EPWForge — Data Sources & Methods

EPWForge generates simulation-ready weather files for building energy modeling using physically consistent methods based on CMIP6 climate projections, ERA5 reanalysis, and ASHRAE-aligned workflows.

The result is distribution-aware, physically coherent weather data that improves confidence in peak load estimation, HVAC sizing, and resilience analysis.

What makes it different

System overview

Baseline Weather (TMY / ERA5)
            +
      CMIP6 Climate Deltas
            ↓
   Morphing + QA/QC + Ensembles
            ↓
EPW Files • Design Conditions • Extreme Events

EPWForge generates simulation-ready weather files for building energy modeling using physically consistent methods based on CMIP6 climate projections, ERA5 reanalysis, and ASHRAE-aligned workflows.

Key Differentiators

EPWForge departs from conventional weather file workflows in several important ways:

Physically consistent percentiles

Percentiles are assigned at the model level. All variables at a given percentile originate from the same CMIP6 model realization, preserving internally consistent relationships between temperature, humidity, solar radiation, and wind. Approaches that rank each variable independently can produce physically impossible weather states (for example, very high temperature paired with very low solar radiation drawn from different models).

Design conditions from morphed hourly data

Future ASHRAE design values are derived directly from the full 8760-hour morphed dataset, ensuring consistency between extreme values and coincident conditions. This is materially different from delta-only methods that simply offset historical design values.

Per-model climate ensembles

Each CMIP6 model produces a distinct EPW file, allowing simulation across the full ensemble to quantify uncertainty in loads, energy use, and thermal comfort.

In contrast to traditional delta-only approaches, EPWForge operates on full hourly distributions rather than adjusting summary statistics — preserving extremes, coincident conditions, and inter-variable relationships throughout the morphing process.

Weather File Sources

Typical Meteorological Year (TMY)

TMY data is sourced from Climate.OneBuilding.Org, derived from ASHRAE IWEC2, TMY3, and TMYx datasets. Using the de facto reference for the energy modeling community ensures compatibility and reproducibility of simulation results across projects and firms.

Infrastructure: EPWForge maintains a local mirror — GuzzStations — of 74,693 EPW files covering 17,137 stations, audited monthly against the upstream OneBuilding spreadsheet so requests never depend on third-party uptime. Catalog and library are kept in 1:1 sync; dead entries from OneBuilding's own spreadsheet drift are pruned from our station picker so users never click into a file that doesn't exist.

Actual Meteorological Year (AMY)

AMY data is generated from the ECMWF ERA5 reanalysis, which provides hourly, gap-free global coverage from the mid-20th century to present. ERA5 integrates millions of observations — surface stations, radiosondes, satellites, aircraft, and ocean buoys — into a physically consistent atmospheric state at every timestep, so all variables are internally consistent rather than spliced from disparate sources. This eliminates the missing-data and instrument-gap issues common to traditional station-based AMY files, and extends coverage to any location on Earth.

As with all gridded datasets, ERA5 represents regional atmospheric conditions and may not fully capture localized effects such as complex terrain, coastal microclimates, or dense urban environments. EPWForge addresses the urban case through its UHI adjustment system.

Multi-Source Refinement (in progress)

ERA5's coarse grid resolution (25 km) is a known weakness for solar radiation and near-surface wind, both of which benefit from satellite-derived or higher-resolution sources where they exist. EPWForge runs a multi-source router that prefers the best available source per variable per location, falling back to ERA5 when no superior source is available.

Active sources: ERA5 (global, all variables), ERA5-Land (global land surfaces, 9 km), NSRDB (US satellite-derived solar, 4 km). Sources in acquisition: CM SAF SARAH-3 (Europe / Africa / Middle East satellite-derived solar, 5 km — order placed), NREL WIND Toolkit and NEWA (high-resolution wind). Lineage of which source supplied each variable is recorded in the EPW header COMMENTS field.

Site-Specific TMYx Generation

EPWForge generates TMYx weather files for arbitrary locations using the Finkelstein–Schafer statistic in accordance with ISO 15927-4 — the same methodology used by OneBuilding for their published TMYx datasets — with multiple period variants available to match published TMYx products.

Validation against several hundred OneBuilding TMYx stations across all major climate zones shows strong agreement in annual temperature and solar radiation. Deviations are primarily observed in complex terrain and high-latitude regions, consistent with the known limitations of gridded reanalysis datasets.

