TAFFNet: Time‐Aware Adaptive Feature Fusion Network for Very Short‐Term Precipitation Forecasts

• Published in: Geophysical research letters Vol. 50; no. 15
• Main Authors: Wang, Jingnan, Wang, Xiaodong, Guan, Jiping, Zhang, Lifeng, Zhou, Jie
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 16.08.2023; Wiley
• Subjects: Accuracy; Deep learning; Dissipation; Echoes; Emergency management; Extrapolation; Lead time; Mathematical models; Modelling; Nowcasting; Numerical prediction; Numerical weather forecasting; Physics; Precipitation; Precipitation forecasting; precipitation forecasts; Public safety; Radar; Radar data; Radar echoes; Safety management; Storms; Weather forecasting
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2023GL104370

Deep learning models based on radar echo extrapolation have been widely used in precipitation nowcasting. However, they face the challenge of insufficient input information when extending the forecast lead time, requiring the incorporation of physics‐based numerical weather prediction (NWP). Given that the strengths of radar and NWP data vary depending on the forecast time, effectively fusing these two data sources in a unified deep learning model remains an open research problem. In this study, we propose a Time‐aware Adaptive Feature Fusion Network (TAFFNet) for very short‐term precipitation forecasts up to 12 hr. TAFFNet fuses features adaptively according to their relative contributions to forecast skill at different times. Experimental results demonstrate that TAFFNet performs the best for very short‐term precipitation forecasts. The case studies show that adaptively fusing NWP with radar can improve the accuracy of precipitation forecasts, especially for predicting the initiation and dissipation of storms at longer lead times. Plain Language Summary Very short‐term precipitation forecasting aims to predict precipitation for a couple of hours, which is crucial for public safety and emergency management in cities. Recently, deep learning models have been widely used for precipitation nowcasting (0–2 hr) based on observation extrapolation. However, as the forecast time increases, extrapolating historical observations alone is inadequate without incorporating physical constraints. On the contrary, traditional NWP can provide essential physical background, but the prediction accuracy still needs improvement for early forecasting periods. In this study, we proposed a novel deep learning model named TAFFNet to combine both advantages of radar and NWP for very short‐term precipitation forecasts up to 12 hr. The features of radar and NWP are first extracted, respectively. Their features are fused adaptively in the forecasting part through a time‐aware adaptive feature fusion (TAFF) module, according to their relative contributions to forecast skill at different lead times. The feature fusion approach can be optimized adaptively through model training. The proposed model is evaluated on the radar and NWP model data set covering southeast China. Results demonstrate that incorporating NWP into a radar‐based deep learning model can improve very short‐term precipitation forecasts. TAFFNet can better capture the storm's initiation and dissipation during forecasting. Key Points A novel deep learning model termed Time‐aware Adaptive Feature Fusion Network (TAFFNet) is proposed for very short‐term precipitation forecasts up to 12 hr Features from radar and numerical weather prediction are fused adaptively in deep levels based on their relative contributions to forecast at different lead times TAFFNet can improve very short‐term precipitation forecasts, especially for predicting storms initiation and dissipation at longer times