基于雷达反射率因子和雷电定位数据的深度学习雷电预报模型

Research of Lightning Forecasting Based on Deep Learning Model with Radar Reflectivity Factors and Lightning Location Data

利用卷积神经网络和门控循环单元(Gated recurrent units)神经网络,基于雷达反射率因子和雷电定位数据开展了雷电预报研究。首先构建了引用注意力机制的基于卷积神经网络和门控循环单元神经网络的深度学习模型(Attention-ConvGRU);然后将雷达反射率因子数据和对应时间段(6 min)的雷电定位数据处理成图像数据后输入深度学习模型,训练出可预报雷电的模型,包括3种模型:单雷电数据模型、单雷达数据模型和雷电-雷达双数据模型;最后开展了预报试验和定量评估。综合评估表明,本文建立的雷电预报模型综合预报准确率达到96.74%,虚警率35.83%,关键成功指数(Critical Success Index,CSI)为0.2072。个例分析表明,预报模型对于具有明显移动趋势的雷暴过程(A类雷暴)的预报效果优于不具有明显移动趋势的雷暴过程(B类雷暴),且随着B类雷暴强度减弱模型预报能力逐渐减弱。

In this paper,the convolutional neural network and gated recurrent units neural network are used to conduct lightning forecasting research based on radar reflectivity factors and lightning location data.First,a deep learning model(Attention-ConvGRU) based on the convolutional neural network and gated recurrent unit neural network that introduces the attention mechanism is constructed.Then,the radar reflectivity factor data and the lightning location data of the corresponding period(6 minutes) are processed into image data,and input into the deep learning model to train the models that can predict lightning,including three models:single lightning data model,single radar data model and lightning-radar dual data model.Finally,forecasting experiment and quantitative evaluation are carried out.The comprehensive evaluation shows that the forecasting model has a comprehensive forecasting accuracy of 96.74%,a false alarm rate of 35.83%,and a Critical Success Index(CSI) of 0.2072.The case study shows that the forecasting model has better lightning forecasting skills for thunderstorms with obvious moving trends(type A thunderstorms) than those without obvious moving trends(type B thunderstorms),and the forecasting skill of the model gradually weakens as the intensity of type B thunderstorms weakens.