Future Climate Projections

Future weather files are generated using a Belcher et al. (2005) mean-shift methodology paired with an ERA5-derived climatological diurnal anomaly profile for the stretch term — a hybrid consistent with how modern production tools handle morphing (UKCP18 local projections, NOAA Atlas 14 for sub-daily disaggregation, WeatherShift commercial). Pure Belcher-classic stretches against the specific baseline EPW's hourly deviation from its monthly mean, which can over-amplify outlier days and over-distort low-DTR tropical baselines. EPWForge replaces that stretch target with the 45-year ERA5 climatological diurnal anomaly profile for the same grid cell, giving a smoother and more physically defensible diurnal redistribution. Climate change deltas remain CMIP6-derived and are applied across all variables to preserve inter-variable relationships.

Supported scenarios:

• SSP1-2.6 — low emissions
• SSP2-4.5 — moderate emissions
• SSP3-7.0 — high emissions (recommended high-end for design; SSP5-8.5 is available as an opt-in extreme stress test)

Each scenario is available across ten time horizons (2030, 2035, 2040, 2045, 2050, 2060, 2070, 2080, 2090, 2100) and seven percentile bands (P5–P95). The near-term 5-year cadence through 2050 supports authorities-having-jurisdiction (AHJ) starting to require 5-yr future-projection submittals for net-zero compliance. The 2100 horizon uses a 10-year averaging window (2091–2100) because the underlying CMIP6 SSP runs end at 2100.

Variables morphed

The CMIP6 climate-deltas dataset extracts 11 variables across 34 models per scenario, per horizon, per cell. Ten of these flow into the EPW through morphEPW with the convention noted; the eleventh (DTR derived from tasmax/tasmin) drives the diurnal-stretch term:

• Dry-bulb temperature (tas) — additive monthly shift + diurnal stretch (Belcher + ERA5 anomaly template).
• Relative humidity (hurs) — additive %-point shift, clamped 1–100%. Dew point recomputed from morphed T and RH via Magnus.
• Shortwave radiation (rsds) — multiplicative ratio applied uniformly to GHI / DNI / DHI. Luminance fields (illuminance, zenith luminance) scale by the same factor to preserve photometric consistency.
• Surface wind (sfcWind) — multiplicative ratio clamped to ±5% per the "global stilling" note above.
• Longwave radiation (rlds, new in v5) — additive W/m² shift applied to the horizontal infrared field. Replaces what would otherwise be a Brunt/Berdahl-Martin sky-emissivity reconstruction; the direct CMIP6 delta is cleaner physics.
• Precipitation (pr, v5) — multiplicative ratio applied to liquid precipitation depth. Duration in hours is left unchanged (multiplying duration is wrong dimensional analysis).
• Sea-level pressure (psl, v5) — additive Pa shift applied to atmospheric pressure, clamped to physically plausible bounds (50–110 kPa).
• Total cloud cover (clt, v5) — additive %-point shift applied to both total and opaque sky cover (same shift, preserving opaque ⊂ total), clamped 0–10 tenths.
• Snowfall (prsn, v5) — multiplicative ratio applied to snow depth as a proxy. See Snow physics limitation in the Limitations section for the caveat on this approach.
• Precipitable water — derived from morphed T and RH via the Clausius-Clapeyron / Magnus saturation curve. CMIP6 doesn't publish a column water vapor delta directly; this is the standard parameterization.

Per-variable ensemble sizes vary because not every CMIP6 model publishes every variable in Pangeo. The API surfaces this transparently via model_counts: for a typical mid-latitude SSP2-4.5 query you'll see tas: 33,hurs: 29, prsn: 19 — narrower ensembles for the harder-to-publish fields. The wider variables drive the mean climate shift; the narrower ones still represent a credible multi-model signal.

SSP5-8.5 as an extreme stress-test scenario

SSP5-8.5 — peak fossil-fuel use, no policy response, ~4.4 °C of warming by 2100 — is not recommended for standard design work. The IPCC AR6 and the upcoming CMIP7 generation place this trajectory at low likelihood; observed emissions and policy momentum have moved well below it, so treating it as a routine "worst case" would inflate the design envelope. We nonetheless offer it as an explicit, opt-in extreme stress-test scenario — a "what breaks first" upper bound for resilience and worst-case analysis, clearly labelled as such throughout the UI. For defensible design conditions, use SSP3-7.0 paired with P90 or P95 percentile bands. Historical comparison data that referenced SSP5-8.5 (e.g. the EnergyModelIQ V2 building-energy simulation results) remains available as a frozen reference.

