Research On Severe Convection Nowcasting Based On Improved Prediction Recurrent Neural Network And Graph Convolutio

Severe convective weather has the characteristics of short life cycle,rapid effective evolution and strong destructive ability.Therefore,in order to reduce casualties and property losses,it is very important to accurately forecast the severe convection weather.The precipitation nowcasting provides strong support for the severe convective nowcasting.With the development of meteorological technology,the forecast of precipitation based on radar observation data has become a widespread concern.Although many achievements have been made in the research on the severe convective nowcasting,there are still challenges in the highprecision prediction of severe convective weather due to its low frequency and variable manifestations.The purpose of this paper is to improve the accuracy of the precipitation nowcasting by studying two deep learning methods using Doppler weather radar data.The main work of this paper is as follows:(1)Aiming at the problem that the existing methods lack the structure that can learn the global spatiotemporal correlation information from the input radar image when training the radar data,which makes it difficult to capture the characteristics of the deep level network,a strong current proximity prediction method based on the Self-Attention mechanism and predictive recurrent neural network is proposed.This method increases the depth of transition between adjacent states by using a novel cycle unit: the long and short memory network(Causal LSTM)with cascaded double memory structure.This paper proposes a combination structure of gradient highway structure and Self-Attention mechanism,which provides an alternative short route for gradient flow from output to input.This architecture works seamlessly with Causal LSTMs,enabling predictive recurrent neural networks to adaptively capture short-term and long-term correlations.This paper verifies that the model can effectively improve the prediction accuracy through comparative experiments.(2)In order to solve the problem of low accuracy and low definition of radar echo extrapolation strong echo,a method of strong convection prediction based on spatio-temporal graph convolutional networks is proposed.The spatio-temporal graph convolution block is formed by combining the graph convolution block with the time convolution block.Spatiotemporal graph convolution blocks can automatically learn deep network structures without any learned knowledge,thus capturing hidden spatial relationships.This model addresses the problem that the graph structure does not necessarily reflect the real dependencies and real relationships may be lost due to incompleteness in the graph structure.It avoids the resulting severe convective nowcasting errors.The model can effectively guarantee the accuracy of severe convective nowcasting results.Experiments have proved that this algorithm has significantly improved image accuracy compared with other algorithms,and it is more accurate in positioning strong echo areas.