Scholarlink

<p>Accurate rainfall nowcasting plays a vital role in flood control, agricultural planning, and climate research. However, the inherently nonlinear, highly variable, and spatiotemporally complex nature of precipitation poses considerable challenges for achieving reliable short-term forecasts. Conventional deep learning approaches, such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs), often fail to capture the intricate temporal and spatial dependencies in precipitation data, particularly in regions with complex topography. To address these limitations, this study proposes AE-ConvGRU, an enhanced ConvGRU-based architecture that integrates a multi-head attention mechanism with feature enhancement strategies for early fusion of multimodal meteorological data, specifically Precipitable Water Vapor (PWV) and rainfall observations. The model adaptively assigns dynamic weights to highlight key spatial features and improve the representational capacity of fused data. Using ERA5 reanalysis datasets from 2019 to 2024 over Zhejiang Province, China, the AE-ConvGRU model was trained on multi-year data and evaluated on independent 2024 data. Six standard metrics, including the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination R<sup>2</sup>, were used for performance assessment. Experimental results demonstrate that AE-ConvGRU achieved a one-hour rainfall prediction RMSE of 0.339&#xa0;mm, representing a 27.10%, 24.16%, 13.96%, and 5.04% improvement over the CNN, GRU, XGBoost, and ConvGRU models, respectively. Moreover, the proposed method achieved average improvements of 23.11% for MAE and 72.18% for R<sup>2</sup> compared with four traditional methods. Additionally, AE-ConvGRU exhibited robustness in capturing rainfall patterns under both heavy and light precipitation scenarios, demonstrating superior ability to model spatial distribution and temporal evolution. These findings confirm the effectiveness of attention-guided multimodal fusion and highlight the potential of AE-ConvGRU as a powerful tool for accurate and stable short-term rainfall nowcasting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A new rainfall nowcasting method by introducing the attention mechanism into the ConvGRU deep learning model

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Mingkun Su,
  • Wenjie Yuan,
  • Zhihao Lai,
  • Cong Chen,
  • Junyao Du,
  • Yourui Fan,
  • Wenqin Deng

摘要

Accurate rainfall nowcasting plays a vital role in flood control, agricultural planning, and climate research. However, the inherently nonlinear, highly variable, and spatiotemporally complex nature of precipitation poses considerable challenges for achieving reliable short-term forecasts. Conventional deep learning approaches, such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs), often fail to capture the intricate temporal and spatial dependencies in precipitation data, particularly in regions with complex topography. To address these limitations, this study proposes AE-ConvGRU, an enhanced ConvGRU-based architecture that integrates a multi-head attention mechanism with feature enhancement strategies for early fusion of multimodal meteorological data, specifically Precipitable Water Vapor (PWV) and rainfall observations. The model adaptively assigns dynamic weights to highlight key spatial features and improve the representational capacity of fused data. Using ERA5 reanalysis datasets from 2019 to 2024 over Zhejiang Province, China, the AE-ConvGRU model was trained on multi-year data and evaluated on independent 2024 data. Six standard metrics, including the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination R2, were used for performance assessment. Experimental results demonstrate that AE-ConvGRU achieved a one-hour rainfall prediction RMSE of 0.339 mm, representing a 27.10%, 24.16%, 13.96%, and 5.04% improvement over the CNN, GRU, XGBoost, and ConvGRU models, respectively. Moreover, the proposed method achieved average improvements of 23.11% for MAE and 72.18% for R2 compared with four traditional methods. Additionally, AE-ConvGRU exhibited robustness in capturing rainfall patterns under both heavy and light precipitation scenarios, demonstrating superior ability to model spatial distribution and temporal evolution. These findings confirm the effectiveness of attention-guided multimodal fusion and highlight the potential of AE-ConvGRU as a powerful tool for accurate and stable short-term rainfall nowcasting.