Harmonics, Sags, and the
Costs That Never Reach
the Model

An energy model is a story about kilowatt-hours: how many, at what hours, at what price. Power quality is a story about the shape of the waveform that delivers them. The two are measured by different instruments, owned by different departments, and almost never reconciled. That gap is where some of the largest recurring energy costs at industrial sites live, invisible to the analysis that decides the site’s energy future.

Section 01The event that costs the most is the one nobody logs

A fault occurs on a transmission line thirty miles away. Protection operates correctly and clears it in four cycles. At your site, roughly sixty-seven milliseconds pass during which service voltage falls to perhaps seventy percent of nominal. The lights flicker. Nothing trips at the utility. Nothing trips on your main breaker.

Three variable-frequency drives on the coating line detect an undervoltage on their DC bus and shut down. The process is continuous, so the material in the line is scrap. The restart takes two and a half hours: purge, reheat, requalify, rethread. Maintenance logs a drive fault. Production logs a downtime code. Nobody logs an outage, because there was no outage.

This is not an unusual event. Pacific Gas and Electric defines a voltage sag as a decrease in voltage magnitude below ninety percent of nominal that is not a complete interruption, typically lasting three to ten cycles, or fifty to one hundred sixty-seven milliseconds. Citing the Electric Power Research Institute’s Distribution Power Quality project, the utility states that a typical customer could on average experience twelve voltage sags per year from the utility, and names the causes plainly: severe weather, vehicle-pole accidents, utility equipment operations or failures, and adjacent customers.¹

Twelve times a year, something happens to the waveform that never becomes a number in the energy analysis. The financial system records scrap and overtime. The energy system records nothing at all.

Section 02The vocabulary, and where the responsibility line sits

Power quality has a precise vocabulary, and owners lose arguments with utilities and vendors mostly by using it loosely. Five terms carry nearly all the weight.

Steady-state voltage

The only national standard for voltage regulation in the United States is ANSI C84.1, which sets two bands for each nominal system voltage. In PG&E’s own summary, the Range A service voltage range is plus or minus five percent of nominal, and the Range B utilization voltage range is plus six percent to minus thirteen percent. Range A limits should be exceeded only infrequently, and sustained excursions into Range B require corrective measures within a reasonable time.²

Two details in that document matter more than the percentages. First, these limits apply to sustained voltage levels and explicitly not to momentary excursions from switching, fault clearing, or motor starting. Sustained, the utility notes, usually means a period greater than two minutes. Second, the standard separates service voltage, measured at the point of delivery and regulated by the utility, from utilization voltage at the equipment terminals, which the owner’s own wiring affects. The National Electrical Code permits up to a five percent drop inside the building. A service fully compliant at the meter can therefore deliver marginal voltage at a motor, and that difference is the owner’s responsibility.²

Sags, swells, and the missing immunity standard

Below two minutes, ANSI C84.1 stops applying and there is no equivalent. What exists instead is the CBEMA curve, published in 1977 and widened in 1996 by its successor organization to become the ITIC curve: an envelope of voltage-deviation tolerances representative of power-conscious electronics manufacturers, and a reference rather than a requirement. PG&E states the position directly. There are no general equipment susceptibility standards for momentary voltage variations.²

One industry declined to accept that. Facing very high losses from process interruptions, the semiconductor sector wrote its own specification. SEMI F47 requires that semiconductor processing, metrology, and automated test equipment tolerate sags to fifty percent of equipment nominal voltage for up to two hundred milliseconds, to seventy percent for up to half a second, and to eighty percent for up to one second, without interruption.¹ Why it exists is instructive. An analysis of thirty monitor-years of disturbance data at major semiconductor sites found 15.4 percent of recorded events fell below the 1996 CBEMA curve, averaging 5.4 such events per site per year.² The industry concluded the general curve was not a high enough bar and specified above it.

Every other industry buys equipment with whatever immunity the manufacturer happened to design in, and finds out what that is during the first bad month.

Harmonics

Nonlinear loads — variable-frequency drives, rectifiers, uninterruptible supply front ends, electronic drivers, battery chargers, induction heating, electrolysis — draw current in pulses rather than sinusoids. Those currents flow back into the distribution system, and across the system impedance they produce voltage distortion every connected customer then shares.

