利用深度学习预报美国东北部日降水分布

现阶段降水预报主要依靠数值天气预报模式。但受物理参数化、计算资源等因素的影响,基于数值模式的降水预报还存在非常大的不确定性。近年来,深度学习在天气预报领域显示出巨大优势和潜力。本文通过构建神经网络预报美国东北部日降水分布,探讨神经网络模型基于低分辨率气象场(ERA-Interim,0.7°)预报高分辨率降水(CPC,0.25°)的能力,并比较3种主流网络框架(VGG,ResNet,GoogleNet)在该任务中的表现。结果表明,3种网络框架都对美国东北部日降水分布具有一定的预报能力(VGG框架表现最优),但三者的均方根误差(RMSE)均高于ERA-Interim 24-h(ERA24)的降水预报。3种神经网络的集合预报结果优于ERA24预报,且这三者与ERA24预报结果的集合平均能够显著提高ERA24对不同季节、不同强度降水的预报。

Using deep learning to predict daily precipitation distribution in the northeastern United States

At present,precipitation forecasting mainly relies on numerical weather forecasting models.However,due to factors such as physical parameterization and computational resources,there remains significant uncertainty in precipitation forecasting based on numerical models.In recent years,deep learning has shown great advantages and potential in the field of weather forecasting.The present study constructs neural networks to predict daily precipitation distribution in the northeastern United States,to explore the capabilities of neural-network models in predicting high-resolution precipitation (CPC,0.25°) using low-resolution meteorological fields (ERA-Interim,0.7°).Next,the study compares the performance of three mainstream network frameworks (VGG,ResNet,and GoogleNet) in the aforementioned task.The results indicate that all three frameworks have certain capabilities for predicting the daily precipitation distribution in the northeastern United States,with VGG performing the best,but their root mean square error (RMSE) is higher than that of the ERA-Interim 24-hour (ERA24) prediction.The ensemble-mean results of the three neural networks are all superior to the ERA24 prediction,and combining these three with the ERA24 prediction results can significantly improve ERA24 prediction in different seasons and intensities.It is thus concluded that deep learning has great potential in improving the resolution and accuracy of precipitation prediction.