Edinburgh Napier University

  In recent years, the phenomenon of severe convective disaster weather has increased significantly in the world. Due to the characteristics of small spatial scale, short occurrence period, great destructiveness and drastic changes, severe convective disaster weather shows a high degree of randomness and uncertainty in its generation and extinction process. Machine learning is advancing rapidly, and made great strides in weather radar echoes extrapolation, but the existing radar echo extrapolation methods are mainly based on the finite observation dynamics of the flawed echo motion extrapolation paradigm, disregarding the echo sequence itself has a more intricate evolutionary process. High intensity echo region is easy to lose details in the convolution process, and the generated image is often fuzzy and distorted, resulting in low prediction accuracy. Consequently, we believe that the task of radar echo extrapolation is one of spatiotemporal series prediction. Based on the GAN model, which can generate high-resolution graphics, we propose a GL-GAN model structure, which is made up of the generator and discriminator, two competitive learning systems. Its generator is GL-PredNet, which is based on stack generation structure, and its discriminator is made up of one fully linked layer, one pooling layer, and five convolution layers. The loss function is optimized and the gradient difference error GDE is proposed. This improves the model's capacity to record and process high-intensity radar echoes' information, the fuzzy distortion of radar echoes is improved, and the model's performance in future frame prediction is improved. The findings demonstrate that, on average, the suggested model outperforms current deep learning models.