Generator paralleling switchgear commissioning for data center standby power

Generator paralleling, and why you tie sets to one bus

Generator paralleling is connecting two or more generator sets to a common bus so they run together and share the load as if they were one larger machine. The paralleling switchgear is the gear that makes that safe: it synchronizes each set to the bus before its breaker closes, then keeps the sets sharing real and reactive power in proportion once they are tied. Without it, you have several engines that can each feed a load but cannot feed it together.

You parallel for three reasons that pull in the same direction. Capacity, because four 2 MW sets on a bus deliver 8 MW that no single available frame may reach. Redundancy, because an N+1 plant carries the load on N sets and keeps one in reserve, so a unit can fail or be down for service and the critical load never knows. And maintainability, because you can take one set off the bus, service it, and put it back without dropping the rest. A data center lives on all three at once.

The thing that fails in the field is rarely the engines. It is the paralleling, the part nobody fully commissioned. A plant of sets that each pass their own acceptance can still fail to share load, fail to close the second breaker to a live bus, or trip a healthy unit on reverse power the first time the real outage hits. Single-set acceptance is covered in its own guide. This one is about what happens when you put them on the same bus.

What are the conditions to parallel two generators?

Before a generator breaker closes onto a live bus, four conditions have to match between the incoming set and the bus: voltage, frequency, phase sequence, and phase angle. Miss any one and closing the breaker is violent. The machine slams into step with the bus, draws a current surge that can trip protection, twist the coupling, and damage windings. The switchgear exists to confirm all four before it lets the breaker close.

Voltage has to match in magnitude, because a set running higher than the bus tries to dump reactive power the instant it ties, and one running lower absorbs it. Frequency has to match, because any difference means the two waveforms are sliding past each other and the phase angle is never still. Phase sequence, the rotation order of the three phases, must be the same A-B-C on both sides; a reversed rotation can never be synchronized and is caught at installation, not at every close. Phase angle is the live one: at the instant of close the incoming voltage and the bus voltage should be as near zero degrees apart as the window allows.

Frequency is set a hair high on purpose. The incoming set is brought to just above bus frequency so it crosses zero phase angle moving in the right direction and lands slightly ahead, picking up load rather than motoring. Bring it in low and the set can close as a motor and trip on reverse power before it ever carries a watt.

Voltage match

Incoming set and bus equal in magnitude before close, so no reactive surge on tie

Frequency match

Incoming set brought to bus frequency, set slightly high so it lands picking up load

Phase sequence

Rotation order of the three phases, A-B-C, which must agree on both sides

Phase angle

Angular difference between incoming and bus voltage at the instant of close, near zero

The sync-check relay and the auto-synchronizer

Two devices stand between you and an out-of-sync close. The synchronizing control, the auto-synchronizer, actively matches the incoming set to the bus by trimming its governor for frequency and phase and its voltage regulator for magnitude, then issues the close command when all four conditions land inside the window. The sync-check relay, ANSI device 25, is the independent permissive that watches the same two voltages and only allows the breaker to close when the difference in voltage, frequency, and phase angle is inside its preset limits.

The two are not the same and they are not redundant in the lazy sense. The auto-synchronizer does the work. The 25 relay is the supervisor that says yes or no to the close, and it is wired in series with the close circuit so a software glitch, a miswire, or an operator pushing a manual close cannot put a breaker across an out-of-step bus. The synchroscope, the rotating-pointer instrument or its digital equivalent, shows the angle so a human can watch what the controls are doing, slow at twelve o'clock meaning in phase.

Never bypass the sync check to make a stubborn breaker close. That is the bypass that welds a breaker and shears a coupling. If a set will not synchronize, the cause is real: a voltage or frequency offset, a sensing problem, or a phasing error at the gear. Find it. The 25 relay refusing to close is the relay doing exactly its job.

Isochronous or droop: how do paralleled generators share load?

Paralleled generators share real power one of two ways, and which one the plant uses changes what acceptance has to prove. Droop lets engine speed, and so frequency, sag slightly as load rises, commonly 2 to 4 percent from no load to full load, and the sets naturally divide load along their matched droop curves. Isochronous holds frequency dead constant and uses a load-share line or a network between the controllers to tell each set how much of the total to carry.

