Methodology & limitations

Connection Difficulty Score (CDS)

A single 0–100 number answering one question every developer asks: "is this place hard or easy to plug in?" Higher = harder. The score is sampled at the click point you select.

Formula

Each sub-score is on the 0–100 scale and oriented so that higher = more difficult. The weights default to the values above; the JavaScript API accepts a weights override for callers with a different project profile (e.g. a BESS investor may weight curtailment higher).

Sub-scores

1. Voltage headroom difficulty (40%). Computed as 100 − headroom_score, where headroom_score is the existing CEAtlas voltage-headroom heuristic (see §Voltage Headroom). Inputs: the highest-voltage substation within the search radius, the reference asset MW (default 100 MW), and the required voltage class from CIGRE bands.

2. Substation distance difficulty (25%). Distance to the closest viable-voltage substation, mapped through a hand-tuned curve:

3. Constraint exposure (25%). We sample the nearest GB ETYS boundary at the click point and take its mean utilisation (NESO day-ahead forecast flow ÷ limit) over the data window shown on the GB Constraint Boundaries layer — computed from the data, never a fixed number of days; the CDS detail line prints that window. Mapping:

Outside GB (or in tiles where no ETYS boundary is rendered), we use a neutral fallback of 30 and note this in the notes array returned with the score. NESO publishes day-ahead flow data for ~10 of the 34 ETYS boundaries so blind spots exist even inside GB.

4. Curtailment activity (10%). Count of BM-active units — Balancing-Mechanism units with accepted bids or offers (Elexon settlement acceptance volumes, BOAV, with BOALF system flags) in the data window shown on the GB BM Actions layer — drawn within the search radius. Units the layer cannot place from an evidenced location are not counted. Mapping:

Inputs & data sources

• Substations: PyPSA-Eur (EU) + OpenStreetMap (global, incl. 75k US via chunked Overpass)
• GB constraint boundaries: NESO ETYS boundaries + NESO Day-Ahead Constraint Flows and Limits (NESO Open Data Licence) — the data window shown on the layer, computed from the data
• GB BM actions: Elexon settlement acceptance volumes (BOAV) + acceptance flags (BOALF) (Elexon BMRS data licence) — the data window shown on the layer, computed from the data
• Voltage class thresholds: CIGRE / IEC standard bands

Known limitations

CDS is best used for relative comparison across sites, not as a go/no-go threshold. A real connection feasibility study still requires a TSO conversation and a power-flow run.

Versioning

Current version: CDS v1.0 (May 2026 launch). Weight changes and new sub-scores will be released as versioned updates with a brief changelog. If you cite a CDS value in a deck or filing, cite the version too.

24/7 CFE Study (real 8760-hour dispatch)

The study pulls a full year of actual hourly weather for the site from CEAtlas's private ERA5 hourly weather store (store-first; the Open-Meteo archive, ERA5 reanalysis family, is the transparent fallback and every result records which one served it), converts it to hourly wind and solar capacity factors, and runs a chronological hour-by-hour dispatch across all 8,760 hours with a stateful battery. That sequencing — not a 12×24 month-hour average — is what makes the 24/7 CFE %, the grid-import residual, and the storage sizing real: it sees multi-day weather lulls that an averaged model cannot.

Resource models. Wind and solar CFs come from cf_model.py v2.0.0 — the one canonical, CompoundVision-parity physics engine shared by the map, the API, the studies and the forecasts (see Capacity factor & the weather store for the full derivation). Both fuels return net expected-farm CFs: wind is an IEC 61400-12 smootherstep curve (300 W/m² reference turbine, rated 10.36 m/s at 100 m) run through Bastankhah wake, 0.97 availability, thermal, storm-hysteresis and electrical losses; solar is a full SPA → Erbs → Perez → King chain with an ILR 1.30 AC clip, for a single-axis tracker (default) or fixed tilt (solar_array). The battery round-trips at 90% (RTE), applied as a per-leg square-root. These are single-reference-plant CFs, not a specific turbine or inverter — a project's own kit shifts the absolute number.

Weather years. Any year from 1940 to the last complete calendar year is selectable; pre-1950 is ECMWF's preliminary ERA5 back-extension and is labelled as such. The multi-year mode fixes the generation mix once and re-runs the dispatch across up to 86 real weather years (1940–2025) to report a P50 with an inter-annual band — so a single calm or sunny year can't flatter the result. Percentiles follow the exceedance convention (as on the maps): P90 is the year-in-10 downside — the CFE beaten in nine years of ten, the number an offtaker or lender should underwrite — and P10 is the one-in-ten good year.

Resource droughts (worst-window analysis)

Every multi-year study also reports resource droughts: the longest contiguous windows in which combined wind + solar output stayed below 10% of the load, with storage deliberately ignored — a drought is a property of the weather and the chosen mix, not of battery sizing. The study returns each year's longest window, the P50 and P90 of those annual maxima ("plan for an N-hour drought nine years in ten" — the sentence a long-duration-storage conversation starts from), and the record's five worst windows with their actual dates. Events of the size that drives seasonal-storage economics occur a handful of times per generation; they are only measurable because the record spans 86 years.

