| --- |
| license: cc-by-4.0 |
| tags: |
| - electricity |
| - energy |
| - power-markets |
| - weather |
| - nordic |
| - time-series |
| - forecasting |
| pretty_name: Power Price Forecasting Inputs (Nordics + NL) |
| --- |
| |
| # Power Price Forecasting Inputs — Nordics + Netherlands |
|
|
| Every input behind the price-forecast models at |
| [elpriser.org](https://elpriser.org): weather, prices, hydro reservoirs and |
| interconnector capacity, for 13 bidding zones. |
|
|
| Models trained on this data: |
| [nordic-price-forecast](https://huggingface.co/Elpriser/nordic-price-forecast) (13 zones) |
| and [denmark-price-forecast](https://huggingface.co/Elpriser/denmark-price-forecast) |
| (DK1/DK2, higher resolution). |
|
|
| ## The part that is hard to get elsewhere |
|
|
| `weather_forecast_as_issued/` holds the weather forecast **as it looked N days |
| before the delivery hour**, for N = 1 to 7 — not what the weather turned out |
| to be. |
|
|
| This matters more than it sounds. Train a price model on *actual* weather and |
| serve it a *forecast* at inference time, and the model has never seen the |
| noisier input it must actually work with. It will look strong in backtest and |
| underperform in production, with the gap widening at longer horizons — exactly |
| where a multi-day forecast earns its keep. Derived from Open-Meteo's |
| previous-runs API, which only covers 2024-03-15 onward. |
|
|
| `weather_actual/` holds observed weather over a longer window, for fitting |
| weather-to-production relationships where using actuals is legitimate (a |
| static physical mapping, fed forecast weather at both train and serve time). |
|
|
| ## Contents |
|
|
| | Path | What it is | Coverage | |
| |---|---|---| |
| | `weather_forecast_as_issued/{zone}.parquet` | Wind 100m, direct radiation, temperature, precipitation — each at leads of 1-7 days. One representative point per bidding zone | 2024-03-15 → present, hourly | |
| | `weather_actual/{zone}.parquet` | Same four variables, observed | 2022-01-01 → present, hourly | |
| | `weather_denmark_highres/` | 14 points for the Denmark-only model: offshore wind clusters (Horns Rev, Anholt, Rødsand), German wind north and solar south, plus neighbouring zones | as above | |
| | `market/prices_hourly_13zones.parquet` | Day-ahead prices, EUR/MWh, all 13 zones from one source so no currency boundary | 2024-01-01 → present, hourly | |
| | `market/hydro_reservoir_weekly.parquet` | Reservoir energy content, 10 hydro zones. The driver behind Nordic price levels — producers hold water back when reservoirs are low | 2022 → present, weekly | |
| | `market/transfer_capacity_daily.parquet` | Month-ahead forecast transfer capacity per border and direction, for Denmark's six interconnectors | 2024-03 → present, daily | |
| | `training_dataset_nordic.parquet` | The assembled, leakage-guarded training set: one row per (zone, hour, horizon 2-9), 30 features | 2.1M rows | |
|
|
| ## Zones |
|
|
| DK1, DK2 · NO1-NO5 · SE1-SE4 · FI · NL |
|
|
| ## Leakage discipline in the training set |
|
|
| If you use `training_dataset_nordic.parquet` directly, the rules already |
| applied are: |
|
|
| - one row per (zone, target hour, forecast horizon h = 2..9) |
| - price lag `k` is populated **only when `k >= h-1`** — at h=5 you do not yet |
| know yesterday's price, and filling it would leak |
| - `level_30d` covers the 30 days ending at the last *known* day, not the 30 |
| before the target |
| - weather comes from the as-issued archive at lead `min(h, 7)` |
| - reservoir readings are lagged 14 days, deliberately more conservative than |
| ENTSO-E's real weekly publication lag |
|
|
| ## Sources & license |
|
|
| CC BY 4.0. Prices, reservoir levels and transfer capacity: ENTSO-E |
| Transparency Platform. Weather: [Open-Meteo](https://open-meteo.com) (free for |
| commercial and non-commercial use with attribution). Collection code: |
| [elpriser.org on GitHub](https://github.com/x2q/elpriser.org), |
| `scripts/forecast_model/`. |
|
|
