Missing-weather recovery and audit trail

Policy and defaults

EPWConfig.weather_policy="automatic" is the default. The compiler aims to return a complete requested-year EPW while retaining usable observations and making reconstruction visible. Supported requests are completed calendar years from 1999 onward. Required meteorological variables follow this order:

1. Accepted-quality primary observations.

2. Interpolation for gaps up to three hours, circularly for wind direction.

3. Assessed same-year nearby observations for longer gaps or unfillable short wind gaps.

4. Bounded daily-profile reconstruction for remaining gaps up to 48 hours.

5. Same-year ERA5 at the requested coordinates, supplied through Open-Meteo.

Solar recovery remains donor-free: usable primary radiation, bounded solar filling or GHI decomposition, then same-year ERA5 for unresolved GHI.

weather_policy="strict" retains source and gap failures instead of enabling reanalysis. In strict mode, metobs_gap_fallback=True permits assessed donors; its default remains False. Automatic policy already enables donors and does not require this legacy flag. Donor-enabled strict mode uses the same meteorological precedence through daily profiles, with no ERA5 stage. Without that flag, strict mode uses primary observations and bounded temporal filling. Both modes preserve valid primary observations. Invalid or nonfinite inputs become missing cells in automatic recovery.

Temporal filling remains bounded. Scalar gaps of at most three hours use endpoint interpolation; a single endpoint permits a constant boundary estimate. After assessed donors are tried, remaining gaps of up to 48 hours use endpoint-adjusted daily profiles, searching independently on each side within seven days. Usable sides receive equal weight; one side is sufficient. Profiles can include earlier reconstructions. Wind direction uses sine/cosine components. Solar gaps use day-aware profiles without donor transfer. Longer gaps are filled from another source, never by extending interpolation.

The frozen sensitivity campaign covered 17 weather-years, six regions and six building profiles. For its 3–12-hour gaps, the largest group-level 95th percentile annual-load error was 0.893%. Four week-long seasonal solar gaps reached 12.728% at the 95th percentile for annual sensible cooling. In its 48-hour event-centred meteorology cases, the winning peak moved more than 24 hours in 13/41 heating and 14/44 cooling cases. Annual energy and event timing therefore need separate checks.

These results compare reconstructions with the campaign’s own reference weather; 4,146 automatic comparisons are not 4,146 independent climate samples. Both reference and candidate used the same solar finalizer, so a shared error can cancel in a paired comparison. The original campaign is frozen at source commit 5930c07, before the low-sun and historical-DNI corrections and the pvlib implementation. It supports the reported conditional comparisons, not site-weather accuracy or national error bounds. Regenerated references and targeted paired runs are needed when the solar method changes. The archived completeness evidence below is a separate generation check, not additional validation of sensitivity results.

Automatic station selection prefers full coverage but can use a partial station. Source failures are reported while other requests continue. If station discovery fails, supplied coordinates allow reanalysis recovery. Supplying only a station ID still requires enough metadata to resolve a target location. Invalid configuration, unknown coordinates, incomplete future years, unavailable weather from every source, unrecoverable physical inconsistencies and filesystem errors remain explicit failures. “Complete” does not mean inventing weather when no usable source exists.

Nearby-observation assessment

Donors are searched within gap_fallback_max_distance_km (default 75 km). At most gap_fallback_max_stations candidates (default three in total) are attempted, ordered by distance and station ID. Metadata is resolved for the parameter being borrowed; a donor need not have temperature or all five required variables. Year-specific positions are checked against the radius again. Only same-hour SMHI observations with accepted quality grades G or Y are used. Invalid donor values and transport failures are recorded and skipped.

For each gap and variable, assess original paired primary/donor observations within 30 days on either side. Reconstructed primary values are excluded. Validate on three complete, distinct UTC overlap days: the earliest, middle and latest. Each fit excludes its test day and the adjacent days, and requires at least 168 paired training hours spanning seven dates. Insufficient overlap rejects a donor rather than treating its proximity as evidence of accuracy.

