An improved prediction model of dynamic normal stress on silo wall based on gradient boosting and parameter optimizations

Dynamic normal stress on silo wall is one of the major issues in prediction of silo structure failure. An improved predict model is proposed to predict dynamic normal stress on silo wall combined with gradient boosting algorithm and parameter optimization. First, a database of 864 datasets was established, and highly correlated features were eliminated using the Pearson correlation coefficient to reduce dimensionality. Next, hyperparameters were optimized using grid search and the Whale Optimization Algorithm (WOA). The optimized model's fitting ability and generalization performance were evaluated to identify the best model. Finally, SHAP (SHapley Additive exPlanations) was used to interpret the WOA-XGBoost model and quantify feature contributions. Results showed the R2 of the WOA-XGBoost model is 0.933 increasing by 4.3%, and the training time is 230 s, which is increased by 20 times compared by grid search method. It can be concluded that the dimension of discharge port for silo structure and the coefficient of internal friction angle for granular materials play important role in dynamic normal stress prediction, which should be considered firstly during silo structures design.