Quantitative Precipitation Nowcasting Based On Deep Learning And Weather Radar Data

With the rapid development of Air-Space-Ground integrated weather monitoring technology and the continuous innovation of artificial intelligence technology,new opportunities have been brought to improve the accuracy and timeliness of precipitation nowcasting.It is of great scientific value and practical significance for promoting the application of meteorological radar and improving the accuracy of adjacent precipitation forecast to study the deep learning precipitation near forecast method based on radar detection,evaluate the influence of different methods on the accuracy of echo prediction,quantitatively estimate the accuracy of echo forecast and quantitatively forecast the approaching precipitation.Based on the corrected radar echo data,this paper studies the application of Deep Learning Models in nowcasting precipitation.Constructed Convolutional Gated Recurrent Unit(ConvGRU),Convolutional Long Short-Term Memory(ConvLSTM),Long Short-Term Memory(LSTM),and Optical flow(Of)two traditional nowcasting prediction models,using radar and rain gauge data in Xining area to compare the echo prediction accuracy of different models was compared and analyzed.The effect of single-layer and multi-layer radar echo sampling on the prediction accuracy is comparatively studied.Combined with the observation data of the ground rain gauge,the quantitative estimation of the site radar precipitation is realized.The main conclusions of this paper are as follows:(1)Combining the multi-dimensionality of radar data and the self-learning ability of Neural Networks,.ConvGRU and ConvLSTM precipitation prediction models based on Deep Learning have better structural performance.Compared with Optical flow and LSTM model,the prediction accuracy is higher,which can effectively extend the extrapolation time.(2)The prediction result of ConvGRU at 1.5-2h is more stable than that of ConvLSTM.In the prediction time of 12min-2h,the correlation coefficient of radar echoes predicted by ConvGRU decreased by 16.3%(single-layer)and 12.6%(multilayer),and ConvLSTM decreased by 29%(single-layer)and 24.1%(multi-layer).(3)The single-layer data at 3000 m height has a better prediction effect than the sampled data in 3000m-5500 m height range.Compared with the single layer data at3000 m height,the false positive rate of ConvGRU,ConvLSTM,LSTM and Optical flow method is 9.92%,14.18%,11.39% and 6.17%,respectively.Because the peak value of radar echo will be greatly affected after multi-layer sampling of data,it is not suitable to be used in practical service forecast.(4)Setting different thresholds has influence on the evaluation of forecast results.Evaluation index strength of more than 10 d Bz prediction score of the radar echo of the highest,with the increase of the threshold value,the score to reduce gradually,to more than 30 d Bz strong echo,minimum skill score,should not be applied in the prediction of radar reflectivity factor higher threshold.(5)ConvGRU and ConvLSTM have high accuracy in the estimation of rainfall intensity at a single point.By comparing the measured rain intensity with the estimated value,the correlation coefficients of ConvGRU and ConvLSTM are 0.80 and 0.74,respectively,and the correlation coefficients of ConvLSTM and Optical flow are 0.53 and 0.42.The former has a small overall dispersion and high stability,and has a certain ability to estimate the rainfall intensity of a single point.The correlation of the latter is poor,and the degree of underestimation is obviously not good enough to predict the rainfall intensity of a single point.The deep learning model based on radar echo data with high temporal and spatial resolution has significant advantages in the application of precipitation nowcasting,such as high prediction accuracy,fast response speed,and effective extension of extrapolation time,and has great potential for further improving the prediction accuracy.Compared with the numerical weather forecast system based on long time practice and business experience,it still needs further research and development in all aspects.At present,the application of Artificial Intelligence in the near nowcast is still in the primary stage.Combining with Numerical Weather Prediction,improving the accuracy of nowcast and enhancing the interpretability in the evolution,development and dispersion of the weather process will be the development trend of the future nowcast service.