Data center generator sizing and selection for standby power

Generator sizing, and the balance it has to strike

A data center generator is the source that carries the entire facility when the utility fails, holding the critical load up until power returns or the site is brought down in an orderly way. Size it from the load it actually has to carry, not from the size of the building or a watts-per-square-foot habit. The whole sizing problem is a balance between two failures.

Undersize the set and it cannot hold the load. Frequency and voltage sag when the block transfers, the engine bogs, and the plant trips on overload or low frequency right when the load needs it, which collapses the bus instead of saving it. Oversize it and you have spent capital on capacity you never use, and worse, a lightly loaded diesel wet stacks. Unburned fuel fouls the exhaust and the engine smokes and loses performance over time.

The middle is the engineered size. It carries the running load plus the largest transient the load throws at it, with enough margin for site derating and growth, and no more. Acceptance testing and load bank work prove that size after the fact, and paralleling multiple sets is how the plant gets redundancy. Both have their own guides. This one is about arriving at the number before the gear is on the order.

What load does a data center generator carry?

The data center generator carries the total coincident running load of the site, which breaks into four blocks: the IT load through the UPS, the cooling, the life-safety load, and the house load. Add them as the real load that runs together, not the sum of every nameplate, because not everything runs at once.

Cooling is the block people underestimate. The chillers, pumps, cooling towers or fans, and CRAC or CRAH units that reject the heat the IT makes can run 30 to 40 percent of total facility power, and in dense air-cooled and AI deployments that fraction climbs. The IT load is steady and well behaved on the UPS. The cooling is where the big motors live, and it is often the dominant non-IT block and the source of the worst starting transients.

Life safety is the emergency egress lighting, fire alarm, and smoke control the building code puts on emergency power. House load is the lighting, receptacles, controls, and building management system that keep people working and the plant monitored. Build the load list from the actual equipment schedules with the real running and starting characteristics of each motor, then let that list, not a rule of thumb, drive the size.

Do you size a generator in kW or kVA?

You size a generator in both, because kW and kVA are not the same number. kW is real power, the work the engine does, and it is what the IT and the cooling actually consume. kVA is apparent power, what the alternator and the conductors carry, and it is the larger number whenever the load is not purely resistive. The two relate through power factor, kW divided by kVA.

Most engine-generator sets are rated at 0.8 lagging power factor by convention, which means the alternator is built to deliver about 25 percent more kVA than the kW rating implies. A set rated 2000 kW at 0.8 PF carries 2500 kVA. Size to both numbers. A load that draws its kW at a worse power factor than 0.8 can run the alternator to its kVA limit before the engine reaches its kW limit, so the alternator, not the engine, becomes the constraint.

Confirm the genset rated power factor and the actual load power factor against the data sheet and the load study, because the 0.8 figure is a rating convention, not the power factor of your load. A modern UPS at near-unity input changes the picture, and reactive cooling load pulls it the other way. Size on the kW the engine has to make and the kVA the alternator has to carry, and check which one runs out first.

What do the ISO 8528 standby, prime, and continuous ratings mean?

ISO 8528-1, the standard manufacturers rate engine-generator sets against, defines several power ratings, and the rating sets how much of the nameplate you are allowed to use and for how long. The same physical set carries a different kW on its nameplate depending on the rating applied. Pick the rating before you compare sizes, or you are comparing different things.

Emergency standby power, the ESP rating, is the traditional standby number: power available during a utility outage, at a varying load, with no sustained overload and a limited run, commonly held around 200 hours per year at an average load factor near 70 percent. Prime power, PRP, allows unlimited hours at a varying load with a defined short overload, for sites that run the set as a working source. Continuous, COP, is for a constant load for unlimited hours, the duty of a set that runs base load all the time.

The traditional data center spec was ESP, because the set only runs on a utility failure. That is where the debate sits now, because data centers run their sets harder and longer than a classic standby load, so many specs have moved up. Confirm the rating against the manufacturer's published ratings for the exact set, because the kW differs by rating and a number quoted at one rating does not transfer to another.

