Research On Radar Echo Extrapolation Algorithm Based On Spatiotemporal Convolutional LSTM And Adversarial Networ

The precipitation nowcasting aims to provide high spatiotemporal resolution forecasts of precipitation in local areas for the next 0-2 hours,with a particular emphasis on small to medium-scale weather systems that are characterized by strong suddenness,short life cycles,and rapid evolution.It is of great significance for severe convective weather warning and disaster prevention and mitigation.Radar echo extrapolation is one of the commonly used methods for precipitation nowcasting in meteorological applications.However,traditional radar echo extrapolation methods suffer from low data utilization and poor extrapolation accuracy.In recent years,the rise of deep learning technology has provided new ideas and methods for the prediction of near-term precipitation,thanks to its excellent advantages in nonlinear mapping,massive information extraction,and spatiotemporal modeling capabilities.A series of deep learning-based radar echo extrapolation models have been proposed.Although the predictive indicators have been improved,these models still have shortcomings such as poor accuracy of extrapolation results,poor timeliness,and blurry extrapolation images,which cannot meet the requirements of business needs.In order to further improve the accuracy and availability of radar echo extrapolation,we conduct research on radar echo extrapolation methods to address existing problems.The main innovative work is summarized as follows:(1)To address the issues of missing short-term dynamic features and insufficient retention of long-term features in the extrapolation of radar echoes,this study proposes using spatiotemporal convolution(3D convolution)instead of conventional convolution operations to enhance the model’s ability to perceive local dynamic information in radar echo data.This enables the model to fully learn the spatiotemporal characteristics of radar echo data,resulting in a more accurate understanding of the trends in radar echo evolution.In addition,a timechannel-based attention mechanism is designed to assign different weights to high-dimensional features at different time points,preserving important feature information extracted from the radar echo data and improving the accuracy and timeliness of radar echo extrapolation.(2)In response to the issue of radar echo images blurring and echo attenuation in in radar echo extrapolation,this study proposes a multi-scale spatial feature fusion module to replace the simple multi-layer convolution stacking feature extraction method.Furthermore,the temporal modeling capability of recurrent neural networks is integrated into the generative adversarial network,this paper constructs a generator network with convolutional long shortterm memory as the recurrent unit to enhance the model’s spatiotemporal information representation capability.In addition,a weighted loss function based on radar echo intensity is designed to address the distribution characteristics of radar echo intensity,improving the global averaging strategy of the mean squared error loss function.This allows the model to better learn the radar echo details in the stronger echo regions,which helps to alleviate the echo attenuation problem and generate clearer echo images,further improving the usability of the extrapolation results.