The utility supply and power quality

Backup power exists because the utility supply is imperfect. To design UPS, battery and generator systems sensibly, an engineer needs to know what the utility actually delivers: how a large data center is connected, how often the supply is interrupted and for how long, what voltage sags, harmonics and transients occur, and what the utility in turn expects of a large load. Lesson 1 introduced the ITIC curve and the disturbances a UPS protects against. This lesson looks at the supply side: service arrangements for data centers, reliability statistics, the causes and frequency of sags, power quality monitoring, the harmonics and other effects data centers impose on the grid, and the growing interconnection requirements for very large loads.

How data centers connect to the grid

The connection depends on size. A small enterprise data center may take a single low- or medium-voltage service like any commercial building. Large facilities, from a few megawatts to hundreds, connect at higher voltages and with redundancy:

• Single radial feed: one utility circuit; simplest and cheapest, but any fault or maintenance on it interrupts the supply, leaving everything to the backup systems.
• Dual feeds from one substation: two circuits from different transformers or bus sections, with automatic transfer between them.
• Dual feeds from separate substations: protects against the loss of a whole substation; often required by hyperscale operators.
• Dedicated on-site substation: campuses of 100 MW or more take service at transmission voltages (69 kV to 500 kV) into their own substation, with multiple transformers and lines.

Even dual feeds share some common elements (the regional transmission grid, rights-of-way, weather) and do not remove the need for on-site backup. They do reduce how often the backup must run, and they allow maintenance on one feed without relying on generators.

Reliability statistics

Utilities report reliability with standard indices (IEEE 1366):

• SAIFI (system average interruption frequency index): sustained interruptions per customer per year.
• SAIDI (system average interruption duration index): minutes of sustained interruption per customer per year.
• CAIDI (customer average interruption duration index): SAIDI/SAIFI, the average length of an interruption.
• MAIFI (momentary average interruption frequency index): momentary interruptions (typically under 1 to 5 minutes, depending on the definition) per customer per year.

Typical North American values are a SAIFI of about 1 to 1.5 and a SAIDI of about 100 to 250 minutes, excluding major storms, but values vary greatly between urban and rural systems and between utilities. A data center on a dedicated transmission connection usually sees far fewer interruptions than an average customer. Momentary interruptions and voltage sags, however, are much more frequent than sustained outages, and they are what the UPS handles daily.

Voltage sags and momentary interruptions

A voltage sag (dip) is a short reduction in voltage, typically lasting from half a cycle to a few seconds, while a fault elsewhere on the system is cleared. When lightning, a tree or an animal causes a fault on one feeder, every customer connected to the same substation sees the voltage drop until the protection clears the fault, typically in 50 to 300 ms. Customers on the faulted feeder itself may also see a momentary interruption while an automatic recloser opens and recloses the line, often several times.

Sags vastly outnumber outages. A service may see a few dozen sags a year, most to 70% to 90% of nominal voltage, while experiencing only one or two sustained outages.

import numpy as np

# A year of recorded voltage events at a data centre service: (residual voltage %, duration in s)
rng = np.random.default_rng(5)
events = [(rng.uniform(55, 89), rng.choice([0.05, 0.08, 0.1, 0.15, 0.3])) for _ in range(28)]
events += [(rng.uniform(20, 50), rng.choice([0.05, 0.1, 0.2])) for _ in range(6)]
events += [(0.0, d) for d in (0.4, 1.2, 2.5)]                 # recloser operations
events += [(0.0, 3600.0)]                                     # one long outage

def itic_ok(v, d):
    # simplified lower part of the ITIC curve for undervoltage
    if d <= 0.02:
        return True                 # up to 20 ms of zero voltage is tolerated
    if d <= 0.5:
        return v >= 70
    if d <= 10:
        return v >= 80
    return v >= 90

