气象2026，Vol.52Issue(3)：325-336,12.DOI:10.7519/j.issn.1000-0526.2025.091601
基于机器学习的四川盆地雷暴大风格点预警
Machine Learning-Based Grid-Point Warning of Thunderstorm Gale in Sichuan Basin
摘要
Abstract
Based on thunderstorm gale cases in Sichuan Basin from March 1 to September 30 in 2018-2022,combined with three-dimensional radar mosaic data and surface maximum wind observations,this paper constructs a thunderstorm gale sample dataset and develops a grid-point thunderstorm gale warning model.Independent validation is performed on thunderstorm gale events in 2023 and the warning perform-ance of four models is evaluated.The results show that the LightGBM model achieves the highest proba-bility of detection(POD),reaching 0.536 at a 15 min lead time and a 10 km evaluation radius,but it also exhibits the highest false alarm rate(FAR).The random forest model demonstrates the optimal compre-hensive performance,with the highest critical success index(CSI)being 0.306 at a 30 min lead time and a 10 km evaluation radius.Both CSI and POD decrease significantly with prolonging warning lead time or de-creasing evaluation radius,with a particularly notable decline in CSI when the lead time extends from 30 to 45 min.Synoptic conditions significantly influence the warning performance.Under pronounced cold air influence,factors such as echo intensity,echo top height,and 45 dBz echo top height are more likely to have high values,favoring the development of severe convection.However,newly initiated storms at con-vective fronts often lead to the increase in missed detections.In the absence of strong cold air,thunder-storm gales mainly occur at the leading edge of convective systems,resulting in higher POD.The temporal variation of vertically integrated liquid water content contributes most to the decision-making of models,followed by vertically integrated liquid water content density,echo top height,and maximum reflectivity factor.This highlights the central role of deep convection in the generation of thunderstorm gales.In the scenarios without cold air intrusion,downdrafts play a dominant role in thunderstorm gale warnings.Analysis of key feature values and high SHAP values reveals that temporal variations in convective echoes are critical for effective warnings.Samples with high echo-tracking wind speeds often correspond to positive SHAP values,indicating an increasing probability of convective wind events when echo motion accelerates.关键词
雷暴大风/机器学习/回波特征/预警Key words
thunderstorm gale/machine learning/radar echo characteristic/warning分类
天文与地球科学引用本文复制引用
罗辉,杨康权,向筱铭,苟阿宁,张武龙,王彬雁..基于机器学习的四川盆地雷暴大风格点预警[J].气象,2026,52(3):325-336,12.基金项目
高原与盆地暴雨旱涝灾害四川省重点实验室科技发展基金(SCQXJZD202102-09、SCQXKJYJXZD202402)、四川省科技计划重点研发项目(2022YFS0542、2024YFFK0408)、中国气象局创新发展专项(CXFZ2024J013、CXFZ2025J014)、湖北省自然科学基金联合基金重点项目(2024AFD205)和四川省气象局重点创新团队(SCQXZDCXTD202401)共同资助 （SCQXJZD202102-09、SCQXKJYJXZD202402）
