Research Of ERA5 Daily-scale Precipitation Prediction Method Based On LSTM Algorithm

Water is an indispensable part of human life,and one of the most important sources is precipitation.Accurate prediction of precipitation can not only effectively prevent floods,droughts,typhoons,mudslides or other natural disasters,but also forecast weather to make it more convenient for people to travel.With the increasing abundance of satellite remote sensing information,there are more and more areas where satellite precipitation inversion products are applied.How to use satellite precipitation products for precipitation prediction and thus meet the needs of meteorology,hydrology and other fields has become an important research direction.The impact of human activities and global climate change have led to an increasing frequency of extreme drought events in mainland China,so that how to use satellite precipitation products to forecast precipitation in mainland China has become an urgent problem in the field of hydrology and meteorology.Due to the vast geographical area of mainland China,the variable climate and the large altitude differences between the east and west,there are many meteorological factors affecting precipitation,and the amount of daily-scale precipitation data is large and the processing process is complex.Currently,precipitation forecasts in mainland China are mostly based on annual and monthly time scales,so that there is less research on precipitation prediction based on daily-scale satellite precipitation products.This thesis firstly assesses the accuracy of the latest global observation satellite precipitation data ERA5 in mainland China,compares and analyses the long time series(1979-2018)satellite precipitation product data and gridded measured data by accuracy assessment index,calculates the interannual and seasonal correlation coefficients between ERA5 precipitation products and measured precipitation,analyses the performance of precipitation data across China showing spatially intrinsic characteristics,study the accuracy of satellite precipitation products in different regions,combine the topographic distribution and climatic zones,and divide the Chinese mainland into three precipitation prediction research regions of good,better and average precipitation data accuracy.On this basis,long short-term memory network algorithms in deep learning suitable for processing long time series data are investigated,meteorological factors affecting precipitation prediction are analysed,and data-driven LSTM daily precipitation prediction models for different regions are developed respectively.Through the experiments,correlation coefficients of the precipitation prediction models for the three regions improved by 0.13,0.08 and0.09 respectively compared to the original data,and the LSTM precipitation prediction model for the Tibet Plateau region,which has the worst accuracy,improved the correlation by 2%-9%compared to the traditional time series prediction algorithm in all cases.The results show that the data-driven LSTM precipitation prediction model has good predictive power for ERA5 reanalysis daily precipitation data in mainland China,which is more suitable for long time series datasets than traditional deep learning algorithms dealing with time series.Moreover,this study provides a reliable data assessment for the application of ERA5 reanalysis precipitation data in hydrological simulation and precipitation trend analysis in China,which also provides the accuracy reference for further use of precipitation satellite data for hydrological calculations and numerical climate simulations.