RFC: SMHI STRÅNG irradiance + Lantmäteriet roof geometry for PV prediction, performance scoring & map-drawn arrays

[HuggeK]

Status — 2026-09-02. Implemented. #718 (plane-of-array wiring) merged;
open as green drafts based on master: #734 (STRÅNG + source coverage +
vendored MapLibre) · #735 (roof geometry + building picker, now with
catalog-agnostic STAC Basic auth) · #826 (draw your array on the map).
Optional vostok shading (#736) is closed and parked as idea #1048 — the
operator would have to build GPL C++ themselves, which is too much to ask.

This post was rewritten on 2026-09-02 to consolidate the original proposal
and its three update comments into one text; implementation details live in
the PR bodies.

Summary

Two Swedish open-data sources sharpen FTW's PV modeling:

1. SMHI STRÅNG — modelled solar irradiance
   (hourly, ~2.5 km, 1999→present, CC BY 4.0, no key), used to score how real
   panels actually performed against a weather-expected baseline and to
   calibrate the forward PV predictor.
2. Lantmäteriet open geodata — building
   footprints + LiDAR, used to derive roof tilt/azimuth/rated W automatically —
   plus a map-drawing fallback that needs no account at all.

Everything is additive and optional; nothing changes the control tick,
dispatch, or the optimizer contract.

The finding that shaped the design

STRÅNG is an analysis product with no forward horizon — it is historical
truth, not a forecast. So it scores and calibrates; forward forecasting stays
on the existing providers. Its parameter set was probed against the live API
(data for exactly codes 116–122; the measured table and the two checks that
pin the identification are in #734). It publishes no cloud cover, but sunshine
duration derives it as 1 − minutes/60, with an explicit unknown when the
sun is too low for the number to mean anything.

Its domain was probed too, and the grid is rotated relative to lat/lon, so the
declared coverage box is honestly advisory: covers: false is definitive,
covers: true means "worth asking". The same logic handles Sweden's diagonal
border for Lantmäteriet — the box admits Oslo, and the STAC search returning no
tiles is the authority, not geometry.

The pieces

PR	What	State
#718	Project irradiance onto each configured panel plane (Erbs split for GHI-only providers)	merged
#734	STRÅNG client, performance scoring + calibration, GET /api/data-sources coverage registry (answers part of #726), mpc.PVBand, Leaflet → vendored MapLibre GL 6	draft, green
#735	roofmodel/ optional module: pick your building → clip LiDAR to its footprint → RANSAC/DBSCAN segmentation → pre-filled pv_arrays. Credentials are the operator's own Geotorget account (HTTP Basic — Lantmäteriet provides no OAuth for the STAC download APIs), and the client is catalog-agnostic: stac_base_url + collection keys point it at any STAC catalog	draft, green
#826	Draw an angled rectangle over the map → area, azimuth and both azimuth candidates; tilt typed, slope-corrected by 1/cos(tilt)	draft, green
#1048	Per-face shading factors via vostok — parked as an idea	idea issue

Decisions worth recording (and two corrections)

Picking a building is a correctness fix, not a nicety. RANSAC fits
infinite planes: a south roof is z = f(y) with no x term, so any second
building sharing pitch and ridge orientation lies on the same plane — measured
as byte-identical contamination at every separation from 3 m to 40 m
(distance-invariance is the signature of a global fit). Clipping the cloud to
the picked footprint restores the isolated segmentation exactly. The honest
caveats: in one synthetic scene the contaminated area was coincidentally
closer to truth, and the clipped pipeline systematically under-reads area
(~−12 % at 8 pts/m², worse at Laserdata Skog's real 1–2 pts/m²) because DBSCAN
drops boundary points — both reasons a derived rated W is a proposal, never
a quote.

Derived arrays fill the editor, not the config. A successful derive puts
real numbers in the form with the 3D preview updating, then stops at "Review
them and press Save". Config never changes behind the operator's back, but the
proposal lands where it can be seen and edited. (The original text promised
fully automatic writes; this is the better shape.)

Ingest matches what Lantmäteriet actually publishes. Byggnad-vektor
delivers GeoPackage, Laserdata-skog delivers LAZ organised as COPC; assets are
selected by declared media type, not key names. GeoPackage is decoded with
~100 lines of sqlite3 + struct against the published OGC layouts rather
than a GDAL dependency, and COPC's octree index turns "which building" into a
kilobytes-sized range request instead of a hundreds-of-MB tile — the difference
between a derive being feasible on a Pi and not. (An earlier claim here that
GeoJSON assets were already handled was wrong; both paths now exist and are
tested.)

A roofmodel (Python) CI job now exists. The module's suite was invisible
upstream (only optimizer/tests ran); wiring it in immediately caught two
shell-quoting bugs that Windows was hiding.

The GPL boundary for vostok (now in #1048): separate process, files and a
command line, never bundled, never auto-installed; absent vostok,
shading_factor is omitted, never defaulted to 1.0 — "we did not look" and
"we looked and it is clear" are different claims.

The original open questions, answered

1. Third optional module? Yes — roofmodel/ mirrors the optimizer:
   arm's-length subprocess, one versioned roof_model.json, safe-absent.
2. MapLibre: CDN or vendored? Vendored (web/vendor/maplibre/,
   web/vendor/terra-draw/), per the ship-it-on-the-box policy fix(web): escape loadpoint HTML and ship Leaflet on the box #910
   established after this RFC opened. Byte-identity to the npm artifacts is
   verified; nothing loads from a CDN.
3. Right first slice? feat(forecast): project irradiance onto each PV array's plane #718 was — smaller than the scoring endpoint, and it
   improved every radiation provider with no new surface.
4. irradiance_history table? Yes; forecasts stays forward-only.

Still needed

A live derive against Lantmäteriet's real service with an operator's own
Geotorget credentials (Byggnad-vektor + Laserdata-skog ordered on the
account). Every layer is tested against fakes and synthetic tiles, and the
HTTP/STAC layer against a local server speaking the real formats — the one
thing that has never run is the production endpoint end-to-end.

Attribution: SMHI STRÅNG and Lantmäteriet data are licensed CC BY 4.0.