Short-term Precipitation Forecast Correction in Beijing-Tianjin-Hebei Region Based on Deep Learning

• Published in: Ying yong qi xiang xue bao = Quarterly journal of applied meteorology Vol. 36; no. 3; pp. 257 - 267
• Main Authors: Wang Yuhong, Kong Dexuan, Zhu Shoupeng, Zhang Nan, Ji Yan, Zhang Shan, Xu Huan
• Format: Journal Article
• Language: Chinese; English
• Published: Editorial Office of Journal of Applied Meteorological Science 01.05.2025
• Subjects: deep learning; forecast correction; influence factors; loss function
• ISSN: 1001-7313
• DOI: 10.11898/1001-7313.20250301

To enhance the quality of short-term precipitation forecast in Beijing-Tianjin-Hebei Region, a U-Net deep learning model is employed to correct short-term precipitation forecasts of 3-12 h based on the forecasting data of INCA (Integrated Nowcasting through Comprehensive Analysis) System and observation data of automatic weather stations during the flood season from 2022 to 2024. The model employs a weighted combination of threat score (TS) and mean square error (MSE) as its loss function. The weight assigned to TS directly influences the model's performance. Through independent experiments, weighting coefficients assigned to TS within loss function are optimized for hourly correction models with lead times ranging from 3 to 12 h, with the following results: 1.2, 1.0, 1.0, 1.0, 1.2, 1.2, 1.2, 1.2, 1.1, and 1.1, respectively. Comparative experiments are conducted to explore the impact of various influencing factors, including temporal and spatial characteristics of precipitation, as well as the thermal and dynamic conditions that affect short-term heavy precipitation. Results indicate that incorporating temporal and spatial characteristics of precipitation, as well as thermal and dynamic circulation conditions influencing precipitation, into the deep learning correction model significantly enhances its correction capability. A single-factor deep learning model only improves the forecast skill for precipitation rates of 5 mm·h-1 and above, while multi-factor deep learning model enhances the forecast skill for precipitation rates of 0.1, 5, 10, and 20 mm·h-1 and above. Multi-factor deep learning model significantly enhances the 3-12-h forecasting capabilities of INCA, with varying degrees of improvement based on precipitation intensity. TS can be increased by up to 0.07, 0.06, 0.06, and 0.03 compared to INCA forecast when the precipitation intensity is at least 0.1 mm·h-1, 5 mm·h-1, 10 mm·h-1, 20 mm·h-1. Multi-factor deep learning model enhances the forecasting accuracy of most stations in Beijing-Tianjin-Hebei Region without increasing precipitation forecast errors. The spatial extent of improvement in TS varies with the intensity of precipitation. For precipitation intensities of at least 0.1 mm·h-1 and 5 mm·h-1, TS can be improved in most areas of Beijing-Tianjin-Hebei Region. For precipitation intensities of at least 10 mm·h-1 and 20 mm·h-1, TS can be improved in the eastern part of Beijing-Tianjin-Hebei Region.