Two scales of resolution: mean climate vs. event tails

CMIP6 morphing in EPWForge operates at two distinct temporal resolutions, and it's worth distinguishing them when reading the methods that follow:

• Mean climate shift uses a monthly mean ΔT field from the CMIP6 ensemble. The Belcher morph applies that monthly delta plus a diurnal-range stretch term to the hourly baseline, so the output is hourly but the underlying climate signal is monthly. This handles seasonal warming cleanly; it does not, on its own, reshape the tails of the distribution.
• Event-tail amplification uses daily-extreme metrics from the IPCC AR6 Atlas — annual peak daily max temperature (TXx) for heat, annual minimum daily min (TNn) for cold, annual peak wet bulb (TWx) for hot-humid. These factors are applied hour-by-hour to the anomaly above climatology when extreme events are stitched into the file. See AR6 Phase 4 SSP scaling below for the math.

The split is deliberate. Daily-extreme metrics empirically embed both the thermodynamic background warming and the synoptic-regime contribution (per the Coen 2026 framework) — applying them to the tails avoids over-correcting the mean climate, which is already handled by the monthly delta. Sub-monthly delta fields for the mean climate path are on the research roadmap; see Diurnal-cycle climate deltas in Ongoing R&D.

Wind: conservative clamping

Surface wind speed is the most uncertain variable in the CMIP6 ensemble — inter-model spread on future mean wind change typically exceeds the ensemble-mean signal, and observed mean wind has been declining over much of North America, Europe, and East Asia (the “global stilling” trend documented by Zeng 2019). To avoid amplifying this noisy signal, EPWForge clamps the CMIP6 wind multiplicative factor to ±5% per month regardless of the raw ensemble value. The clamp is symmetric so it doesn't bias the morph in either direction, and acknowledges that wind under warming is genuinely an open question rather than something CMIP6 ensembles resolve cleanly.

Design Conditions, Ground Temperatures & UHI

Design conditions

Design day values (heating 99.6%/99%, cooling 0.4%/1%/2% dry bulb and wet bulb, evaporation, dehumidification) are computed directly from the 8760-hour dataset using ASHRAE Fundamentals methodology. Mean coincident values are derived from percentile exceedance subsets, ensuring physically meaningful pairings between dry-bulb temperature and coincident variables.

Handbook preservation for unmodified TMY downloads: When you download a baseline TMYx file with no UHI, no events, no smoke, and no future-climate morphing applied, the published ASHRAE Handbook of Fundamentals (Chapter 14 Climatic Design Information) values are passed through verbatim into both the EPW DESIGN CONDITIONS header and the .ddy file. This matches what auditors expect for ASHRAE 90.1 code compliance — the Handbook values come from a 30-year reanalysis at each station, which is more authoritative than computing percentiles from a single 8760-hour TMY. For any modified download (UHI on, events, morphed scenarios, AMY), values are recomputed from the modified hourly data — Handbook values no longer apply once the climate has been modified. The .stat file's Comment 1 line states which path was taken.

Custom-pin and off-station design conditions: For locations not in the OneBuilding catalog (custom-pin lat/lon, or stations between published OneBuilding stations), design conditions are computed from a corrected long-term hourly record. Two corrections are applied before the percentile math:

• Elevation lapse — surface temperatures are adjusted along a standard atmospheric lapse rate between the model grid-cell average elevation (HSURF) and the actual pin elevation (ETOPO1). Important in mountains and along sharp terrain.
• Per-station diurnal correction — ERA5's grid-cell averaging dampens the daily temperature swing relative to single-point observations. We apply a small per-station scaling around each day's mean to recover the observed swing, calibrated against 10,000+ OneBuilding stations.

The percentiles themselves (heating 99.6%/99%, cooling 0.4%/1%/2%, etc.) are always computed from the full multi-year corrected hourly record, not from the 8760-hour TMY composite. This matters because typical-year synthesis optimizes the bulk distribution, not tail extremes.

