdamped.
If the process variable initially remains almost unchanged after the output is moved, the loop contains dead time. If the process then approaches a new value and eventually settles, it behaves as a self-regulating process. If the process variable continues rising or falling instead of finding a new equilibrium, the process has an integrating characteristic. If the response repeatedly crosses its final value before settling, there may be significant second-order or oscillatory dynamics.
The measured trend also includes more than the physical process itself. Sensor filtering, valve deadband, actuator travel, VFD acceleration ramps, PLC scan time, network updates and measurement conversion can all add delay. For that reason, time delay and process dynamics in PID control should be evaluated as part of the complete loop rather than as isolated controller parameters.
1- Self-Regulating Processes
A self-regulating process naturally moves toward a new steady operating point when its input is changed and then held constant. A simple thermal process is a good example: increase heater power and the temperature rises, but heat losses eventually increase until the system reaches another approximate equilibrium.
Three quantities are particularly useful when describing this response. Process gain describes how strongly the process variable changes for a given output change. Time constant describes how quickly the process approaches its new condition. Dead time represents the delay between the controller action and the first meaningful response seen in the measurement.
For many ordinary self-regulating industrial loops, PI control is entirely adequate. Proportional action provides immediate correction while integral action removes the remaining offset. Increasing proportional gain generally makes the response faster, but there is a limit: once the controller becomes too aggressive for the process dynamics, overshoot and cycling increase.
The aim is therefore not to find the highest Kp the loop can survive. A useful tuning must balance response speed, disturbance rejection and stability margin. This matters because the real plant rarely behaves exactly the same at every production rate, valve position, temperature or load.
2- Integrating Processes
An integrating process behaves fundamentally differently. A typical example is tank level control. If inflow is greater than outflow, the level continues to rise. If outflow is greater than inflow, the level continues to fall. The process does not automatically move toward a new stable level simply because the manipulated variable has stopped changing.
This distinction has an important consequence for PID tuning: the process itself already contains an integrating action. It is therefore too simplistic to say that an integrating process merely needs stronger integral control to remove steady-state error. Excessively aggressive integral action can produce long, slow oscillations and large actuator movements.
Proportional action also behaves differently here. On a self-regulating loop, reducing controller gain often makes the loop less aggressive. On an integrating loop, reducing proportional gain too far can actually make the closed-loop behaviour poorly damped and more oscillatory. This is one reason why tuning rules intended for self-regulating processes should not be copied blindly into level or other integrating loops.
Integral action may still be required, particularly to remove persistent offsets caused by load imbalance or plant bias, but its purpose should not be confused with simply making the level return faster. The reset action must be coordinated with the integrating rate and the loop dead time.
There is also a wider process-control question. In many surge or buffer vessels, the objective is not to hold the level perfectly at one number. Allowing level to move within an acceptable band may absorb upstream variability and produce a smoother downstream flow. In that case, a deliberately slower level controller can improve overall plant performance even though the level trend itself appears less tightly controlled.
For integrating processes, good tuning therefore depends on integrating rate, dead time, allowable process deviation, expected disturbances and acceptable actuator movement, not simply on minimizing instantaneous setpoint error.
3- Dead-Time-Dominant Processes
A dead-time-dominant process creates a difficult situation for any feedback controller: the controller changes its output but receives no useful information about the result for a significant period of time.
Transport through a long pipe, material travelling on a conveyor before reaching a measuring point, analyzer sampling cycles and some thermal systems can exhibit this behaviour. The control problem is not simply that the process is slow. The deeper problem is that the controller is temporarily blind to the effect of its own previous action.
If the controller makes another large correction before the previous correction becomes visible, it can easily overdrive the process. On an operator trend, the familiar pattern is a long period of apparently little response followed by a much larger movement than expected.
Dead-time-dominant loops usually require a more conservative closed-loop response. Proportional and integral action must be chosen so the controller does not repeatedly react to information that is already out of date.
A common tuning myth is that large dead time automatically means that derivative gain should be increased. This is not a reliable general rule. Derivative action cannot predict information that has not yet reached the sensor. In a process dominated mainly by pure transport delay, particularly when the measurement is noisy, increasing derivative action may add little benefit while producing a nervous controller output.
Derivative action becomes more useful when the process contains a genuine lag or underdamped dynamic whose rate of change carries useful information. Long dead time by itself is a different problem.
When dead time fundamentally limits achievable feedback performance, continually changing PID gains may no longer be the best answer. If an important disturbance can be measured before it affects the process, feedforward control may respond earlier. For suitable process models, dead-time compensation techniques such as a Smith Predictor may also be considered.
