Extrapolation and effect analysis of weather radar echo sequence based on deep learning

Weather radar data is the main reference for nowcasting of severe convective weather. To address the problems of insufficient data utilization and limited extrapolation time in the radar echo extrapolation method widely used in China, the neural network is applied to radar echo extrapolation. The predictive neural network model gives 2 h prediction results of radar echo variation. The essence of radar echo extrapolation problem is the spatiotemporal sequence prediction problem. The network has long and short time memory unit (Long Short-Term Memory, LSTM) to solve the time memory problem and convolutional layers to extract spatial features. The training and testing datasets are constructed using radar data from Fujian, Jiangsu and Henan for several years. Considering the fact that the frequencies of different rainfall levels are highly imbalanced, the network is trained by weighted loss function to improve the prediction accuracy of strong echoes. The test set and individual case evaluation show that CSI (Critical Succes Index) and POD (Probability Of Detection) of predictive neural network are higher than that of optical flow method and FAR (False Alarm Ratio) is lower than that of optical flow method under the same extrapolation aging and reflectivity threshold. Among different precipitation types, the SSIM (structural similarity) value of the predictive neural network is higher than that of the optical flow method, and the SSIM value of the stratiform-cloud precipitation is larger than that of the convective-cloud precipitation. Therefore, the predictive neural network has a better ability to predict strong echoes than optical flow. In terms of the timeliness of forecasting, the predictive neural network model has certain advantages, and it is more accurate in forecasting stratiform-cloud precipitation than forecasting convective-cloud precipitation.