Cooling 1% Dry Bulb
  Historical TMY:        32.1 °C
  EPWForge (morphed):    35.4 °C   ← derived from morphed hourly distribution
  Delta-only method:     33.6 °C   ← historical + ΔT only

Ground temperatures

Undisturbed ground temperatures at any depth are computed using a published two-harmonic analytical model that extends the classical Kusuda–Achenbach approach with improved seasonal asymmetry. Inputs are derived from TMY surface air temperature data.

Urban Heat Island (UHI)

Most weather stations are located at airports or rural outskirts, which can be several °C cooler than dense urban cores at night. EPWForge offers UHI presets aligned with the Stewart & Oke (2012) Local Climate Zone framework. Diurnal profiles are applied to dry bulb temperature, dew point, and surface wind, with magnitudes consistent with observed ranges in the urban climatology literature. For known urban locations (major metros) we auto-apply a default UHI preset; the chosen preset and all applied modifications are disclosed in the downloaded .stat and .pvsyst headers so the data lineage stays auditable.

Output file formats

Every download is a zip bundle containing five companion files that mirror what climate.onebuilding.org ships alongside their static TMYx files — with the critical difference that ours are regenerated from the morphed weather data, so a future-climate or UHI'd scenario gets matching design conditions, statistics, and PV-tool inputs that reflect the modified climate. Nobody else does this for any of these formats.

• .epw — EnergyPlus Weather format, 8760 hourly rows with all 35 fields per row. The DESIGN CONDITIONS header line preserves Handbook values for unmodified TMY (see above) or is recomputed from morphed data otherwise.
• .ddy — EnergyPlus Design Day file withSizingPeriod:DesignDay blocks for ASHRAE 99.6% / 99% heating and 0.4% / 1% / 2% cooling conditions. Consumed by EnergyPlus, OpenStudio, IES VE, eQUEST.
• .stat — EnergyPlus statistics summary file. Location header, ASHRAE design block, monthly statistics tables (dry bulb, relative humidity, wind), monthly HDD/CDD totals at multiple bases, ASHRAE 169-2021 climate zone (derived from CDD10/HDD18 thresholds), Kusuda-model undisturbed ground temperatures at three depths, 24×12 hourly average tables for dry bulb and wind speed, and a 16-bin annual + per-month wind frequency table.
• .pvsyst — PVsyst-compatible hourly CSV for photovoltaic system design simulation. 8760 rows with GHI / DHI / DNI / ambient temperature / wind speed / wind direction at mid-interval (30-minute) timing. Imports directly via PVsyst's Databases → Meteorological tables → Import file. For future-climate scenarios this is the only credible way to assess PV array performance under projected conditions — degraded modules, hotter cell temps, future-shifted irradiance distributions.
• .csv — flat CSV of the hourly EPW data for direct spreadsheet inspection or downstream pipelines that don't consume the EPW format directly. Imperial or metric units, controlled by the unit toggle.

Climate Ensembles & Extreme Events

Climate ensembles

A single morphed weather file is deterministic and cannot capture inter-model climate uncertainty. EPWForge generates per-model climate ensembles using real CMIP6 model deltas — multiple individual model EPWs per scenario, each representing a distinct projection. Engineers can run simulations across all members to understand the range of possible outcomes (peak loads, annual EUI, overheating hours) rather than relying on a single point estimate. This approach aligns with emerging best practice from NREL and LBNL for climate-resilient design.

Ensemble generation may take several seconds depending on location and scenario.

Extreme event injection

Typical meteorological years deliberately exclude extreme events, but resilience analysis requires understanding building performance under heat waves, cold snaps, and humidity events. EPWForge maintains a global extreme events database derived from ERA5 reanalysis, with statistically fitted return-period intensities at every grid cell. Events can be stitched into baseline or future weather files for direct analysis of questions like “if grid power fails during a 25-year heat wave in 2050, does the building remain habitable?”

Intensity scale

Event intensity is exposed as a 1–10 slider per event type, mapped piecewise to an anomaly multiplier on the historical event:

• Slider 5 = 0.6× — typical extreme event (default)
• Slider 7 ≈ 0.96× — severe historical (~50-yr return period)
• Slider 10 = 1.5× — stress test, exceeds observed extremes

Sliders 8–10 are gated behind an explicit stress_test=true flag on the API and the MCP server. These produce events more extreme than anything in the observational record and are appropriate for resilience studies (“what fails first?”) but not for HVAC sizing or code compliance — downstream simulators may also exceed their psychrometric bounds at extreme intensities.

