Why Two Compressors Fight Each Other on the Same Header

Why this matters

A plant adds a second compressor because the first one could not keep up, and the bill goes up by more than the air did. Nothing is broken. Both machines run, the header holds pressure, and every gauge in the room reads normal. What has happened is that two machines with independent local controls are both reading the same pressure and both responding to it, so instead of one machine working and one machine off, the plant gets two machines each doing part of the job at the worst point on their own curves. The correction is usually a setpoint change and a contactor, not equipment.

Two machines, one variable

A local pressure switch or transducer on a compressor is a controller with one input, and on a shared header that input is the same physical variable for both machines. Neither machine knows the other exists. Both are trying to hold the same number by acting on the same quantity, and the outcome depends entirely on how their bands are arranged relative to each other.

If the bands are widely separated, the machines behave as a crude cascade and only one works at a time, which is close to correct by accident. If the bands overlap or coincide, both machines respond to every excursion together, which is the fight. And if the machines modulate rather than load and unload, they do not even take turns: they share continuously, each throttled to part of the demand, permanently.

Two general relationships get used below and both carry the condition they were derived under. A commonly published rule of thumb puts compressor power change at about 1 percent per 2 psi of discharge pressure, derived for a lubricated rotary screw near 100 psig discharge; it is not a constant and the machine's curve governs. And every fixed restriction in the system, meaning every leak and unregulated point of use, passes choked flow at plant pressures, so its mass flow varies in direct proportion to upstream absolute pressure; that relationship belongs to the pressure and flow control card in this library.

Worked case: the plant whose air got 80 percent more expensive

Two machines, take each as 100 units of capacity. Both are inlet-modulating, which was common when the older one was specified and was carried over when the newer one was matched to it. Both are set to hold the header at the same pressure. Plant demand runs near 60 units for most of the shift.

What the machines do. With the same setpoint and similar droop, they share roughly equally, so each throttles to about 30 units, which is 30 percent of its own capacity. That is deep into the flat region of a modulation curve.

What that costs. Using the modulation shape published for machines of this type, power runs from about 0.70 of full load at zero output to 1.0 at full, so power equals 0.70 plus 0.30 times output. At 0.30 output each machine draws 0.79 of its full-load power. Two of them is 1.58 machine-equivalents of full-load power, and the plant is getting 60 units of air.

What one machine alone would cost. A single machine at 0.60 output on the same curve draws 0.70 plus 0.18, which is 0.88. Same 60 units of air, 0.88 against 1.58, so the shared arrangement costs about 80 percent more.

The part that surprises people. Now separate the bands so the second machine is only called when the first cannot hold pressure, but leave it running unloaded rather than stopping it. A modulating machine at zero output still draws about 0.70 of full load. The total becomes 0.88 plus 0.70, which is 1.58: exactly what the fight cost. Cascading the setpoints changed which machine was doing the work and saved nothing at all, because the saving was never in the sequencing. It was in the second machine being off.

That is the load-bearing point of this card. On a plant of modulating machines, the fix is not a wider band or a smarter setpoint. It is whether the lag machine's motor is running.

The same case with load and no-load machines is milder but the same shape. Sharing at 30 percent each, on a curve of 0.25 plus 0.75 times output, gives 0.475 each, so 0.95 total. One machine at 0.60 gives 0.70, and its partner genuinely unloaded gives 0.25, so 0.95 again. Stopping the lag machine gives 0.70. Sequencing alone still saves nothing; stopping the lag machine saves about 26 percent against the fight, on a base of 0.95.

Why separating the bands still matters, even though it does not save the energy

Cascading is not worthless, it just does a different job. It stops the two machines cycling against each other, which is a wear and a stability problem rather than an energy one, and it makes the plant's behaviour predictable enough to automate.

A correct cascade puts the lead machine at the highest setpoints, so it responds first as pressure falls, and each lag machine at successively lower setpoints, so a lag only comes on when the lead alone cannot hold the header. With 10 psi bands, lead at 105 to 115 and lag at 95 to 105 gives a clean handover: the lag's unload point is the lead's load point, with no overlap.

