Samgöngustofa (Iceland Transport Authority) — bifreiðatölur

Vehicle-registration statistics from https://bifreidatolur.samgongustofa.is/. The site is a thin SPA that embeds a separate Power BI report per section, each with its own resource key. scripts/samgongustofa.py extracts tidy CSVs from two of them.

Quick start

# What's available (both reports, their keys, every groupable dimension)
uv run python scripts/samgongustofa.py list

# Current fleet on the road, by fuel — the EV-transition read
uv run python scripts/samgongustofa.py fetch --report onroad --dimension fuel

# New registrations (imports) by brand, per year
uv run python scripts/samgongustofa.py fetch --dimension make --years 2020-2026

# Same, month by month, for a YoY view
uv run python scripts/samgongustofa.py fetch --dimension make --years 2025,2026 --monthly

# Only brand-new (exclude imported-used)
uv run python scripts/samgongustofa.py fetch --dimension fuel --years 2024-2026 --import-state new

Output lands in data/processed/samgongustofa/<report>_<dimension>[_by_year[_month]][_new|_used].csv as tidy long format, e.g. make,year,count or fuel,count.

Two reports: flow vs stock

--report	Section	Meaning	Time
nyskraningar (default)	#nyskraningar	New registrations = imports. First Icelandic registration — brand-new and imported-used. The flow into the fleet.	year / month / new-vs-used slicers
onroad	#tolfraedi ("Tölfræði ökutækja")	Current fleet on the road ("í umferð") — every vehicle with an active registration. The stock.	snapshot (no year)

Four dimensions (--dimension)

Every report groups by the same four axes. Confusingly, the brand column is named Tegund ("kind") in the model.

--dimension	PB column (nyskr. / onroad)	Example values
make	Tegund	TOYOTA, KIA, VOLKSWAGEN, BYD, TESLA, MG, XPENG, …
fuel	Orkugjafi / Orkugjafi (groups)	Bensín, Dísel, Rafmagn (electric), Tengiltvinn (PHEV), Hybrid, Metan, Vetni, Vélarlaus (engineless=trailers), Annað
class	Ökutækisflokkur / Ökutækjaflokkur	Fólksbifreið M1, Sendibifreið N1, Vörubifreið N2/N3, Hópbifreið M2/M3, bifhjól, dráttarvél, eftirvagn, torfæruhjól, …
model	Undirtegund	MODEL Y, DUSTER, ID.4, RAV4, YARIS, … (top ~1000)

fuel is the cleanest read on Iceland's EV transition: e.g. the current fleet snapshot is ~171k Bensín, ~151k Dísel, ~44k Rafmagn, ~30k Tengiltvinn.

Slicers & cross-filters

Temporal slicers (nyskraningar only):

• --years 2020-2026 or --years 2025,2026 — the Ár - ísl. slicer (values like 2023L).
• --monthly — break each year into months (Mánuður - ísl., text labels 01-janúar … 12-desember); --through N stops at month N.
• --import-state new|used|all — the Innflutningsástand slicer ('Nýtt'/'Notað'). Default all = both.

Cross-filter either report by any model column with --where 'COL=VALUE' (repeatable = AND across columns; ; inside a value = OR). This turns the single-axis visuals into crosstabs the dashboard never shows directly:

# BEV imports by brand, 2026 — cross make × fuel
uv run python scripts/samgongustofa.py fetch --dimension make --years 2026 --where 'Orkugjafi=Rafmagn'

# Electric + PHEV vans on the road, by make
uv run python scripts/samgongustofa.py fetch --report onroad --dimension make \
    --where 'Orkugjafi (groups)=Rafmagn;Tengiltvinn' --where 'Ökutækjaflokkur=Sendibifreið N1'

Column names are the raw Power BI properties — list prints them all. Values are matched as text; the year axis has its own --years path.

Note the current calendar month is partial — the newest month reflects registrations to date, not a full month.