Time-warp placement (day-level)

When a user requests an event of duration N days, EPWForge maps the historical event's available days across the requested duration via a day-level time-warp: stitched day k draws from source event day floor(k × source_days / stitched_days), preserving the natural rise-peak-fall shape and each day's diurnal cycle. The peak day still appears for ~1–2 stitched days; surrounding days carry the natural shoulder pattern.

This replaces an older “peak-day-cycling” approach that repeated the single hottest / coldest day for the full event duration. The cycling produced sustained extremes that were physically implausible and inflated stitched temperatures by ~10°F relative to a natural event shape — particularly noticeable when stacked with SSP morphing and high percentile bands.

Compound events

Two compound pairings are supported with co-driven physics:

• Heat wave + Hot-humid — heat sets the temperature anomaly, hot-humid contributes its humidity profile. Secondary anomaly layered at 0.5 × the secondary's own intensity factor (so the hot-humid slider scales humidity contribution independently).
• Cold snap + Cold-windy — cold sets the temperature anomaly, cold-windy contributes the wind profile at the same 0.5 × secondary-intensity blend.

The 0.5 physical blend reflects that compound extremes are not fully additive — the atmospheric circulations driving them are partially correlated rather than independent.

AR6 Phase 4 SSP scaling

When an SSP scenario is active, event intensity sliders auto-populate using factors from the IPCC AR6 Atlas (Phase 4): TXx (annual peak temperature) for heat events, TNn (annual minimum) for cold. The factor reflects how much extreme events scale faster (or slower) than the mean climate signal at each location and horizon.

Cold-family floor. AR6 evidence suggests cold extremes warm faster than mean (dampening future severity), but recent observational record (Texas 2021, polar-vortex disruption events documented in Cohen 2026) doesn't yet support this cleanly. EPWForge defaults cold-family events to slider 5 under SSP (no future dampening, no false amplification) — users can manually override either way.

Improbability indicator. Each generated file carries an improbability score (1–10) computed from the joint rarity of active dials (SSP × year × percentile × UHI × event intensity × smoke). The UI surfaces this as a header warning when the combined settings drift into stress-test (≥4) or exploratory (≥7) territory.

Where the event is placed in the year

Events are stitched at the climatologically peak window of the baseline file, not at the historical event's calendar date. Heat-family events (heat waves, hot-humid events) are inserted at the hottest 14-day window of the baseline file; cold-family events (cold snaps, cold-windy) at the coldest. Smoke overlays auto-align to whichever event drives the season.

We do this for three reasons:

• Design-condition relevance. ASHRAE 0.4% / 1% / 2% percentile design conditions are by definition the worst hours of the year — they land in peak season. Layering an extreme event on top of those hours is the natural extension; placing the event off-season produces a smaller peak temperature than the design-day already sits at, which has no equipment-sizing implication.
• AR6 scaling alignment. The Phase 4 SSP intensity factor is derived from AR6 TXx (annual peak temperature) and TNn (annual minimum). Those metrics are by construction in the peak month. Scaling the event anomaly by a TXx factor while placing the event in a shoulder month would mix two different season references; both the factor and the placement should refer to the same season for the result to be physically coherent.
• Statistical honesty. The historical event database returns the single worst event over ~75 years of ERA5 at each cell. Whether that specific event happened in May or July is partly atmospheric-circulation luck — one stochastic realization. Treating that one calendar slot as definitive isn't methodologically meaningful; the physically robust part of the historical record is the shape of the anomaly (ramp / plateau / decay), which we preserve.

What we keep from the historical event: the anomaly profile shape and the relative magnitudes across temperature, humidity, wind, and solar. What we replace: the calendar slot. The historical reference dates remain visible in the UI tooltip as provenance.

Wildfire Smoke Overlay

Wildfire smoke increasingly drives building HVAC sizing in fire-prone regions but is absent from conventional EPW files. EPWForge models smoke impact via a CAMS-AOD climatology (aerosol optical depth, 86,896 global grid cells, 2003–2025 record) and applies it physically to the relevant EPW fields when active.

