Changelog

The same ten countries and territories and the same years. This release adds context around the numbers rather than changing the picture: every official reporting area is a named place with its published figure beside what the map holds, the downloads carry per-person and per-built-area values, the color scales come from the data, and every pixel says whether the model was trained on its own area. South Korea’s non-residential figures are corrected. Pin 2026.10.4 if you need these files; 2026.10.3 stays published unchanged.

• Official reporting areas as places. The areas the official statistics are published for are now searchable places in their own right: local authority districts in Great Britain, intercommunal areas (EPCIs) in France, communes in the overseas departments, forward sortation areas in Canada and local government areas in Australia. Each uses the polygon its statistics office publishes and is named as the statistics name it, so its pixels are exactly the pixels the official figure was spread over. South Korea keeps its geoBoundaries districts. Where an OpenStreetMap place covers the same ground the two are linked (same_as), and a geoBoundaries outline that duplicated an official area is dropped; Great Britain keeps its 23 shire counties.
• The official figure beside the map. Each official area now carries the published residential and non-residential figure next to the map’s comparable sums (official_* and served_* columns in the region files, an official block in the place summaries). In Great Britain, France and the overseas departments the two are equal in every area, to within a thousandth of a GWh, and the file says so (official_relation = equal). In Ontario, Australia and South Korea the published figure only splits a coarser total, so both values ship (shape). Rooftop solar used on site and large-facility use that the statistics do not count are listed separately, so a reader can say exactly how a place total relates to the official number: Sheffield in 2024 is the published 1,898.7 GWh, plus 87.1 GWh at large sites outside the published figure, plus 11.7 GWh of rooftop solar on the gross basis.
• South Korea’s non-residential figures corrected. Where a district’s large sites exceed half of its non-residential figure, the site load the map carries is capped at that half. The cap was applied to the total and to the non-residential figure separately, so the total kept more site load than the sector did and the difference landed on non-residential: six districts (Pohang, Gwangyang, Samcheok, Incheon Dong-gu, Yeongwol and Danyang) read 19 to 418 GWh above KEPCO’s figures, up to 4.3 %. The two now agree, and every district’s non-residential sum equals KEPCO’s figure. Seven Korean sites on reclaimed land just outside the generalized coastline (74 m to 1.5 km) are now assigned to the nearest district rather than left off the map, so their estimates, about 4.2 TWh in 2024, appear as facility pixels inside the Korean totals.
• Per-person and per-built-area values in the downloads. The pixel files carry ten new columns on both bases: electricity per person for the total and for homes, and electricity per square meter of built area for the total, for homes over the residential built area and for everything else over the non-residential built area (elec_tot_kwh_per_person, elec_res_kwh_per_person, elec_tot_kwh_per_built_m2, elec_res_kwh_per_res_built_m2, elec_non_res_kwh_per_nres_built_m2 and their net_ twins). A pixel with fewer than 5 residents, or under 100 square meters of the relevant built area, has no value. They equal the map’s ratio layers exactly. A ratio GeoTIFF per basis and year joins the downloads (<country>_electricity_<basis>_ratios_<year>.tif), five bands in the same order with the column name in each band’s name tag. Population and built area themselves are not republished.
• Color scales from the data. Each map layer’s scale now rests on the 2nd, 50th and 98th percentiles of its values, weighted by population or built area for the per-person and per-area layers, and the tiles manifest publishes them (value_p02, value_p50, value_p98), so a handful of extreme pixels no longer stretches a scale. In Great Britain in 2024 electricity use per person across all sectors runs 1,330 kWh at the 2nd percentile, 2,570 at the median and 14,200 at the 98th, against a maximum of 5.7 million on a pixel with almost no residents; for homes alone it is 940, 1,380 and 2,820 kWh.
• Where the model was trained. Every pixel now says whether it lies in an area whose own figures trained the model (training_tier): an official reporting area (admin_label), a smaller training unit such as British Columbia’s communities, Hydro-Québec’s service areas, Australian postal areas or Ausgrid suburbs (unit_label), or neither (none). An area counts as trained if its figures were used in any year. The provenance file names each country’s trained regions with their share of its electricity: in Canada British Columbia 100 %, Ontario 97 % and Quebec 69 %; in Australia Victoria 88 %, Queensland 81 % and New South Wales 48 %. The map carries the tier as a layer from zoom 8, and the place summaries give the trained share weighted by electricity and by population.
• Facility points all on the map, with unique names. Points are published only where the map has a modeled pixel to hold them, year by year. Northern Ireland is modeled for 2021 to 2023 and not for 2024, so its five sites and eleven districts ship for those years and leave 2024; 16 French communes and 2 in French Guiana with no consumption in any year leave the region files. Sites that share a name within a country get their municipality appended, as in “Tereos France (Artenay)”: 35 French and 2 Canadian sites. A new scaled_to_fit flag marks a site where the map holds less than half of its estimate (13 in France, 5 in South Korea, 3 in Australia and 1 in Great Britain); it does not lower the confidence. In the grid-operator comparison a figure that is itself the source of the estimate is marked is_input and counted apart from the independent checks.
• Plainer validation text, and the data lead for Europe’s largest sites cited. The validation page’s descriptions are rewritten without internal vocabulary. The EU Emissions Trading System data used to find and size European sites is cited by its Zenodo DOI (10.5281/zenodo.20509230; the version used is 10.5281/zenodo.21414185).
• One build, one commit. Every published file that names the code it was built from names the same commit, with no uncommitted changes, and the release tag marks that commit.