The data center continuous rating

Many data centers now spec a data center continuous or mission-critical rating instead of the traditional emergency standby rating, and the reason is duty. A standby rating assumes the set runs a few hours a year at a varying load. A data center can sit through a long utility outage, run extended tests, and in some markets run for grid programs, which is more hours and a steadier load than ESP assumes.

The data center continuous rating, offered by the major manufacturers under their own names, sits between prime and standby in how it is defined. It permits an unlimited number of hours at a high average load factor, commonly cited around 85 percent of the nameplate with a varying load, with the run hours and the load factor set by the manufacturer's own definition. It is the rating used where the design treats the generator, not the utility, as the primary source on an extended event, which is the intent behind the higher Uptime Institute tiers.

The practical effect is that the data center continuous nameplate kW is lower than the emergency standby nameplate kW for the same physical set, because the set is allowed to run harder for longer. Size to the rating the spec calls for. Do not accept a set sized at the standby number for a load that will be run at the continuous duty, because the standby nameplate assumes hours the load will not honor.

Why does generator sizing consider transient load?

Generator sizing considers transient load because the set has to accept the load as a block the instant the transfer switch moves to it, and that step, not the steady kW, is often what sizes the machine. When the utility fails, the engine starts, comes up to speed, and the transfer switch hands it a large chunk of load in one or a few steps. The engine speed dips and the voltage sags while the governor and the voltage regulator catch up.

ISO 8528-5 frames generator transient performance, the frequency and voltage dip on a load step and the time to recover, and the project spec sets the band the design has to stay inside. A common data center target holds the frequency dip within a few percent and recovers in a second or two, but the numbers come from the spec and the data sheet, not from memory.

A set sized only for steady kW can carry the load all day and still fail the transfer, because it cannot swallow the block without dipping out of the allowed band and tripping the UPS to battery or stalling a motor start. Size the engine for the steady kW and the worst step, and size the alternator for the inrush, because the transient is frequently the governing case and the steady number alone hides it.

Why does the UPS make you oversize the generator?

The UPS load is the reason data center generator sizing is not a simple kW addition. The UPS rectifier is a non-linear load. Older six-pulse rectifiers pull current in pulses rich in the fifth and seventh harmonics, with total current distortion that can reach roughly 30 percent, and that distorted current makes the alternator work harder than the kW alone suggests and raises the voltage distortion on the generator bus.

The classic guidance was to oversize the generator well beyond the UPS kVA to hold the bus distortion down, sometimes by a factor approaching two for a heavy six-pulse load, while accepting a higher distortion limit on the load bus. Modern UPS designs changed the math. A twelve-pulse rectifier, an input filter, or an active front end at near-unity input power factor cuts the harmonic content sharply and shrinks the oversizing the generator needs.

Two more UPS behaviors matter for sizing. The walk-in feature ramps the UPS load onto the generator gradually after the set is up to speed, commonly over a window around 30 seconds, instead of slamming it on as a block, which eases the transient the generator has to ride. And the UPS battery bridges the start, so the generator gets the start time to come up before it has to carry the IT load. Confirm the UPS rectifier type, its input power factor and distortion, and its walk-in behavior, because they swing the generator size more than almost anything else on the load list.

How does motor starting size the generator?

Motor starting sizes a generator because a motor started across the line draws six to eight times its running current for the first moment, and that inrush hits the alternator as a sudden block of mostly reactive kVA that sags the voltage. On a data center the big motors are in the cooling plant: the chiller compressors, the condenser and chilled-water pumps, and the tower or condenser fans.

The starting method changes the size more than the motor does. Across-the-line starting throws the full inrush at the set and is the worst case for sizing. A reduced-voltage soft starter cuts the inrush, and a variable frequency drive cuts it the most, holding the starting current close to the running current so the generator barely sees the start. Much of the cooling load runs on VFDs anyway for efficiency, which softens the starting transient the generator has to carry.

