Abstract: AI-based forecasting models for urban flooding face the challenge of selecting an appropriate sample size: an excessively large sample size may result in unnecessary computational costs, whereas an insufficient sample size may lead to overfitting. In this study, hydrological and hydrodynamic modeling was integrated with AI techniques, and four algorithms, namely ridge regression, k-nearest neighbors (KNN), random forest (RF), and backpropagation (BP) neural network, were employed. By identifying the performance change points, saturation points, and optimal sample sizes of the four models, the variation in forecasting performance with increasing sample size was investigated. The results showed that moderately sized datasets were sufficient for the models to achieve high and stable forecasting accuracy. Under the prescribed accuracy criteria (E_NS≥0.85 and E_RMS≤0.02), the optimal sample sizes for ridge regression, KNN, RF, and BPNN were 100, 90, 50, and 120, respectively, accounting for 59.5%, 53.6%, 29.8%, and 71.4% of the full dataset. Compared with the use of the full dataset, the dataset construction time at the optimal sample sizes was reduced by 5.00, 5.71, 8.61, and 3.39 h, respectively. Among the four models, RF exhibited the greatest potential for sample-size optimization. These results provide a basis for balancing dataset construction costs and forecasting accuracy in AI-based urban flood forecasting.