IEEE Std 519-2022, IEEE Standard for Harmonic Control in Electric Power Systems, approved May 13, 2022 and published August 5, 2022, structures this as a shared responsibility. Its steady-state voltage and current distortion limits apply at the user’s point of common coupling to facilities containing harmonic-producing loads, and the standard is explicit that transient conditions can exceed those limits.³ The division is this: the user is accountable for the distorted current it injects, the supplying system for the resulting voltage distortion at the shared bus. Neither party can be judged in isolation, which is why harmonic disputes are so often circular.

Flicker and power factor

Flicker is rapid, repetitive voltage variation from fluctuating loads such as arc furnaces, large compressors, welders, and crushers. It is a nuisance parameter until it becomes an interconnection condition, which in California it is. Power factor, the relationship between apparent and real power, is the one power-quality parameter already on the bill. Section 05 returns to it.

Section 03Why your reliability statistics do not contain your worst events

Owners commonly reason about site reliability from published utility indices. That reasoning fails for process loads, and it fails for a structural reason rather than a data-quality one.

Electric utilities report reliability to the federal government on Form EIA-861. The instructions for that form define the frequency index as counting sustained interruptions of over five minutes, and state directly that momentary outages are not included in determining SAIDI and SAIFI. The form also asks each utility to indicate how it defines a momentary outage, that is, below how many minutes, which means the threshold separating the two categories is not uniform across reporting entities.⁵

Three consequences follow. A circuit can post excellent duration and frequency statistics while being a poor host for a continuous process. The momentary index that would capture short interruptions, MAIFI, is collected and published far less consistently. And a voltage sag is not an interruption at any duration, so it can never appear in these indices by construction. The instrument that would record it is a power quality monitor, and most industrial services do not have one.

Section 04What a short event is actually worth

The most defensible public estimates of interruption cost in the United States come from Lawrence Berkeley National Laboratory’s customer-survey meta-analysis, funded by the Department of Energy and published in January 2015. For medium and large commercial and industrial customers, the weighted-average estimated cost of a momentary interruption is $12,952 per event in 2013 dollars. The same table gives $15.90 per average kilowatt of load and $190.70 per unserved kilowatt-hour.⁴

Compare that last figure with its counterparts at longer durations. Estimated cost per unserved kilowatt-hour is $190.70 for a momentary event, $37.40 at thirty minutes, $21.80 at one hour, and $12.10 at four hours.⁴ The per-kilowatt-hour cost falls by a factor of roughly fifteen as the event lengthens, because almost none of the cost of a short event is the energy that was not delivered. It is the restart. The report’s own explanation is that cost per unserved kilowatt-hour is relatively high for a momentary interruption because the expected amount of unserved energy over a five-minute period is small.

Two further findings shape how the number should be applied. A summer interruption is estimated at roughly twenty-one to forty-three percent higher than a non-summer one, with the gap narrowing as duration grows. And sector matters more than duration: at every duration studied, the estimated cost for manufacturing customers is more than double that for non-manufacturing customers.⁴

These are sourced figures from a national survey population and should be read as exactly that. They are appropriate for order-of-magnitude screening. They are not an estimate of what an event costs at a specific site, and a study that presents them as one is misusing them. A site’s own scrap value, restart labor, requalification time, and contractual delivery exposure produce a defensible site number. The distinction between a sourced fact and a site estimate does not soften because the source is a national laboratory.

What the LBNL data does establish firmly is the shape of the problem. If cost per unserved kilowatt-hour is fifteen times higher for a momentary event than for a four-hour one, a decision model denominated in kilowatt-hours is structurally blind to the most expensive class of event on the site.

Section 05The one place power quality already reaches the bill

Harmonics and sags do not appear on a utility invoice. Power factor does.

Under PG&E Electric Schedule B-19, Medium General Demand-Metered Time-of-Use Service, bills are adjusted based on power factor for all customers except those selecting voluntary B-19 service, with power factors rounded to the nearest whole percent. The schedule states that its rates are based on a power factor of eighty-five percent. If the average power factor is greater than eighty-five percent, the total monthly bill is reduced by the product of the power factor rate and the kilowatt-hour usage for each percentage point above. If it is below eighty-five percent, the bill is increased on the same basis. On the current sheet the power factor adjustment rate is $0.00005 per kilowatt-hour per percentage point, and the adjustment is assigned to distribution for billing purposes. Advice 7846-E, effective March 1, 2026.⁶

Per point, that is a small rate. On a site consuming several million kilowatt-hours a year, several points of movement becomes a real recurring number most owners have never located on their own bill. It is also the one power-quality line that moves when power electronics are added or removed, which makes it the easiest place to start: already measured, already billed, already sitting in twelve months of statements the owner has on file.