Droop is the simpler, more forgiving mode, and it is how sets used to be paralleled with nothing but matched governors. The cost is that frequency moves with load, which a data center bus does not want, and sharing accuracy is only as good as how well the curves are matched. Isochronous holds 60 Hz flat regardless of load, which is what the critical load expects, but it only works because the controllers are actively talking. One set runs the bus and the rest follow its load-share signal, each holding its proportional share.

Modern paralleling switchgear runs isochronous load sharing across a network of genset controllers, not a single master holding the frequency alone. Whichever mode the gear uses, prove it under load: the test is whether the sets divide real load in proportion to their ratings and stay there as total load changes, not whether the frequency looks right at one operating point.

kW sharing and kVAR sharing are two separate jobs

A paralleled set has to share two things, and they are controlled by two different systems. Real power, kW, is shared through the governor, because real power is the torque the engine puts on the shaft and the governor controls fueling and speed. Reactive power, kVAR, is shared through the automatic voltage regulator, because reactive power is set by excitation and the AVR controls field current. Get kW sharing right and one engine can still be carrying all the reactive load while the others loaf, overheating one alternator.

This is why a resistive-only load bank cannot fully commission a paralleling plant. Resistive load is kW only. It works the governors and proves real-power sharing, but it never asks the AVRs to share reactive load, so the kVAR loop sits untested. A resistive and reactive load bank at the rated power factor, commonly 0.8, loads both and lets you watch both sets of meters divide.

On a parallel bus a kVAR sharing fault shows up as one set running a high power factor and pushing VARs while another absorbs them, the cross-current that the AVR droop or the reactive load-share line is supposed to stop. Watch the per-unit kW and kVAR at every load step. Sets that match on kW and diverge on kVAR have an excitation or AVR sharing problem, not a governor problem, and chasing the wrong one wastes a commissioning day.

The controls: genset controllers, the master, and the network

The brains of a paralleling plant live in two layers. Each set has its own genset controller that runs that engine: cranking, the governor, the AVR, the protective functions, and the close command for its own breaker. Above them sits the plant-level control, a master PLC or a distributed scheme, that arbitrates which set takes a dead bus, sequences the breakers, and manages load add and shed across the transfer switches. The two layers talk over a network or over dedicated lines.

How they talk matters for reliability. Older and simpler plants use hardwired load-share lines, an analog signal looped between controllers that carries the sharing reference. Modern plants network the controllers over a field bus, commonly a CAN bus between gensets and Modbus or a similar protocol up to the master and the building system. A network gives you metering, alarms, and remote control the hardwired loop never could, but it adds a failure mode: lose the network and the sets have to fall back to a safe sharing mode, usually droop, rather than fight each other.

The architecture decision that controls dead-bus reliability is where the first-close logic lives. For the fastest, most reliable close to a dead bus, the arbitration and breaker control should sit in the genset controllers themselves, not depend on a master PLC that adds a layer and a single point of failure. Confirm where that logic lives, because it sets whether the plant can still make the bus if the master is down.

What is a dead-bus close?

A dead-bus close is the first generator energizing a de-energized bus by simply closing its breaker once it reaches rated voltage and frequency, with no synchronizing required, because there is nothing on the bus to synchronize to. After that first set makes the bus live, every other set has to synchronize to it the normal way before it can parallel on. The first close is free; every close after it goes through the sync check.

The risk a dead bus creates is two sets closing to it at the same instant, which is an out-of-sync collision between two live machines. The switchgear prevents that with dead-bus arbitration, also called first-start logic. The sets race to be ready, and the control logic picks one winner that closes to the dead bus while signaling all the others to hold their breakers open until the bus is live and they can synchronize. A common scheme staggers each set's first-start backup timer to different values so two sets can never claim the bus together.

On a real outage the first set typically closes to the dead bus within a few seconds, often quoted around 3 to 6 seconds for the arbitration and close, which is what makes a Type 10 start achievable. Test it deliberately. Force the dead-bus condition more than once and vary which set is fastest, because the failure you are hunting is two breakers closing together or no breaker closing at all, and it only shows when you make the sets compete for the bus.

