A new forecasting method based on machine learning and numerical model for rainstorm event management

Abstract In the context of global warming, heavy precipitation is a frequent and impactful hazardous event. Timely identification of meteorological conditions associated with heavy rainfall is critical for urban flood risk management. This study proposes a machine learning–based approach for rapid detection of atmospheric fields favourable for heavy precipitation. The method involves three steps: (1) dimensionality reduction of national centre for environment prediction (NCEP) simulations (2015–2021) using local linear embedding to create a historical dataset capturing key features—humidity, wind speed, and temperature at 1000, 850, and 500 hPa; (2) applying the same reduction procedure to 2022 NCEP GFS outputs to form a test dataset; and (3) performing similarity comparisons between test and historical datasets to identify analogues, allowing inference of heavy rainfall likelihood from historical cases. Validation with 2022 ERA5 reanalysis over an island in the South China Sea (ISCS) (16°50′3″N 112°20′15″E) achieved 74.8% identification accuracy, 24.9% false alarms, and 33.3% missed alarms. It should be noted that heavy precipitation events account for only a small fraction of the annual samples (approximately 3%), resulting in a strong class imbalance that increases the difficulty of rare-event detection and partially explains the relatively high false alarm proportion. While the method demonstrates promising identification capability, the elevated false alarm rate indicates that further optimisation is required before operational implementation. A generalisation test over Shenzhen, a major coastal city in southern China, shows that the method performs well in both island and mainland environments, with higher accuracy in data-rich urban regions, highlighting its potential for heavy rainfall identification and disaster risk management.