Machine learning-based prediction of peak outflow of embankment dam break with data augmentation

This work proposes an intelligent framework to predict the peak outflow of embankment dam break. To address data scarcity, a Copula-based data augmentation method is applied, which generates synthetic samples that preserve the statistical characteristics and dependency of the original dataset. The extreme gradient boosting (XGBoost) model, optimized via grid search cross-validation, is adopted to predict the peak outflow, with 80% of the dataset for training and 20% for testing. Three metrics including the root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) are utilized to evaluate model effectiveness. Comparative analyses with light gradient boosting machine (LightGBM), adaptive boosting, and random forest demonstrate the superior accuracy of the present model. The results show that (1) data augmentation method effectively improves the data diversity while preserving the distribution and statistical characteristics; and (2) the XGBoost achieves the optimal accuracy (RMSE = 0.0064 m3/s, MAE = 0.0035 m3/s, R2 = 0.988) in the test set, outperforming other models. Finally, Shapley additive explanations provide both global and local explanations for the present model, offering impacts of each input variable on the prediction outcomes. The framework combines robust data augmentation, high-accuracy modeling, and transparent interpretability, which offers a reliable tool for dam break risk assessment.