Energy use benchmark: methodology

Where the numbers come from

Everything on this page is derived from OpenSynth / Faraday, published by Centre for Net Zero under CDLA-Permissive-2.0: synthetic household profiles generated from around a billion smart meter readings, sampled 1 March 2024 to 1 March 2025. We read 3,000,000 of them and cut them by the conditioning labels the dataset carries. The derivation is scripts/derive-cohorts.py, and the aggregates it produces are published in full.

The cross-check that makes it worth trusting: the baseline comes out at 9.737 kWh/day from 1,786,768 profiles. SERL Statistics Report 1 — an entirely independent source, around 13,000 real metered GB households — publishes a mean of 9.8 kWh/day. One synthetic dataset and one metered one, 0.6% apart.

Why there is no percentile

Because we do not have one. The aggregates are means. Turning a mean into “you are in the 78th percentile” requires a distribution we did not derive, and inventing one would be exactly the sort of thing this site exists not to do.

So the comparison is a ratio against a cohort average, and the bands around it are wide on purpose. Eight per cent above a population mean is not “high”; it is noise dressed up as a finding.

What each property type averages

* Bungalows are 3,401 profiles against 319,763 detached, and come out below flats — which is not credible for British bungalows. Small-sample noise in an under-represented category. Shown, flagged, not leaned on.

These cohorts exclude households with a heat pump or an EV, so the technology uplifts below add to them cleanly rather than double-counting.

Why we don’t ask for your EPC band

Because in this data it makes no difference, and asking for something that changes nothing wastes your time and implies a precision that is not there.

For a gas-heated home this is defensible: EPC measures fabric and heating, and your electricity is lights, appliances and hot water. It should not move much.

For heat pump households it is not defensible, because there insulation should dominate, and the data shows the same flatness. EPC band moves consumption by under 0.1% within a property type, including for heat pump households where it should dominate. We treat that as a limitation of the synthetic data rather than a finding, and build nothing on it. We have asked Centre for Net Zero directly whether the model conditions on EPC at all, because they are the only people who can settle it. We publish the EPC cuts anyway, because a null result is still a result and somebody may want to argue with it.

The technology uplifts

• Heat pump: +15.49 kWh/day (5,655 kWh/year), from 45,602 profiles. This is the whole difference between the two groups, not the heating alone — and the difference is a good deal more complicated than it looks. We take it apart properly here.
• Electric vehicle: +13.13 kWh/day (4,794 kWh/year), from 532,002 profiles. That implies roughly 15,339 miles a year of home charging, which is a lot of driving. If it is not your driving, enter your own mileage: we convert at 3.2 miles per kWh measured at the meter, which is lower than the figure on a car’s display because AC charging loses around a tenth on the way in.

Storage heating has no label in the dataset, so it is not a cut and not in the expected figure. We ask about it only so we can tell you that.

The optional postcode, and what it actually buys you

The dataset carries no region. What it carries instead is a k-means cluster of each household’s LSOA, computed on socio-demographic features — so it groups neighbourhoods by who lives in them, not by where they are. Two neighbourhoods in the same cluster can be three hundred miles apart.

That makes it an area type, and the page says so rather than saying “homes near you”, which would be a different claim and a false one. What it is not is negligible. Across the 29 area types the mean runs from 8.26 to 11.45 kWh a day, a spread of 1.39×. Property type — the thing every comparison tool asks for — spans 1.19×. The cut nobody asks for moves the answer more than the cut everybody asks for.

That property figure leaves bungalows out, and we should say so rather than let you discover it. Bungalows come out at 8.05 kWh a day on 3,401 profiles, below flats, which is not credible and is flagged as noise everywhere else on this page. Put it back in and property type spans 1.32× — still less than the area spread, so the point survives either way. It is a smaller gap than the headline, and quoting only the wider one would have been the easy thing to do.

We do not fold it into the expected figure. Property type and area type are correlated — semis are not spread evenly across area types — so multiplying the two ratios would double-count by an amount we cannot measure. It sits next to the main result instead, and where the two disagree, that disagreement is the useful part.

Your postcode does not reach our server. Your browser sends it to postcodes.io, gets back the LSOA code, and sends us only that. A postcode is around fifteen households; an LSOA is around fifteen hundred. The lookup table from LSOA to cluster is roughly 42,000 rows and stays on the server. There is a longer version of this on the privacy page.

Three limits worth stating. The mapping holds 33,755 in England, 6,972 in Scotland, 1,917 in Wales. Northern Ireland is not in it, and a Northern Irish postcode is told so rather than quietly matched to something wrong. Second, those English and Welsh figures are the 2021 boundary set — we assumed 2011 while building this and the counts said otherwise, which matters because the code a postcode lookup hands back by default is the 2011 one. We ask for the 2021 code first and fall back, so either works. Scotland is on its 2011 data zones, which is correct rather than inconsistent: Scotland has never re-cut them.

Third, one of the thirty area types had too few profiles to publish a mean — 0.34% of neighbourhoods fall in it. A postcode there gets told there is no figure, rather than being given the nearest one.

Seasonality

Monthly expectations come from the monthly cuts, expressed as an index against the annual mean. Heat pump households get their own index, because theirs is a completely different curve: their December runs around 1.7× their July, against roughly 1.2× for everyone else. Applying the general index to a heat pump home would understate winter by hundreds of kilowatt hours.

The reason to look at the monthly table at all is that it catches something the annual total hides. A household whose yearly figure is ordinary but whose January bill is double the expectation is almost always heating with electricity somewhere — an immersion heater doing more than it should, a plug-in heater in a cold room, or electric heating nobody mentioned.

Why this is higher than Ofgem’s typical figure

Ofgem’s typical domestic consumption value for electricity is a medium household benchmark, revised down in July 2026. The mean here runs around 40% above it. Both are correct; they answer different questions. A mean is pulled upward by the minority of homes that heat with electricity and by the long right-hand tail of heavy users, and a typical value is deliberately not.

This matters more than it sounds. Measuring your household against a median-like figure and concluding you are wasteful is one of the most common mistakes people make about their own energy use, and it leads to money spent on the wrong things.

Known limitations

• Synthetic data. Faraday is a generative model, not a meter. Its outputs match published metered statistics closely in aggregate, which is the basis for using it, but no individual profile is a real household.
• Octopus Energy customers. The training population over-indexes on smart tariffs and low-carbon technology. Centre for Net Zero resample a non-LCT subgroup to broaden the demographics, and our baseline filter lands in that subgroup, but it is not a clean national sample.
• Means only, so no percentiles, no spread, no confidence intervals.
• No real geography, no occupancy, no floor area. The area-type cut above is the nearest thing to the first and it is not the same thing. Occupancy and floor area are absent entirely, and both would probably explain more than property type does.
• Cohorts under 500 profiles are dropped entirely rather than published thin, which is why some combinations are simply absent.

Back to the benchmark, or take the underlying aggregates.