Physics

• Solar attenuation — Beer-Lambert law applied to GHI / DNI / DHI using the AOD profile of the active event.
• Asymmetric surface temperature impact — −3 °F daytime cooling per AOD unit (shaded mornings, reduced solar gain) vs. +1 °F nighttime warming per AOD unit (heat trapping under aerosol layer).
• Humidity bump — +3 % relative humidity per AOD unit, with downstream dewpoint recomputation.
• AOD field — written into EPW field 29 (aerosol_optical_depth) for downstream tooling that reads it.

Coefficients are mid-points of published literature ranges; site-specific tuning (by albedo, vegetation, aerosol chemistry) is not currently applied. Smoke events are auto-aligned in time to whichever event drives the season (smoke onset overlapping the peak of an active heat event, for instance).

SSP scaling — v1 biome-weighted TXx proxy

Smoke severity auto-amplifies under SSP scenarios using a v1 proxy methodology, pending acquisition of wildfire-specific projection data (Touma 2021, Burke 2023, Wang 2025). For each of 46 AR6 reference regions, the peak-AOD intensity factor is derived as:

smoke_anomaly  = heatwave_anomaly × biome_factor
                 (from AR6 TXx)   (region-specific, ∈ [0.6, 1.2])

intensity_factor ≈ biome_factor × 1.7
                   ⇒  ∈ [1.0, 2.0]  (floor: no decrease; ceiling: 2× peak AOD)

Biome factor tiers (centered on a baseline temperate response):

• 0.60 — rainforests, deserts, polar (no real amplification; fuel- or humidity-limited)
• 0.70 – 0.85 — tropical mixed, humid temperate, savanna, steppe (moderate amplification)
• 0.90 – 1.00 — boreal taiga, Mediterranean Basin, well-documented fire-prone temperate
• 1.10 – 1.20 — Western N. America fire belt, S. Australia, Mediterranean Chile, E. Australia (highest amplification)

The floor=1.0 guarantees smoke severity never decreases below historical climatology even in fuel-limited regions. The ceiling=2.0 caps peak AOD at 2× the historical worst-case event for the location (e.g., Bay Area 2020 AOD ~7 → max ~14 under SSP3-7.0 2090 P95 in Western N. America), well within observed physical bounds.

Sources informing tier assignment: Abatzoglou & Williams 2016 (PNAS) on fire-weather TXx supra-linearity; Bowman et al. 2020 (Nat Rev Earth Env) global fire-climate review; IPCC AR6 WG1 Ch. 12 regional fire-prone projections.

This is a v1 proxy methodology — explicit because we'd rather show our work than silently apply a flat global multiplier. Plan: replace per-region biome factors with wildfire-projection-specific intensity factors from published CMIP6 smoke-day / burned-area projections once we've sourced and audited those datasets.

Validation & Quality Control

EPWForge undergoes ongoing automated validation across the data pipeline, the Belcher morphing implementation, and the EnergyPlus simulation outputs of pre-computed reference cases. Recent audits confirm:

• CMIP6 deltas align with IPCC AR6 assessed warming within expected bounds.
• Physical consistency — Arctic amplification, land-vs-ocean contrast, and SSP ordering all match observed climate patterns.
• Belcher morphing produces exact temperature shifts and physically valid humidity, solar, and wind output across every model in the ensemble.
• EPW outputs satisfy physical bounds and have valid 8760-hour structure across all tested locations and scenarios.

Validation includes cross-checks against published TMYx datasets and internal consistency checks across all generated variables.

Bias correction validation

The lapse and per-station diurnal corrections used for custom-pin design conditions are validated against the published OneBuilding 2025 ASHRAE Handbook values across 16,906 reference stations:

• Cooling 0.4% dry bulb — median bias against published values reduced by ~50% (from −0.80 °C to −0.34 °C).
• Heating 99.6% dry bulb — mean bias reduced by ~45% (from +0.47 °C to +0.20 °C).
• Residual standard deviation across all 16,906 stations reduced 6–11% on standard heating/cooling DC values.
• 60–68% of stations show improvement, 31–37% show small degradation, 2–3% unchanged.

Generated EPW and DDY files are runtime-tested in EnergyPlus 25.2 to confirm they load and simulate without errors.

Detailed audit reports are maintained internally and are available to enterprise customers on request.