Portfolio diversification (multi-site)

The portfolio mode takes 2–4 sites with per-site nameplates, sums their hourly generation in every weather year, and dispatches the combined profile against one load. As the baseline, each site alone carries the same total nameplate (the portfolio's aggregate wind:solar split) — so the comparison is like-for-like on capacity, and the difference is geography alone. The headline, diversification value, is the portfolio's P90 CFE minus the best single site's: what spatial spread across weather regimes is worth in the bad years, measured on real weather rather than asserted from correlation matrices. Sites in anti-correlated regimes can be worth tens of percentage points at P90; sites in the same regime are honestly reported as worth little.

Validation (published, not asserted). Solar is validated against PVGIS seriescalc across six globally-stratified sites (2020): hourly correlation 0.902 (tracker) / 0.952 (fixed), DC-basis physics bias +0.006 / +0.011 CF, GHI input agreement 0.969. On the served AC convention the CFs sit +0.058 CF above PVGIS's 1 kWp DC reference — that offset is the ILR 1.30 design (AC nameplate = DC / 1.30), reported separately, not an error. Wind is validated against independent measured data — the MERRA-2 cross-validation and the EIA-923 metered-generation study below, plus the GB settlement-meter check; the earlier PVGIS wind comparison was a same-family consistency check and is no longer the basis of the claim.

MERRA-2 cross-validation (wind, independent reanalysis). To bound reanalysis-family risk we cross-checked against NASA's MERRA-2 (GEOS — an assimilation system independent of ECMWF's) at 36 globally-stratified windy on-land sites, chain-symmetric: both sides as raw 100 m reanalysis speeds (ours native from the ERA5 store; MERRA-2's 50 m winds lifted with a per-site empirical shear exponent from its own 10/50 m ratio). The claims our studies actually rest on hold up: inter-annual variability agrees — median annual-speed correlation 0.89 over 1981–2024 (CF-domain 0.88 over 2001–2024) — and the drought years agree: at 28 of 36 sites our worst wind year of the 44 falls inside MERRA-2's worst three (~7% would by chance). Where the families part is absolute magnitude: MERRA-2 runs +28% windier than ERA5 on this sample, a documented divergence over land where published tall-tower comparisons generally favour ERA5. We publish that number rather than hide it: it is the honest scale of between-model uncertainty on absolute levels, it is exactly what per-site met-data calibration exists to collapse, and measured-generation validation (below) closes the loop. Full per-site results: merra2_crosscheck.json. Sampling frame: windy on-land regions — these statistics are for that sample, not map-wide.

EIA-923 measured-generation validation (wind, metered reality). The decisive test: 583 US onshore wind plants (87 GW, 34 states, 2019–2024) with monthly net generation from EIA-923, each re-modeled twice — once at the EIA plant point with the standard 100 m class-curve chain, and once turbine-true: actual USGS-USWTDB turbine positions (45,437 turbines), capacity-weighted hub heights, and each plant's real specific power setting the power curve. One honesty detail most comparisons skip: EIA's monthly values from annual-frequency respondents are allocations, not meter readings — so every monthly-shape statistic below uses only true monthly-metered plant-years. What holds up: the shape and the rankings — median metered-month correlation 0.88 (570 plants; p90 0.97), per-plant annual correlation 0.61, and both metered reality and the model agree 2023 was the fleet's worst wind year (the fleet-aggregate annual series correlates at 0.84, though with only six years that worst-year agreement is the sturdier claim). What runs hot: absolute levels — and the gap is mostly the old fleet. Metered generation lands at a median 82% of the turbine-true model overall, but split by build year: pre-2012 plants 74%, 2012–2015 86%, and 2016–2018 plants 92% — for plants resembling new builds (what a siting screen actually predicts), the model runs ~8% high on the turbine-true chain, and the gap varies strongly by region — ERCOT 0.75 vs SPP/CAISO 0.91–0.94 — a bundle of curtailment (forced and economic), regional model bias, and fleet losses that the data does not let us separate by name. The rest of the old-fleet gap is degradation and availability we deliberately model as a fixed modern-farm loss. Two upstream conclusions already actioned: this validation exposed the auto class layer assigning low-wind machines too widely (its class thresholds are now fleet-calibrated against the machines actually bought), and MERRA-2 is adjudicated — metered reality sits below our chain while MERRA-2 sits +28% above it, so the ERA5 family is the closer one, and still warm. Treat absolute wind CFs as indicative, use per-site met-data calibration to collapse site-level bias, and for the vintage/curtailment gap itself switch on the correction this validation now feeds: the measured-anchored wind basis (wind_basis=measured_us, next section). Full per-plant table: eia923_validation.json. Sampling frame: US onshore ≥50 MW plants fully in service before 2019, gross-of-storage.