The number that sizes the set is the largest motor starting against the load already on the bus, because the worst case is starting a big motor while the IT and the rest of the cooling are already running. The voltage dip on that start has to stay inside the band the sensitive load and the contactors tolerate. Reduce the inrush with soft starters or drives and you can often hold a smaller set, which is a sizing decision and a cost decision at once.

Sizing the alternator for harmonics and inrush

The alternator, the generator end, is sometimes sized larger than the engine, and motor starting and harmonics are why. The engine is sized for kW. The alternator is sized for kVA and for the heat that distorted current and reactive inrush put into its windings, so a load that is fine on kW can still drive the alternator up a frame.

Harmonic current from non-linear loads like the UPS rectifier and the VFDs does no useful work, but it heats the alternator windings and distorts the bus voltage. The alternator has a relatively high internal reactance, higher than a transformer of the same rating, so the same harmonic current produces more voltage distortion on a generator bus than on a utility-fed bus. The fixes are a larger alternator with more thermal and reactance margin, a winding pitch chosen to suppress the troublesome harmonics, or load-side mitigation that cleans the current.

Size the engine and the alternator as two questions, not one. A set quoted at a kW number can still be the wrong machine if its alternator cannot carry the reactive inrush of the largest motor start or the harmonic heating of the UPS load. The manufacturer's sizing program models both, which is why the engineered size beats the rule of thumb on a plant with heavy motor and non-linear load.

How much does altitude and temperature derate a generator?

A generator derates at altitude and at high ambient temperature because the engine has less dense air to burn fuel, so the kW you can actually use on site is less than the data sheet shows at standard conditions. Size to the derated number, because the engine does not know what the catalog says.

As a rough field frame, a naturally aspirated diesel loses on the order of 3 to 3.5 percent of power per 1000 ft of elevation, and a turbocharged engine less, often nearer 2 to 2.5 percent, with the engine maker's curves controlling the exact figure. High ambient temperature takes another cut, on the order of 1 to 2 percent per 10 degrees F above the rating temperature, and it gets steeper at the high end. The two stack. A set on a hot day at elevation can lose a meaningful chunk of its rated kW.

The derate is the manufacturer's number, not a constant you carry in your head, because it depends on the engine, the aspiration, the fuel system, and the cooling. Pull the derate curves for the exact engine at the site elevation and the worst-case design ambient, then size against that derated rating. A set that meets the load at sea level and 25 degrees C can fall short of it on the August afternoon the utility actually fails.

Sizing for redundancy: N, N+1, and 2N

Redundancy decides how many sets you buy and how big each one is, and it is a sizing decision before it is a controls decision. N is just enough capacity to carry the load with nothing to spare. N+1 carries the load on N sets and keeps one more in reserve, so a set can fail or be down for service and the load never knows. 2N is two complete independent plants, the highest redundancy and the highest cost.

That redundancy is where the choice between one large set and several smaller paralleled sets comes from. A single large set is simpler and cheaper to buy, but it is a single point of failure and you cannot service it without dropping the load. Several smaller sets on a paralleling bus give you N+1 or 2N, let you take one off for maintenance while the rest run, and let the plant grow, at the cost of paralleling switchgear and the load-sharing controls that have to be commissioned. The paralleling guide covers synchronizing, isochronous load sharing, and the unit-fail test in depth.

Size each set so the surviving sets still carry the critical load when one is lost. In an N+1 plant of equal sets, the load has to fit on N of them at their derated, rated-duty number, not on all of them, because the plus-one is the one you are allowed to lose. Size to all of them and the redundancy is a label, not a behavior.

Fuel and the on-site runtime

The fuel on site sets how long the plant can run, and the runtime is a sizing requirement of its own. NFPA 110 classifies the system by Class, the minimum hours the set runs at rated load without refueling, and a data center often carries a Class of 48, 72, or even 96 hours depending on the site and how long it has to ride an extended outage before a fuel truck can reach it.

Size the fuel system to the burn rate at full load for the required hours, plus margin, not to a round tank size. The runtime is the on-site bulk storage divided by the full-load consumption, and the day tank, the transfer pumps, and the return all have to keep up at full burn for the whole run, not just at the start. The fuel guide covers storage, day tanks, polishing, and the consumption math.

