How We Prove It: The On/Off Test

1. Two ways to prove a saving

There are fundamentally two ways to prove that an energy-saving device works. One removes the confounding factors with mathematics. The other removes them with time.

The mathematical route is weather-normalised regression: measure the facility for a year before and a period after, build a statistical model of how consumption responds to weather, occupancy, and production, and use it to predict what the bill would have been without the intervention. The difference is the saving. It is rigorous, it is the right tool for whole-facility claims over long periods, and we cover it in full in a dedicated article on weather normalisation. Its weakness is that it is indirect — it depends on the quality of the model and on months of accumulated data.

The other route is the subject of this article: the on/off test. Instead of statistically modelling away the things that change between "before" and "after", it simply refuses to let them change. It switches the device on and off so quickly — minutes apart — that the weather, the production schedule, and the occupancy have no time to move. The only thing that differs between the two measurements is the device itself. The saving falls straight out of the comparison, with no model in between.

The core idea

Regression models away the confounders. The on/off test holds them still. If the same load is measured with the device on and then off within a few minutes, weather and production cannot have changed — so any difference in power is the device, full stop. It is the closest thing in energy work to a controlled laboratory experiment, run on a live plant.

2. Holding the world still

Every objection to a simple before-and-after comparison is really an objection about something else changing at the same time as the intervention. It was a milder year. A production line was added. A shift pattern changed. Each of these is a variable moving in the background, and regression exists to estimate and subtract those movements.

The on/off test attacks the same problem from the opposite direction. Over a window of a few minutes, the outdoor temperature is effectively constant. The same machines are running the same product at the same rate. The same lights are on. If you measure power with the correction device active, then bypass it and measure again moments later, essentially nothing in the facility has changed except the device. There is nothing to model away because nothing else moved.

In the language of the international M&V frameworks, this is a form of retrofit isolation — IPMVP Options A and B — where the effect of a measure is measured directly at the affected system, rather than inferred from the whole-facility meter (Option C, the home of the regression method). The two articles are, deliberately, the two halves of the same standard.

3. The protocol

A credible on/off test is not "turn it off once and see." Doing it properly means defeating the one enemy of a short comparison — drift, the slow change in load that happens even over minutes — through repetition and structure.

Stabilise and define the window

Let the site reach a representative, steady operating state. Fix a cycle length — typically a few minutes on, a few minutes off — short enough that weather and production hold constant across a full on-off pair.

Allow settling time after each switch

When the device is switched in or out, discard the first several seconds of data. Correction equipment and the loads themselves need a moment to settle; measuring the transient instead of the steady state corrupts the comparison.

Repeat for many cycles

One on-off pair is an anecdote. Dozens of pairs, back to back, turn it into a measurement — averaging out the small random fluctuations in load and revealing whether the difference is consistent or noise.

Interleave to cancel drift

Alternate on-off-on-off rather than testing all the “on” periods first. If load is slowly trending up or down through the test, interleaving makes that trend affect the on and off states equally, so it cancels out of the difference instead of masquerading as a saving.

Reject contaminated cycles

Flag and discard any pair where something genuinely changed mid-cycle — a large motor started, a line stopped, a chiller staged. These non-routine events break the “nothing else changed” assumption and must be removed, not averaged in.

Compute the paired difference

For each clean cycle, subtract the “on” power from the matched “off” power. Average those paired differences and put a confidence interval around them. That average, with its uncertainty, is the measured effect of the device.

Why interleaving matters

Imagine a plant slowly ramping up through the morning. If you measure every “device on” period before lunch and every “device off” period after, the natural afternoon rise in load would show up as the device increasing consumption — a complete artefact. Tight on-off-on-off interleaving makes the trend hit both states alike, so it subtracts away. It is the single most important discipline in the whole protocol.

4. What we measure, and with what

The headline number is real power, in kilowatts — the actual rate of energy use, not apparent power and not current. But a power-quality correction device changes the shape of the current as well as its size, so a proper on/off test captures the full picture on both sides of every switch:

Real power (kW) and energy (kWh) over each matched window — the saving itself.
True power factor — distinguishing genuine displacement and harmonic correction from a simple change in load, as set out in our article on power factor.
Total harmonic distortion, voltage and current — the direct evidence that the device is cleaning the waveform, not just shifting power around.
RMS current per phase — showing the reduction in conductor loading that translates into recovered network capacity.

The instrument matters as much as the method. A revenue-grade power analyser — Class 0.2S for energy accuracy, meeting IEC 61000-4-30 Class A for the power-quality parameters — logs all of the above simultaneously at the same measurement point, time-stamped so each reading can be assigned cleanly to an "on" or an "off" interval. Accuracy is not a nicety here: if the device saves a few percent, an instrument with a few percent of its own error cannot see it. The measurement chain must be an order of magnitude tighter than the effect it is trying to resolve.