The measured-anchored wind basis

The validation above is now a product control. The single-year and multi-year 24/7 studies (and the API) accept wind_basis=measured_us: an opt-in basis that scales the modeled wind CF by the empirical cohort median from the EIA-923 validation, fitted on the IEC-2 chain the studies actually run. The headline is what it does not change: for 2016–2018-built plants, metered generation lands at ×0.99 of this chain (95% CI 0.96–1.02) — the default study chain is measured-true for modern builds, within confidence, at national scale. The toggle's value is honesty at the edges: older-asset analysis (×0.88 for 2012–2015 builds, ×0.71 pre-2012 — via the API's basis_vintage), and the few regions that stay statistically distinct after a CI gate (SPP ×1.05 — metered above model; ERCOT ×0.58 for the pre-2012 fleet; non-plains US ×0.95). Four honesty rules govern it. Cohort-anchored, never extrapolated: every factor is the median of an observed vintage-by-region cell with a bootstrap 95% CI; a regional cell ships only when its CI excludes the cohort median, and 12 plants with ratios outside (0.3, 1.3) — capacity-attribution and repowering artifacts — are excluded, by a published, re-runnable recipe. All-in, not decomposed: a cell bundles forced and economic curtailment, regional model bias, and real fleet losses — provably inseparable by name (LBNL reports SPP curtailed more than ERCOT in 2022, 9.2% vs 4.7%, yet SPP's modern plants deliver more of the modeled energy). Annual, not hourly: the factor anchors annual energy and is applied uniformly across hours, while the losses it bundles concentrate in high-output hours — 24/7 CFE results carry roughly ±1–2 pp of extra shape uncertainty. Default-off and labeled: the resource basis remains the default, every measured-anchored result carries the factor, cell, sample size and CI in its inputs, and the maps are never silently re-baked. Known frame limits, stated plainly: the fleet is 2023 survivors (factors may flatter plants that retired); partial repowers keep their original vintage; and "new build" means 2016–2018 machines metered at ages one to eight, applied unchanged to newer builds — potentially optimistic for lifetime P50, potentially pessimistic for genuinely better machines. The full table with CIs is at wind_measured_anchor.json; non-US use falls back to labeled US-fitted cohorts — except GB, which now has its own cells: basis_vintage=gb_offshore (×0.78, CI 0.69–0.82) and gb_onshore (×0.58, CI 0.56–0.62; Scottish-weighted, as-delivered including constraint curtailment — a revenue model bears that risk, and the curtailment decomposition is published alongside). Those two factors were fitted on the August 2026 GB check (Feb 2019–Dec 2024); re-deriving them from the September 2026 rebuild below the same way (the median; CI a bootstrap of it) would give ×0.75 (CI 0.69–0.81) offshore and ×0.57 (CI 0.52–0.61) onshore — ×0.77 / ×0.57 taking the upper of the two middle sites, the convention the check's own summary medians use (both groups now have an even count). The factors in use stay ×0.78 / ×0.58 (decided 30 September 2026): each lies inside the rebuild's bootstrap CI, and the rebuild's value inside theirs, so the rebuild confirms them rather than moving them. The rebuilt file states both, the rebuild's for reference.

The GB check (settlement meters, second continent). The same comparison run against Elexon B1610 settlement-metered generation — 80 GB wind sites (36 of them offshore), 18.3 GW, Feb 2019–Dec 2025, rebuilt in September 2026 on the Fleet Census's metered joins (the reviewed BMU crosswalk) and its corrected site points — and reported the way it deserves: as a negative result with structure. What replicates is temporal shape at a fixed site (median monthly correlation 0.89; 0.93 per-site offshore, 0.83 capacity-weighted). What fails is everything a per-site correction would need: cross-site correlation is weak (offshore +0.27, onshore +0.06) — the model does not reliably rank these GB sites against each other — and per-site ratios disperse widely (offshore IQR 0.68–0.82; onshore 0.50–0.69 with outliers to 1.8). Fleet-level, metered output is ×0.71–0.77 of the IEC-2 chain offshore and ×0.57–0.62 for the onshore sample — 43 of whose 44 sites are Scottish (the other is Pen y Cymoedd in Wales), mostly transmission-metered (8 are embedded, distribution-connected units), the Scottish ones behind the B6 constraint boundary, so it reads as a constrained-fleet result, not a GB-onshore truth. On matched vintages the US-GB gap narrows by about a third (GB 2016–18 builds ×0.77 vs US ×0.99, against ×0.67 for the whole GB sample); ratios use per-month REPORTING capacity (Elexon registered), which runs up to 16% above DUKES nameplate at some sites. And unlike every other anchor, this one IS partially decomposed: reconstructing constraint curtailment from Elexon bid-offer acceptances (FPN minus accepted level, the latest acceptance at each instant — within 0.3% of Elexon's settled bid-offer volumes) shows the onshore fleet lost 18.4% of its would-be output to the grid saying no (capacity-weighted) — pre-curtailment its ratio rises to ×0.76 capacity-weighted, converging on the GB 2016–18 cohort — while offshore curtailment is ~6% capacity-weighted and concentrated in a few farms (the median offshore site lost under 0.5%), so the offshore gap (×0.75 pre-curtailment, capacity-weighted) is mostly dense-array wakes, availability and residual model bias. Commissioning ramps and B1610 reporting gaps are handled in these figures (full calendar months from the 2019–2025 archive). Consequences, stated plainly: the US-fitted measured basis must not be borrowed for GB, and no GB per-site correction is justified by this data — treat GB absolute wind CFs as indicative, fleet ratios as context (gb_wind_validation.json, dispersion stats included), and per-site truth as the job of met-data calibration. What the data does justify has shipped: a GB-fitted fleet anchor, fitted on the August 2026 build of this check (as-delivered ratios, its curtailment decomposition published beside them) — basis_vintage=gb_offshore (×0.78) and gb_onshore (×0.58), the GB cells of the measured-anchored basis above, fleet-scale by construction and deliberately not per-site corrections.

