Abstract: With the development of urban construction, the frequency and spatial distribution of precipitation are constantly changing. The traditional design of rainfall patterns inference based on single or representative station is difficult to reflect the spatiotemporal comprehensive characteristics of rainfall processes, and cannot meet the requirements of accurately identifying the spatiotemporal distribution of rainfall. This article is based on 20 years of rainfall monitoring data（1999-2018） and takes different watersheds in the hilly and plain areas of Beijing as the research scope. Machine learning methods are used to extract typical rainfall features within each watershed range and analyze the rainfall characteristics of different watersheds. Overall, the spatiotemporal distribution characteristics of rainfall in plain areas are complex. Compared with rainfall in hilly areas, the spatiotemporal distribution characteristics of rainfall over different durations in plain areas vary greatly over time, and the rainfall characteristics are related to the underlying terrain. The research results are integrated and applied to the official query system of rainstorm design results in Beijing. It can provide basic technical support for the rational allocation planning and refined management of urban water resources, and flood control planning in Beijing.