How to size and select an industrial or medical UPS

A 100 kVA UPS that trips on a pump start is not undersized by accident. It was sized from a label, against a load nobody had measured, and the first time the real load showed up was the day it was commissioned.

Sizing runs in one direction. Measure what the load actually draws, choose the protection topology that load needs, then fix the capacity, the redundancy, the runtime and the room — in that order. Reverse it and every later step inherits the first mistake.

1. Start with three measured numbers, not the old nameplate

Every load has three numbers that matter, and they answer three different questions. Real power in kW is what the equipment consumes. Apparent power in kVA is what the UPS has to deliver, and it is larger than the kW figure whenever the power factor is below one — divide kW by power factor to get there. Peak current is what the load pulls for the cycles or seconds when a contactor closes, and it decides overload behaviour more often than either of the other two.

The kVA stamped on the unit you are replacing is the weakest of these numbers. It was itself a guess made years ago, then the installation grew: two machines added, one panel extended, a compressor moved onto the same feeder. Dividing one nameplate by another propagates the whole history.

Log the feeder instead. A week at one-minute resolution covers a full operating cycle including weekends and shift changes, and it will show whether the plant genuinely runs everything at once or simply has the capacity to. Then capture the switching events separately, because a one-minute average smooths away the peaks that trip an inverter.

The worked example below is arithmetic from three inputs, not a recommendation. It shows how much the demand on the same 62 kW load changes as the power factor moves.

Then take the third number seriously. Pumps and compressors behave very differently under starting conditions: direct-on-line starting pulls multiples of running current for several cycles, while a variable-frequency drive starts softly but can push harmonics back into the output. The starting method is part of the load description, and it belongs in your notes next to the kW figure.

2. Choose the topology before you fix the capacity

Capacity answers "how much". Topology answers "what kind of protection", and getting the second wrong makes the first irrelevant.

In a double-conversion online unit the rectifier converts the incoming AC to DC and feeds the DC bus, the inverter continuously supplies the load from that bus, and the battery sits on the same DC system and takes over without interruption when the input fails. Line-interactive units regulate the incoming supply and switch to inverter when it fails. Both are legitimate; they protect different things.

Rather than trusting marketing labels, read the classification code printed in the datasheet. IEC 62040-3 expresses it in three characters describing whether output frequency and voltage depend on the input supply, such as VFI, VI and VFD. The code tells you what the topology does with mains variations, harmonics and frequency drift — none of which you can infer from a product photograph.

Where the input fails completely, the question becomes how long the load stays unpowered when the transfer happens. For a line-interactive unit on a sensitive load, those few milliseconds are the whole design case. Moving from normal operation to battery operation is a different matter: the inverter never stopped carrying the load, so the output sees no interruption. That is not the only transfer in an online unit — inverter to static bypass, bypass back to inverter, and the transfer made under an overload each carry their own timing, and each belongs on the datasheet.

Choose line-interactive where the load tolerates brief interruptions and the mains is reasonably stable. Move to double-conversion where it does not: imaging equipment, precision instruments, process controllers that latch a fault on a missing half-cycle, and anything running behind a generator whose frequency wanders under load.

3. Read the datasheet numbers that decide whether the load survives

Two quotes look identical until you check the basis they were quoted on.

Output power factor. A unit's kVA rating and its kW capability stopped meaning the same thing when modern loads started drawing current in phase with the voltage. Compare kW capability against your measured kW, and compare kVA against your measured kVA. Some units accept any power factor up to unity; others step down their kW capability at leading load. The datasheet says which.

Overload capability. This arrives as pairs of figures — a percentage, and the seconds the unit sustains it. One machine holding 150% for 60 s and another holding it for 10 s look identical on a capacity line and behave completely differently on a motor start. Put your measured peak next to both numbers.

Input-side figures. Input current distortion and input power factor matter for the generator and the upstream breaker, not for the load. Read them with the optional input filter fitted, or compare two offers with neither fitted — mixing the two bases is how an upstream generator ends up undersized.

Battery figures. A runtime quoted at one load and one temperature describes that condition, not yours. New batteries deliver more than aged ones, warm rooms shorten life and cold ones reduce available capacity, and the corrections live in the battery maker's own tables. Take ageing and temperature from there, not from a rule of thumb.

4. Match the machine to the application

Three buyers can take the same 100 kVA unit and ask it to do three unrelated jobs. A hospital imaging room, a bank branch and a factory line differ in the supply feeding them, the environment around them and what has to happen the moment mains fails — none of which appears on the capacity line.