Not modelled: curtailment from grid limits (the Connection Study covers that separately), panel/turbine degradation over time, and any specific commercial turbine or inverter. Three physics terms are also deliberately omitted, each matching CompoundVision's own ensemble fallback — icing, spectral losses (~1–3%), and snow soiling (a flat 2% soiling is used instead). Wake, availability, thermal, storm and electrical losses are now modelled (they were not in the retired resource-curve model). Every result stamps its weather source — the CE ERA5 store or the Open-Meteo archive fallback — and the cf_model v2.0.0 engine version.

Connection Study (grid screening)

A bidirectional grid-connection screen at a point: for a generation, load or storage asset of a given MW, it returns the required voltage class, the nearest adequate line and its N-1-secure usable headroom, any published DNO headroom, a GB curtailment-risk read, and an indicative upgrade tier + cost.

Multi-year curtailment band (weather-driven)

Supply the generator's wind and/or solar nameplate and the screen adds a physical curtailment distribution: the hourly generation profile is dispatched against the connection's own usable (or published) headroom as a firm export limit, across 20 real weather years (default) from the 86-year archive. The result is an exceedance band — the median year's spilled share and the 1-in-20 bad year (95th percentile; for curtailment the high tail is the downside; with fewer than 20 years the response labels it a near-worst sampled year instead) — plus the median year's monthly spill pattern and a peak-window CF: the fleet's mean deliverable output share (capped at the export limit) during the Nov–Feb 16:00–18:59 UTC winter-evening stress window, a v1 capacity-value proxy for the GB/EU system only (sites outside a GB/EU box return no proxy rather than a meaningless one; it is a documented convention, not a full ELCC study). The band is deliberately un-mitigated — battery and flexible-load mitigation live in the 24/7 generation study. This marries the network screen with real weather: the same connection can spill 3% in a median year and 9% in a windy one, and the band is what a revenue model should carry, not a single draw.

Required voltage (heuristic by MW, by region). Great Britain: ≤20 MW → 33 kV, ≤60 → 66, ≤150 → 132, ≤500 → 275, >500 → 400 kV. Northern Ireland: ≤20 MW → 33 kV, ≤150 → 110, >150 → 275 kV. Continental Europe and elsewhere: ≤20 MW → 20 kV, ≤150 → 110, ≤500 → 220, >500 → 400 kV. The contiguous United States: ≤68 MW → 46 kV, ≤102 → 69, ≤171 → 115, ≤205 → 138, ≤240 → 161, ≤343 → 230, ≤750 → 345, >750 → 500 kV.

Usable headroom. A connection must survive the loss of one circuit, so firm headroom is well below a line's nameplate thermal rating. We apply an N-1 screening derate of 0.55 to the thermal rating and show both numbers. Where measured DNO/DSO published headroom exists at a nearby substation it overrides the derate estimate (with its RAG rating, licence and constraint note). Line thermal ratings are measured where available, else estimated (line_ratings.py).

Indicative connection cost (a screen, not a quote): the study adds up itemised, sourced unit costs from the cost book for the upgrade tier: a bay at an existing substation, then local reconductoring and a reinforcement allowance, then a new substation, plus the new line per km and any step-up transformer. It carries the local factor Site Development Cost uses at the point: in Great Britain, Northern Ireland and Europe the European price × the country's price level × the local labour factor; in the contiguous United States the local labour factor only, as the region's price table already carries the region; never on a step-up transformer, and on the reinforcement allowance only in the contiguous US. The line is the straight-line distance, with no steepness or access tracks: the route finder prices those.

Not modelled: the NESO TEC connection queue, power-flow / load-flow, detailed DNO works, bilateral TSO costs, or consents. This is a first-pass screen to rank and kill sites, not a connection offer.

Route finder (least-cost line routes)

The route finder draws the least-cost route for a new transmission line — from a site to the grid, or between two points — and prices it with low, central and high estimates: the line for its voltage and region, terrain and land cover, sensitive areas, ground conditions, easements, water crossings by cable, converter stations for HVDC and an allowance for each significant road, railway, waterway or line crossing. Protected land, military land and airports are avoided; on the 5 km grid, of the airport points and outlines only those with an IATA code count, as the guide says. Routes are found and priced at 100 m in the contiguous United States, Great Britain and Europe, however long the line, with water at 1 km; at 2 km on other continents; and at 5 km for a line between continents or too long for the 2 km grids.

The full method, every cost item with its source, the obstacles it weighs, the line types, coverage, limits and a step-by-step guide are in the Route finder guide.

Deliverability Pre-screen (can an area's wind, solar and batteries cover a load?)