The sequence of operation, start to retransfer

The paralleling sequence is a chain, and every link has to fire in order. Normal power fails. The transfer scheme senses it and signals every set to start. The sets crank and run up to rated voltage and frequency. Dead-bus arbitration picks the first set, which closes to the dead bus and makes it live. As each remaining set reaches readiness it synchronizes to the now-live bus and parallels on, and capacity climbs as breakers close.

Load does not all land at once. As the bus builds capacity, the load add and shed scheme sequences the transfer switches and steps load on in priority order, highest-priority first, each step behind a short time delay so the bus is not asked to swallow the whole building in one block. The critical load comes on first within the transfer time; lower-priority load follows as more sets tie on and capacity allows.

Restoration runs the chain in reverse, and how it returns depends on the design. The simplest plants wait out a retransfer delay, move the load back to the utility in an open transition with a brief interruption, then run the sets unloaded to cool down and shut off. Plants designed to parallel with the utility do a closed-transition return: the sets briefly parallel the utility, soft-unload by ramping their output down to near zero, then open their breakers so the load rides through without a blink. Confirm which return the gear performs, because it changes the test and the utility's involvement.

Load add, load shed, and the priority scheme

Load management is what keeps a paralleling plant from being asked to carry more than it has online. Every load, usually each transfer switch or breaker, is assigned a priority. The control adds load in ascending priority as capacity comes online and sheds load in descending priority if capacity is lost, so the system always protects the highest-priority load with whatever sets are running.

Load add is the startup side. As sets parallel on and the available online capacity rises, the control connects loads in priority order, each step preceded by a time delay that is commonly adjustable in a small range of seconds, so the bus stabilizes between steps instead of taking a stack of blocks at once. The critical load is the first priority and comes on as soon as the first set or the required sets are tied.

Load shed is the failure side, and it is the half people skip in commissioning. If a running set fails or is taken off the bus, the remaining capacity may not cover all the connected load, so the control sheds the lowest-priority loads fast enough to keep the bus from overloading and collapsing. This is the heart of N+1 behavior. The plant gives up the non-critical load to save the critical load. Test it by pulling a set under load and confirming the right loads drop in the right order and the critical load stays up, not by reading the priority table and trusting it.

Why do paralleled generators need reverse-power protection?

Paralleled generators need reverse-power protection because on a common bus a set can stop making power and start consuming it, running as a motor driven by the other sets. The reverse-power relay, ANSI device 32, watches the direction of real power at the generator breaker and trips the set off the bus when power flows into it instead of out. The classic cause is a prime mover problem: a set loses fuel or governor control, its engine can no longer carry its share, and the bus spins it like a motor through its own alternator.

Motoring is bad in both directions. The motored engine is being driven by the others and can be damaged, and the sets still making power are now also carrying the motored unit's drag, eating capacity the load needs. The 32 relay catches it and sheds the dead weight. This is protection that a single set running alone does not need and never has; it only exists because the set is paralleled, which is exactly why a single-set acceptance never exercises it.

Reverse power is not the only paralleling-specific function. Loss of field, sometimes called reverse-VAR or device 40, catches a set whose excitation has collapsed and which is now pulling reactive power from the bus instead of supplying it. The generator breaker itself, its protective relaying, and the bus protection round out a paralleling-bus protection package that goes well beyond a standalone set. Prove the 32 by motoring a set during commissioning and confirming it trips, not by reading the relay's settings.

Grounding the paralleled system: the single-point neutral problem

Paralleling creates a grounding problem a single set never has. If each generator neutral is bonded to ground separately, the bonds tie the neutrals together through the equipment grounding system, and normal current imbalance plus harmonics now have multiple parallel paths to flow between sets. The result is circulating neutral current that does no work, heats conductors, and can read alarmingly high. Reported circulating currents on poorly grounded parallel plants run from a meaningful fraction of rated up toward rated current in the worst cases.