Limitations & Assumptions

EPWForge uses methods that are standard practice in the field, but all modelling tools carry inherent limitations users should understand when applying results to design decisions.

• Spatial resolution. CMIP6 deltas are computed on grids consistent with established practice. Mountain regions and immediate coastlines may show larger interpolation uncertainty.
• Within-cell uniformity (mean climate). A single monthly delta value is applied uniformly across all hours in a month for the mean climate shift — this is the standard Belcher approach and does not capture intra-month variability in the rate of climate change. Note that the extreme-event tails are not bound by this limitation: AR6 Phase 4 amplification uses daily-extreme metrics (TXx / TNn / TWx) applied hour-by-hour to the event anomaly above climatology.
• Baseline file quality. Morphed future files inherit any biases or gaps present in the underlying baseline, whether TMY or AMY.
• UHI presets are approximations. Urban microclimate is site-specific and varies by morphology, albedo, and vegetation cover. Presets are LCZ-based estimates and should not substitute for site-specific urban climate measurements where precision is critical.
• Snow physics limitation. The snow-depth field is morphed as a multiplicative ratio derived from the CMIP6 snowfall-flux delta (prsn) — a useful proxy but not a full accumulation × melt model. A proper treatment would track snow water equivalent through a degree-day melt model and couple albedo to snow depth. In practice this rarely matters for building energy simulation: most BEMs read ground temperature (which we compute correctly via Kusuda) and surface-specific albedos defined in the IDF rather than the EPW's air-side snow fields. For applications where the snow signal materially matters — slab-on-grade with snow coverage effects, ground-source heat pumps with frozen-soil regimes — treat the morphed snow depth as indicative rather than authoritative.
• Surface albedo passthrough. CMIP6 doesn't publish a directly comparable surface albedo delta in the standard output set. The EPW albedo field is passed through unchanged from the baseline. Most BEMs ignore this EPW field anyway in favor of surface-by-surface albedo specified in the IDF, so the practical impact is small — but worth noting for tools that do consume it.

These limitations are inherent to global-to-local downscaling. For critical design decisions, results should be validated against primary data sources and applicable local code requirements. EPWForge is designed to make these limitations explicit while providing the most physically consistent weather inputs practical for simulation workflows.

Ongoing R&D

EPWForge is under active development. We continuously evaluate new datasets, climate-science advances, and modeling techniques as they emerge — refining the underlying methodology when peer-reviewed work or new operational data warrants. Specific roadmap items shift as the science evolves and customer use cases sharpen what's worth investing in.

If your project depends on a specific capability — a particular climate scenario, geographic coverage, validation against a dataset you trust — reach out and we'll prioritize accordingly.

Citations

Belcher, S.E., Hacker, J.N. & Powell, D.S. (2005). “Constructing design weather data for future climates.” Building Services Engineering Research and Technology, 26(1), 49–61.

Hersbach, H. et al. (2020). “The ERA5 global reanalysis.” Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049.

Stewart, I.D. & Oke, T.R. (2012). “Local Climate Zones for urban ecosystem studies.” Bulletin of the American Meteorological Society, 93(12), 1879–1900.

Eyring, V. et al. (2016). “Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization.” Geoscientific Model Development, 9(5), 1937–1958.

IPCC. (2021). “Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change” (AR6 WG1) — including the AR6 Atlas regional CMIP6 indices used for Phase 4 event amplification factors.

Van Vuuren, D.P. et al. (2026). “The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP7.” Geoscientific Model Development, 19, 2627–2656.

Cohen, J. et al. (2026). Polar-vortex disruption and cold-extreme trends under continued Arctic warming. Used as the basis for our conservative cold-family floor in AR6 Phase 4 auto-fill (no future dampening applied).

EPWForge composes each hourly EPW field from the dataset best suited to that variable. The exact per-variable provenance is recorded in every output file's COMMENTS 1 line and its companion .stat file's Comment 1.

8,784 hours of overlap between our multi-source synthesis and NOAA SURFRAD ground-truth measurements. Lineage active for this run: ERA5 (pr/pressure/rlds/sfcWind) + ERA5-Land (hurs/tas) + NSRDB (dhi/dni/ghi)

Validated against 11 SURFRAD sites across the continental US (BON / TBL / DRA / FPK / GWN / PSU / SXF). Pick a station above to see its 2020 error stats.