Sizing (since November 2026). For a drawn area the pre-screen fetches, once, the area's hourly wind and solar capacity factors for a range of weather years (any years from 1940): wind from the windiest tenth of the area's points (scanned about every 25 km, the weather's own scale) and solar from the sunniest tenth (at most two of each), each hour the mean of those points; the newest years load first and earlier ones follow in the background — as if power flows freely across the area, with no network limits. You set the demand (a load type's daily shape, a flat load factor or a pasted hourly shape) and the mix (wind MW, solar MW, battery MW and hours); every change is judged at once against every year with the same hour-by-hour rules as the 24/7 study: wind and solar meet the demand first, a battery (90% round trip, empty at the start of each year) charges on a surplus and discharges on a deficit, and what is left is short or spilled. The result gives the average and the worst year's share of demand met, the hours fully covered, the biggest shortfall (the backup capacity a firm supply would need), the energy short and spilled, each year's coverage, the worst year hour by hour and its worst week. Size it for me searches for the cheapest mix (indicative unit costs) whose worst year meets a coverage target: for each battery size and wind size, the least solar that reaches it, searched on the hardest years and checked on every year.

Where in the area? Given a drawn area (bbox or freeform polygon) and a load, the pre-screen lays a ~5 km lattice of candidate cells over it, runs the same real 8760-hour 24/7 dispatch (above) at each cell against the load shape, and ranks every cell by achievable 24/7 CFE %. The full ranked list is exportable as CSV; the map highlights cells that clear your target %.

Bounds. The live solve is capped (order of 2,000–4,000 cells and a ~25 s wall budget) so a warm-cache screen stays interactive; a larger or colder area returns a partial result flagged as such rather than timing out silently. Each cell's weather year is fixed and stamped on the result and the CSV.

Interpretation: the pre-screen answers "where in this area could a 24/7 load be met best by on-site renewables?" — it is a resource + dispatch screen, not a grid-deliverability guarantee. Hand a promising cell into the Connection Study for the grid read.

Site Suitability Score

The Site Suitability layer renders 0–100 scores per technology (wind, solar, gas, nuclear, datacenter) on a hex grid. Scores combine resource quality (wind speed / GHI / population proximity), land-use exclusions, grid proximity, and policy zones. The full sub-component breakdown is shown when you click a hex.

Suitability scoring is currently CEAtlas v1.4 — see the in-app Layer info → Site Suitability for sub-component weights and band cutoffs.

Voltage Headroom

Heuristic mapping from required asset MW to required voltage class (CIGRE bands), then a check whether any substation within the search radius has the required class. Returns a VIABLE / UPGRADE / INFEASIBLE verdict with a 0–100 confidence score.

Voltage classes used: MV 11–33 kV (< 10 MW), HV 66–132 kV (10–100 MW), EHV 220–275 kV (100–500 MW), SEHV 345–400 kV (500–1000 MW), UEHV 500 kV+ (> 1000 MW).

Local Electricity Price sample

The Local Electricity Price panel samples the LMP zones polygon at your click point and returns 30-day mean / peak / min / volatility plus an annualised cost estimate at the asset's MW and a user- selectable utilisation factor (default 90%). LMP zones are GB bidding zones (NESO) for the GB region and ISO LMP zones for the US.

Capacity factor & the weather store

The wind and solar capacity-factor (CF) explorers are built on CEAtlas's own hourly weather archive — ERA5 reanalysis, 86 weather years (1940–present), stored globally at the native 0.25° grid. Every value comes from real hourly weather, not a single vendor climatology, so you can see not just the long-term average but the spread, the downside, the trend, and the shape of a typical year.

How CF is derived

Every CEAtlas CF surface — the map COGs, the /api/v1/cf endpoint, the 24/7 studies, and the CompoundVision-linked forecasts — runs through one canonical module, cf_model.py v2.0.0, a line-faithful physics port of CompoundVision's production engine. Both fuels return net expected-farm CFs (energy at the meter, after farm and system losses), not a gross resource curve. Inputs are the store primitives (100 m wind, GHI, hour-ending UTC) plus two climatology sidecars — a 1991–2020 monthly air-density field (0.644–1.532 kg/m³) and a month×hour 2 m temperature field.

Wind models a modern-onshore reference farm: a 300 W/m² reference turbine (rated 10.36 m/s at the ERA5-native 100 m hub, cut-in 3, cut-out 25 m/s), an IEC 61400-12 smootherstep curve applied in density-equivalent wind, then a Bastankhah–Gaussian wake for a 5-row 7D×5D farm, 0.97 base availability, temperature derates (×0.85 at ≥38 °C, ×0.60 at ≤−25 °C), a storm taper with shutdown hysteresis (taper 23–25 m/s, re-cut-in below 22), and load-dependent electrical (1.5–2.8%) and blockage (0.5–3.5%) losses.

Solar offers two selectable arrays — a single-axis N–S tracker (backtracking, GCR 0.35) and a fixed array at optimal tilt clip(|lat|·0.87 + 3.1, 8°, 45°) facing the equator. The chain runs the NREL SPA sun position at the −30 min accumulation midpoint (hour-ending labels verified empirically, corr 0.9978 vs 0.9513), Erbs GHI decomposition, Perez 2002 plane-of-array transposition, King IAM (0.955/0.832 diffuse factors), a King–Sandia cell temperature (climatology temperature + store wind cooling) at −0.35%/°C, flat 2% soiling, and an ILR 1.30 inverter chain (0.985 inverter × 1.5% cabling × 0.99 transformer) with a hard AC clip and a −2° night gate. The result is a net AC CF against AC nameplate.

