<p>Urban flood simulations typically ignore rainfall spatial heterogeneity, leading to inaccurate risk assessments. In this study, we developed a rainfall spatial distribution model that generates spatially nonuniform rainfall based on the multilayer perceptron (MLP) neural network and adaptive moment estimation (Adam) optimizer. Rainfall scenarios considering rainfall spatial heterogeneity and patterns were subsequently designed by the proposed model and the Chicago approach. The impact of rainfall spatial heterogeneity and patterns on urban flooding was quantitatively analyzed through a hydrodynamic–hydrological model. Using the central urban area of Zhengzhou, China, as a case study, the results revealed that the rainfall spatial distribution model exhibited excellent performance, with a coefficient of determination (R²) of 0.968. Neglecting the spatial heterogeneity of rainfall led to a systematic underestimation of flooding in the study area, resulting in average reductions of 3.97% in the inundation volume and 2.77% in the inundation area. Furthermore, the underestimation of the inundation volume caused by neglecting rainfall spatial heterogeneity intensified with increasing peak coefficient. Specifically, when the rainfall peak coefficient increased from 0.2 to 0.8, the magnitude of underestimation increased by 6.85%. This study provides a new approach for considering the impacts of spatial variations in rainfall on flooding and offers an important reference for future urban flood warning, disaster prevention and mitigation efforts.</p>