0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transfor...0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transformer利用双级路由注意力机制定位特征中的几个最相关的键值对来提高计算效率。模型同时采用MaxPool和AvgPool提取目标对象的局部特征,并将局部特征和全局特征进行有效融合,进而充分提取目标对象的特征用于临近预报。提出的模型与几个模型在一个公共的临近预报数据集上进行实验验证。实验结果充分证明了我们提出的模型的有效性和准确性。The nowcast of 0~2 hour is an important toolth for forecasting severe convective weather. Accurate nowcast is an important barrier to prevent from weather disaster. Therefore, nowcast is a very important field in the meteorological society. This paper proposes a nowcast model which based on Encoder-Forecasting architecture and temporal Transformer. The transformer utilizes bilevel routing attention to improve the computational efficiency. At the same time, the proposed model also use the operations of Maxpool and AvgPool to extract the local features. The global features and local features are effectively fused to represent the objects’ features. The comprehensive experiments based on the proposed model and four state-of-art models are conducted on a public datasets. The experimental results effectively show the correctness and effectiveness of the proposed model.展开更多
文摘0~2小时的临近预报是预报强对流天气的重要手段,准确的临近预报是气象防灾的重要屏障。临近预报是气象预报的一个重要业务。本文提出了一个基于时序Transformer的全局特征提取的Encoder-Forecasting临近预报模型结构。该模型的Transformer利用双级路由注意力机制定位特征中的几个最相关的键值对来提高计算效率。模型同时采用MaxPool和AvgPool提取目标对象的局部特征,并将局部特征和全局特征进行有效融合,进而充分提取目标对象的特征用于临近预报。提出的模型与几个模型在一个公共的临近预报数据集上进行实验验证。实验结果充分证明了我们提出的模型的有效性和准确性。The nowcast of 0~2 hour is an important toolth for forecasting severe convective weather. Accurate nowcast is an important barrier to prevent from weather disaster. Therefore, nowcast is a very important field in the meteorological society. This paper proposes a nowcast model which based on Encoder-Forecasting architecture and temporal Transformer. The transformer utilizes bilevel routing attention to improve the computational efficiency. At the same time, the proposed model also use the operations of Maxpool and AvgPool to extract the local features. The global features and local features are effectively fused to represent the objects’ features. The comprehensive experiments based on the proposed model and four state-of-art models are conducted on a public datasets. The experimental results effectively show the correctness and effectiveness of the proposed model.