Fuel quality is part of the runtime on a standby plant. Diesel that sits in a tank for months grows water and microbial growth and plugs filters at the worst time, so the design needs fuel polishing and the runtime number assumes clean fuel that will actually burn. A plant rated for 72 hours that starves on dirty fuel in hour six did not have 72 hours.

Diesel, natural gas, or dual fuel

Diesel is the default fuel for data center standby because it stores on site, starts fast, and accepts load in one step, which is exactly what a ten-second standby start needs. The whole NFPA 110 framework and the on-site Class runtime assume a fuel you can store in a tank in the yard, and diesel fits that.

Natural gas is the alternative, and it has a real pull where emissions or fuel logistics favor it. A gas set produces lower emissions and never needs a fuel tank or a fuel truck, because the gas comes from the utility pipe. The catch for standby is that the pipe is a utility too, and an event that takes the electric utility down can take the gas pressure with it, so a gas standby plant depends on a supply it does not store. Gas sets also generally accept block load more slowly than diesel, which matters for the transient.

Dual-fuel and bi-fuel sets run on a blend or switch between fuels to get some of both. The choice is set by the emissions rules, the fuel reliability at the site, the transient the load demands, and the runtime the spec requires. Diesel still wins most data center standby specs on the strength of on-site storage and fast block-load acceptance, but the emissions pressure is moving the line.

EPA Tier 4 emissions and the standby exemption

Emissions rules can size and select the generator as hard as the load does, because they decide which engine and aftertreatment you are even allowed to install. The EPA Tier system limits diesel engine emissions, and Tier 4 Final is the strictest, cutting NOx and particulate sharply, but reaching it generally requires a full aftertreatment stack: selective catalytic reduction with diesel exhaust fluid for NOx, and a filter for particulate.

The standby case has an exemption worth knowing. The EPA generally treats true emergency standby engines, which run very few hours a year, under the less strict tiers without the full Tier 4 aftertreatment, so a genuine standby set can often be a Tier 2 or Tier 3 engine. The moment the set runs non-emergency hours beyond a tight limit, commonly around 100 hours a year for testing and maintenance with the real number set by the local air district, that exemption narrows.

This is where the rating and the emissions tier collide. A set spec'd to run at a data center continuous duty for many hours may not qualify for the standby emissions exemption and may need Tier 4 aftertreatment, a DEF supply, and a more involved air permit. Confirm the engine tier, the aftertreatment, and the permitted hours with the local air authority early, because a permit problem found late re-selects the whole plant.

Enclosure and siting

The enclosure and the site are part of selection because the set has to live where you put it and breathe, cool, fuel, and exhaust from there. Most data center sets sit outdoors in a weatherproof, sound-attenuated enclosure or a walk-in housing, in a yard sized for the sets, the fuel storage, and the clearance to service them and to connect a load bank.

Cooling and combustion air drive the enclosure. The set has to pull enough air for combustion and for the radiator to reject full-load heat at the worst-case ambient, and the intake and discharge louvers have to move that air without recirculating hot discharge back to the intake. A tight enclosure or a crowded yard that lets the discharge feed back to the intake quietly derates the set on the hot day it matters.

Sound, exhaust, and fuel finish the siting. The enclosure carries the sound attenuation the site and the neighbors require, the exhaust terminates clear of building and enclosure air intakes so the plume does not feed back, and the fuel storage sits where it can be filled and contained. Get the siting wrong and a correctly sized set underperforms, which is a sizing problem the yard created, not the math.

What is wet stacking, and why it bounds oversizing

Wet stacking is the fouling a diesel suffers when it runs too lightly loaded to get hot enough to burn all its fuel, so unburned fuel and soot collect in the exhaust, foul the turbo and the valves, and the engine smokes, loses performance, and slobbers oily residue. It is the reason oversizing is a real cost, not just wasted capital. A diesel held below about 30 percent of nameplate for long runs is the classic case.

