Research On Short-term Precipitation Forecasting Method Based On Deep Learnin

With rapid developments in meteorological observation and forecasting technology in China,including the widespread application of ground-based automatic weather stations and meteorological radars has provided significant support for monitoring,warning,dispatching,and disaster prevention and mitigation of extreme weather,climate,and environmental events.However,due to the influence of factors such as climate conditions,topography,and land-sea location,precipitation has strong randomness and uncertainty,making it difficult for traditional manual or machine learning methods to directly extract deep features,thus limiting the accuracy of precipitation forecasting.Therefore,to improve the performance of precipitation forecasting,it is necessary to integrate observation data from multiple sources,effectively extract and mine spatiotemporal features related to precipitation,and discover the nonlinear relationship between multiple source features and precipitation intensity.In this paper,deep learning technology is used to conduct research on short-term precipitation prediction during the summer in Jiangsu province.The main research contents are as follows:(1)To address the issue of learning complex spatiotemporal features of precipitation through observational data,a short-term precipitation prediction model based on improved spatiotemporal graph convolutional network is proposed.The spatiotemporal graph convolutional network is used to learn the nonlinear mapping relationships between the current station and its surrounding stations,as well as between the current precipitation and past precipitation.Firstly,the concept of ensemble empirical mode decomposition is introduced to analyze the complex spatiotemporal relationships in precipitation data.Then,considering of the types of observational data from automatic stations and the topological structure of their distribution,the traditional graph convolutional network is improved by introducing the wind direction in the past time and the geographic location of the station to guide the model learning,thus capturing the spatial correlation of the topological structure.Additionally,the temporal memory flow module and attention module are used to fuse low-level and high-level features,thus prioritizing the target feature.Experimental results demonstrate that the model is consistent with the physical characteristics of precipitation and achieves better performance in precipitation prediction tasks.(2)To address the problem of learning precipitation trends from radar echo sequences,a precipitation prediction model based on video Transformer and sparse attention is proposed.The 3D convolution-based encoding module is used to extract the global context information and time information of the sequence.Meanwhile,the multi-head spatiotemporal fusion sparse attention module is used to learn the spatiotemporal relationships between different regions in the image sequence,and reduce the computational complexity.In addition,a non-recurrent output decoder module is used to output the results of all time steps in one step,overcoming the problem of long-term dependence in traditional recurrent neural networks.Experimental results show that the model has good performance in both spatiotemporal sequence prediction tasks and short-term precipitation prediction tasks.(3)To address the demand of effectively integrate information from multiple meteorological data sources,a precipitation prediction model using multimodal data fusion based on cross-attention mechanism is proposed.The Inception structure,an encoder module based on video Transformer and an encoder module based on spatiotemporal graph convolutional network to extract spatiotemporal sequence features from heterogeneous data sources.The cross-modal feature fusion module is used to exchange and align features from two modalities to learn joint cross-modal representations and construct internal connections of modal features.Experimental results demonstrate that compared to models with single-data inputs,the proposed model successfully improves prediction performance,thereby verifying the effectiveness of the multi-modal fusion strategy in short-term precipitation prediction tasks.