The third harmonic makes it worse. Each alternator produces some third-harmonic voltage set by its winding pitch, often a few percent, and because no two machines produce exactly the same harmonic, the difference drives a third-harmonic current between the neutrals through the multiple bonds. Match-pitch the alternators and the difference shrinks; mix machine designs on one bus and it grows.

The fix is a deliberate grounding scheme, not separate bonds at each set. Common approaches are a single-point ground for the paralleled system, so there is one neutral-to-ground bond for the whole bus rather than one per set, or switched neutrals where four-pole transfer switches and neutral contactors keep only one neutral path connected at a time, or neutral grounding reactors that limit both the third-harmonic circulating current and the line-to-ground fault current while keeping the system effectively grounded. The scheme has to be designed for the whole plant. Get it wrong and the symptom is hot neutrals and nuisance ground-fault trips that no single set would ever show.

The commissioning tests for a paralleling plant

Commissioning a paralleling plant runs in layers, and you do not start at the top. First, accept each generator on its own: the cold-start, the transfer time, the load bank hold, the alarms, all of it, which is the single-set acceptance covered in its own guide. A set that has not passed its own acceptance has no business going onto a shared bus, because a problem there is buried once it is one of four.

Then comes the paralleling itself, and the tests stack. Prove synchronizing and the close: each set syncs to a live bus and closes through the 25 relay inside the window. Prove load sharing: the sets divide real and reactive load in proportion across the range. Prove capacity: a resistive and reactive load bank across the paralleled bus loads the whole plant to rated kW and kVA at the rated power factor, which is the load bank work the load bank guide details. Then prove the failure behaviors: the dead-bus close and its arbitration, the reverse-power trip by motoring a set, the load shed by pulling a set under load, and the N+1 unit-fail ride-through.

The order is deliberate. You confirm the sets are individually sound, then that they cooperate, then that they degrade gracefully. Skip a layer and you find the gap during an outage instead of during commissioning, which is the difference between a finding and an incident.

How do you verify generators are sharing load?

You verify load sharing by loading the paralleled bus and reading the real and reactive power each set carries at several load points, then checking that each set's share tracks its rating across the range. Two equal sets on the bus should each carry about half the kW and about half the kVAR; if one is at 60 percent and the other at 40 percent at the same instant, they are not sharing, and the gap will only get worse as total load rises.

Take readings at steps, commonly 25, 50, 75, and 100 percent of plant rating, with a load bank that includes reactive load so both the kW and kVAR loops are exercised. Log the per-set kW, kVAR, and current at each step, plus bus voltage and frequency. The acceptance is not a single snapshot. It is that the sets stay proportional as load is added and removed, including when load steps on and off, because a sharing loop can look fine at steady state and diverge on a transient.

When the sets do not share, sort kW from kVAR before you touch anything. A kW imbalance is a governor or speed-bias problem, often mismatched droop settings or a load-share line fault. A kVAR imbalance is an AVR or excitation problem, the reactive cross-current the voltage droop is meant to kill. The meters tell you which loop to chase, and chasing the wrong one is the most common way a load-share test eats a day.

N+1 redundancy: the plant and the unit-fail test

N+1 is the reason most data center plants parallel at all. The plant is sized so the critical load is carried on N sets with one extra, so any single set can fail or be out for service and the remaining sets still cover the load. A 2N plant goes further with two independent paralleled systems, often a system-plus-system or a catcher arrangement where one block of capacity catches the load another block drops. The redundancy is only real if the plant actually behaves that way when a set is lost, and that is a test, not a calculation.

Maintainability is the quieter half of the value. Because the load rides on N sets, you can open one set's breaker, take it off the bus, service it, and parallel it back, all while the plant keeps running. A plant that cannot do that without dropping load is not maintainable, and you find that out the first time a set needs an oil change during production, which is the worst possible time.

The unit-fail test is how you prove N+1 for real. With the bus loaded to a representative load, kill a running set, by tripping its breaker or simulating a fault, and watch what the plant does: the remaining sets pick up the lost share, load shed drops the right low-priority loads if capacity demands it, and the critical load never drops. Then put the set back on the bus under load and confirm it resynchronizes and reloads cleanly. Skip this test and N+1 is a number on a one-line, not a behavior of the plant.