This bites a standby data center plant two ways. An oversized set carrying a light early load wet stacks in service, and any set tested under a load too light to get the engine hot wet stacks during exercise. The standard answer is a load bank, a resistive or resistive and reactive bank that loads the set high enough to burn clean and clear the deposits. The acceptance and load bank guide covers sizing, connection, and the test criteria.

The sizing lesson is to not buy more set than the load needs on the theory that bigger is safer. A set sized with a sensible margin and exercised under real or load bank load stays healthy. A set sized two frames too big for a load that will not grow into it spends its life lightly loaded and smoking, and the load bank becomes a permanent crutch instead of a commissioning tool.

The sizing study, not the rule of thumb

Size a critical plant with an engineered sizing study, not a rule of thumb, because the interactions between steady kW, the step load, the motor inrush, the harmonic and reactive load, the power factor, and the site derate are exactly what a watts-per-square-foot guess gets wrong. The major engine manufacturers publish sizing software that models the specific engine and alternator against the actual load list, step by step, and reports the frequency and voltage dip on each step.

Feed the program the real load list: each motor with its horsepower and starting method, the UPS with its rectifier type and input characteristics, the cooling sequence and what starts when, the power factor of each block, and the site elevation and design ambient for the derate. The output is a set, an engine kW, and an alternator kVA that carries the load and rides every transient inside the spec band, with the governing case identified, which is often a motor start or the UPS step rather than the steady kW.

The rule of thumb has its place for a first pass and a sanity check, the same way it does for any sizing. It does not have a place on the submittal for a plant the building cannot lose. The study is the record that the set was sized for the load it actually has, and it is what the acceptance test and the load bank later confirm in the field.

Sizing for AI density and growth headroom

Data center load is climbing, and the rack densities driving it change how you size the generator and how much headroom you leave. AI and high-performance compute have pushed rack power from a few kilowatts to tens of kilowatts and beyond, and the cooling that rejects that heat has grown with it, often to liquid cooling that shifts where the load sits but not the fact that it grew. Size for the load the site will actually carry, including the build-out the owner intends, not just the day-one fit-out.

Headroom is a judgment, not a free upgrade. Leaving room to grow means a larger set or a paralleling bus you can add sets to, which costs capital up front and, if the growth is slow, risks the lightly loaded set wet stacking before the load arrives. A modular paralleling plant that adds sets as the load fills in usually beats one oversized set bought for a load that is years away, because each added set runs loaded instead of one big set running light.

Plan the path, then size the first step to the load that is real. Confirm the growth plan and the staging with the owner, because the difference between sizing for the densities of today and the densities the owner is buying for is the difference between a plant that fits and a plant that is wrong within a year.

Putting it together: selecting the set

Selecting the set pulls the whole study into one machine, or one bus of machines, and the order the decisions fall in matters. Start from the load: the real coincident running load with the cooling and the UPS, in kW and kVA at the actual power factor. Then the rating, standby or data center continuous, because it sets how much of the nameplate you can use. Then the transient, the largest step and the worst motor start, which often governs the size. Then the site derate for elevation and ambient.

With the size settled, the plant decisions follow. The redundancy, N+1 or 2N, sets how many sets and whether you parallel. The fuel type and the on-site runtime set the tankage and the Class. The emissions tier and the permitted hours set the engine and the aftertreatment. The enclosure and the yard set whether the set can breathe and be serviced where it sits.

None of these is independent. A data center continuous rating can collide with the standby emissions exemption. A heavy motor start can drive the alternator up a frame. A tight yard can derate a set sized at standard conditions. The selection is the machine that satisfies all of them at once, which is why the manufacturer's sizing study and the project spec, not a single rule, control the choice.

What to document

Write down what drove the size, because the next engineer who looks at the plant needs to know why the set is the size it is, and the operator who runs it needs to know what it was sized to carry. Size a set with no record of the load list and the governing case and you are left with a bare kW figure that holds up to nothing once the load grows or the plant gets questioned.

