With global climate change and accelerating urbanization， urban waterlogging disasters have become increasingly frequent and severe， making rapid waterlogging forecasting a key research focus. Compared with traditional numerical simulation methods， deep-learning-based artificial intelligence （AI） models can significantly improve computational efficiency. However， they often encounter training bottlenecks due to limited GPU memory. To address this， this study proposes an efficient AI urban waterlogging forecasting model named as UNet-KAN-SR. This model first employs the UNet-KAN module to efficiently simulate the spatio-temporal evolution of waterlogging over low-resolution grids， and then leverages the SR （super-resolution） module， along with high-resolution surface information， to progressively map the low-resolution waterlogging distribution to high-resolution distribution. This spatiotemporal decoupling strategy can ensure simulation accuracy while substantially reducing the computational resources required for training AI models. Experimental results demonstrate that the UNet-KAN-SR model can simulate a 3-hour waterlogging distribution within 3 minutes， achieving a root mean square error （RMSE） of 9 cm and a probability of detection （POD） of 0.84， demonstrating high accuracy and computational efficiency. Further analysis reveals that the integration of the KAN module can significantly enhance the model’s capability to capture nonlinear flood dynamics when compared with common CNN modules， reducing RMSE by 10%. Furthermore， this study finds that incorporating high-resolution features， such as surface topography， building coverage ratio， and land use， can significantly improve the simulation performance but performance improvement is similar under different feature combinations. This indicates that by optimizing the combination of input features during AI model construction， training speed can be enhanced， modeling costs controlled， and efficient intelligent forecasting achieved.