The same ten countries and territories and the same years, with the sectors rebuilt around the large sites that dominate electricity use where they sit. Pin 2026.10.3 if you need these files; 2026.10.2 stays published unchanged.

• Three sectors: residential, non-residential and large facilities. The commercial / industrial split of 2026.10.2 spread a region’s industrial consumption over its built-up area, so factory load appeared over homes and offices. It is withdrawn. Non-residential is now everything that is not a home and not a large facility: shops, offices, schools, hospitals, light and medium industry, farms, transport and street lighting. Residential and the total are unchanged in meaning.
• Large facilities, one by one. Smelters, steelworks, cement, lime, glass and chemical plants, refineries, paper mills, mines, and large food and textile plants that use 10 GWh or more in at least one year are placed on their own pixels in every year, about 780 sites in all. A site stays a facility once it has reached that size, including years it ran below it or was closed; smaller sites of the same kinds stay inside non-residential. On the map each site is a clickable point with its name, operator, operating status, estimated use with a low-to-high range, how sure we are, and the sources behind the estimate. The raster spreads each site over about 2 km around it; the points are the concentrated view. The raster carries each site’s mapped value; where the official area total can’t hold the full estimate, the raster holds the smaller figure and the point shows both.
• Large sites we cannot size yet are still shown. About 60 Australian mines have production figures but no reviewed electricity-per-tonne value yet; they appear as points marked “no electricity estimate yet” and add nothing to the raster.
• Net means what the grid sees. For homes and other buildings, net is gross minus the rooftop solar they use on site; for a large facility it is gross minus its own on-site generation, so a smelter or steelworks with its own power plant shows only what it draws from the grid. Where we have no data on a site’s own generation, net equals gross and the file says so (facility_basis = self_gen_unknown).
• Regional totals can exceed the official figure. Some large-site use is not in the official regional consumption statistics. Where those statistics count what sites draw through their meters, the electricity a site makes for itself is outside them: on the gross basis it adds about 4.0 TWh in Great Britain, 5.9 TWh in France and 14 TWh in South Korea in 2024 on top of the official figures. In Great Britain the steelworks and the aluminum smelter connected straight to the transmission grid are outside them too. On the net basis these sites add little, because net is what they draw from the grid. In Australia the Tomago, Boyne Island, Portland and Bell Bay smelters are inside the official state totals but outside the local billing figures, so the local government areas that hold them read far above their billed totals. Each facility point says how the official statistics count it, and the metadata file gives the extra amount per country and year.
• New downloads. Per country, <country>_facility_points.parquet (GeoParquet, one row per site and year: the estimate on both bases with its range, on-site generation and grid export where known, the amount the raster carries for the site, status with years, method, confidence and sources); an empty file for a territory without large sites. For the release, facility_grid_side.json (and a CSV twin) with the grid-operator comparison below.
• Checked against grid operators’ own figures. Large-facility estimates are made site by site, so the hidden-area validation does not score them. The validation page now compares the largest sites with figures that grid operators publish: metered imports, the output of on-site power plants and the peak demand of the local substation. A mismatch is flagged, never used to change an estimate. Great Britain has the widest coverage. In Australia the four aluminum smelters are checked against the loads the market operator reported during grid events and against the energy of the transmission zone that holds one of them; most other Australian sites have no published grid-side figure yet.
