PID Control: Principles, Terms and Industrial Examples

Write the controlled variable and manipulated variable as an engineering sentence: “This loop changes heater demand in percent to control outlet temperature in degrees Celsius,” or “This loop changes valve position in percent to control flow in litres per minute.” Record the measurement range, output range, normal operating window and physical final element. If a display says CV, clarify whether it is the raw controller request, a limited demand, the analog output command or measured actuator position.

This reference uses e = SP − PV. Under that convention, a heating process normally needs an increase in output when the error becomes positive. A cooling valve, pressure-relief valve or reverse-acting final element can require the opposite relationship. Vendors express this choice as direct/reverse action, controller gain sign, error equation or output-direction settings. The names are not reliable across products, so derive the required sign from the process and verify it at low authority in manual mode.

A direction test is not “put it in Auto and see.” Under the approved commissioning method, establish a safe operating point, make a small bounded manual output change, wait long enough for the process response, and record whether PV rose, fell or did not respond. Confirm the same direction from controller output through selector, I/O, drive or positioner to the physical element. A reversed measurement scaling and a reversed actuator can temporarily cancel each other, leaving a dangerous latent defect.

Document normal, failed and substituted measurement behavior. A stale PV that remains numerically plausible is more difficult than an obvious bad value. The loop should have a defined reaction for bad quality, implausible rate, range violation, frozen value and communication loss. That reaction may be manual hold, tracked fallback, a bounded safe process response or an orderly shutdown, but it comes from the process hazard and operating design—not from a universal PID rule.

A common continuous parallel teaching form is u(t) = bias + Kp·e(t) + Ki·∫e(t)dt + Kd·de(t)/dt. A common ideal or dependent form is u(t) = bias + Kc[e(t) + (1/Ti)∫e(t)dt + Td·de(t)/dt]. These expressions can describe equivalent behavior after a correct conversion, but the numeric parameters are not interchangeable. Some systems use proportional band rather than gain; some express integral as repeats per minute, seconds per repeat or integral gain; and some place proportional gain around the other terms.

Units expose mistakes. If error is degrees Celsius and output is percent, Kp may be percent per degree Celsius. In the parallel expression, Ki may be percent per degree Celsius per second and Kd may be percent-seconds per degree Celsius. If a vendor uses reset time Ti in seconds, Ti is a time rather than an independent integral gain. Changing seconds to minutes without conversion changes integral action by a factor of sixty.

Industrial controllers are sampled, so software implements a discrete approximation. Integration may use rectangular, trapezoidal or another update. Derivative may use a filtered difference and commonly acts on PV rather than error to avoid a sharp derivative response to a setpoint step. Position and velocity algorithms also store and update state differently. These choices matter during migration, restart and manual transfer even when steady-state trends look similar.

Never migrate a loop by copying Kp, Ki and Kd labels alone. Capture old and new equations, independent or dependent form, proportional band convention, error sign, derivative source, derivative filter, sample interval, scaling, output limits, anti-windup, bias, setpoint weighting and initialization. Convert on paper or in a controlled model, then execute a bounded acceptance protocol. The old loop trend is evidence, not proof that a copied number has the same meaning.

Proportional action creates an immediate output contribution related to present error. In a proportional-only loop, the process may need a nonzero error to sustain the output required by the load; that operating offset depends on bias, process gain and disturbances. Increasing proportional action can reduce that offset and speed the initial correction, but it also increases sensitivity to measurement movement and can expose dead time, interaction or final-element problems as oscillation.

Integral action changes output as error persists. That makes it powerful for rejecting a sustained load disturbance and removing offset, yet the integrator cannot create physical capacity. If a valve is fully open or a heater is at its limit while PV remains below SP, continuing to accumulate integral state stores a recovery problem. The correct response starts with capacity, limits and anti-windup evidence rather than automatically increasing integral action.

Derivative action responds to rate. An ideal derivative reacts strongly to high-frequency noise, which is why practical blocks filter it and why instrumentation quality comes first. Derivative on error can respond to a setpoint step; derivative on PV avoids that particular setpoint kick but still responds to measurement changes. The selected source, filter and discrete algorithm must be known before interpreting a D setting.