The integrated systems test: pull the utility

The integrated systems test is where the paralleling plant stops being tested in isolation and has to perform inside the whole power chain. The classic version is the black-building or pull-the-utility test: open the utility source for real, with the building on the line or on a representative load bank, and watch the entire sequence run on its own. Generators start, the first set takes the dead bus, the rest parallel on, load steps in priority, the UPS rides the gap, and the cooling stays alive, all without anyone touching a control.

By the time the plant reaches the IST it should already be proven on its own. The generators passed their individual acceptance, the paralleling passed its load-share, dead-bus, reverse-power, and load-shed tests, and the transfer switches were commissioned. That means a failure during the IST is a coordination or timing fault between systems, not an unproven machine, which is exactly what the IST should be left to find. Bring an unproven set into the IST and you spend the test debugging a machine instead of the integration.

The integrated test belongs to the broader power-QA scope of data center commissioning and is run as part of that pillar. The handoff from this guide is clean: the paralleling plant arrives at the IST already proven to synchronize, share, and ride a unit failure, so the IST can focus on whether it does all of that in concert with the UPS, the ATS scheme, and the mechanical plant under a real loss of utility.

NFPA 110 timing when the plant has to parallel

The transfer-time requirement does not relax because the plant parallels. NFPA 110, the standard for emergency and standby power systems, classifies the emergency power supply system by Type, the maximum seconds the load may be without acceptable power after normal fails. A Type 10 system has 10 seconds, and the life-safety and critical load case is almost always Type 10. Paralleling does not buy more time; it has to deliver power to the load inside the same window.

Paralleling makes hitting that window harder, because synchronizing and closing multiple sets takes time the single set never spent. With engine start times often in the 5 to 7 second range, getting one set to a dead bus and the load energized inside 10 seconds is achievable, but it leaves little slack, which is exactly why the dead-bus close exists. The first set energizes the bus and picks up the critical load within the Type time without waiting to synchronize, and the rest parallel on afterward to build capacity. The design has to account for the arbitration and close time in the transfer-time budget.

Confirm the timing against the adopted NFPA 110 edition, the project specification, and the authority having jurisdiction, and time the real sequence during commissioning rather than trusting the design intent. The headline number is when the load terminals see acceptable power, and on a paralleling plant that is when the first set closes to the dead bus and the load steps on, not when the last set ties in.

Closed-transition return and paralleling with the utility

How the plant hands the load back to the utility splits into two designs. Open transition breaks before it makes: it drops the load from the generators, waits, and reconnects to the utility, accepting a brief interruption. Closed transition makes before it breaks: the sets briefly parallel the live utility, soft-unload by ramping their output down to near zero, then open their breakers so the load never sees a blink. Closed transition is the gentler return, and it is common where the load cannot tolerate even a momentary open.

The moment the generators parallel the utility, even for the hundred milliseconds of a closed transition or for a longer soft-load window, the rules change. The installation now falls under the interconnection requirements for parallel sources, NEC Article 705 for interconnected electric power production sources, by topic, and the utility usually requires its own protection at the point of common coupling. That utility-grade interconnect protection, the relaying that trips the tie if the generators try to back-feed or run on after the utility is gone, is on top of the generator protection, and the utility has to agree to the parallel operation.

Soft loading and soft unloading are the same idea in both directions. On the way out, the sets parallel the utility and ramp load on gradually rather than dropping it as a block; on the way back, they ramp it off before opening. The ramp avoids the transient a hard transfer would throw at both the generators and the utility. Confirm which transition the gear performs, whether the utility permits the parallel, and what interconnect protection the utility requires, because that last item is set by the utility, not by the switchgear vendor.

What to document

A paralleling commissioning record has to prove not just that the plant ran, but that it synchronized, shared, and degraded the way the design promised. The record is what a future operator trusts when a set fails at 2 a.m. and the question is whether the plant was ever proven to ride it.

