Windspeed Dataset

The Windspeed dataset is constructed using data from ECMWF’s ERA5 over a region of west Kazakhstan (lat 46-54, long 46-54). Hourly values for wind (u and v components), temperature (at 2m) and surface pressure were collected between 1st January 2020 to 31st Dec 2024. The task is to predict the average wind speed in the central 50% of the region from an input of 6 hours of temperature and pressure data covering the whole region. Each input data point has shape [12, 32, 32] (6 hours x 2 variables for the channels, and 32x32 pixels forming the spatial region). The label is the mapped onto a fine-grained version of the Beaufort wind scale (120 instead of 12 classes). To avoid classes with no samples, we filter out data to only include samples where the label sits between class index 20 and 40 (corresponding to the range 2.36 - 6.69 m/s). Data from 2020-2022 form the training set, while 2023 is the validation set and 2024 the test set.