Competitive Powertrain Benchmarking

Benchmark competitor powertrains — on free, recurring data

Give R&D and the Benchmarking Team an evidence base on engine specs, driveline configuration and certified whole-vehicle CO₂ — without paid market-intelligence subscriptions.

• Frame — turn "how do we compare?" into measurable questions across OEM × segment × year.
• Mine & validate — recurring ingest from an authoritative EU source, reproducible and auditable.
• Analyse & model — honest statistics; a CO₂ predictor that says what it does and does not know.
• Deliver — an interactive dashboard R&D can drive, plus a static, offline record.

EEA HDV CO₂ monitoring — queried live over HTTP

Regulation (EU) 2018/956: Member States and manufacturers report every certified truck to the European Environment Agency. Exposed through Discodata — a SQL Server reachable by plain HTTP, no authentication.

Five stages, offline-first, one shared core package

powerbench/ holds paths, the Discodata client, the schema, data-IO, features and evaluation — imported by every script and by the Streamlit app.

Work around the endpoint, don't fight it

• Pagination is broken — Discodata's p parameter repeats page one. Instead, bound every request with a WHERE on year × manufacturer and pull each chunk whole.
• Two source tables, one schema — CO2_HeavyDutyVehicles (2019–20, full VECTO detail) and HDV_2023_viewer (2023, + registration country); a mapper reconciles them.
• Every pull is a record — one JSON snapshot per source with a SHA-256 sidecar; retry with backoff; unpublished years skipped with one clear message.

Reject the row's mistakes, not the row

# powerbench/schema.py
class HDVRow(BaseModel):
    Engine_RatedPower_kw: float | None
    GrossVehicleMass_t:   float | None
    # value outside (0, hi] -> None,
    # never a rejected record

The manifest is the authority

Clean rows land in DuckDB (powerbench.duckdb) and a Parquet mirror. manifest.json records every source snapshot, its SHA-256 and row count — so the active dataset is always known and reproducible.

Three different "CO₂", three different jobs

Who is lowest, and who is improving fastest

The dashboard renders this as OEM box plots plus a 2019→2023 change ranking — narrowed to comparable vehicle groups.

Engine-cycle CO₂ is not a league table

Correlation analysis shows engine-bench CO₂ (g/kWh) falls as engine displacement rises — r ≈ −0.5. Bigger diesels run closer to their efficient point.

• The consequence — ranking OEMs on raw g/kWh flatters whoever sells the heaviest long-haul tractors.
• The rule — benchmark on CO₂v (whole vehicle); use g/kWh as a model feature, never as a scoreboard.

Predict whole-vehicle CO₂ from the spec sheet

Target: CO₂v (g/km), restricted to the VECTO classes with a comparable figure. Two feature sets, so coverage and ceiling are both explicit:

• base — mass, vehicle class, powertrain, fuel, flags. Works on every reporting year.
• rich — base + engine ratings + axle configuration. 2019–2020 only.

The evaluation method is the result

• Out-of-fold headline — shuffled 5-fold cross-validation; every row scored by a model that never saw it.
• Baseline in view — "predict the median" MAE shown next to every model. A useful model must beat it.
• Optimism gap shown — the naive train-on-test R² is printed beside the CV number, not instead of it.
• Leakage guard — assert_no_leakage() bans every CO₂ / fuel-consumption column; g/kWh feeds the VECTO calc that makes the target.

Spec sheet explains about half; engine specs push it to ~60%

Curb mass and vehicle group carry most of it; engine displacement and power add the rich set's lift. The naive↔CV gap is small — not overfit.

A number that admits what it can't see

• Inputs — sliders for GVW, curb mass, vehicle group, powertrain, and (rich) engine power / displacement / speed.
• Output — predicted CO₂v ± the model's cross-validated MAE.
• Scope flag — any input outside the training min–max raises an extrapolation warning.
• Labelled illustrative — the model sees only these fields; real CO₂v also depends on aero, tyres, gearbox, auxiliaries.

An eight-tab Streamlit app R&D can drive

• Pipeline — mine → validate → load → train, live from the browser, streamed progress. Nothing pre-downloaded.
• Overview · Distributions · Correlations · Benchmark — the analysis, filterable by year / OEM / powertrain / segment.
• ML — both feature sets side by side, out-of-fold scatter, feature importance, what-if.
• Metrics · Documentation · Provenance — definitions, this write-up, and the SHA-256 lineage. Light / dark.