PID Tuning: Gain, Reset, and Rate

Back to Blog

What Gain, Reset, and Rate Each Do

Gain (Proportional, P): This sets how much the output moves per unit of error right now. Higher gain means a faster, more aggressive response, but too much causes overshoot and oscillation. Note the terminology: gain and proportional band are inverses - a wider proportional band means lower gain and a gentler loop.

Reset (Integral, I): Proportional action alone leaves a residual offset - the loop parks slightly off setpoint. The integral term sums error over time and keeps adjusting the output until that offset is gone, so the process actually reaches setpoint. Reset is expressed as repeats per minute or, inversely, minutes per repeat. Too much reset (too fast) causes slow oscillation and windup.

Rate (Derivative, D): This term responds to how fast the error is changing, applying a braking action that anticipates where the variable is heading. It reduces overshoot and speeds recovery on slow processes like temperature. But because it amplifies noise, derivative is often left off flow and pressure loops and reserved for sluggish, clean signals.

How the Three Terms Interact

The terms are not independent. Adding derivative lets you push gain higher without oscillating; too much integral forces you to back off gain to stay stable. A well-tuned loop is a balance among all three, matched to the process dynamics - its gain, time constant, and dead time.

In practice, most oil and gas loops are tuned P and I only. Fast, noisy loops (flow, liquid pressure) use modest gain, brisk reset, and no rate. Slow loops with dead time (temperature, some levels) benefit from lower gain, slower reset, and a touch of rate. Level loops on surge vessels are often tuned loosely on purpose - averaging level control - so the loop absorbs swings instead of passing them downstream.

Getting the Process Model From a Bump Test

Before adjusting anything, learn the process. With the loop in manual and the process steady, step the controller output by a small, deliberate amount - call it delta-OP - and record the process variable until it settles. Three characteristics fall out of that single trend. The process gain is the ratio of the PV change to the output change, delta-PV over delta-OP, both expressed in percent of span. The dead time is how long the PV does nothing after the output moves - see process dead time for why this is the enemy of tight tuning. The time constant is how long, once the PV starts moving, it takes to cover about 63 percent of its total change, a property of any first-order response.

Those three numbers describe most self-regulating loops well enough to tune from. The bump should be small enough not to upset the process and large enough to stand clearly above the noise; running two bumps in opposite directions and comparing them exposes nonlinearity or a sticking valve before it corrupts the tuning. Note everything down: conditions change, and a model taken at one operating point may not hold at another.

A Worked Symbolic Tuning Calculation

The lambda (internal model control) method turns the bump-test model directly into settings. Choose lambda, the desired closed-loop time constant - in effect, how quickly you want the loop to settle - expressed as a multiple of the open-loop time constant. Slower lambda buys robustness; faster lambda buys speed. The method then sets the integral time equal to the process time constant, and the controller gain to the time constant divided by the product of the process gain and the sum of lambda plus the dead time. Symbolically: Kc = tau / (Kp x (lambda + theta)), with Ti = tau.

Read the structure of that formula rather than plugging in numbers. A larger process gain calls for a smaller controller gain - the process is already sensitive. Longer dead time forces the gain down, which is why dead-time-dominant loops are doomed to be slow no matter how bravely they are tuned. And lambda appears only in the gain, as the single knob trading speed against robustness, which makes the method easy to defend at a management-of-change review: the choice is explicit rather than buried in trial and error.

Is It the Tuning or the Valve?

Oscillation is not always a tuning problem, and retuning a loop with a mechanical fault wastes the afternoon. The discriminating test is simple: put the controller in manual. If the oscillation stops, the loop itself was the cause and tuning is a fair suspect. If the PV keeps cycling with the controller frozen, the disturbance is external - another interacting loop, or an upstream process swinging - and no tuning change will fix it.

Valve stiction has its own signature: the PV traces a square-ish wave while the controller output ramps steadily up and down as the integral term winds against the stuck valve, then the valve jumps. Backlash shows up as a limit cycle mainly around load reversals. Both are maintenance problems, not tuning problems, though loosening the tuning can sometimes calm the symptom until the valve is serviced. Trends of OP and PV together, not PV alone, are what make these patterns visible.

Recording Tuning Like the Configuration It Is

Tuning constants are configuration and deserve the same discipline as any setpoint change: record the values before and after, the date, the reason, and the response observed. Two administrative traps catch engineers moving between systems. First, units - gain versus proportional band, repeats per minute versus minutes per repeat - invert the meaning of a number, so the same digits under the wrong convention produce a wildly different loop. Second, algorithm structure: controllers implement ideal, series, and parallel forms of PID, and the same three constants behave differently in each. When replacing a controller or migrating a loop to a new platform, verify the form and the units before copying numbers across, then re-verify the loop with a small setpoint step, watching for a clean response and a smooth handoff in and out of manual - the same concern that motivates bumpless transfer.

Cascade structures add one more rule: tune the inner loop first, then the outer loop with the inner in cascade, as covered under cascade control. Tuning an outer loop around an untuned inner loop bakes the inner loop's sins into the outer settings, and both will need doing again.

Frequently Asked Questions