4- Oscillatory and Underdamped Processes
Some systems naturally tend to oscillate even before the controller is tuned aggressively. Mechanical loads with elasticity, high-performance servo systems, coupled masses and processes with several significant dynamic lags can behave this way.
If the process variable crosses its target repeatedly after a disturbance and the oscillations gradually decay, the plant is underdamped. Increasing proportional gain may improve speed but can also excite the existing dynamics. Fast integral action adds another source of phase lag and can make the oscillation worse.
Derivative action can be valuable in this class of process because it responds to the rate at which the measured value is changing and may add useful damping. However, the derivative term is also highly sensitive to measurement noise, so industrial implementations normally use a filtered derivative rather than an ideal mathematical differentiator.
Even here, adding Kd simply because a trend is oscillating is poor troubleshooting. The oscillation may come from valve stiction, mechanical backlash, actuator saturation, an excessively fast integral term, poor sensor resolution or another loop interacting with the controller.
Before changing Kp, Ki and Kd, first determine whether the controller is exciting the process or merely responding to a problem elsewhere in the loop.
5- Near-Integrating and Lag-Dominant Processes
Not every industrial process fits neatly into one category. A particularly important case is the near-integrating or very lag-dominant process. Technically the process may eventually settle at a new equilibrium, but its time constant is so long that during the normal observation period it appears to behave almost like an integrator.
Large thermal masses, large mixed vessels and some composition-control systems can exhibit this behaviour. Treating such a process as an ordinary fast self-regulating loop can result in weak or poorly damped tuning.
As the dominant process time constant becomes much larger than the dead time, tuning concepts used for integrating behaviour may become more useful than aggressively applying conventional self-regulating rules. This is one reason process classification should be based on the actual trend rather than the equipment name alone.
What Does Lambda / IMC Tuning Add?
Statements such as "use a low gain" or "make the controller less aggressive" are not very useful unless they are tied to a measurable performance target. Lambda or IMC-oriented tuning provides a more structured way to make that decision.
The basic idea is to identify an approximate process model and then choose the desired closed-loop response time, commonly represented by lambda (λ). Controller parameters are calculated to produce approximately that response rather than being chosen only by trial and error.
A shorter lambda gives a faster and generally more aggressive response. A longer lambda produces a slower controller with greater tolerance for modelling errors, process variation and uncertain dead time. In practical terms, lambda creates an explicit trade-off between performance and robustness.
For a self-regulating loop, this means deciding how quickly the controlled variable should return toward its target relative to the natural plant dynamics. For an integrating loop, it may mean deciding how quickly a disturbance-induced slope should be arrested while still allowing enough process movement to avoid unnecessary actuator activity.
This is especially useful for applications such as surge vessels, where the engineering objective may not be tight setpoint tracking at all. The correct objective may instead be to use available vessel capacity to absorb disturbances while keeping outlet flow comparatively smooth.
What About Ziegler-Nichols and Cohen-Coon?
Ziegler-Nichols remains historically important and can provide useful starting values, but it should not be treated as a universal final tuning recipe. Many Ziegler-Nichols settings are intentionally aggressive and may produce more overshoot and actuator movement than a modern industrial process requires.
Cohen-Coon uses an estimated process reaction curve and can also provide practical starting parameters for suitable process models. However, both methods depend on the process behaviour observed during the test. If process gain, dead time or time constant changes substantially with operating point, a set of parameters identified at one production condition may not provide the same performance elsewhere.
These methods are therefore best treated as starting-point generators. The final tuning should still be validated against actual disturbance rejection, setpoint behaviour, output movement and stability over the expected operating range.
Kp, Ki and Kd Do Not Mean Exactly the Same Thing in Every Controller
A surprisingly common commissioning problem occurs when PID values are copied from one PLC, drive or process controller to another. Two controllers may display similar-looking parameters but implement different PID equations.
One controller may use proportional gain, integral gain and derivative gain. Another may use proportional gain together with integral time (Ti) and derivative time (Td). Integral gain and integral time are not interchangeable values; depending on the controller definition, increasing one may strengthen integral action while increasing the other may make integral action slower.
Parallel, standard/ideal and series/interacting PID structures can also produce different behaviour from apparently similar numerical settings. For that reason, PID values should never be transferred between different controllers without checking the manufacturer's algorithm and parameter definitions.
Sampling Time Is Part of the Control Loop
A digital PID controller does not calculate continuously. It reads the process, performs its calculation and updates the output at discrete intervals. If that interval is too slow compared with the important process dynamics, the digital controller effectively adds additional delay to the loop.
This can become important in fast pressure-control, current and motion loops, where PLC scan time, I/O update time and communication cycles may represent a significant part of the process response.