Scalar candidates are raw transfer and a median primary-minus-donor offset. Select an offset only if held-out mean absolute error (MAE) improves by at least 10 percent, then refit on all eligible overlap. Wind direction uses shortest-arc errors and raw transfer only. The selected candidate must satisfy these default MAE ceilings in source units:

Variable	Maximum held-out MAE
Temperature	3 degrees Celsius
Relative humidity	15 percentage points
Wind speed	3 m/s
Wind direction	45 degrees
Sea-level QFF	5 hPa

These are engineering defaults, not scientifically established accuracy thresholds. Corrected values outside physical bounds are ignored. Remaining gaps proceed to another eligible donor, then bounded daily profiles where possible, and ERA5 in automatic mode. Entirely missing variables cannot be calibrated against that year’s primary record and therefore normally require reanalysis. No ML, wind rotation, multi-variable prediction or future-climate morphing is applied.

Primary and donor subhourly winds share circular hourly averaging. A cancelling direction vector remains missing. Donor solar data is never copied; solar recovery remains tied to the requested coordinates.

ERA5 recovery and physical consistency

Automatic policy uses the Open-Meteo Historical Weather API with ERA5 selected explicitly, UTC timestamps and the buffered requested-year window. The adapter checks units, array lengths, calendar coverage and usable values. Returned grid coordinates, elevation, model and response fingerprints accompany the report.

Temperature, humidity, wind and surface pressure recover unresolved required hours. Missing cloud cover is filled from ERA5 only where ERA5 actually replaces required meteorology or GHI. Usable existing cloud values and all other hours remain unchanged; fetching ERA5 only to resolve metadata does not add clouds. reanalysis_filled_hours includes cloud_cover when these optional replacements are used. SMHI parameter 9 is sea-level QFF and requires the inverse SMHI reduction; ERA5 surface pressure is used directly after unit conversion, without a second reduction. See Data provenance and acknowledgment for elevation assumptions.

Reanalysis radiation already represents the preceding-hour mean and is not averaged again like instantaneous STRÅNG samples. Solar replacements use complete component groups; when only usable GHI is available, the continuous Erbs-Driesse fallback derives DNI/DHI using the interval conventions in Pipeline and scientific methods. Missing direct/diffuse solar components alone do not request ERA5 when existing GHI can be decomposed. Dew point and longwave radiation are derived from the final meteorological inputs.

Hourly solar inputs outside 0–2,000 W/m² are treated as missing before recovery. This is a broad engineering plausibility guard, not a validated local solar threshold or a limit on plausible heatwave temperatures.

Final checks require complete ordered hours, finite required values, physical ranges, dew point consistent with dry bulb, nonnegative solar fields, solar closure and the existing extraterrestrial DNI cap. Positive solar data during fully dark intervals is treated as inconsistent. Darkness checks use interval geometry so sunrise and sunset hours are not mistaken for complete darkness. Reconstructed humidity, wind and cloud estimates obey representable limits; valid observations are not clipped merely because they are unusual.

When ERA5 is fetched, reanalysis_metadata["overlap"] compares it against available original primary observations. It records paired-hour counts, mean difference and mean absolute difference; wind direction uses shortest angular arcs. Pressure comparisons use Pa after original QFF is converted to target surface pressure with original valid temperature. Substantial disagreement produces a warning rather than rejecting the fallback or suppressing an event. These comparisons are descriptive and do not provide independent validation.

Abrupt changes at source boundaries also generate warnings. Boundary warning thresholds are 10 degrees Celsius for temperature, 40 percentage points for humidity, 15 m/s for wind speed and 1,000 Pa for surface pressure. They trigger review, not automatic rejection of extreme weather. Usable primary extremes are preserved, and another year is never substituted. Reconstruction may change a missing event’s peak intensity, timing or duration; these checks cannot guarantee heatwave preservation or building-site accuracy. Review source extent and uncertainty when extremes matter.

Usage

Python:

from smhi2epw import EPWConfig, compile_epw

result = compile_epw(EPWConfig(
    year=2018,
    output_path="weather-2018.epw",  # parent directory must exist
    latitude=59.3,
    longitude=18.0,
    weather_policy="automatic",  # default; also enables assessed donors
    gap_fallback_max_distance_km=75.0,
    gap_fallback_max_stations=3,
))
print(result.report.warnings)
print(result.report.required_reconstructed_fraction)
print(result.report.gap_fallback_sources)
# Default audit sidecar: weather-2018.epw.json

CLI:

smhi2epw 2018 weather-2018.epw --lat 59.3 --lon 18.0 \
  --gap-fallback-max-distance-km 75 --gap-fallback-max-stations 3

For strict compilation with optional assessed donors:

smhi2epw 2018 weather-2018.epw --lat 59.3 --lon 18.0 \
  --weather-policy strict --metobs-gap-fallback --provenance weather-2018.json

Settings are validated before requests. Explicit provenance_path (CLI --provenance) overrides the automatic OUTPUT.epw.json destination. Strict mode creates no sidecar unless one is requested. The output and sidecar paths must differ and their parent directories must exist.

