Research On Precipitation Prediction Based On Monthly Rainfall Series And Radar Echo Data

As a common and important hydrological phenomenon in nature,rainfall is vital to the earth’s water resources cycle,life-sustaining activities,and industrial and agricultural production of human society.However,within the framework of global warming,the escalation of extreme precipitation and recurrent inundation-related calamities in recent times have posed considerable hazards and inflicted significant harm upon both the natural ecosystem and the broader human populace.Therefore,accurate forecasting of precipitation is of considerable importance for numerous facets of human society,including the enhancement of social productivity,safeguarding of the natural environment,optimization of regional water resource allocation,as well as the maintenance of human life and the protection of property.This thesis focuses on two aspects: medium and long-term rainfall forecasting,as well as precipitation nowcasting.Medium and long-term rainfall forecasting involves anticipating the cumulative precipitation levels for the subsequent month or longer by analyzing their value and trend,with the aim of comprehending variations in regional water resources.This practice provides essential data support to facilitate the optimal allocation of water resources.Precipitation nowcasting refers to the prediction of rainfall within 0 to 2 hours in the future.Accurate precipitation nowcasting can help minimize the damage caused by sudden extreme rainfall weather events.The main work and innovation of this thesis are as follows:(1)For medium and long-term rainfall forecasting,this thesis employs a dataset comprising monthly rainfall series and defines this research problem as a time series prediction task.Due to the strong nonlinear and non-stationarity of rainfall time series,in order to improve the accuracy of forecasting,a medium and long-term rainfall time series forecasting model based on the idea of "combined forecasting" is constructed in this thesis,the model consists of two parts: a sequence decomposition and reconstruction module based on CEEMDAN-SE and a sequence prediction module based on MLR-GRU.The experimental results show that the model can mine the sequence features well,improve accuracy,and restore local features effectively.(2)For precipitation nowcasting,this thesis takes the radar echo sequence as the research dataset and defines this research problem as the prediction problem of spatiotemporal sequences.Due to the large amount of spatiotemporal information in radar echo data,traditional radar extrapolation methods have several limitations,including insufficient utilization of data spatiotemporal correlation information and weak ability to capture spatial dependence.In order to further improve the prediction accuracy,this thesis constructs a precipitation nowcasting model based on self-attention and spatiotemporal LSTM.By introducing high-speed network units,it effectively solves the problem of gradient disappearance when spatiotemporal information is transmitted between deep networks.Experiments show that the model can effectively mine the spatiotemporal dependence between sequences,and more accurately predict the trajectory and spatial shape of the target object.To sum up,the research in this thesis covers the fields of medium and long-term rainfall forecasting and precipitation nowcasting,and a brand-new rainfall forecasting model is constructed in a targeted manner.The above two models can meet most important rainfall forecasting needs,and provide new exploration ideas and solutions for further improving the accuracy of rainfall forecasting and building an intelligent rainfall forecasting system,which have good feasibility and industrialization value.