基于深度可分离组卷积的临近预报

Precipitation Nowcasting Based on Depthwise Separable Group Convolution

摘要: 强对流天气具有突发性、局地性等特征，它同时又可能造成各种重大损失。精准、实时的强对流天气预报对风暴预警、航空运行及大型集会组织等活动的规划均具有重要价值。临近降水预报作为强对流预报的一个重要工具，因此越来越受到气象预报专家与研究人员的高度关注。过去数十年间，科研人员主要通过求解大气运动方程组，借助数值天气预报开展临近降水预报工作。该方法虽取得一定成效，但庞大的计算资源需求与复杂的方程求解过程，导致其在临近预报中的应用面临较大困难。近年来，研究人员采用自注意力机制、扩散模型等深度学习方法，有效提升了预报性能，却普遍受困于模型尺寸过大、轻量化程度不足等问题，使得模型难以在预报性能与参数规模之间实现平衡。为在模型参数与预报性能之间实现一定平衡，本文提出一种基于深度可分离组卷积的编码器–转换器–解码器模型。该模型聚焦于强对流天气中强对流降水的及时预报，其内部通过多层深度可分离组卷积实现降水信息的多尺度特征提取，同时通过时空特征融合模块实现多层级特征的融合。SEVIR数据集上的实验结果表明我们提出的模型预报效果优于其他对比方法，量化实验与可视化实验结果进一步验证了该模型的有效性与实用性。

Abstract: Severe convective weather, characterized by its abrupt onset and high localization often leads to significant losses. Accurate and real-time forecasting of such events is therefore great importance for storm warnings, aviation operations, and the planning of large-scale gatherings. As a key tool for predicting severe convective events, nowcasting of precipitation has attracted increasing attention from meteorologists and researchers. Over the past several decades, most studies have relied on solving atmospheric dynamic equations through numerical weather prediction (NWP) methods. Although this approach has achieved notable progress, its application in short-term forecasting remains challenging due to the intensive computational demands and the complexity of solving these equations. In recent years, deep learning techniques, such as self-attention mechanisms and diffusion models, have been introduced to enhance forecasting performance. However, these models often suffer from excessive parameter sizes and insufficient lightweight design, making it difficult to balance predictive accuracy with computational efficiency. To address this issue, this study proposes an encoder-transformer-decoder architecture based on depthwise separable group convolution. The model focuses on timely forecasting of convective precipitation associated with severe convective weather. Within the architecture, multi-scale precipitation features are extracted through multiple layers of depthwise separable group convolutions, while a spatiotemporal fusion module integrates multi-level features. Experiments conducted on the SEVIR dataset demonstrate that the proposed model outperforms other baseline methods. Both quantitative and visualization results further confirm the model’s effectiveness and practical value.