Circulation Classification And Prediction Of Persistent Heavy Precipitation In Jianghuai Based On Residual Neural Network

Based on 1981-2019 historical case data of regional persistent heavy precipitation process,daily precipitation data of 2474 observational stations in China,NCEP/NCAR global reanalysis data and 2016-2019 ECMWF forecast data of precipitation and circulation,cases of persistent heavy rainfall in Jianghuai region are analyzed.Firstly,typical rain patterns and circulation fields are refined based on extreme persistent rainfall cases in Jianghuai region.Then single-layer transfer learning CNN,R and COS are used to classify the circulation of 1981-2015 cases of heavy rainfall,and compare the classification effects of different methods.The heavy rainfall classification model database are established and the circulation characteristics of heavy rainfall are investigated.Further,an improved three-layer transfer learning CNN classification model is established,and an classification and forecast experiment is carried out on the 2016-2019 summer precipitation.The conclusions are as follows:(1)The first three patterns of EOF decomposition show the trend of uniform change in the whole area,reverse change from north to south,and reverse change from central to north-south.According to the spatial distribution characteristics of regional persistent heavy precipitation,three types of typical daily samples were extracted for network training modeling.The precipitation centers of three typical patterns are located in the middle,northeast,northeast and southwest.In the training process of transfer learning CNN classification model,over-fitting is avoided by increasing the sample size,adding the pre-training model and setting up three migration training processes.Compared with R and COS,transfer learning CNN has a higher accuracy on the test set.(2)Using transfer learning CNN,R,and COS to classify the circulation of regional heavy rainfall.The persistent heavy rainfall model database is established and the typing effect of transfer learning CNN is better than that of R and COS.There are obvious differences in the500 h Pa circulation characteristics of three types heavy rainfall obtained by transfer learning CNN.The middle and high latitudes of Type I are two troughs and one ridge,and the subtropical height extends slightly to the west.The high latitudes of Type ? are two ridges and one trough,with a strong subtropical high,which can extend westward to southern China.The mid-latitude of Type ? has two ridges and one trough,and the subtropical height is weak.In each type,Jianghuai region is affected by low-value systems.All types are affected by high-altitude jets,with high-altitude divergence and low-altitude convergence,resulting in ascending motion.Water vapor conditions are all sufficient.The precipitation spatial distribution of each type obtained by three methods is all similar to typical patterns.While the result of transfer learning CNN is more in line with the characteristics of typical patterns,and the height field variance between different types and the precipitation correlation coefficients between each type and typical patterns are all greater than the results of R,COS.The analysis of samples with inconsistent types shows that only the three types precipitation distribution of transfer learning CNN can show characteristics of typical patterns,and its precipitation correlation coefficients is much higher than the results of R and COS.(3)Compared with the single-layer transfer learning CNN,the three-layer transfer learning CNN classification model established by adding the 200 h Pa zonal wind field and the850 h Pa meridional wind field has higher accuracy on the test set.The height field variance between different types and the precipitation correlation coefficient between each type and typical patterns is larger,and the classification effect is better.In the 25 mm and 50 mm or more precipitation forecast of 2016-2019 summer daily samples and samples of persistent heavy precipitation cases,compared with ECMWF,H-R(statistical forecast of R method based on single-layer transfer learning CNN precipitation database),H-R-EC(correction forecast of R method based on single-layer transfer learning CNN precipitation database and EC circulation field),HUV-R(statistical forecast of R method based on three-layer transfer learning CNN precipitation database),and HUV-R-EC(correction forecast of R method based on three-layer transfer learning CNN precipitation database and EC circulation field)change more slowly with time,and the results are not much different.The results of HUV-R and H-R-EC are slightly higher than those of H-R.In precipitation forecasts over 4 days,the threat scores of four methods are all higher than ECMWF,and the results of HUV-R-EC are slightly higher than HUV-R,but in short-term precipitation forecasts,the scores of HUV-R-EC are lower.