Capture the load list and the coincident running load, the rating used and its source, the steady kW and the governing transient case, the alternator kVA and what sized it, the site derate inputs and the derated rating, the redundancy scheme, the fuel type and the Class runtime, and the emissions tier and permitted hours. Record the sizing study output and its assumptions, because the assumptions are what change when the load does.

Common mistakes

• Sizing on steady kW alone and missing the step load and the largest motor start, so the set carries the load but fails the transfer.
• Ignoring the UPS rectifier harmonics and reactive load, so the alternator overheats or the bus distortion runs too high.
• Skipping the altitude and ambient derate, so a set rated at standard conditions falls short on a hot day at elevation.
• Undersizing the alternator for motor starting inrush, so the voltage dips out of band even though the engine kW is fine.
• Buying one large set with no redundancy, so a single failure or a service outage drops the critical load.
• Oversizing on the theory that bigger is safer, so the lightly loaded diesel wet stacks and smokes in service.
• Using a watts-per-square-foot rule of thumb instead of an engineered sizing study on a plant the building cannot lose.
• Spec'ing a data center continuous duty without checking it against the standby emissions exemption and the permitted hours.
• Sizing in kW and forgetting the kVA and the load power factor, so the alternator is the constraint nobody checked.

Field checklist

Standards and references

Several bodies govern different parts of generator sizing and selection, and naming the right one for the point is the credibility. ISO 8528 is the engine-generator rating standard. ISO 8528-1 defines the power ratings, ESP, PRP, and COP, that set how much of the nameplate is usable, and ISO 8528-5 covers transient performance, the frequency and voltage dip on a load step. Cite it by topic and confirm the rating definitions and the transient limits against the manufacturer's published ratings, because the data center continuous rating is a manufacturer rating built on the ISO framework, not an ISO rating itself.

NFPA 110, the standard for emergency and standby power systems, classifies the system by Type, Class, and Level, which set the transfer time, the on-site fuel runtime, and the consequence class the sizing has to meet. 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, 702 for optional standby, and 708 for critical operations power systems. The EPA Tier standards set the engine emissions and whether aftertreatment is required, and the local air authority sets the permitted hours and the permit.

Above all of these sit the engine and generator manufacturer's instructions and the project specification, which set the actual numbers: the rating, the power factor, the transient band, the derate curves, and the sizing study itself. Cite the controlling document by topic, confirm the section and the numbers against the adopted edition and the AHJ, and when a standard and the spec disagree, the stricter controlling document wins.

Units and terms

The numbers on a generator sizing come in a few forms, and reading the wrong one sizes the wrong machine. kW is real power, what the engine produces and the load consumes. kVA is apparent power, what the alternator and the conductors carry, the larger number on a non-resistive load. kVAR is reactive power, the part the inrush and the lagging load add. They relate through power factor, kW over kVA, commonly 0.8 at the genset rating.

The rating shorthand repeats across vendors. ESP is emergency standby power, PRP is prime, COP is continuous, and the data center continuous or mission-critical rating is the manufacturer rating for the heavier data center duty. The NFPA 110 Class is the runtime in hours at rated load. Frequency is hertz, 60 Hz in North America and 50 Hz in much of the world, held by engine speed. Derate is the percentage of nameplate kW lost to elevation and high ambient.

kW (real power)

The power the engine produces and the load consumes, what the engine is sized for

kVA (apparent power)

The total the alternator and conductors carry, what the alternator is sized for

Power factor (PF)

Ratio of kW to kVA; gensets are commonly rated at 0.8 lagging

ESP / PRP / COP

ISO 8528 ratings: emergency standby, prime, and continuous power

Data center continuous rating

Manufacturer rating for the heavier data center duty, between prime and standby

Step load

A block of load applied at once, the transient that often governs the size

Derate

The percent of nameplate kW lost to site elevation and high ambient temperature

Class

NFPA 110 minimum hours at rated load without refueling

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is an NFPA 110 acceptance test?
• What are the conditions to parallel two generators?
• How much fuel does a generator need?
• What is a rack unit (U)?
• What is the difference between single-mode and multimode fiber?