• Hourly preview for Great Britain. The hourly files use the three new sector names. The large-facility series is a standard industrial profile applied to each site’s yearly estimate, not a measured shape; its net series is lower than gross wherever a site generates its own power. The national comparison with the grid operator’s demand now has the transmission-connected sites inside it.
• File changes in one line: elec_com_gwh / elec_ind_gwh (and their net_ twins) become elec_non_res_gwh / elec_facility_gwh, com_anchor_basis becomes non_res_anchor_basis, ind_split_basis is gone, facility_basis is new, and the map layers and place summaries use the same names.
• Corrections to these notes and files (28 October 2026). The files are unchanged; read them with these in mind.
  • An earlier version of these notes said the facility points add up to the raster. They do not wherever the official area total has no room for a site’s full estimate: there the raster holds the smaller mapped value (mapped_elec_facility_gwh) and the point carries both.
  • The region metadata files of France, Martinique and South Korea say “published gross == net” in basis_note. That holds for residential and non-residential, which have no rooftop-solar model there, but not for large facilities: their own generation makes gross higher than net.
  • Five facility points in Northern Ireland (about 130 GWh in 2024) sit outside the area this release publishes: there is no Northern Ireland consumption on the map, and their mapped_elec_facility_gwh is blank. Leave them out of any total.

Ten countries and territories: Australia is new, and every pixel of the other nine was rebuilt with three real sectors. Pin 2026.10.2 if you need these files; 2026.10.1 stays published unchanged.

• Australia. Every 1 km cell of Australia for the financial years ending 2017 to 2025 (a year labelled 2025 runs July 2024 to June 2025). National and state totals follow the Australian Energy Statistics; within a state the map is pinned to local-government-area billing totals where the distributors publish them (Queensland, most of Victoria and the Ausgrid area of New South Wales), and estimated elsewhere. Residential totals are shown on the gross basis: the electricity that rooftop solar supplies on site is added back, so the numbers describe what homes use, not only what they draw from the grid. Large identified plants (smelters, refineries, mines, gas processing) are placed at their sites from an audited facility register and count as industrial. No hourly preview yet.

• Validation chart centered on the mean. The “RMSE by scale” chart on the validation page now centers each error bar on the mean known demand of the scored areas, the same quantity nRMSE divides by, instead of the median. The validation JSON gains mean_actual_gwh next to median_actual_gwh. Validation results were recomputed for every country; the France comparison against the JRC atlas reads 11.2 % median error (9.5 % in 2026.10.1) after the sector rebuild.

• The sectors mean what they say. Until now the “large industrial” sector held only the handful of big plants we could identify one by one, and everything else that was not a home — including most factories — sat in “commercial & other”. The three sectors are now residential, commercial and industrial: industrial is manufacturing, mining and construction (the big identified plants included), commercial is every other non-residential use — shops, offices, schools, hospitals, public buildings, farms, transport, street lighting. The files rename elec_large_ind_gwh / net_large_ind_gwh to elec_ind_gwh / net_ind_gwh; residential, the non-residential sum and the total are unchanged by the split itself.

