Study On The Retrieval Of GNSS-PWV And Its Application Into Forecasting Rainfalls Based On Machine Learning

As one of the most active elements in the troposphere,water vapor is not only an important greenhouse gas causing climate change,but also a major basic material for the formation of weather events such as rainfalls.Accurate monitoring of atmospheric water vapor contents and its rapid change is of great significance for the study on the formation,evolution and extinction of major weather events.At present,the atmospheric Precipitable Water Vapor(PWV)retrieved from Global Navigation Satellite Systems(GNSS)is an important refence to measure the total water vapor content on the signal propagation path.Moreover,it has the advantages of high spatial and temporal resolution,high precision and all weather,which provides high-quality data support for the monitoring of water vapor content and its spatio-temporal variation rules during a rainfall event,as well as research on rainfall forecast methods.This thesis mainly studies the state-of-the-art methods and models of PWV derived from ground-based GNSS and its applications in rainfall forecast.First,constructing a high-quality model plays an important role in solving the problem of low-accuracy PWV due to missing critical meteorological data.Then,on the basis of PWV’s advantages of high quality and high spatial and temporal resolutions,response of PWV to the formation,evolution and extinction of rainfall events is investigated,the primary of which is to improve the in-depth understanding of weather events.Finally,the PWVbased methods for forecasting rainfall events are studied by combining multiple meteorological parameters and deep learning algorithm.The main research contents and conclusions of this thesis are as follows:(1)Retrieval of real-time PWV from GNSS requires high-precision Zenith Hydrostatic Delay(ZHD).However,most stations are not equipped with meteorological sensors,so high-precision real-time ZHD cannot be obtained,which affects the accuracy of real-time PWV.To solve the problem,the ratio between ZHD and Zenith Total Delay(ZTD)is investigated through the profiles from global sounding stations.It is found that the ratio has obvious variation characteristics of annual and semiannual periods.Then,based on this discovery,a global ZHD model is constructed by using error back propagation(BP)neural network.This model takes into account not only the spatial and temporal information of the stations,but also the real atmospheric information by taking ZTD as the input.Next,the mean bias and Root Mean Squared Error(RMSE)of the new model are 12.9 mm and 23.5 mm,respectively,over 137 global sounding stations,which are better than those derived from traditional methods.Moreover,the new model is compared against GPT3 by using ERA5 data from European Centre for Medium-Range Weather Forecasts(ECMWF)as the reference.The results show that the advantages of the new model are more obvious in the middle-and high-latitude regions.Finally,the benefits of the new model in realtime retrieval of PWV are verified on 41 GNSS stations worldwide.Compared with a traditional calculation method,the accuracy of real-time PWV based on the new model is improved by 21%.(2)In view of the difficulty for determining the optimal threshold in traditional threshold-based models for forecasting rainfalls,this thesis introduces the maximum True Skill Statistic(TSS)standard to solve the problem in this field,and proposes an improved threshold-based model.In the new model,PWV value,PWV increase and the rate of PWV increase are used as forecasting factors,and the maximum TSS standard is used to determine the optimal threshold for each of the three forecasting factors.Meanwhile,the optimal criterion for recognizing a rainfall event is determined based on the analysis of various strategies.The improved model is evaluated over 66 meteorological stations.The results show that the model can accurately forecast 87%of rainfall events,and its false alarm rate is about 53%,which is better than the other three-factor models.What’ more,real-time PWV is retrieved from GNSS by using ultra-fast orbit product provided by International GNSS Service(IGS)and ZHD model constructed in(1),for evaluating its effect on rainfall forecast.Over 66 meteorological stations,a real-time model for forecasting rainfall events is constructed and evaluated based on the real-time PWV.The real-time model can accurately forecast 84% of rainfall events,and its false alarm rate is 54%,which means that real-time PWV is able to be applied into rainfall forecast.(3)Threshold-based methods for forecasting rainfall events only considers the short-term variation of PWV,but does not take into account the multiple meteorological parameters.Therefore,this thesis proposes a method for forecasting rainfall events based on Gated Recurrent Units(GRU)that integrates short-term sequences of multiple meteorological parameters,and studies the effects of input parameters and their time series on the accuracy of the method.In terms of parameters,increasing input meteorological parameters can effectively improve the number of successfully forecasting rainfall events and reduce false alarm rate.And PWV derived from GNSS plays an important role in this method.In terms of time series,the optimal time series for a station is different from each other.This thesis determines the time series for a station according to the maximum TSS standard.Over 54 meteorological stations,the GRU-based method can accurately forecast 93% of rainfall events,and the false alarm rate is 45%,which is significantly better than the threshold-based method.(4)In this thesis,a 1°×1° region in the northeastern United States of America is taken as an example.The four-dimensional water vapor density in the region is obtained by using the tomography technology,and the spatio-temporal variation of water vapor during the occurrence,evolution and extinction of rainfall is analyzed.During the occurrence of a rainfall event,the water vapor density has obvious horizontal migration,and its direction can be detected.Based on the analysis,the Convolutional Neural Network(CNN)algorithm and the GRU algorithm are used to construct the M3D-GRU,M3D-CNN and M3D-CNN-GRU models for forecasting rainfall events,which all take the four-dimensional water vapor density as the main input data.It takes into account the spatial and temporal distribution of water vapor.In the verification data set,these three models can accurately forecast more than 80% of rainfall events,but with different false alarm rates.M3D-GRU model is considered as the best model by the maximum TSS standard,which is better than the model constructed in(2)and(3).In this these,aiming at the applications of GNSS water vapor into meteorology,the algorithms for retrieving real-time PWV from GNSS and its application into forecasting rainfall events are investigated in depth,including improving the accuracy of PWV,determining the relationship between PWV and rainfall events,and constructing the methods for forecasting rainfall events.It is believed that the research outputs are able to provide theoretical method supports and application references for short imminent warning of rainfall and other weather events.This thesis includes 50 figures,23 tables and 195 references.