And that arrangement has a cost you have to state out loud. The total band across the pair is now 20 psi, not 10. The bottom is set by what the plant actually needs, so if the worst point of use plus the measured distribution drop puts the minimum acceptable header at 95, the top of the range is 115. The plant now runs at up to 115 whenever the lead machine is near the top of its band, and at the rule of thumb above that is roughly 10 percent more compressor power than running at 95, plus loss flow through every leak and unregulated point that scales with absolute pressure, which between 115 and 95 psig is a ratio of about 129.7 to 109.7, so about 18 percent more air lost at the top of the range than at the bottom.

Add a third machine and the total band becomes 30 psi, and the top of the range climbs again. That escalation is exactly why plants with more than two machines move to a network controller that holds one narrow band and sequences on time and flow rather than on separated pressure bands. Sequencing without widening the band is a controls job, and this library covers the procedure separately.

The second mechanism: the header is not one pressure

A shared header is treated as a single number and it is not. Each machine senses pressure at its own discharge, upstream of the run that connects it to the header, so what each machine reads is the header pressure plus the drop through its own connection at whatever flow it is passing.

The consequence is a load split nobody set. The machine with the shorter, larger connection reads closer to true header pressure. The machine at the end of a long run of smaller pipe reads lower than the header actually is when it is delivering, so it believes the plant is starving and loads harder.

The direction of this effect is worth holding onto, because it runs the way that makes it hard to catch. At low total flow the drop through both connections is small and the two sense points agree closely, so the split looks fine when a technician checks it during a quiet period. Pressure drop through pipe in turbulent flow rises with roughly the square of flow, so at high flow the two readings diverge sharply and the imbalance is worst exactly when the plant is working hardest. A load split checked at low demand tells you nothing about the split at peak.

Fix the piping asymmetry where you can. Where you cannot, sense header pressure at a common point downstream of both connections and feed that signal to both machines, which removes the connection drop from the control input entirely.

Two faults that look like a control problem and are not

A failed discharge check valve. When a machine unloads or stops with a check valve that no longer seats, air from the header back-feeds through it. The machine's sump repressurises, the second machine sees the leak as demand, and the pair never settles. The tell is a stopped machine whose sump gauge does not fall, or an unloaded machine whose separator vessel stays at header pressure. Treat that vessel as stored energy: before opening anything on it, lock the disconnect open under 29 CFR 1910.147, close the isolation valve, bleed through the drain, and confirm the indicating gauge that 29 CFR 1910.169 requires the receiver to carry reads zero before breaking a joint.

A drifted pressure switch. Mechanical switches drift, and a pair that was commissioned as a cascade can walk into overlap over a few years without anyone changing a setting. Verify the actual load and unload points against a calibrated gauge on the header rather than reading the dial on the switch, because the dial is the thing that drifted.

Confirming which arrangement you actually have

Do this with a gauge and a stopwatch, at a stated plant demand, and record the demand with the result because the answer changes with it.

Watch the header pressure and both machines through several cycles. Write down, for each machine, the pressure at which it loads and the pressure at which it unloads, and whether the motor stops or keeps running when it unloads. Four readings and two yes-or-no answers.

If the two load points are within a couple of psi of each other, you have a fight. If the lag machine's motor keeps running when it is not delivering, you have the expensive version of a cascade, and the contactor or the controller output that stops that motor is worth more than any setpoint you could change. And if the load points look correctly separated but the machines still both run, check the sense points: the connection drop may be putting them into overlap at working flow even though the bench settings are clean.

References

• Manufacturer documentation for each machine's published power-versus-capacity curve, its unloaded power, and its permitted stop and restart behaviour
• 29 CFR 1910.147 for lockout and stored air energy before opening a separator vessel or receiver; 29 CFR 1910.169 for the receiver's required drain and indicating gauge
• See related: How to Sequence Multiple Compressors Without Short Cycling; What a Compressor Control Strategy Is Choosing Between; Why Part Load Behaviour Decides the Energy Bill; What Pressure Drop Through Distribution Actually Costs