Three loss terms are deliberately omitted, each matching CompoundVision's own ensemble fallback: icing (needs humidity/precipitation), spectral (~1–3%, conservative for c-Si), and snow soiling (a flat 2% is used instead). The old GHI / 1000 × 0.91 proxy has been retired. These are single-reference-plant CFs — a specific turbine, hub height, tracking choice, or loss stack shifts the absolute number, while the relative geography and inter-annual behaviour stay robust.

What the dropdown shows

• Statistics across 86 years: Mean, Median (P50), P90 — the bankable year-in-10 downside a lender underwrites on — Variability (inter-annual σ), and Trend (CF change per decade).
• By year — the annual-mean CF for any single year 1940–2025, with a ▶ animation that sweeps the whole record.
• By season / by month — DJF/MAM/JJA/SON and Jan–Dec climatologies (seasonal pooled by hours, leap-aware).
• Source — the CE weather store (ERA5, 86 yr), the ECMWF IFS HRES ~9 km dataset (2017–present, its own independently computed CF set at 0.1°), or the external DTU Global Wind Atlas / Global Solar Atlas long-term mean as a cross-check.
• Turbine class — IEC Class I / II / III reference machines, plus Auto: per-cell selection of the fleet-equivalent class from the record-mean 100 m wind speed, with class boundaries calibrated to the USWTDB specific-power distribution of the EIA-923-validated cohort rather than the nominal IEC 61400-1 V_ave values. The nominal bins put most of the operating US fleet's capacity on the low-wind Class III curve while the machines actually bought average a Class II specific power — installed rotor loading tracks machine vintage, not site wind, so the boundaries are an aggregate fit to the validated fleet, not a per-site suitability rule — fitted on the US fleet's 5–8 m/s band and applied globally, so outside that band the class choice is extrapolated. Auto therefore tracks the fleet-equivalent machine (modern low-SP builds trend toward the Class III curve — pick iec3 to model those explicitly). Shown as its own map and used to resolve wind_class=auto API queries.

Per-site read-out, 12×24, and downloads

Click anywhere for a full read-out (mean / P50 / P90 / variability / trend) plus:

• a 12×24 month-hour heatmap — mean CF for each calendar month × hour of day, the "8760 at a glance" that reveals the diurnal and seasonal signature (solar's midday-summer peak, wind's night/winter lift);
• CSV downloads of the site's full record — monthly (1940–2025) or hourly (paged into API-sized windows and merged), plus the 12×24 grid.

250 m downscaling (optional map overlay)

Map layers can overlay a 250 m ratio texture on the 0.25° field, w(x) = fine-scale value / ERA5-cell mean. Wind uses the Global Wind Atlas CF-domain ratio on the map; the API's hourly path instead applies the GWA speed-domain ratio through the full power chain (the same mechanism as the EIA-923 ws_mult calibration). Solar uses the Global Solar Atlas PVOUT ratio. Two further correction layers are baked into every ERA5 surface and applied identically in the hourly API: a monthly mesoscale ratio (ECMWF IFS 9 km ÷ ERA5, smoothly month-interpolated with the monthly means preserved exactly) that restores lake/coastal/convergence structure ERA5's 0.25° grid cannot resolve, and a 12×24 solar diurnal-shape table (IFS ÷ ERA5 month-hour climatology, normalised to mean 1 per month so it reshapes the day without moving monthly energy).

IFS-informed cell anchoring (textured wind maps & print posters, since Aug 2026). The anomaly-ratio construction anchors between-cell magnitude entirely to the 0.25° store means; over extreme relief (Himalaya, Pamir, Andes) that lattice was visible beneath the 250 m texture. Displayed wind-mean surfaces therefore blend the cell-scale anchor with the ECMWF IFS HRES ~9 km means: per cell, IFS supplies the local between-cell pattern while the smoothed store field sets the regional level. The blend weight is the per-cell complexity of the 250 m texture — plains cells are unchanged, ~13% of land cells blend, and the global land mean shifts by +0.16% (relative). Raw store means remain served unmodified for untextured and temporal views; the /api/v1/cf time-series paths are unaffected.

Known limitations

• ERA5 is 0.25° (~28 km) — its hourly dynamics are cell-scale even when a 250 m overlay refines the magnitude, so mountain valleys, ridgelines, and near-shore gradients share the cell's temporal shape. For micro-siting, treat CF as a regional estimate.
• Hub height is fixed at 100 m and the reference turbine/farm is generic; solar is one of two reference arrays (tracker or fixed) — a real project's kit differs.
• Solar physics is validated to ~1 pp against PVGIS. Wind shape and rankings are validated against metered US generation (583 plants: metered-month r 0.88, worst-year 2023 agreement) and cross-validated against MERRA-2 (annual r 0.89, drought agreement 28/36) — but absolute wind levels run hot, mostly for the old fleet: metered CF is 92% of the turbine-true model for 2016–2018-built plants but 74% for pre-2012 ones (degradation/availability), with curtailment visible on top (ERCOT 0.75 vs SPP/CAISO 0.91–0.94). Treat absolute wind CFs as indicative of a modern new build before curtailment, collapse the site-level uncertainty with your own met-data calibration, or switch the studies and API to the shipped measured-anchored basis (wind_basis=measured_us, with basis_vintage cohort cells) when older-fleet or regional realism matters.
• The store covers land plus all sea within ~300 km of a coast (mask v2, no Antarctica); open ocean beyond that margin is not sampled.

