ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
Machine learning-based grid-point warning for thunderstorm gales in the Sichuan Basin
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Using historical thunderstorm gale cases in the Sichuan Basin from March 1 to September 30 during 2018–2022, combined with 3D radar mosaic data and surface maximum wind observations, we constructed a thunderstorm gale sample dataset and developed a grid-based gale warning model. Independent validation was conducted for 2023 thunderstorm gale events to evaluate the model’s warning performance. The main conclusions are as follows:(1) The LightGBM model achieved the highest probability of detection (POD, 0.536 at 15-minute lead time with a 10 km evaluation radius), but also exhibited the highest false alarm ratio (FAR). The random forest (RF) model demonstrated the best overall performance, with the highest critical success index (CSI, 0.306 at 30-minute lead time and 10 km radius). Both CSI and POD scores decreased significantly with longer lead times or smaller evaluation radii, particularly when the lead time extended from 30 to 45 minutes.(2) Synoptic background significantly influenced warning effectiveness, Under strong cold air influence, factors such as composite reflectivity (CR), echo top (TOP), and 45dBZ echo top height (H45) were more likely to reach high values, favoring intense convective development. However, newly initiated storms at convective fronts tended to increase missed detections. In the absence of strong cold air, thunderstorm gales primarily occurred at the leading edge of convective systems, resulting in higher POD.(3) The temporal variation of vertically integrated liquid water (VIL) contributed most to model decisions, followed by vertical liquid water density (VILD), echo top height, and maximum reflectivity (maxZ), highlighting deep convection as the core mechanism of thunderstorm gales. In cold air-absent scenarios, downdrafts dominated the warning process. Analysis of key feature values and high SHAP values revealed that temporal changes in convective echoes were critical for warnings. Samples with strong TREC wind fields often corresponded to positive SHAP values, indicating increased probability of convective gales when echo motion accelerates.
Keywords:
Project Supported:
Clc Number:
Mobile website