Capture it per unit and per test. For each set, record that it synchronized and closed inside the window, the real and reactive load it carried at each load step and the share against its rating, that its reverse-power trip operated on motoring, and that it took part in the dead-bus arbitration. For the plant, record the load-shed behavior and the priorities that dropped, the N+1 unit-fail ride-through result, the as-left controller and protection settings, and the integrated test outcome. Note anything adjusted and re-tested, with the before and after, because the next technician needs the baseline to see whether a setting drifted.

Common mistakes

• Bypassing the sync check to force a stubborn breaker closed instead of finding the real voltage, frequency, or phasing fault.
• Accepting load sharing on kW alone and never confirming the sets divide kVAR, leaving the reactive loop untested.
• Bonding each generator neutral to ground separately, creating multiple paths and circulating neutral and third-harmonic current.
• Shipping a plant with no reverse-power protection proven, so a motoring set is carried by the bus instead of shed.
• Never forcing the dead-bus condition, so the first-start arbitration is untested until a real outage.
• Skipping the N+1 unit-fail test, so redundancy is a number on the one-line, not a proven behavior.
• Reading the load-shed priority table and trusting it instead of pulling a set under load to watch the right loads drop.
• Treating the paralleling transfer time as relaxed because the plant parallels, when the NFPA 110 Type window still applies.

Field checklist

Standards and references

Several bodies govern different parts of a paralleling plant, and naming the right one for the point is the credibility. NFPA 110, the standard for emergency and standby power systems, sets the Type transfer time the plant has to hit, the Class runtime, and the Level, and it governs the on-site acceptance and the ongoing testing; cite it by topic and confirm the section and the numbers against the adopted edition and the AHJ, because they shift between cycles.

NFPA 70, the National Electrical Code, governs the installation by how the load is classified, with Article 700 for emergency systems, 701 for legally required standby, and 702 for optional standby. When the plant parallels with the utility, even briefly in a closed transition, the interconnection requirements of Article 705 for interconnected electric power production sources come into play, by topic, and the utility imposes its own interconnect protection. IEEE covers generator paralleling and protection practice, and the ANSI/IEEE C37.2 device-number standard is where the relay numbers come from: 25 for sync check, 32 for reverse power, 40 for loss of field.

For field acceptance testing of the electrical gear, ANSI/NETA ATS gives the inspection and test requirements before energization. Above all of these sit the genset and switchgear manufacturer's instructions and the project specification, which set the actual numbers: the synchronizing window, the sharing tolerance, the protective settings, the load-add delays, and the sequence of operation itself. The controls manufacturer governs the sequence; when a standard and the spec disagree, the stricter controlling document wins, and the AHJ has the final say on what is enforceable.

Units and terms

The numbers on a paralleling plant come in a few forms, and reading the wrong one accepts a plant that should not have passed. kW is real power, the work the engines do and what the governors share. kVAR is reactive power, set by excitation and shared by the AVRs. kVA is apparent power, what the alternators and conductors carry, and the three relate through power factor, commonly 0.8 at rating. Frequency is hertz, 60 Hz in North America, held by engine speed.

The control vocabulary repeats across vendors. Synchronizing means matching the four conditions before a close. Isochronous means constant frequency with active load sharing; droop means frequency sags with load. A dead bus is a de-energized bus the first set closes to without synchronizing. Load shed is dropping low-priority load to protect the critical load. The ANSI device numbers are shorthand: 25 is the sync-check relay, 32 is reverse power, 40 is loss of field.

Synchronizing

Matching voltage, frequency, phase sequence, and phase angle before closing a breaker to the bus

Isochronous / droop

Constant-frequency sharing via active controls, versus frequency that sags with load

Load share

Dividing real (kW) and reactive (kVAR) power between paralleled sets in proportion to rating

25 sync-check

ANSI device 25, the permissive relay that only allows a close when sync conditions are met

32 reverse power

ANSI device 32, which trips a set that is motoring, drawing power from the bus instead of supplying it

Dead bus

A de-energized bus the first set closes to without synchronizing, under first-start arbitration

Load shed

Dropping low-priority load when capacity is lost so the critical load stays up

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is an NFPA 110 acceptance test?
• What is the difference between open and closed transition transfer?
• What does it mean to parallel generators?
• How long is a generator load bank test?
• How do you size a data center generator?