The terms interact. A loop can appear slow because output is rate-limited, because the measurement filter is excessive, because the process has dead time, because proportional action is low, because integral action is slow or because another loop owns the actuator. A disciplined trial changes one justified parameter at a time and retains the same initial condition, setpoint or disturbance, limits and observation window. Comparing unmatched trends is not a tuning experiment.

P-only control can be appropriate when some offset is acceptable, when another outer loop supplies the setpoint, or when integral behavior would be undesirable. PI is widely useful because it combines immediate correction with removal of persistent offset. Full PID can help when the rate signal is trustworthy and anticipatory damping improves the required response. Other structures—on/off control, feedforward, cascade, ratio, split range, override or model-predictive control—may address a problem that three feedback terms alone cannot.

Fast flow loops often have a responsive measurement and final element, making PI a common starting structure. Temperature loops can be slow and may have substantial dead time; PI is common, while carefully filtered derivative can help in some processes. Level can be intentionally slow or integrating, and aggressive tight control may pass disturbances downstream. Pressure spans very fast gas or hydraulic dynamics and slower vessel behavior, so the word “pressure” does not select a tuning or structure.

Use a structure decision record. State the controlled objective, acceptable deviation, manipulated authority, dominant disturbances, process gain sign, time constant, dead time, measurement resolution/noise/delay, actuator dynamics, interactions, constraints and required mode behavior. Then state why the selected structure is sufficient. This is more defensible than applying a favourite tuning rule to every loop.

A first-order-plus-dead-time model can summarize some process responses using process gain K, time constant τ and dead time θ. It is useful for communicating dynamics and supporting certain tuning methods, but it is not a universal truth. Nonlinearity, integrating behavior, inverse response, multiple lags, changing operating points and interacting loops can make one model inadequate. Record the test range and model fitness rather than presenting K, τ and θ without context.

Identify where the loop runs: cyclic task, timed interrupt, periodic organization block, function-block scheduler or a block that internally schedules its calculation. Measure the actual interval and jitter under representative controller load. A main scan that “usually takes about 20 ms” is not equivalent to a 20 ms periodic control task, especially when communication, diagnostics or conditional routines change execution.

The interval should be short enough to observe relevant process dynamics but not selected by a simplistic fixed ratio. Very fast execution can add noise sensitivity, CPU load and repeated calculations without adding process information. Very slow execution delays correction and distorts discrete integral and derivative action. Instrument update, I/O conversion, network refresh, filtering and actuator command rate can dominate even when the PLC block executes quickly.

Decide whether the block receives an explicit cycle-time parameter, measures elapsed time or assumes a configured update. Check what happens on a late call, skipped call, task overrun, controller mode change and simulation. Siemens, Rockwell, Beckhoff and CODESYS families use different mechanisms, and even versions in one family can differ. Treat the installed help and controller diagnostics as evidence.

Trend scheduled interval, measured interval, maximum jitter, overrun status and block timing fault alongside SP, PV and output during acceptance. If timing changes after adding communications or logic, the loop should fail the same regression tests used for a gain or firmware change. A tuning is tied to its execution environment.

Limits exist at several layers: the PID block, an output range, a rate limiter, high/low selector, equipment permissive, local hand station, drive current limit, valve travel stop and process capacity. A 0–100 percent internal limit says nothing about a valve restricted to 65 percent by an override controller or a heater disabled by an interlock. Inventory every authority boundary and decide which value returns to the controller for tracking.

Integrator windup occurs when integral state continues moving in a direction the actuator cannot deliver. After the error reverses or the limit clears, the stored state can keep the output pinned and delay recovery. Common anti-windup approaches include conditional integration, integrator clamping and back-calculation. Vendor blocks implement these differently, and an internal method may only know the block’s configured limit.

External reset feedback or tracking becomes important when the effective applied output is determined downstream. Feed the controller the value and ownership state specified by the block’s documentation. If the actuator position is itself slow or unreliable, distinguish commanded applied output from measured position and decide which signal belongs in anti-windup, diagnostics and maintenance. One signal cannot silently serve all meanings.

When a loop remains saturated, ask whether the SP is reachable, whether the final element moved, whether supply energy is available, whether another constraint owns the output, whether the PV is valid and whether the process load changed. Saturation is diagnostic evidence. Retuning a capacity failure can make normal operation worse while leaving the actual limit unresolved.

