Research On Weather Nowcasting Based On Deep Learning

Climate change affects the production of society and people’s daily life.As part of the information industry,accurate and timely weather forecasting is an important prerequisite for ensuring the normal operation of all walks of life in society and the safety of human life and property.Radar echo images,as important observation data,are a hot research topic in the field of meteorology to accurately and timely predict short-term weather using echo images.Methods based on deep learning have become one of the mainstream methods for radar echo extrapolation.However,due to the complexity of the spatiotemporal relationships between radar echo sequences,it is still very challenging to use deep learning methods for accurate and efficient radar echo extrapolation.In order to fully utilize the spatiotemporal feature information between radar echo sequences,we propose two improved deep learning echo extrapolation methods,which improve the accuracy of short-term weather forecasting and the efficiency of model extrapolation,and also develop a short-term weather forecasting system based on radar echo extrapolation.The work presented in this article can be summarized into the following three aspects:(1)To address the issue of inaccurate prediction of convolutional recurrent neural networks in high and medium echo intensity,as well as the severe class imbalance problem in radar echo sequence data,we propose the MMA-Net model for radar echo extrapolation.Specifically,in order to capture the long-term temporal dependencies of sequence images and have better spatial feature representation,a multi-scale LSTM is constructed by introducing an auxiliary branch,and then a novel cross attention block(CAB)is designed to better integrate global and local features.Finally,we proposed a radar echo extrapolation method based on multi-scale mixed attention LSTM in this paper.Experiments on the CIKM Analyti-Cup 2017 dataset show excellent performance,with HSS metrics achieved at 71.83%,51.57%,and 23.53% for 5d Bz,20 d Bz,and 40 d Bz,and CSI metrics achieved at77.45%,42.77%,and 13.45% respectively.(2)To address the issues of slow extrapolation speed due to the difficulty of parallelizing LSTM units and the complex spatiotemporal relationships between radar echo sequences,inspired by U-Net,we construct a fully convolutional network with an encoder-decoder structure,and then a non-local module was introduced to supplement global information for shallow networks to better capture long-term dependencies and establish global information correlation.To address the problem of low prediction accuracy caused by strong spatiotemporal dependencies and dense distribution of echo sequence data,we design a spatiotemporal feature extraction module(STFE),which focuses on medium and high echo intensity information using a sparse network structure and a dual-attention mechanism.Finally,a non-local neural network model called QFA-Net based on the dual-attention mechanism was proposed.Experimental results show that the computational complexity of the model is less than that of a convolutional recurrent neural network,and the accuracy is higher than that of a convolutional neural network structure,achieving an excellent balance between accuracy and speed.(3)A short-term meteorological forecasting system based on radar echo extrapolation has been developed.The system is built using the C# programming language and.NET framework,and integrated with two core technologies of deep learning-based radar echo extrapolation methods proposed in this paper.It provides users with functions including data input,radar echo extrapolation,visualization of prediction results,and result saving,which can assist forecasters in making short-term forecasts of radar echo images more conveniently and visually observe extrapolation results.