深度学习方法预测阿克苏地区冰雹云雷达回波个例分析

利用2021—2022年4—9月阿克苏地区冰雹云的雷达回波资料，基于轨迹GRU模型和GAN模型共同构建一个深度学习的回波外推模型，应用于强对流（冰雹）天气监测预警。采用分阈值和预报时效的评估方法，对深度学习的回波外推模型预测回波的效果进行分析，结果表明：（1）在30 min预测时间内，随反射率阈值增加，临界成功指数（CSI）和命中率（POD）逐渐降低，虚警率（FAR）先降低后升高，FAR在反射率阈值为35dBZ时最低。（2）在反射率阈值为35 dBZ和相同外推时效的情况下，基于深度学习的回波外推模型和光流法相比，CSI提高0.05~0.15，POD提高0.05~0.15，FAR降低0.05~0.12。（3）在预测反射率阈值为35 dBZ的强对流单体移动路径方面，基于深度学习的回波外推模型与TITAN法相比，预测的单体移动路径会更接近实况单体移动路径。

A Case Analysis of Deep Learning Methods for Predicting Radar Echoes of Hail Clouds in Aksu Region

Using the radar echo data from the hail cloud in the Aksu region from April to September in 2021-2022, a echo extrapolation model of deep learning was jointly constructed based on the trajectory GRU model and generative adversarial networks (GAN) model. This model is applied to the monitoring and early warning of severe convective (hail) weather. By adopting the assessment method of different thresholds and forecast time effectiveness, the effect of the deep learning-based echo extrapolation model''s echo prediction was analyzed. The results show: (1) Within a 30-minute prediction period, as the reflectivity threshold increases, both the critical success index (CSI) and the probability of detection (POD) gradually decrease, while the false alarm rate (FAR) first decreases and then increases. The FAR is at its lowest when the reflectivity threshold is 35dBZ.When the reflectivity threshold is set to 35dBZ and the extrapolation time period is kept consistent, the deep learning-based echo extrapolation model shows improved performance compared to the optical flow method. The model can increase the CSI by 0.05 to 0.15, increase the POD by 0.05 to 0.15, and decrease the FAR by 0.05 to 0.12. (3) In predicting the movement paths of severe convective cells, the path predicted by the deep learning-based echo extrapolation model is closer to the actual movement path of the cells compared to the TITAN method.