How extraction works (and why the naive way fails)

Each visual POSTs a SemanticQueryDataShapeCommand to

https://wabi-europe-north-b-api.analysis.windows.net/public/reports/querydata?synchronous=true

with header x-powerbi-resourcekey: <key> and no bearer token. It is a public report, but the anonymous grant is session-, origin- and rate-bound:

• a cold httpx client works for a few requests, then returns 401 PowerBINotAuthorizedException (rate limit — the API even exposes retry-after);
• a POST from any origin other than the app.powerbi.com iframe is rejected outright with 401 (access-control-allow-origin: * on responses is a red herring — the grant is origin-bound).

The reliable method (what the script does): drive the SPA with Playwright, then replay each query with fetch() executed inside the app.powerbi.com iframe via frame.evaluate — reusing the report's own live session, origin and pacing. Every year/month/state variant then returns 200.

Three more gotchas:

1. The POST body must keep its top-level modelId / version / cancelQueries, or the API answers 400 "ModelId must be between 1 and 9.2e18".
2. Resource keys and the model id rotate — never hardcode them. The script decodes the key from the active iframe's embed token (?r=<base64> → {"k": …}) and reads templates off the section's own live requests.
3. Response rows come in two shapes: nyskraningar uses C: [dimension, count]; onroad uses G0 + X[0].M0 (in-traffic count) + X[1].M0 (new-this-period). The parser handles both.

Beyond the CLI — Playwright one-offs

list, --dimension, --years/--monthly, --import-state and --where cover almost everything by composition. For anything they don't — a bespoke aggregation, a multi-measure visual, driving a slicer in the UI — drop to the generic scripts/powerbi.py primitives (see the powerbi skill) rather than starting from scratch. They already solve discovery, the iframe-replay auth, the modelId gotcha and the DSR decompression.

import asyncio, sys; sys.path.insert(0, "scripts")
import powerbi as pb
from playwright.async_api import async_playwright

async def main():
    async with async_playwright() as p:
        b = await p.chromium.launch(headless=True)
        page = await b.new_page()
        d = await pb.discover(page, "https://bifreidatolur.samgongustofa.is/", anchor="#nyskraningar")
        payload = pb.where_in(d.templates["Tegund"], "Ár - ísl.", ["2026L"], text=False)  # make × 2026
        payload = pb.where_in(payload, "Orkugjafi", ["Rafmagn"])                            # ...BEV only
        rows = pb.group_counts(await pb.replay(d.frame, d.key, payload, retries=1))
        await b.close()
    print(sorted(rows.items(), key=lambda kv: -kv[1])[:10])

asyncio.run(main())

d.templates is keyed by each visual's first group-by column (Tegund, Orkugjafi, Ökutækisflokkur, Undirtegund). To learn a new column or literal format, drive the slicer once with a page.on("request", …) listener and read the post_data — how every constant here was found (pb.capture_requests helps). Full helper reference is in the powerbi skill.

GEO-FENCE — must run from Iceland

bifreidatolur.samgongustofa.is answers Icelandic IPs in ~50 ms and ConnectTimeouts from datacenter address space (GitHub runners, cloud VMs, most hosted notebooks cannot reach it). Run the scraper and the health probe from an Icelandic connection. The daily health probe therefore runs on the self-hosted mac-mini in Iceland (see AGENTS.md).

Caveats

1. Real-time only. The reports show current state; there are no historical snapshots beyond what the year/month slicers expose. Persist CSVs if you need a time series of the stock.
2. make and model visuals are top-N (200 makes / ~1000 models), so their column totals fall slightly below the all-inclusive class/fuel totals.
3. Structure can drift. Keys/model id rotate (handled). If a section is renamed or a slicer column changes, list will show what actually exists — start there.
4. Pacing. The script sleeps ~2.5 s between queries; don't hammer it.

Example: Chinese-brand car imports (2026)

Filtering make to Chinese-owned brands (BYD, MG, XPENG, Polestar, Leapmotor, Maxus, FAW, …) over nyskraningar_make_by_year.csv shows imports rising from <1 % of new registrations (2020) to 11 %+ in 2026 YTD, with Jan–Jul 2026 Chinese registrations up ~108 % YoY — while they are still only ~1.6 % of the on-road fleet (onroad_make.csv). The stock lags the flow.

Alternative sources

• Hagstofan Umhverfi/5_samgongur/…/SAM30120.px — fleet by fuel type, not by make.
• Bílgreinasambandið — industry-association registration statistics.