Urban waterlogging disasters occurred frequently. It is of paramount importance to conduct accurate and efficient forecasting, early warning systems, and rehearsals for urban waterlogging prevention and control strategies along with urban flood drainage planning. However, the hydrodynamic model-based simulation of urban stormwater faces challenges such as low computational efficiency and extensive modeling data requirements that hinder the realization of "four pre-forecasts". In this paper, Xincheng River area of Yangzhou was selected as the research area, a prediction model for urban waterlogging based on a two-dimensional convolutional neural network driven by temporal and spatial data (rainfall and terrain) was established to realize time-by-time simulation of the global grid in the study area. The results demonstrate the model's exceptional spatio-temporal prediction performance. During the simulation of multiple rainstorm-waterlogging events, all spatial performance indicators, including the Kappa coefficient, exceed 0.80, with half surpassing 0.95. Moreover, the Nash efficiency coefficient for most water accumulation points in the study area ranges from 0.80 to 0.99 in terms of time series analysis of water depth. Compared with the physical process model, the training (calibration) and prediction efficiency are improved by 77.7 times and 285.2 times, respectively. These research findings serve as valuable technical references for real-time forecasting, immediate warning systems, and rapid predictions related to urban waterlogging.