Sea Winds Predictions

I. Data Phase 1 The objective of this challenge is to develop probabilistic forecasting models for offshore wind speed and direction across two maritime regions, supporting wind energy planning and offshore operations. The provided dataset includes multi-level atmospheric reanalysis data, operational forecasts, with prediction horizons of +1, +7, and +14 days ahead. Spatial dimension: Covers two offshore maritime regions: the North Sea (51-62°N, -4-10°E) and the East China Sea (24-35°N, 118-132°E) Regional gridded data at 0.25° resolution (~25 km) Worldwide daily reanalysis at 1° resolution f Ground truth observations from 15 stations Predictions required at 7 vertical levels: 10m, 100m, and pressure levels 1000 / 925 / 850 / 700 / 500 hPa Temporal dimension: Training data spans from 2019 to 2021 at 6-hourly resolution (3 years) and daily resolution (worldwide) Station observations available at 6-hourly resolution for the same period Evaluation/inference data spans 2022, organized as 8 independent forecast windows distributed across the year to capture seasonal variability Each window provides 14 days of context data and requires predictions at three horizons (+1 day, +7 days, +14 days), each at four hours of day (00, 06, 12, 18 UTC) Participants must produce probabilistic forecasts: Wind speed: three quantiles (q05, q50, q95) forming a 90% prediction interval Wind direction: median (dir_50), optionally with interval endpoints (dir_05, dir_95) Auxiliary data: ECMWF forecasts at surface and pressure levels (2019-2021) Elevation from ETOPO 2022 Land-sea masks for both regions Station metadata (coordinates, measurement heights, nearest grid point) All features are provided as Parquet files for efficient loading and direct integration with standard machine learning pipelines. For more details about the hackathon see Here --> https://www.codabench.org/competitions/13821/ Mini Challenge — Caribbean Buoy dataset (companion dataset) A short, standalone side-challenge: post-process raw weather forecasts into calibrated probabilistic wind forecasts at 6 offshore buoys, at horizons +1, +7 and +14 days. Spatial dimension: 6 stations in the Gulf of Mexico / Caribbean (~14-29°N, 90-58°W) Operational forecasts interpolated to each buoy, plus a local reanalysis patch (3°×3° at 0.25°) per buoy Ground truth from the 6 buoys Temporal dimension: Training: 2020-2021. Evaluation: 2022, in 8 independent forecast windows across the year Each window: 14 days of context, predictions at +1/+7/+14 days × 4 hours (00, 06, 12, 18 UTC) Forecasts reach +10 days (cover +1/+7); +14 days must be predicted from buoy history / climatology Phase 2 This companion dataset extends the challenge to the scale an offshore turbine actuallyoperates at. It adds a high-resolution target — Météo-France AROME wind at 1.3 kmresolution and 125 m hub height over the North Sea. Spatial dimension: North Sea only (≈51-62°N, -4-10°E) AROME native grid at 1.3 km (479×433), 125 m above sea level Low-resolution reanalysis and ECMWF forecasts inputs at 0.25° Scored footprint: the 43,715 AROME sea points inside the reanalysis domain Bathymetry (water depth) provided for fixed-bottom wind-farm siting Temporal dimension: Training: 2016-2020, AROME at 8 steps/day (3-hourly) Inference: 2021, organized as 8 independent forecast windows across the year; each window provides 14 days of context and requires predictions at +1 / +7 / +14 days The HRES forecast driver spans 2016-2018 here and 2019-2021 in the Phase-1 dataset — use both together Outputs: same probabilistic format as Phase 1 — wind-speed quantiles (q05, q50, q95) and adirection interval (dir_05, dir_50, dir_95), at +1 / +7 / +14 days. Contents (phase2_dataset): train/ — arome (1.3 km target), arome_coarse125 (forecast target space), reanalysis, reanalysis_extra (MSLP, 2 m temp, pressure-level u/v/z/t), hres inference/window_1..8 — reanalysis + HRES context to forecast from static/ — arome_static.nc, reanalysis_static.nc, bathymetry, footprint_points.parquet Final-evaluation inference windows (2022):The file `inference_2022.zip` contains the eight forecast windows used for the final leaderboard evaluation, re-cut on the withheld year 2022. It mirrors the `inference/window_{1..8}/` folder of the main archive: each window holds 14 days of reanalysis context, the HRES forecast issued at `context_end`, and a `metadata.json` giving the scored days (+1 / +7 / +14). II. Acknowledgments This dataset contains modified data from Copernicus Climate Change Service information [2026]. Neither the European Commission nor ECMWF is responsible for any use of the Copernicus information or data it contains. This dataset is derived from meteorological data provided by Météo-France, available underthe Licence Ouverte Etalab 2.0. Source: Météo-France. Post-processing (including spatial and temporal aggregation) was performed by FranceEnergies Marines. This dataset constitutes a derived product and does not representofficial Météo-France data.