# Wind-fleet demo — data sources & provenance

Every wind-fleet figure the MCP returns traces to the real, permissively-licensed public
dataset below. Raw files are **not** committed — only the curated, derived slice under
`public/data/wind/`, rebuilt with `scripts/build-wind-kelmarsh.py`.

> Representative public data — not connected to any live wind farm. SCADA ingestion for a
> customer fleet is a per-deployment step; see `describe_connector_requirements`.

## Kelmarsh wind farm, 2020 — `kelmarsh-2020.json`

- **Source:** *Kelmarsh wind farm data*, Cubico Sustainable Investments Ltd. 6 × Senvion MM92
  (2.05 MW, 12.3 MW site), Northamptonshire, UK.
- **URL:** https://zenodo.org/records/8252025 (concept doi:10.5281/zenodo.5841834) ·
  **License:** CC-BY-4.0 · **Retrieved:** 2026-07-07 (`Kelmarsh_SCADA_2020_3086.zip`,
  SHA-256 prefix `1cf3ad08`; plus `Kelmarsh_WT_static.csv`).
- **Slice taken:** derived features from the calendar-2020 10-minute SCADA and status logs of
  all 6 turbines — per turbine: a 0.5 m/s-binned power curve (bins with ≥1 h of data; mean
  power from `Power (kW)` vs `Wind speed (m/s)`), monthly `Energy Export` / `Energy
  Theoretical` (kWh) + mean wind, a daily series (energy, mean wind, max
  `Front bearing temperature`), and a downtime summary from the status log (hours + event
  counts by IEC category, top downtime causes by message). Turbine static data (model, rated
  power, hub height, coordinates) from `Kelmarsh_WT_static.csv`.
- **Note:** "energy availability" = Energy Export ÷ Energy Theoretical, the production-based
  availability Greenbyte reports; it can exceed 100% in months where the turbine beat the
  theoretical model. 2020 is a leap year (366 daily rows).