Percentile context & the annual-summary API

/api/v1/cfannual (free; rate-limited without a token) returns each weather year's annual and monthly mean CFs for a point across the archive — the distribution behind the climatology maps. CEAtlas uses it to place live CompoundVision production forecasts in context: the forecast-window mean CF is compared with the same calendar month across the 86-year record and labelled with an exceedance percentile — "P12 vs the 86-yr July record" means a July this strong occurs in 12% of years. It is an indicative anomaly read (a days-long forecast window against a monthly climatology), designed for owner reports, not settlement. The forecast panel also now surfaces CompoundVision's own P10–P90 uncertainty band alongside the ensemble mean.

Calibrate with your own met data

A met mast or an operating asset knows what a reanalysis chain cannot. Upload hourly measurements (timestamp,value CSV, UTC) — wind speeds at a stated mast height, or production MW with a nameplate — and CEAtlas aligns them with the modeled series over the overlap (minimum 1,000 aligned hours, at least 60% of in-archive rows matching, up to 15 archive years per upload) and returns the evidence: hourly correlation, mean bias, and a bounded calibration factor. Mast speeds yield a speed-domain multiplier against the full modeled 100 m site chain (cell weather × mesoscale × 250 m terrain, bilinear-blended over the surrounding cells exactly as the studies compute it; measurements shear-lifted with a power-law α=0.14 from a stated 10–300 m mast height, with extrapolation-heavy lifts flagged), so it reshapes hourly CFs through the power curve; production data yields a CF-domain scale. As a diagnostic, the hourly correlation is also scanned at lags of ±3 hours — a peak away from zero is the signature of a timezone or hour-labelling offset in the upload (the mean-ratio factor itself is shift-robust), and the response says so. Factors are clamped to [0.5, 1.5] — a truthful site outside that band usually signals a units or capacity mismatch, and the response says so. Nothing is stored server-side: the factor rides your subsequent study runs explicitly and appears in their inputs, so every calibrated result is labelled. This is a bias correction, not a resource assessment — low hourly correlation is reported and the mean ratio remains valid, but hourly-shape conclusions deserve care.

Weather-year ensembles for expansion planning

Capacity-expansion models are notoriously biased by their single weather draw: a plan optimised against one calm winter under-builds firm capacity, one windy year over-builds wind. CEAtlas selects a small, defensible ensemble of real weather years from the 86-year record for a planning region: representative years anchored at the P25/P50/P75 of a combined resource index (equal-weight z-scores of annual wind + solar CF, averaged over sample sites across the region), carrying the bulk of the probability mass, plus extreme years — the record-worst annual wind, annual solar, and winter (Jan/Feb/Dec) wind years — at fixed stress weights. The selection is deterministic and quantile-anchored (no clustering hyper-parameters), and exports directly in the CENovaSage bundle's weather_weights schema, so expansion runs execute as weather ensembles and their build-outs return to the map as Run Results layers. The selection is driven by the /api/v1/cfannual distribution endpoint.

Fleet Census

/census is a demographic portrait of the power fleet — an age pyramid by fuel, additions and retirements through time, and the developer-reported build pipeline. It is deliberately per-region, best-registry rather than one blended global source: each region uses the most authoritative registry available and inherits that registry's own frame. Plant popups on the map draw on the same per-plant lifecycle records, with metered-generation sparklines where a public meter exists (EIA-923 in the US, Elexon B1610 in GB, ENTSO-E A73 in the EU).

World data: Global Energy Monitor, Global Integrated Power Tracker (August 2026 release), used under CC BY 4.0. Full source licences and attributions: /licenses. The aggregates and per-plant lifecycle records behind the page are published as JSON — census: US · GB · EU · World · US pipeline + what changed (EIA-860M); lifecycle: US · GB · EU · World.

CONUS price forecast accuracy

The CONUS daily cycle publishes a day-ahead price forecast for every US market region each morning (the CENovaSage Run overlay shows the nodal product). Its accuracy record is public and lives in two places drawn from the same numbers: /scorecard, the persisted record re-scored every morning against what actually happened, and the free Price Forecast Accuracy (CONUS) map layer (Markets & Economics) — one marker per scored hub / zone.

