In order to accurately predict the peak discharge of earth-rock dam failure, a database containing 156 cases of earth-rock dam failure was established. Based on the failure process and correlation analysis, the dam type, failure mode, the height of water above the breach, and the volume of water stored above the breach were selected as control variables. Considering the constraints of limited available data and the difficulty in obtaining dam failure data, a conditional tabular generative adversarial network (CTGAN) was used to augment the earth-rock dam failure data, so as to increase the diversity of samples, enrich the model training information, and improve the generalization ability. Based on the augmented data, a peak discharge prediction model was constructed using the CatBoost algorithm. The results show that the model based on the augmented data achieves higher prediction accuracy, with a coefficient of determination reaching 0.93, demonstrating excellent prediction performance. Combined with the shapley additive explanation (SHAP) analysis, it is found that the volume of water stored above the breach has the most significant impact on the peak discharge, followed by the height of water above the breach.