Dynamic spatial-temporal precipitation distribution models for short-duration rainstorms in Shenzhen, China based on machine learning

Abstract

Abstract Short-duration rainstorm, the main cause of urban waterlogging and mountain torrents, is characterized by sudden, intense, and highly destructive rainfall. Understanding the dynamic temporal and spatial distribution patterns of short-duration rainstorm can help to predict their development processes. In this study, the dynamic temporal and spatial distribution models of various types of short-duration rainstorm events were established by using machine learning (ML) method based on the rainfall data of the recent decade in a Chinese coastal megacity, Shenzhen. The dynamic characteristics of these rainstorm events were extracted by using ML method in conjunction with the Locally Linear Embedding algorithm, which shows a potential capability to predict the developmental trend of a heavy rainstorm before it occurs. Based on the method put forward in the current study, characteristic rainfall process models consistent with the local temporal and spatial distribution characteristics of rainstorms can be designed, which is important to understand the risks of the rainstorms and consequently helpful for the assessment of urban flood insurance, the scientific design of drainage systems and the forecasting and warning of urban waterlogging.