Definitions

• Skill vs yesterday's price — the headline: 1 − MAE_model / MAE_persistence, where persistence is the market's own price at the same hour of the previous market day — the cheapest honest forecast a trader already has — computed on the same hours for both sides. 0 = no better than yesterday's price; −1 = twice its error; +0.5 = half. A negative headline is the record working, not a display error — it is shown as measured, with the number of regions, days and the date span beside it.
• Lead — day-ahead is the run made the day before the scored day (the forecast a subscriber could have bid with — the page's headline view); same-day is the run made on the scored day (a nowcast; the daily calibration gate evaluates it).
• Scored against — the market's day-ahead price (da) or its real-time price (rt: hourly, or hourly means of 5-minute intervals). Only day-ahead rows are ever gated.
• Before gate — every day-ahead row carries the run's UTC solve stamp and a flag against that market's day-ahead gate closure, measured per row and never asserted: PJM 10:30 EPT, ISO-NE 10:00 EPT, MISO 10:30 EST, SPP 09:30 CPT, ERCOT 10:00 CPT, CAISO 10:00 PPT, NYISO 05:00 EPT. NYISO's gate (09:00 UTC in summer, 10:00 UTC in winter) closes before the 09:30 UTC cycle's solves land, so NYISO runs in its own early cycle at 06:00 UTC and solves at about 06:15 UTC, before the gate; its rows scored before 13 September 2026 were solved after the gate and stay flagged. Days published after the gate are still scored, and flagged.
• Hours / day — how many hours of the 24-hour local market day were scored. The published hours run through 07:00 UTC of D+2 so every local day completes; days scored before that extension show the truncated evening (19 of 24 hours in the Eastern zones, 18 Central, 16 Pacific) — reported, never silently dropped. Since 26 September 2026 each run also looks a further day ahead (a 96-hour model horizon) so the end of the day-ahead day is not priced at the edge of the model; those look-ahead hours are neither published nor scored.
• Artefact bus-hours — bus-hours inside a hub / zone aggregate priced at a solver construct: the flat $250 export-relief price for 6 or more hours of a day, a price below −$400 (a dual, not a price — every ISO's bid floor is higher), a price above the ISO cap, or a must-run unit backed below its floor where the engine flags it. They are counted on the face and never removed from the score; an ex-artefact MAE is shown beside it as a diagnostic only, because a construct we created is ours to fix, not to exclude.
• Era / system — windows never pool engine eras (the July workstation era vs the daily runner from 2026-08-25), and every row carries a system label (era | topology revision | inputs revision) so model generations are never averaged together.
• Hub / zone level, on purpose — markers and rows are the market's published pricing points, not individual buses: the review of 2026-09-10 measured that the model's intra-zone price structure carries no skill (correlation ≈ 0 with the observed intra-zone deviations), so per-bus markers would draw a precision the numbers do not have.

What it is scored against

• ISO-NE — 8 load zones; DA and RT hourly LMP (public histRpts files).
• NYISO — zones A–K; DA zonal LBMP and hourly-integrated RT LBMP (public MIS archive).
• PJM — 20 transmission zones + the PJM-RTO aggregate; DA and RT hourly LMP (Data Miner 2 public feed).
• MISO — 8 trading hubs; DA ex-post and RT LMP (public market reports).
• SPP — SPPNORTH_HUB / SPPSOUTH_HUB; DA LMP (Marketplace portal daily file, else the CompoundVision price-store export) and RTBM 5-minute LMP averaged to hours.
• WECC — CAISO trading hubs TH_NP15 / TH_ZP26 / TH_SP15 only (the rest of WECC has no organised market); DA from OASIS PRC_LMP, RT from hourly means of OASIS PRC_INTVL_LMP 5-minute prices.
• ERCOT — DAM settlement point prices at HB_NORTH / HB_SOUTH / HB_WEST / HB_HOUSTON via gridstatus.io (keyed) or the CompoundVision price-store export; no real-time source without a key.
• SERC and FRCC — no organised market, so price columns are null. SERC border sub-areas are compared with the neighbouring PJM DOM / PJM AEP / MISO MS.HUB day-ahead prices as a reference only; FRCC is validated on generation fuel shares against EIA-930 net generation by fuel.

Freshness: the record is re-scored every morning by the CONUS rolling cycle (06:00 UTC for NYISO, 09:30 UTC for the other eight regions); eastern real-time prices publish late and enter the record on a later re-score. The posted prices are public ISO data: they are shown beside the forecasts and used to score and calibrate them, and the feeds themselves are not redistributed — see /licenses.

Plant deduplication

Wind and solar plants are merged across sources (EIA-860M, MaStR, REPD, PowerPlantMatching + ENTSO-E, WRI Global Power Plant Database, GEM Global Integrated Power Tracker) using a source-priority + spatial-bucket + name-similarity pipeline. Priority order:

1. EIA-860M (US, monthly authoritative)
2. MaStR (Germany, Marktstammdatenregister)
3. UK REPD (GB, authoritative)
4. National registers: DK-ENS (Denmark), Vindbrukskollen (Sweden), NVE (Norway, wind + hydro), SEAI (Ireland), RIVM (Netherlands), ANEEL SIGA (Brazil), Geoscience Australia + AEMO (Australia)
5. PowerPlantMatching + ENTSO-E (EU)
6. PowerPlantMatching (global, less authoritative)
7. France, from the first nightly build after 2026-10-06: the ODRÉ register defines onshore wind and solar (PowerPlantMatching and GEM operating rows dropped, as for MaStR/REPD/DK-ENS/Vindbrukskollen); points are commune centroids except Hauts-de-France wind farms matched to the DREAL turbine register (park turbine centroid); offshore wind keeps PowerPlantMatching/GEM positions.
8. WRI Global Power Plant Database (v1.3, fallback)
9. GEM Global Integrated Power Tracker (ranked last per site, but the primary registry for regions the others don't cover)

At every tier a GEM-enriched record (source tag +gem) outranks its unenriched sibling, so GEM attributes survive even where another registry wins the dedup.

Within a 1 km × 1 km spatial bucket (and 3×3 neighbour scan), a candidate is treated as a duplicate of an already-claimed record if name similarity ≥ 0.85 (difflib SequenceMatcher) or token overlap ≥ 0.75, and capacity ratio is between 0.87× and 1.15×.

Data refresh cadence