Diagnostics and reproducibility

Review ProcessingReport.warnings and the sidecar alongside the EPW. They record source failures, rejected donors, insufficient overlap, corrected or reanalysis-filled hours, and source-boundary concerns. weather_classification is observation_based, mixed_reconstructed or reanalysis_only. Reconstruction of any required meteorological variable or solar GHI makes a result mixed; routine physical darkness zeros do not. Reanalysis-only requires all required meteorological hours from ERA5 and solar GHI from ERA5 or physical darkness. solar_source separately identifies the radiation path. source_fractions maps required meteorological variables and GHI to shares of exported hours using primary, temporal, donor, reanalysis and physical source tags. It does not currently describe optional cloud cover; its recovered-hour count is included in reanalysis_filled_hours. reanalysis_filled_hours and reanalysis_metadata describe the extent, provider details and original-observation comparisons of ERA5 use. EPW comments retain fallback attribution and a quality summary when the EPW is copied without its sidecar.

Recovery diagnostics include:

• primary_missing_hours and invalid_observation_hours by required variable before recovery;

• cross_station_filled_hours and gap_fallback_sources with donor IDs, locations, distances, selected methods, offsets and filled UTC intervals;

• gap_fallback_attempts including rejected candidates, assessment status, overlap period, held-out errors and metadata/parameter errors;

• linear_filled_hours, diurnal_filled_hours, max_gap_hours and bounded_filled_hours for temporal reconstruction;

• required_reconstructed_fraction: missing/invalid primary cells divided by all cells in the five required meteorological variables on the ingestion grid, including calendar boundary hours.

The existing interpolated_fraction reports temporal filling across processed columns. It is not the fraction of all reconstruction; donor and reanalysis estimates must also be reviewed. Donor held-out errors measure agreement with primary observations during overlap, not independent site validation or an accuracy guarantee for the outage.

The atomic JSON sidecar records configuration, result/report, EPW SHA-256, package version, source-code fingerprints and UTF-8 response hashes. Live and cached payloads follow the same receipt contract. Receipts are scoped per compilation even when a client is reused; unrelated earlier requests are excluded. Custom clients without scoped response receipts are marked incomplete.

Receipt writing occurs after successful EPW export. Automatic policy returns the valid EPW with a provenance_write_failed warning when a sidecar filesystem failure prevents saving the receipt; strict policy retains a reported error. Neither case establishes a saved audit trail. EPW replacement remains atomic and follows successful validation of headers, 35-field rows, physical values and annual cardinality. Preserve old artifacts and regenerate when changing reconstruction policy, pressure conversion or source payloads.

Service access and limitations

The packaged adapter uses Open-Meteo’s public noncommercial endpoint. Check the current service access and limits before batch or commercial use; the package does not configure a paid API key or a customer endpoint. Responses are cached to avoid unnecessary repeat requests. Attribute Open-Meteo and Copernicus Climate Change Service ERA5 in studies using this fallback. Model weather is not an independent local observation and can underrepresent microclimates or extreme intensity.

Historical replay evidence

The earlier opt-in, raw-donor policy (commit 4861f9e) was exercised on 5 October 2026 when the EPSM national-cohort workflow replayed its 53 failed municipality/year weather jobs in a separate artifact directory, retaining cached primary observations and fetching needed donor archives. Of these:

• 52 produced EPWs passing export and EPSM calendar validation;

• nine recovered without donors, and 43 used nearby recorded observations;

• 19 reconstructed more than five percent of required meteorological cells;

• all 32 previously successful EPWs retained their original SHA-256 values;

• the remaining job, municipality 2101/year 2018, retained an 86-hour pressure gap on 15–19 October UTC after its only eligible donor within 75 km was used.

