Feedwater Flow Loop Hunting After an Oversized Valve Trim Was Fitted
A 250 t/h boiler feedwater line is controlled by a DCS PID loop driving a 4-inch globe valve with a pneumatic positioner. Flow is measured by an electromagnetic flowmeter with damping set to 0 s, feeding a 4–20 mA signal into a 12-bit analogue input card. During an overhaul the eroded trim was replaced with a larger one, raising valve Cv from 110 to 190; nothing else in the loop was changed.
Within a week the loop shows a sustained oscillation of +/-8% of flow at about 0.05 Hz, with the controller output swinging between 12% and 68%. The operator now runs the loop in manual for most of each shift and trims the valve by hand. A bump test in manual at 50% output gave an 18 t/h change in flow for a 10% output step, settling in about 6 s. Present tuning is Kc = 0.8, Ti = 8 s, Td = 1.5 s. The valve positioner datasheet quotes a deadband of 1.5% of stroke. Normal flow corresponds to about 25% valve lift.
1. Quantify why the loop became unstable and rule out valve stiction as the primary cause.
Read the worked answer
Process gain change. The bump test measures the combined valve-plus-process gain: Kp = 18 t/h per 10% output = 1.8 (t/h)/%. That is with the new trim. Scaling by the Cv ratio, the old trim would have given 1.8 x 110/190 = 1.04.
Loop gain. The dimensionless product Kc x Kp (both in %/%) was 0.8 x 1.04 = 0.83 before the overhaul and is 0.8 x 1.8 = 1.44 now - a 73% increase. With roughly 2 s of apparent dead time and a 6 s settling time, a loop gain of 1.44 at a period of 20 s sits squarely in the unstable region; the observed 0.05 Hz is consistent with a limit cycle driven by gain plus dead time, not by an external disturbance.
Why not stiction. Stiction produces a characteristic sawtooth in controller output with the valve sticking and then jumping, and the amplitude is set by the stiction band. Here the output swings 56 percentage points - 37 times the 1.5% deadband - so the valve is following its command; the loop is simply too fast for the process. Deadband contributes a small residual amplitude floor but cannot explain +/-8% flow.
Secondary contributor. Td = 1.5 s acting on an undamped flowmeter signal adds a noisy component to the output, worsening the hunt and hiding the true period.
2. Propose the tuning and instrumentation changes, with justification for each, and state the trade-off you are accepting.
Read the worked answer
Tuning.
- Halve the gain: Kc ≈ 0.35–0.45. Target a loop gain around 0.6 with the measured 1.8 (t/h)/%, giving margin for load changes.
- Lengthen integral time to 12–15 s. With ~2 s dead time and a 6 s settling time, Ti = 8 s is faster than the process can support and drives the oscillation period down.
- Set Td = 0. Flow loops are fast and noisy; derivative amplifies sensor noise and adds nothing when the dead time is small.
Instrumentation.
- Set flowmeter damping to 0.5–1 s (or add a first-order filter with tau ≈ 1 s in the DCS). This suppresses noise but adds phase lag, so it must be applied before the final tuning pass. Accepted trade-off: a small loss of speed for a much quieter controller output.
- Verify the positioner deadband with a step test; if it exceeds 1%, recalibrate or replace the positioner.
- Address the root cause: at design flow the valve now sits at 25% lift, where rangeability, wear and controllability are all poor. A reduced trim or a different valve sized for 40–60% lift at design flow restores the original process gain and lets the loop be re-tuned rather than permanently detuned.
Trade-off accepted. Detuning alone gives sluggish response to genuine feedwater disturbances; that is why the trim change should be scheduled rather than relying on tuning.
3. The operator objects that the loop 'works fine in manual'. Explain what this masks and what data you would collect before changing anything.
Read the worked answer
What manual operation masks. In manual the feedback path is broken: the operator closes the loop by eye at a far lower bandwidth than the controller, and any valve or loop-gain problem becomes invisible. The loop appears acceptable because human response is slow enough to be stable, but it also means the plant runs with variable flow, the boiler sees repeated small disturbances, and the oscillation restarts the moment the loop is placed in auto. It also removes the very data needed to diagnose the problem.
Data to collect first.
- A trend of flow, controller output, valve position feedback and setpoint at 1 s resolution over at least two oscillation cycles, in auto, with no operator intervention, to confirm the period and phase relationship.
- A valve step test in manual (small, e.g. 2%, then 5%, then 10%) looking for stick-slip versus linear response and measuring actual deadband and stroking time.
- A controller output step in auto with the integral and derivative disabled, to measure the open-loop gain and dead time at the current operating point.
- Flowmeter signal with the transmitter isolated from the loop, to quantify noise amplitude and decide the damping.
Sequence. Restore the loop to auto only after the step test and set the new tuning from the measured model, then monitor for one shift. Changing tuning while the operator continues to fight the loop by hand will produce misleading results.