The opposite mistake is also common. Running a PID calculation every millisecond does not automatically improve a process whose physical temperature changes over tens of seconds or minutes. Sampling must be fast enough to describe the relevant dynamics, not simply as fast as the hardware permits.
Check the Hardware Before Retuning the PID
Sensor behaviour: A noisy measurement encourages controller activity that does nothing useful to the actual process. Heavy filtering can reduce noise, but excessive filtering adds phase lag and delay. Sensor quality and filtering therefore form part of the control design.
Valve and actuator condition: Stiction, deadband, mechanical backlash and poor actuator resolution cannot be repaired with PID gains. A controller may repeatedly increase and decrease its output while a sticky valve remains stationary and then suddenly jumps, creating a limit cycle that looks like bad tuning.
Output limits: When an actuator reaches its physical limit, the integral term may continue accumulating error even though the process cannot respond further. This is integral windup. Anti-windup, output tracking or equivalent controller functions should be used where saturation is likely.
Operating point: Process gain is often nonlinear. Valve characteristics, pump curves, thermal losses and reaction behaviour may cause a loop tuned at low production rate to become much more aggressive or much slower at another operating point.
Cascade loops: If an outer PID controller sends the setpoint to another inner control loop, the inner loop should be commissioned and tuned first. The inner loop must respond substantially faster than the outer loop, otherwise both controllers can begin correcting the same disturbance on incompatible time scales.
If a loop suddenly becomes unstable after months of acceptable operation, repeatedly changing the tuning should not be the first response. The problem may instead be one of the common PID control problems caused by instrumentation, actuator condition, changed process load or another interacting loop.
Why One PID Setting Cannot Suit Three Different Machines
Consider three apparently simple applications.
In a storage tank, level represents the accumulated difference between inlet and outlet flow. The main control challenge is the integrating nature of the process. In a furnace, the controlled temperature may respond slowly and only after considerable thermal delay. In a servo-driven mechanical axis, the system may react almost immediately but exhibit inertia, elasticity and resonance.
All three controllers may display boxes labelled Kp, Ki and Kd. Yet the physics underneath those three numbers is completely different. Good tuning begins by recognising that difference rather than by searching for a universal combination of PID values.
A Practical PID Tuning Sequence
Start by confirming that the sensor, transmitter, actuator and scaling are correct. A controller should not be tuned around a mechanical or instrumentation fault.
Next, observe how the process responds to a safe and controlled change. Estimate the dead time, determine whether the process reaches a new steady value or continues to ramp, and assess whether significant oscillatory dynamics are present.
Then select a controller structure and tuning method appropriate to that process. A simple PI controller may be preferable to a full PID if derivative action provides no real benefit. More parameters do not automatically mean better control.
After the initial tuning, evaluate more than setpoint tracking. Observe disturbance rejection, overshoot, settling behaviour, actuator movement and behaviour at different operating conditions. If the valve is moving continuously to remove insignificant measurement noise, the loop may be technically stable while still being badly engineered.
Finally, leave sufficient robustness for the process to change. Production plants do not remain at the exact operating point used during a tuning test. The fastest possible tuning is rarely the most useful tuning if a modest change in process gain or dead time makes the loop unstable.
Conclusion
PID tuning for different process types cannot be reduced to rules such as "increase Ki for integrating processes" or "increase Kd when there is delay." Those shortcuts ignore the physical behaviour that actually determines whether a loop will be stable.
For self-regulating processes, process gain, time constant and dead time provide the main tuning clues. For integrating processes, the integrating rate, allowable deviation and load behaviour become more important. For dead-time-dominant loops, the fundamental limitation is delayed information. For oscillatory systems, damping and secondary dynamics must be considered.
The purpose of a well-tuned PID loop is not merely to force the process variable onto the setpoint as quickly as possible. A good industrial controller should be stable, repeatable, tolerant of process variation and no more aggressive than the application requires.
When a PID loop performs badly, do not begin by randomly changing Kp, Ki and Kd. First read the trend and determine what the process is telling you.
Related PID Control Topics
- What Is PID Control?
- Proportional, Integral and Derivative Components
- How Are Kp, Ki and Kd Adjusted?
- Time Delays and System Dynamics in PID Control
- Stable and Accurate PID Operation
- Common PID Control Problems
This article explains practical PID tuning principles for different process dynamics. Parameters used on a real machine or industrial process should always be selected with consideration for process behaviour, operating limits, instrumentation, actuator characteristics, safety requirements and the PID algorithm implemented by the controller manufacturer. You may share this article provided that OtomasyonAVM is cited as the source.