基于分层生成对抗网络的短临降水预报方法研究OACSTPCD

This paper attempts to improve the accuracy of precipitation nowcasting by using generative adversarial network(GAN)method in deep learning.The hierarchical generative adversarial network(HGAN)is proposed to generate future radar echo sequences based on historical radar echo sequences.HGAN is composed of a global generator and a local discriminator.The global generator is constructed in a hierarchical structure of multiple subnets,and the model is trained using an upsampling process to capture the evolution trend of radar echoes,which is conducive to generating a clear future radar echo map.The local discriminator distinguishes the predicted radar echo map from the observed radar echo map based on the local area,and introduces a buffer mechanism to save the historical prediction se-quence so that the final prediction is more time-series compliant.Both are trained in an adversarial man-ner,and the resulting model is able to generate sufficiently clear and close to realistic future radar echo sequences for more accurate portrayal of echo intensity extremes and ranges.The test set verification and individual case analysis are performed for HGAN and GAN,and the experimental results verify the effectiveness of HGAN for radar echo prediction.Meanwhile,the critical success index and probability of detection of HGAN are higher than those of GAN,while the false alarm rate is lower than that of GAN under the same test reflectivity threshold,and the structural similarity index(SSIM)of HGAN is better than that of GAN under the same prediction duration.