inside = sum(itic_ok(v, d) for v, d in events)
print(f"events recorded: {len(events)}; within the ITIC tolerance: {inside}; would disturb unprotected IT: {len(events) - inside}")
for lo, hi in ((80, 90), (70, 80), (50, 70), (10, 50), (0, 10)):
    n = sum(lo <= v < hi or (hi == 10 and v == 0) for v, d in events)
    print(f"  residual {lo:2d}-{hi:2d}%: {n} events")
print("every event is carried by the UPS; only the long outage needs the generators")

events recorded: 38; within the ITIC tolerance: 15; would disturb unprotected IT: 23
  residual 80-90%: 7 events
  residual 70-80%: 8 events
  residual 50-70%: 13 events
  residual 10-50%: 6 events
  residual  0-10%: 4 events
every event is carried by the UPS; only the long outage needs the generators

The design consequences follow directly. The UPS must ride through every event; the generators must start only when an outage lasts longer than a short delay (commonly a few seconds), so that they are not started for every sag and recloser operation; and the battery must handle repeated short discharges without wearing out (Lesson 3).

Harmonics and the data center's effect on the grid

A data center is not only a victim of power quality problems but also a source. UPS rectifiers, server power supplies, variable-speed drives for fans and pumps, and chillers draw non-sinusoidal currents, creating harmonic currents that flow back into the utility system and distort the voltage for other customers.

• Modern UPS systems use active (IGBT) rectifiers with input current distortion below about 3% to 5%, and server power supplies with active power factor correction draw near-sinusoidal current, so large modern facilities are much cleaner than older ones with six-pulse rectifiers.
• IEEE 519 sets limits on current distortion at the point of common coupling with the utility, scaled to the ratio of available short-circuit current to load current, and on voltage distortion (typically 5% total, 3% individual, for systems up to 69 kV).
• Harmonic filters (passive or active) are added where needed, for example on legacy UPS systems or large drive loads.

Other effects on the grid include voltage changes when large blocks of load start or stop, low-frequency power oscillations from synchronized AI training loads (Lesson 12), and, in the worst case, the sudden loss of a large load when many facilities transfer to backup at once.

Large-load interconnection and ride-through

A single hyperscale campus can draw as much power as a city of several hundred thousand people. Grid operators have found that data centers can behave in ways that affect the grid itself. In July 2024, a fault on a transmission line in Virginia caused about 1,500 MW of data center load to transfer to on-site backup almost simultaneously, a sudden loss of load that the grid operator had to balance. Such events prompted grid operators and the North American Electric Reliability Corporation (NERC) to study large loads and to develop requirements for them.

Emerging practices include:

• Interconnection studies for large loads, similar to those for large generators, examining their response to faults and voltage disturbances.
• Ride-through settings: UPS and transfer equipment configured so that the facility stays connected through grid faults that are cleared normally, rather than transferring to backup at the first sag, while still protecting the IT load.
• Coordination with the grid operator on when the facility transfers back to the grid after an event, so that large blocks of load do not return all at once.
• Demand response and flexibility agreements, in which data centers reduce grid consumption (by shifting workloads or running generators or batteries) during grid emergencies (Lesson 6).

Transients and surges

Lightning, capacitor bank switching by the utility, and switching of large loads produce fast transients: microsecond-scale voltage spikes that can reach many times nominal voltage. Protection is layered: surge arresters at the utility interface and medium-voltage equipment, surge protective devices (SPDs) at the service entrance and distribution boards, and the inherent filtering of double-conversion UPS units. Good grounding and bonding (DCE-310) are essential, so that surges are diverted without creating voltage differences between equipment.

Power quality monitoring

Large facilities install power quality meters at the utility service, at major switchboards and at UPS inputs and outputs. Meters meeting IEC 61000-4-30 Class A record sags, swells, interruptions, harmonics, flicker, unbalance and transients with timestamps synchronized to GPS or network time, so that events can be correlated across the facility and with the utility's records. These data:

• Show whether problems originate on the utility side or inside the facility.
• Support discussions with the utility about reliability and with equipment vendors about ride-through.
• Verify compliance with harmonic limits and interconnection agreements.
• Provide evidence after incidents, such as whether a UPS transferred to battery correctly during a sag.