Abstract: Radar echo extrapolation is an important tool for short-term precipitation forecasting.The methods used are categorized into numerical and data-driven forecasts.The former relies on mathematical and physical models,while the latter relies on deep learning techniques to summarize historical patterns.Despite the active research in deep learning of short-term precipitation forecasting,practical applications still face challenges,especially the accuracy problem.Therefore,a radar echo extrapolation system was designed by applying the proposed  Residual Convolutional Long Short Term Memory(RC-LSTM)model.Residual connections to each layer of stacked ConvLSTM cells had been added while using planning sampling,which made the model deeper while retaining the learning capability of the original small model,ensuring that the network minimizes the loss of historical information in the spatial dimension,thus improving the accuracy of long-time radar echo extrapolation.

Key words: meteorological business, precipitation now casting, residual convolutional long short term memory, radar echo extrapolation

摘要： 雷达回波外推是降水临近预报的重要手段,所用的方法分为数值预报与数据驱动预报。前者依托数学与物理模型,后者依托深度学习技术总结历史规律。尽管深度学习在气象预报中研究活跃,但实际应用仍面临挑战,尤其是精度问题。因此设计了一种雷达回波外推系统,应用所提出的残参卷积长短期记忆(RC-LSTM)模型,在使用规划采样的同时,为每一层堆叠的ConvLSTM单元添加了残差连接,使得模型在更深的同时保留原有小模型的学习能力,保证网络最大限度降低空间维度的历史信息损耗,从而提高长时效雷达回波外推的精度。

关键词: 气象业务, 降水临近预报, 残参卷积长短期记忆, 雷达回波外推