Research On Weather Radar Echo Extrapolation Algorithm Based On Deep Learnin

The extrapolation technology of weather radar echo is widely used for strong convective weather proximity forecasting.With the rapid development of computer hardware and software,artificial intelligence technology has provided a new direction for radar short proximity extrapolation,and scholars have proposed a variety of spatio-temporal sequence prediction models relying on convolutional neural networks and recurrent neural networks.However,these models have problems such as difficulty in effective extrapolation to strong echoes and low resolution of predicted images.Based on these problems,this thesis improves the convolutional and recurrent neural networks,respectively,and further optimizes them by generating adversarial networks to accomplish the following three aspects of research:(1)In this paper,an adaptive normalization method and a spatio-temporal self-attention LSTM model are proposed.When the radar echo training set contains a large proportion of low-echo samples,the traditional deep learning model tends to favor the prediction of low echoes,while the region of strong echoes will dissipate rapidly with the prediction.To address the problem of imbalance between strong and weak echo samples,the adaptive normalization method is used to preprocess the radar echo data,which effectively improves the prediction ability of the three types of spatio-temporal prediction models for strong echoes.Additionally,in order to make the neural network pay more attention to the growth and dissipation changes of the strong echo regions in the time series,two self-attention structures are jointly used to extract time and spatial feature information and embed them into the time series prediction model.Experimental results show that the time series prediction model with spatiotemporal self-attention structure more accurately captures the strong echo area and slows down the dissipation speed of echoes.Compared with the commonly used spatio-temporal prediction models,this model is more effective for long-term forecasting.(2)In this paper,a radar echo extrapolation algorithm combining differential and residual U-Net is proposed.The fully convolved U-Net model has a strong ability to extract spatial features,but lacks the ability to extract temporal features.And in the fully convolutional radar echo extrapolation model,the temporal dimension is often placed in the channel dimension.Therefore,in this paper,a difference operation on the channel dimension is added to the U-Net model to introduce time-varying features to the model,highlighting the variation of radar echoes on the time scale.Meanwhile,the ability of the model to extract spatial features of radar echo images is enhanced by deepening the number of convolution layers of the model.The residual structure is introduced to enable stable propagation of the gradient.And a portion of ordinary convolution is replaced by depth-separable convolution to release the arithmetic power,and finally a differential-residual U-shaped spatio-temporal prediction model is obtained.This model greatly reduces the GPU computing power required for training,has high training efficiency.Compared with Sma At-UNet,the CSI score of this model under the 45 d BZ threshold is increased by 3.6%,and the FAR score of this model is reduced by 1.2%.(3)In this paper,a GAN-based radar echo extrapolation optimization algorithm is proposed.To address the problem of low prediction resolution of spatio-temporal sequence prediction models in general,it is further optimized by combining GAN.Since the fully convolutional U-network possesses faster convergence speed and better prediction performance compared with the recurrent neural network,the U-network combining difference and residuals is used as the generator,and the discriminator of full convolution is redefined,a penalty module for feature extraction of real and predicted radar echo images is introduced,and the loss function is reset.The parameters of the spatio-temporal prediction module updated in reverse by means of the generator and discriminator playing with each other,thus improving the resolution of the radar echo images and making them more similar to the real situation.