The machine learning dataset of thunderstorm gale in the Sichuan Basin

Forecast predictors were selected from the ECMWF (European Centre for Medium-Range Weather Forecasts) forecast products issued at 08:00 Beijing Time, including 115 atmospheric parameters such as temperature, wind speed, and Convective Available Potential Energy (CAPE), with horizontal resolutions of 0.125° × 0.125° (surface) and 0.25° × 0.25° (upper air).

Data processing included: (1) unifying the temporal resolution to 3-hour intervals; (2) labeling samples as positive if at least one thunderstorm gale occurred within a 3-hour window; (3) feature extraction; and (4) matching gridded forecast data to observation stations using bilinear interpolation.constructs eight negative sample datasets (Prop1–Prop8) with varying positive-to-negative ratios ranging from 1:1 to 1:8.three meteorologically meaningful temporal sampling schemes were further designed: N1 (traditional synchronous sampling), N2 (sampling 12 hours in advance for comparison), and N3 (random sampling within the 12-hour lead time).All data are standardized using the min max method and subjected to random shuffling preprocessing