临近预报旨在预测未来0~2小时的天气情况,这对于减轻强对流天气的影响并保障人们的日常活动至关重要。作为当前天气预报中最具挑战性的任务之一,临近预报需要同时具备高精度和高时效性。为了应对这一挑战,本文提出了一种基于Transformer的编码器–解码器临近预报模型结构。该模型使用基于交叉型窗口的自注意力机制,从而高效捕捉天气雷达回波的全局特征;同时模型结合卷积块和高效的多尺度注意力机制,实现对多尺度局部特征的有效提取。模型通过引入门控机制有效地融合全局与局部特征,从而进一步提升模型预报性能。提出的模型在公开的上海雷达回波数据集上的实验结果有效验证了它的有效性和实用性。Nowcasting aims to predict weather condition within the next 0~2 hours, which is critical for mitigating the impacts of severe convective weather and ensuring the safety of daily activities. As one of the most challenging tasks in modern weather forecasting, nowcasting demands both high accuracy and timeliness. To alleviate this challenge, this paper proposes a Transformer based on encoder-decoder architecture for nowcasting. The proposed model employs a cross-shaped window self-attention mechanism to efficiently capture the global features of radar echo maps. Additionally, it uses convolutional blocks and a multi-scale attention mechanism to effectively extract multi-scale local features. By incorporating a gating mechanism, the model effectively fuses global and local features, further enhancing forecasting performance. Experimental results demonstrate that the proposed model achieves superior performance on the public Shanghai radar echo dataset.