One caution before treating it as a full picture. A displacement power factor measured at fundamental frequency and a true power factor including harmonic distortion are not the same quantity, and a site with heavy nonlinear load can differ meaningfully between them. Confirm how the serving meter computes the figure before designing correction around it. Capacitors added to a harmonic-rich bus can worsen conditions through resonance rather than improve them.

Section 06Every path to power has a power-quality profile

Owners are frequently told that on-site generation solves power quality. It does not, by itself, and the claim is worth taking apart, because the paths behave very differently and each carries a real liability alongside its real advantage.

Rotating synchronous machines, the class that includes reciprocating engines, gas turbines, and steam turbines, contribute physical inertia and short-circuit strength to the local bus. Both are useful: they stiffen voltage against disturbance and help protective devices see and clear faults. The counterweight is that a machine paralleled with the grid does not shield the site from a utility sag. When grid voltage collapses briefly, so does the bus the machine is tied to, unless a transfer scheme or conditioning equipment stands between the disturbance and the sensitive load. Synchronous machines also introduce effects of their own: block-load acceptance produces voltage dips, starting inrush is significant, and excitation systems have to be specified for the harmonic environment they will sit in.

Inverter-coupled sources, including fuel cells, solar, batteries, and most microturbine packages, are precise, fast-acting, and capable of supporting voltage in ways rotating machines cannot, including reactive support on demand. The counterweight is that a conventional grid-following inverter is a controlled current source that behaves during a disturbance exactly as it was configured to behave and no better, and that its front end is itself a harmonic source subject to the same limits as any other nonlinear load. Grid-forming control changes this materially and is a real and improving capability, but it is a specification to verify rather than a property to assume.

The only class of equipment that directly addresses the sag problem is power conditioning: uninterruptible supplies, dynamic voltage restorers, flywheels, active filters, and static transfer switches. These do what generation does not, which is ride the load through the disturbance. Their counterweight is that they buy nothing else. They add no capacity, reduce no energy cost, consume efficiency, and place a component in series with the load that can itself fail.

No row in that table is the answer. Setting them side by side shows that a site whose problem is twelve sags a year and no capacity constraint has a very different best answer than a site facing a multi-year wait for firm capacity, and the two are routinely conflated by whoever is selling.

Section 07On-site generation is a power-quality actor, not only a power-quality fix

The clearest public evidence that ride-through is a configuration question rather than an equipment property comes from the bulk power system, where events are large enough to be investigated and published.

NERC and WECC examined four disturbances in the Western Interconnection during 2024, publishing their findings on November 5, 2025. In each case the initiating fault was normally cleared with proper protection system operation, and in each case a large quantity of inverter-based generation reduced output anyway. On March 25, 2024, a phase-to-ground fault on a 500 kV bus was followed by a 1,046 MW reduction in solar and wind output across 28 facilities. On May 12, a 345 kV line fault brought a 532 MW reduction across eight facilities. On May 20, a transformer differential trip brought a 698 MW reduction across 22 facilities. On June 15, a phase-to-phase fault on a 500 kV line brought a 934 MW reduction across 37 facilities.⁷

The mechanism is the same one that stops a coating line. A fault far away produces a brief voltage depression, and equipment physically capable of riding through it disconnected instead. In its November 2023 performance report, NERC found that fewer than one-third of the inverter settings reported were set based on equipment capability, meaning significant ride-through capability was going unused. WECC then identified 77 facilities in the Western Interconnection whose inverters were not optimized to equipment limits and began contacting them in April 2024, receiving responses from 87 percent by October 2024.⁷

One finding deserves separate attention from any owner planning to operate generation. At one facility, settings changed to improve fault performance reverted to previous values during a firmware update, and those previous settings permitted the facility to trip.⁷ Commissioning a setting is not the same as keeping it. The regulatory response has been to make ride-through mandatory rather than advisory: FERC approved PRC-028-1 and PRC-030-1 on February 20, 2025, and PRC-029-1, frequency and voltage ride-through requirements for inverter-based resources, on July 25, 2025.⁷