• The split follows official statistics. In every region whose statistics publish an industrial / commercial split, our industrial total equals that published share of non-residential consumption: per intercommunality in France, per province in Canada, per state in Australia, per district in South Korea, nationally for Great Britain. A new column, ind_split_basis, says whether a pixel sits in such a region (constrained) or carries our own split (model_only).

• Identified plants only where the statistics contain them. A plant whose electricity the official figures never include (in Great Britain, transmission-connected steel and aluminium works that generate much of their own power) is no longer subtracted from and added back to the map. The metadata file lists, per country, which plant types count.

• Hourly preview for Great Britain. The hourly files use the same three sector names; the industrial series keeps the industrial standard profile.

The same nine countries, years and sectors as 2026.09.2. Every pixel value was rebuilt, the download files gained geometry, and two countries’ numbers were corrected. Pin 2026.10.1 if you need these files; 2026.09.2 stays published unchanged.

• Downloads are GeoParquet. The per-year pixel tables and the region metadata file now carry a geometry column: the pixel’s 1 km cell as a polygon, and the region’s boundary. Both are in longitude/latitude (EPSG:4326) and written as GeoParquet 1.1 with a bounding-box index, so you can open a file in GeoPandas, QGIS, DuckDB or any GeoParquet reader, join it to your own boundaries, or read just one area by bounding box. Nothing else about the tables changed: pixel_id is still the same index in every country and year and still works as a join key, the GeoTIFF rasters stay on the equal-area 1 km grid, and the region file’s CSV version has no geometry. Metadata schema 2.4; everything of 2.3 still reads the same.
• Every pixel value was rebuilt. The allocation model was updated and every country’s pixels were recomputed from the latest inputs, so pixel and place values differ from 2026.09.2 everywhere, usually by small amounts. National and regional totals that are anchored to official statistics are unchanged except where noted below. The accuracy checks on the validation page were rerun for this release; the model’s error against the real values fell at every scale.
• South Korea’s totals now come from the national utility. Korea’s annual totals are now the sales reported by KEPCO, the country’s sole electricity retailer, instead of an international energy balance. The residential total rises by 8 to 14 percent (2022: 81.0 TWh instead of 75.0) because the balance we had used carried an older residential figure than the utility reports; the overall total moves by at most 2 percent. Each district’s share of the country is unchanged, and 2024 no longer has to be extrapolated. One caveat: electricity that industrial sites generate and use on their own premises, about 4 percent of the national total, is not part of utility sales and so is not on the map; every Korean pixel’s metadata says so.
• Ontario’s residential map now includes apartment buildings and a few missing utilities. The postal-area data Ontario’s residential map is built on counts only customers on the residential tariff, which leaves out apartment buildings metered as a single customer, so apartment-dense areas were getting too little. A correction now adds the missing share, about 1.4 TWh a year, to the postal areas where the census counts more homes than the utility data has customers. A second correction gives a few postal areas whose local utility is absent from that data (Cornwall, Fort Frances, Sioux Lookout, parts of Thunder Bay, Kenora and the remote north) the roughly 0.6 TWh a year they were missing. Ontario’s provincial total is unchanged. Every Canadian region’s metadata row says in its basis note which corrections apply to it.
• Everything else stays as it was. Map layers, the place search, the hourly preview for Great Britain (rebuilt so that every profile adds up to its pixel’s new annual value), and both bases are unchanged in form.

The same nine countries, years, sectors and numbers as 2026.09.1: no model change, no new files to download. This cut rewrites the words around the data and adds a few things the website asked for.