Define who may command the final element: operator manual, PID automatic, cascade setpoint, ratio station, override selector, sequence, local hand station or maintenance mode. Record the active owner and requested owner separately so a rejected transfer remains visible. A single Auto Boolean is often too weak to explain why the actuator is moving.

Before manual-to-auto transfer, the automatic controller’s internal state must be consistent with the output currently controlling the process. The setpoint may also need to track PV or adopt a separately approved value. Rockwell documents tieback and set-output behavior for its PID instruction; other blocks provide tracking, synchronization, manual value or initialization interfaces. Map the chosen block rather than implementing a generic “copy manual output once” assumption.

Bumpless does not mean response-free. If Auto takes ownership with SP far from PV, the output can begin changing immediately even if the first sample is continuous. Separate transfer bump from the intentional closed-loop response after transfer. Define maximum first-step output change, rate, mode-confirmation time and acceptable PV excursion for the actual process.

Test Auto-to-Manual, Manual-to-Auto, cascade entry/exit, selector takeover/release, controller restart, communication loss and local/remote transitions. Retain requested, limited and applied output plus mode, owner, SP, PV and quality through every transition. Operators need to see the controlling owner and why an automatic request is not applied.

Validate sensor range, calibration status, installation, damping, update time, signal conversion and quality mapping. Compare the PLC engineering value with an appropriate reference under an approved procedure. A noisy PV can come from process turbulence, electrical interference, insufficient resolution, poor grounding, aliasing, network timing or a failing instrument. Increasing derivative filtering treats one symptom and can add delay.

Validate the complete actuator path. For a valve, compare output request, I/P or positioner command, measured stem position, supply pressure and process response. For a drive, compare speed reference, selected reference source, limits, actual speed/current and process effect. For time-proportioned heating, confirm window time, switching hardware limits and how percent demand becomes pulses. The PID output is not the physical action.

Stiction, backlash, deadband and hysteresis can create repeating sawtooth trends that resemble aggressive tuning. A controller increases output until the element breaks free, PV overshoots, output reverses, and the cycle repeats. Measure output-versus-position behavior or conduct an authorized stroke test before retuning. Maintenance of the final element may produce more benefit than gain changes.

Filtering is a design tradeoff. A filter can reduce noise presented to the controller but also adds lag and can hide a rapid real change. Place filtering deliberately at the instrument, input, PV path or derivative path and document its time constant and purpose. Avoid stacking undocumented filters across transmitter, module, PLC and HMI until the trend looks smooth.

In a flow loop, the PV is flow and the MV may be valve position or pump speed. The process can respond quickly, upstream pressure is a disturbance, and valve stiction or pump limits are common constraints. The acceptance test should include normal operating points, representative pressure disturbance, low-flow measurement quality and transfer between manual and automatic ownership.

In vessel pressure control, the MV may be inlet, outlet, vent or compressor demand. Control direction depends on which stream is manipulated. Gas-volume dynamics, relief protection and equipment limits matter. PID does not replace pressure protection. A pressure loop should be tested within authorized boundaries while independent alarms and protective functions remain active and separately validated.

In temperature control, thermal capacity and dead time can make response slow. The MV may be heater power, steam valve, coolant valve or split-range heat/cool demand. Sensor location and product lag matter. A short observation window can falsely declare success before a slow oscillation appears, while an excessively aggressive test can damage product. Use a model or simulator for learning, then a process-approved protocol for the plant.

In tank level control, the process may integrate: if inlet and outlet do not balance, level continues changing. Tight level control can transmit flow disturbances to downstream equipment, so the objective may be to keep inventory inside a band while smoothing outlet flow. This is why one set of “best PID settings” cannot be applied across flow, pressure, temperature and level.

Assume a teaching-only discrete parallel PI controller, error e = SP − PV, output in percent, temperature in degrees Celsius and sample interval Ts = 1 second. Let Kp = 2 percent per degree Celsius and Ki = 0.25 percent per degree Celsius per second. Derivative is disabled. At sample k, SP = 50.0 °C, PV = 48.5 °C, previous integral contribution I(k−1) = 30.0 percent and bias = 0.0 percent.

