Abstract: Extreme rainstorm is characterized by short duration, strong downpours and widespread destruction-all major causes of flash flood and waterlogging disasters．Exploring spatiotemporal dynamics of extreme rainstorm is helpful to refined management level of urban waterlogging risk．A method of clustering combined with dimensionality reduction algorithm was proposed in this study, to extract spatiotemporal dynamics of extreme rainstorm events．Precipitation data from 308 meteorological stations in Beijing from 2004 to 2016 were used, to construct spatiotemporal dynamics models of various extreme rainstorm events with the HDBSCAN clustering method．This was combined with UMAP dimensionality reduction algorithm to extract 4 types of extreme rainstorm models in Beijing．In mode 1, rainstorm was found concentrated in the main urban area, moving slowly around．In mode 2, rainstorm was found to move from the southwest mountainous area through the main urban area to the northeast．In mode 3, the rainstorm was found to spread from the western mountainous area to the northwest of the city, and to finally move to the north of the city．In mode 4, torrential rain was found to move westward from western mountain area to urban area．These results show that the main type of extreme rainstorm in Beijing is short-duration precipitation, and the rainstorm center has the tendency to move from west to east．Downtown area, southeastern mountainous area and southern plain area all face relatively higher extreme precipitation risks．The physical mechanism of model characteristics of all kinds of extreme rainfall is clear, and the reconstruction characteristics of rainfall events could fully represent actual rainfall events．This study provides important basis for flood forecasting and scientific management in Beijing．