Hospital work brings its own limits: leakage current, earthing and the isolation requirements set by the equipment manufacturer or the site design. Where those requirements call for it, an isolation transformer becomes part of the UPS specification rather than an accessory. A CT scanner does not behave like a steady load while scanning: its power demand changes sharply from one moment to the next, so average kW alone will not size the UPS. That is a different problem from a ward distribution board that sits flat all day. Where the mains is unstable enough to matter upstream, X-ray rooms often also need a stabilizer sized to the feeder — a UPS does not stabilize a bad supply indefinitely, it rides through it and hands the load to its battery.

Financial and data-centre work is about availability tiers and how long the site must run after the mains fails. Runtime here is the interval between losing mains and having the generator carrying the load stably, not a round number someone picked.

Factory work is environmental as much as electrical. Dust, ambient temperature, humidity, vibration and washdown routines set the enclosure rating, and if a unit cannot survive the room, its electrical specification is academic. Motor starting, intermittent loads and long feeder runs are the electrical half of the same problem.

5. Choose how much redundancy you are actually buying

Single-unit capacity is not always the reason for using parallel UPS systems. TAFENG supplies single UPS units up to 800 kVA, and projects that run UPS units in parallel are usually buying redundancy, maintainability or room to expand later. Parallel units let you take one out of service for work and still protect the load. They also require more floor space, additional power connections, and a parallel control and synchronisation scheme that has to be commissioned properly.

Between N, N+1 and 2N, the decision is what actually changes between them: how much of the load each unit carries in normal operation, and what happens the moment one fails.

N+1 is a common choice where maintenance has to happen without a shutdown and a single fault has to be survivable. The price is one extra unit. The same arrangement scales down to a rack: four 10 kVA rack units paralleled into one system share the load between them, and the maintenance question is unchanged — what happens the moment you pull one.

2N belongs where the load has two supply inputs and an outage is worse than the cost of duplicating everything: batteries, distribution, cooling and physical separation included. Half-implemented 2N is a common and expensive mistake — duplicated UPS hardware sharing one battery bank, one static switch or one cable route gives you neither architecture's benefit and both bills.

Scale changes the shape of the problem rather than the logic. A 300 MW campus does not buy one machine; it buys repeated blocks, each one sized, paralleled and maintained like the blocks beside it, with failures expected as routine events rather than emergencies.

6. Size the battery from the sequence you need to survive

Runtime is not a specification, it is a duration someone has to justify. Three situations produce three very different numbers.

Riding through momentary disturbances needs seconds — enough to cover recloser operations and short interruptions that would otherwise drop contactors.

Covering a generator needs minutes, and the minutes are longer than the engine's start time. Add automatic transfer switch operation, engine start, voltage and frequency stabilisation, then load transfer, then a margin for the day something runs slow. Sites routinely size for the engine's quoted start time and are surprised when commissioning takes longer.

Orderly shutdown needs however long the longest controlled stop takes, including the systems that must close down last — not the average, the slowest.

Once the runtime is settled, the battery configuration follows: string voltage, capacity per block, number of strings, and whether they share a common bus. Where a battery string has a midpoint, the monitoring arrangement around that midpoint decides whether a failing block announces itself or quietly takes the string down later.

Then comes the DC path, which is a cable-sizing exercise rather than a distance rule. Maximum discharge current, the one-way distance to the battery cabinet, conductor cross-section, allowable voltage drop and where protection sits: five inputs, and every one of them is available before the order goes in. Get them into the enquiry and the cable size comes back calculated.

7. Plan the room before you plan the order

More installations run late over floor loading, door widths and heat than over electrical specification.

Start with physical access. A cabinet that arrives on a lorry has to travel through a door, along a corridor, possibly in a lift, and stand on a floor carrying its point loads — all with batteries inside, which is where most of the weight sits. Measure the route before signing anything.

Then heat. Every watt lost inside the room becomes air-conditioning load, and the unit needs inlet air at the temperature the manufacturer rates it for, with clear intake and exhaust paths. Equipment that cannot reject heat throttles, ages faster, or trips on a summer afternoon.

Cabling comes next: whether cables run in an overhead tray or through an underfloor void, how they are separated from signal and data runs, the bending radius the large DC conductors need, and accessible isolation so a technician can work safely without killing the whole site. Leave room to pull and replace conductors — the first installation is not the last one those trays will ever carry.

This is also the point at which surprises are cheapest. If the room cannot take the cabinet, changing to a different format costs nothing today and a great deal after delivery.

Finally, send the inputs rather than a kVA figure: measured kW and peak current with timeframe, single or three phase, available routes and floor loading, ambient temperature range, whether the load has one cord or two, the runtime your shutdown sequence needs, and what the mains actually does upstream. TAFENG engineers take that list and size the machine, the battery and the cable route together — and the first thing that comes back is usually a question about whichever input is missing.