Experimental study on convective precipitation nowcasting based on crevnet deep learning model combined with causal-lstm unit
[Objective]Mesoscale convective precipitation prediction is one of the key and difficult objects of weather forecasting.Precipitation data detected by weather radar has high spatiotemporal resolution,which is the main data source for nowcasting in 0～2 hours.It is of great significance for high-resolution rainfall forecasting in small and medium-sized watersheds.The use of ra-dar data to carry out nowcasting of convective precipitation can facilitate peoples travel,agricultural production guidance,disas-ter prevention and mitigation,and has practical application value in the field of meteorology and hydrology.[Methods]Based on the volumetric scan data of the new generation S-band Doppler radar in Guangzhou,this study will explore the prediction perform-ance of the conditional reversible network CrevNet based on the Causal-LSTM memory module in convective precipitation nowcast-ing,and then compare the prediction effect with the model based on ordinary ST-LSTM to verify its superiority.To improve the memory ability of strong echoes,the model was trained with weighted Huber loss function.In this study,the CSI(Critical Suc-cess Index),POD(Probability of Detection,or Hit Ratio)and FAR(False Alarm Rate)were used to evaluate the result of the test dataset under different prediction time and test thresholds,and PSNR(Peak signal-to-noise ratio),SSIM(image structure similarity)and BIAS(bias scores)were used to test the predictive ability of the convective event.[Results]The result show that the CrevNet model based on Causal-LSTM memory unit has higher CSI and POD,and lower FAR during the forecast period.In the prediction of two convection cases,the model has a higher PSNR,SSIM and a BIAS closer to 1 under multiple prediction timeliness.[Conclusion]The CrevNet deep learning model of conditional reversible network has a strong ability to extract spatio-temporal features for spatiotemporal sequences,and the prediction effect will be different with different convolutional recurrent neural units.Therefore,the CrevNet model based on the Causal-LSTM memory unit can better preserve the convective echo mor-phology and is more suitable for convective precipitation nowcasting.
convection nowcastingdeep learningCrevNetHuber loss functionradar reflectivityprecipitationclimate changedisaster prevention and reduction
国家科技期刊平台
NSTL
万方数据
