基于物理信息的堰塞坝峰值流量预测模型OA

To address the issues of overfitting in small-sample peak flow prediction of natural landslide dam failures and the lack of physical mechanisms in purely data-driven models,a grey-box prediction model integrating fluid dynamics priors and heterogeneous ensemble learning is constructed.The study selects 52 natural landslide failure cases,extracting dimensionless features such as volume,energy and Froude number,and injecting the predicted values from classical empirical formulas as prior variables into the feature space as the prediction benchmark.A pipelined architecture is used for feature scaling and model training,combined with Leave-One-Out Cross-Validation(LOOCV)to thoroughly prevent data leakage.Finally,a stacking ensemble model with log residual space is constructed by fusing random forest,gradient boosting tree and support vector regression,using Bayesian ridge regression as the meta-learner.Independent testing results show that the model has a log space coefficient of determination of 0.894,effectively alleviating the black-box limitations under extremely small sample sizes,and providing a physically interpretable and highly accurate decision support tool for emergency early warning of landslide dam disasters.