Investigating temporal scour hole variations: A comparative study of hybrid CatBoost models and experimental data

The scour hole caused by 3-dimensional wall jets (3DWJ) is a serious problem downstream of dams and power stations. The current study utilized a newly created hybrid machine learning (ML) model to predict the changes in the dimensions of the primary scour hole over time caused by a 3DWJ. To fill the gap in the available knowledge, different experiments were done using uniform and non-uniform sediment. ML models and the linear regression were used to derive the prediction models. The results of the current study showed that the ML models have better accuracy than the linear regression model. Among all the developed hybrid ML models, the accuracy of the hybridized Categorical Boosting (CatBoost) with Gray Wolf Optimization algorithm (GWO-CB) yielded superior predictions. Sensitivity analysis confirmed the densimetric Froude number and the scouring time were important predictors. The effects of the expansion ratio on maximum scour depth and ridge height were the less important. However, the expansion ratio effects were larger than the effects of the tailwater depth ratio and sediment size ratio in the development of the scour hole in the streamwise and spanwise directions. The accuracy of GWO-CB models was considerably higher than the models previously applied in the literature. The proposed methodology revealed a robust and reliable model for predicting the scour hole dimensions.