序列结构的RNN模型在闪电预警中的应用

Application of Sequence Structure RNN Model in Lightning Early Warning

应用深度学习模型对强对流天气中发生的闪电事件进行预警。根据闪电发生概率与雷达回波反射率因子的相关性研究,使用具有时间序列特性的历史雷达回波数据集,结合基于循环神经网络(Recurrent Neural Network)改进的神经单元、网络结构与损失函数,通过前2 h的历史回波数据来预测未来1 h内的回波形态与移动轨迹,根据预测结果进行闪电预警。检验结果表明:经过训练后的模型可以有效的预测出后续回波的生消以及dBZ数值的分布变化,在30 min的时效内能够较好的预测40 dBZ以上区域的回波特征,对发生的闪电事件进行预警。将该方法与传统的光流外推算法相比,临界成功指数CSI和HSS预报评分具有20%左右的提升,为在强对流天气下后续闪电的预警工作提供了更为精确的指导。

A deep learning model is applied to early warning of lightning events in severe convection weather.According to the correlation study between lightning occurrence probability and radar echo reflectance factor, using historical radar echo data sets with time series characteristics, combined with improved neural units, network structures and loss functions based on the Recurrent Neural Network, The historical echo data of the previous 2 h is used to predict the echo shape and movement trajectory within the next 1 h, and lightning warning is performed based on the prediction results.The test results show that the trained model can effectively predict the generation and elimination of subsequent echoes and the changes in the distribution of dBZ values.It can better predict the echo characteristics in the area above 40 dBZ within a 30-minute time limit.Be alert.Compared with the traditional optical flow extrapolation algorithm, this method improves the critical success index CSI and HSS prediction scores by about 20%, which provides more accurate guidance for subsequent lightning warning work in strong convective weather.