Prediction of contraction channel scour depth: based on interpretability analysis and PCA-enhanced SVR

ABSTRACT Due to the numerous uncertain factors affecting contraction scour depth, although many traditional empirical formulas have been proposed in past research, their prediction accuracy is generally low. In recent years, with advancements in machine learning (ML) technology, these techniques have been able to accurately capture the nonlinear characteristics of scour-depth data. However, in pursuit of higher prediction accuracy, researchers have explored a wide range of diverse ML models that require various combinations of input parameters. These input parameter combinations often lack reliability, and the models themselves have poor interpretability, increasing the ‘black-box effect.’ Therefore, this study uses a principal component analysis (PCA)-enhanced support vector regression (SVR) model to construct a scour depth prediction model, combined with the interpretability method of SHapley Additive exPlanations (SHAP). The results show that the SVR model's predictions are highly consistent with physical experimental laws, and the model primarily identifies features that are strongly linearly correlated with the dependent variable (scour depth and SHAP values). The application of PCA enhances the correlation, and when using the CC-PCA-4 input parameter combination, the SVR model achieves high accuracy (R2 = 0.971, mean absolute percentage error = 7.54%). Moreover, its comprehensive evaluation in terms of stability, accuracy, and conservativeness surpassed that of other ML models and empirical formulas.