Prediction of local scour depth in bridge piers: Physical information and machine learning based modeling

Local scour is one of the main reasons for bridge collapse. To solve the difficult problem of detecting the local scour depth of underwater pier structures, this paper explores an optimal method for predicting the local scour depth of underwater pier structures based on various ensemble learning methods. Firstly, this paper collects 487 sets of data samples containing nine input parameters with corresponding scour depths from the open-source database in the practical project. Secondly, this paper employs five algorithms commonly used in ensemble learning, that is, Random Forest (RF), Gradient Boosted Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), and Light Gradient Boosting Machine (LightGBM), to build a prediction model of the local scour depth. In addition, the Bayesian hyperparameter optimization method is applied to search for the best hyperparameter combination of the model. Then, eight evaluation indices, including Mean Absolute Error (MAE), Mean Bias Error (MBE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), coefficient of determination (R 2 ), Nash-Sutcliffe Efficiency (NSE), Percent Bias (Pbias), and Willmott Index (WI), were used to compare and analyse the established prediction model, and the importance coefficients of each input parameter were evaluated based on this prediction model. Finally, Conditional Generative Adversarial Network (CGAN) was applied to augment and supplement the samples in the existing database, and the prediction model was used to verify its effectiveness. The results of this paper show that the parameter-optimized LightGBM model achieves the best prediction performance. Moreover, the established CGAN model can effectively solve the problem of insufficient data samples and lack of specific sample data.