The error is 1.5 °C. The proportional contribution is P = Kp·e = 2 %/°C × 1.5 °C = 3.0 percent. The integral increment is ΔI = Ki·e·Ts = 0.25 %/(°C·s) × 1.5 °C × 1 s = 0.375 percent. The new integral contribution is 30.375 percent. The raw request is bias + P + I = 33.375 percent.

If the configured limits are 0 to 100 percent and no downstream constraint is active, the limited request remains 33.375 percent. That arithmetic does not prove the actuator applied 33.375 percent. A rate limiter, selector, interlock, drive or valve can change it. The evidence record therefore retains raw request, limited request and applied output separately.

At the next sample, the calculation uses new PV and the controller’s stored state. A different algorithm may use another integration rule, dependent gain form, normalized error, integer scaling or internal time base, producing different numbers from parameters that look similar. Use the selected block documentation for production calculations. The dedicated worked-example page expands term-by-term response traces and verification.

Start with the question the loop failed to answer: setpoint tracking, load-disturbance rejection, transfer, constraint handling, noise, product variability or energy use. Export a synchronized window containing SP, PV, raw request, limited request, applied command, measured actuator position if available, mode, owner, quality, saturation, interlocks and relevant disturbance. Record timestamps, sample rates and any compression or deadband used by the historian.

If requested output changes but applied output does not, investigate ownership, limits, I/O, communications and final element. If applied output changes but PV does not, investigate process capacity, sensor validity, disconnected equipment and dead time. If PV moves in the wrong direction, return to a safe manual state and verify signal/action direction. If requested and applied output oscillate together, examine process dynamics, gains, noise and interaction after the physical path is proven.

Separate setpoint tests from disturbance tests. A controller can track a setpoint well but reject load poorly, or the reverse. Use consistent initial conditions and observation windows. Compare error measures such as integrated absolute error only when tests are otherwise comparable, and pair numeric measures with limits, actuator travel, overshoot, settling, product quality and operator workload.

Record each change as a controlled experiment: hypothesis, exact parameter before/after, equation and units, initial condition, test input, result, rollback trigger and reviewer. Avoid simultaneous gain, filter, limit and instrument changes because they destroy attribution. A simulator can teach the relationships and help rehearse evidence capture, but its model cannot validate a plant it does not represent.

Complete design review first: loop narrative, P&ID and instrument data, I/O range, output device, action, algorithm form, sample time, limits, modes, tracking, bad-PV reaction, alarms, interlocks, independent protection and acceptance criteria. Confirm controller, firmware, IDE and library versions. Back up the project and parameters under the site change-control process.

With the process in an approved condition, verify raw and scaled PV, quality and timestamp. In manual mode verify bounded output direction and the final element at more than one point where permitted. Confirm commanded versus actual position or speed, output scaling and response. Test local/remote ownership, permissives, limits and fail behavior before enabling automatic control.

Test automatic behavior using an approved setpoint or disturbance change. Retain the complete trend long enough to observe the slowest relevant dynamics. Test saturation and recovery only within process authority. Verify anti-windup and bumpless transfers. Test bad PV, stale data, controller restart and communication recovery using safe methods and representative hardware or an approved test environment.

Close with deviations, final parameters including units and form, trend evidence, task-timing evidence, alarm/interlock results, operator instructions and rollback state. Repeat the defined regression set after firmware, library, task, I/O, instrument, actuator or process changes. A gain sheet without its operating and software context is not a complete loop record.

A PID simulator can provide a repeatable process model, controlled disturbances and synchronized SP, PV and output traces. That makes it useful for learning proportional, integral and derivative effects; comparing setpoint and load responses; seeing saturation and anti-windup; rehearsing manual-to-auto transfer; and practising a one-change-at-a-time evidence method without risking production.

A generic simulator does not reproduce a particular transmitter, valve, drive, heater, PLC task, library block, firmware, network, process nonlinearity, product, hazard or protective function. Similar-looking curves do not certify production gains. Treat simulated performance as a hypothesis and training record, then validate the exact implementation under the site commissioning plan.

For product evaluation, useful proof is not just a moving curve. A credible PID learning surface should disclose its process equation and time base, separate requested/limited/applied output, expose saturation and modes, allow a repeatable disturbance, preserve units and initial conditions, and let the learner export or compare evidence. Those capabilities support troubleshooting habits that transfer beyond one animated demo.