Five percent is a reporting aid, not a validated acceptance threshold. Some recovered jobs substitute entire variables, and one substitutes all primary meteorological inputs. Donor wind differences can be substantial. A complete, range-valid EPW establishes neither local weather accuracy nor archetype qualification. Review replacement extent, distance, overlap and site exposure; consider weather sensitivity before calibrating heavily reconstructed cases. No annual building simulations or calibration were rerun in this replay.

The workflow artifact directory output/weather-robustness-2026-10-05-v1 contains status, source manifest/snapshot, per-job receipts, failure diagnostics, replay script and an HTML report. Large generated EPWs and raw caches are local artifacts, not package source. These artifacts describe the earlier recovery policy and do not validate the automatic ERA5 fallback or donor assessment introduced here. Regenerated receipts will have new code and configuration identities.

Historical regression verification: 97 offline smhi2epw tests passed, four skipped and four live-network tests were deselected. Coverage includes a complete recovered 8,760-hour EPW, a 49-hour donor-filled gap, radius enforcement, circular winds, invalid configuration before requests, preserving valid observations and the 48-hour temporal ceiling. The EPSM workflow suite passed 236 tests with one skip. Live donor retrieval was exercised by the separate replay, not the network-marked test suite.

Current national completion run

On 5 October 2026 the frozen automatic policy at 5930c07 completed all 128 requested municipality/year jobs from 2017–2024. The preceding national run at 4861f9e completed 124 of the same 128 jobs using opt-in raw donors without ERA5. The original temporal-only run completed 32 of 85 jobs; preparation changes expanded the later denominator. The separate 52-of-53 replay above uses another denominator and the earlier policy.

The current 128 EPWs were independently rechecked on 6 October: hashes matched both recorded receipts, ordered calendars and 35-field rows were correct, and required exported values passed finite/range and dew-point checks. There are 112 common-year and 16 leap-year files, totaling 1,121,664 exported hours. Historical v1/v2 EPWs and their per-job receipts are no longer available locally; their retained status manifests support the earlier counts.

All current files are mixed_reconstructed. Donors supply meteorology in 101 jobs, ERA5 in 35, and temporal filling in 126; these groups overlap. Twenty-five jobs exceed 5% reconstructed required meteorological cells on the buffered ingestion grid, ranging from 0.0114% to 100% across all jobs. This fraction excludes solar and differs from exported-hour source fractions. Five percent is descriptive, not an acceptance threshold. The run checks completion and export usability, not local weather accuracy or event preservation.

The portable completion evidence summary includes definitions, source commits, artifact hashes, warning counts and the audit checks. Its original artifact directory is epsm_workflows/output/national-mf-se02-random-baseline-2026-10-05-v3-weather-shadow; the status, quality summary, runtime manifest, verification and per-job weather and source receipts retain the full audit trail. Large EPWs and raw caches are not distributed with the examples. Notebook 08 reads the summary offline.

Corrected-source sensitivity check, 6 October 2026

After the solar correction at 050d195, a small fresh experiment rebuilt three references and six outage cases: Gothenburg 2016, Luleå 2023 and Gothenburg 2024, each with a 168-hour June solar outage and a 48-hour meteorological outage at its original observed temperature maximum. All nine annual simulations had complete calendars and zero engine warnings. The paired compiled models and unmasked weather rows were identical.

For one illustrative 100 m² ideal-load building, annual sensible cooling changed by +0.618–1.259 kWh/m² (+3.04–6.79%) under the solar masks. Meteorological masks changed cooling by -0.166–+0.102%, with hidden-hour temperature MAE of 0.856–1.142 °C. These are descriptive comparisons against corrected references, not national error bounds or independent validation of solar components.

Physical nighttime zeros divided the solar outages into bounded daylight gaps filled by daily profiles. Meteorology used assessed fixed donors, with circular temporal recovery for rejected wind direction. None selected ERA5 in these six outages. The reference solar group includes modelled STRÅNG components and native bounded fills; optional clouds were frozen. The public provider/reanalysis transport is checked separately by deterministic compiler-to-engine fixtures.

The portable summary records assumptions, actual recovery methods, code and model hashes, and output checks. Notebook 08 reads this summary without running a new campaign. The original larger campaign and its inputs remain frozen under their original source identity.