5. Where we measure: the device or the site

An on/off test gives a different answer depending on where the analyser is connected — and both answers are legitimate, as long as you are clear which question each one answers.

At the corrected loads, the test proves the device's local effect — the reduction in power and distortion right where it is working. This is the cleanest, largest, least ambiguous signal, because you are measuring exactly the circuit the device acts on.
At the main incomer, the test proves the device's effect on the whole site as the utility sees it — but that signal is diluted by every other load in the building that the device does not touch, and is harder to resolve against the larger, noisier total.

This is the same locational logic that governs where correction should be installed in the first place, explored in our article on correcting at the meter. A device correcting close to the loads produces a strong, obvious result when measured there, and a real but smaller one at the incomer. Reporting both — and being explicit about which is which — is what separates an honest measurement from a flattering one.

6. Reading the result

A pile of paired differences is not yet a proof. A handful of checks decide whether the result should be believed.

Check	What it asks	What good looks like
Consistency	Do the cycles agree with each other?	The paired differences cluster tightly, not scattered across both signs
Statistical significance	Could this difference be chance?	A paired test gives p < 0.05 across the cycles
Confidence interval	How precise is the saving?	The interval around the mean saving excludes zero comfortably
Direction	Does it always save, never cost?	“Off” power consistently exceeds “on” power
Clean cycles retained	How much data survived?	Most cycles pass the non-routine-event screen

Because each cycle is a paired measurement — the same load on and off, minutes apart — the right statistical treatment is a paired comparison, which is far more powerful than comparing two independent groups. The pairing cancels out the baseline load and leaves only the effect of the device, so even a modest, consistent saving becomes statistically unmistakable across enough cycles.

A consistent small saving beats a large noisy one

A 4% saving that appears in nearly every one of forty cycles is far stronger evidence than a 9% saving that shows up in some cycles and reverses in others. Consistency across paired cycles is the signature of a real effect; scatter is the signature of noise. The on/off test is built to tell the two apart.

7. What the on/off test can and cannot prove

The on/off test is the most direct evidence available that a device reduces power under the conditions tested. That last clause is the honest limit of the method, and it matters.

A test runs over hours, at whatever load states the plant happened to be in. A real facility runs all year, across a full range of loads, seasons, and production levels. The on/off test proves the instantaneous effect at the operating points it sampled; it does not, by itself, tell you the annual saving. To get from one to the other you have to weight the measured effect across the plant's actual load-duration profile — how many hours a year it spends at each load — because a device's effect can vary with loading.

So the on/off test answers "does it work, and by how much, right now?" with near-certainty. It does not answer "what will it save over a year?" on its own. For that annual figure — the one a CFO signs off and an ESG report quotes — the direct test must be combined with a load profile, or run alongside the longer weather-normalised baseline that captures a full year of conditions.

8. How the two methods fit together

The on/off test and weather-normalised regression are not rivals. They are complementary, and the strongest M&V uses both:

The on/off test gives immediate, direct proof at commissioning. Within a single visit, it demonstrates beyond reasonable doubt that the device reduces power and cleans the waveform on this specific site — no waiting, no model, no year of data.
Weather-normalised regression gives the annualised figure. Over the following months, the whole-facility baseline captures the full range of conditions and produces the statistically validated annual saving suitable for guarantees, incentive programmes, and reporting.
Each covers the other's blind spot. The on/off test is direct but short; the regression is comprehensive but indirect. Agreement between the two — an instantaneous effect that scales to the annual figure the model independently finds — is the most convincing evidence there is.

How HarmoniQ proves a saving

We run an interleaved on/off test at commissioning with revenue-grade, Class A instrumentation to demonstrate the direct effect on your site, then maintain a weather-normalised baseline to produce the validated annual figure. The on/off result is the proof you can watch happen in an afternoon; the baseline is the number you can put in front of a board. Both are documented and independently verifiable. We install, the meter proves, then you decide.

Summary

The on/off test is the most direct way to prove a power-quality saving: alternate the device on and off in tight, interleaved cycles, measure the same load both ways with revenue-grade instrumentation, and read the saving straight off the paired difference. Because only minutes separate each "on" from each "off", weather and production cannot change — so there is nothing to model away, and the result is as close to a controlled experiment as a live plant allows.

Its discipline is in the detail: settling time after each switch, many repeated cycles, strict interleaving to cancel drift, and rejection of any cycle contaminated by a non-routine event. Its honest limit is that it proves the instantaneous effect under the conditions sampled, not the annual saving — which is why it pairs naturally with the weather-normalised regression method that captures a full year.

Run together, the two methods leave nowhere for doubt to hide. The on/off test shows the effect is real, today, on this site. The annual baseline shows what it adds up to over a year. When a direct measurement and an independent statistical model agree, a saving has stopped being a claim and become a fact.