• Residential, not “homes”. The residential sector is now called residential everywhere: the map, the downloads’ metadata and the validation page. Nothing about the numbers changed.
• The map says “no data” where it means it. Map tiles now tell apart a pixel inside a country that has no value (drawn as a flat gray swatch) from a pixel outside the modeled area (transparent); exact zeros are transparent. Earlier releases drew all three the same way.
• An overall validation view. The validation page pools every held-out area of every validated country into one “Overall” entry, so the accuracy checks can be read across countries at once; the French overseas départements are labeled as territories of France.
• Pixel distributions per metric. Every named place gets a small companion file with the spread of its pixel values for every map layer (per person, per built area, the hourly statistics), fetched by the map only when asked for; the place files themselves are unchanged in size.
• Metadata schema 2.3 (additive): the tiles manifest and the place files moved to version 2; everything of 2.2 still reads the same.

The same nine countries, years, sectors and files as 2026.09, rebuilt with a new allocation model. Totals anchored to official statistics are unchanged; how each region’s electricity is spread across its 1 km pixels is what changed, so pixel and place values differ from 2026.09 everywhere.

• A new allocation model. The model that spreads a region’s consumption across its pixels was replaced. Residential, commercial & other and the total all changed; large industrial sites did not, because they come from facility reports rather than from the model.
• Measurably closer to the truth. On the check that matters most for countries with no local data, leaving a whole country out and predicting it from the others, the new model’s error is 20 percent below a simple population-based allocation across the headline checks, and it beats that allocation in 13 of the 14 (the previous model was 2 percent below). On the held-out regions inside the countries we have data for, it is 35 percent below (previously 28). Against the JRC EU Energy Atlas on the hidden French areas, half of the areas are now within 10.7 percent of the real value (previously 11.0). The countries page shows the full per-country checks for this release.
• Commercial & other in Canada and South Korea. Neither country has local data on commercial consumption, so each region’s commercial & other total is first estimated from the countries that do, then spread across the region’s pixels, with reported industrial facilities kept on their own footprint. Residential is unchanged in method.
• One id per cut. Release ids now carry a third number, so a pinned id always names exactly one set of files: this is 2026.09.1, and 2026.09 stays published unchanged. Pin the id you use.
• Everything else stays as it was. Map layers, the place search, download formats, both bases and the metadata schema (web schema 2.2) are unchanged. The hourly preview for Great Britain was rebuilt so that every profile still adds up to the new annual value of its pixel.

The first release the website shows. It covers Great Britain, Canada, metropolitan France and its five overseas départements (Guadeloupe, Martinique, French Guiana, Réunion and Mayotte), and South Korea, for every year from 2021 to 2024, at 1 km resolution.

• Three sectors in every pixel. Residential, commercial & other, and large industrial sites. The three add up exactly to the total, so sector shares always sum to 100 percent. Large industrial consumption comes from facility reports rather than from our model, so it lands on each plant’s actual footprint and is not part of our accuracy checks.
• Two ways to count electricity. Gross is everything used, including solar generated and consumed behind the meter; net is what the grid delivers. Where we have no behind-the-meter estimate, net equals gross and the metadata says so.
• Validation you can check. Test areas are whole regions held back from the model, so neighboring districts never straddle the line between what the model saw and what it is tested on. We report our error against a simple population-based allocation at every scale from neighborhood to sub-region, and for countries with no local data we report how well the model does when that country is left out entirely.
• Hourly demand for Great Britain (experimental). For every pixel, hour by hour for an average day, a shoulder day, the pixel’s own peak day and its local authority’s peak day in each season, for residential, commercial & other and large industrial, on both the gross and the net basis. Every profile adds up to the pixel’s annual value on the map. Shown on the map only while the product is in beta; no hourly downloads yet.
• Downloads and the map. Per-year pixel tables and gross and net rasters for each country, with metadata files describing units, sources and licenses; a place search covering more than 70,000 named places with precomputed totals for each; and map tiles that show true 1 km pixels at every zoom level, including per-person and per-built-area views.

Every pixel carries metadata about where its total comes from: the official statistics it is anchored to, whether local data informed the model there, and whether net equals gross there. The methods page explains how to read it.