California already imposes the customer-side version of these obligations at the interconnection. PG&E Electric Rule 21 provides that a generating facility shall not create objectionable flicker for other customers, and that flicker at the point of common coupling should not exceed the limits of IEEE 1547-2018 section 7.2.3. When serving balanced linear loads, harmonic current injection at the point of common coupling shall not exceed the limits stated in IEEE 1547-2018 section 7.3, exclusive of any harmonic currents arising from harmonic voltage distortion already present without the facility connected. The rule also directs that a generating facility’s harmonic distortion be evaluated using the same criteria as the host loads. Advice 7692-E, effective August 29, 2025.⁸

Nor is this only a design covenant. Rule 21’s supplemental review includes Screen O, Power Quality and Voltage Tests, which asks whether voltage regulation on the line section can be maintained in compliance with Commission Rule 2, whether voltage fluctuation falls within IEEE 1547-2018 section 7.2, whether harmonic levels meet the section 7.3 limits at the point of common coupling, and whether the facility will cause voltage impacts given its volt-var settings and the circuit segment.⁸ Power quality is not a refinement applied after approval. It is a screen the project passes or fails, and failing it costs schedule.

Section 08Getting it into the study

The work required to make power quality a priced input rather than a footnote is modest, and mostly measurement rather than analysis.

1. Monitor before you model.Install a power quality recorder at the service and at one or two critical panels, and log sags, swells, transients, harmonic distortion, and power factor for long enough to include a representative weather season. A model built on an assumed disturbance rate is an assumption wearing a number’s clothes.
2. Reconcile the disturbance log against the production log.The value of monitoring is not the event count. It is the joins: which recorded disturbances correspond to downtime codes, scrap batches, or unexplained restarts. Sites routinely discover that a recurring fault has an electrical timestamp.
3. Ask the utility for the circuit’s actual event history, and for its momentary definition.Published indices exclude momentary events, and each utility declares its own threshold on the federal form. The circuit-level record is far more informative than any system average.
4. Establish equipment immunity, not equipment nameplate.The number that matters is the sag depth and duration at which each critical machine actually drops out. For most equipment it is not published, and the honest answer is a test or a documented vendor statement rather than an assumption drawn from a general tolerance curve.
5. Separate what the utility owes from what the site produces.Service voltage at the point of delivery is the utility’s responsibility under a defined standard. Voltage drop through the owner’s wiring is the owner’s, and so is harmonic current injected at the point of common coupling. Determine which side a problem sits on before escalating, or the escalation will fail.
6. Price the restart, not the kilowatt-hours.Build the site’s own cost-per-event figure from scrap value, restart labor, requalification time, lost throughput at the constraint, and contractual delivery exposure. Use published survey averages only to sanity-check the order of magnitude.
7. Check the power factor line on twelve months of bills.It is already measured and already billed. Confirm the threshold and rate on the applicable schedule, how the meter computes the figure, and the effect of any planned power electronics.
8. Put the interconnection power-quality screens on the schedule.If generation is a candidate path, flicker, harmonic injection, and volt-var behavior are review criteria with defined technical references. Treat them as schedule items with owners and dates, and require settings to be documented, verified after commissioning, and re-verified after any firmware change.

Section 09What this changes in the decision

Taking power quality seriously does not push a decision toward generation. It more often does the opposite, which is why it belongs in an owner’s-side analysis rather than a vendor’s.

Three patterns recur. Where the real problem is a dozen sags a year and capacity is adequate, the proportionate answer is usually targeted conditioning on the sensitive loads, at a small fraction of the cost of a plant, and any proposal that leads with generation should be asked why. Where a genuine capacity or timing constraint exists, power quality is not a reason to build, but it is a reason the build has to include the conditioning or transfer arrangements that first prices frequently leave out. And where load is steady and tolerant on a clean circuit, power quality is not a live issue, and a study that manufactures one to justify equipment is doing the thing this series exists to argue against.

What nobody can honestly offer is a guarantee. No party can promise how many sags a circuit will deliver next year, what they will cost, or that a given mitigation will prevent a given trip. What a study can do is state the measured disturbance history, the measured immunity of the equipment that matters, the responsibility boundary between utility and owner, the site’s own cost per event derived from its own records, and the honest power-quality profile of each path under consideration, with every figure labeled as sourced fact or site estimate and neither one dressed as the other.

That is a narrower promise than most owners are offered. It is also the only one that survives contact with the plant floor.