Prediction Investigation Of Dynamic Normal Stress On Silo Wall During Discharge Assisted With Integrated Machine Learning

The pressure generated by the loose particles in the silo on the side walls during unloading is difficult to accurately predict,which is a major challenge in the safety design of the silo structure.The theory of granular pressure is different from the traditional characteristics of solid mechanics and fluid mechanics.The main factors affecting the dynamic lateral pressure of silos are the physical properties of granular particles and the structural dimensions of silos.These factors have complex and nonlinear relationships with the dynamic lateral pressure of silos.The dynamic lateral pressure calculation formula provided by the specification is based on the static pressure calculation formula multiplied by the correction coefficient.However,this calculation method only considers one or several influencing factors,and the calculation accuracy needs to be discussed.Given this,it is particularly important to establish an accurate dynamic lateral pressure prediction model for silos that considers multiple influencing factors.Machine learning algorithms can consider multiple factors for nonlinear fitting.However,single learning algorithms generally have strong randomness,low accuracy and stability,resulting in limited generalization ability of prediction models.Integrated learning algorithms can integrate the outputs of multiple single learners,to some extent preventing overfitting and improving model prediction accuracy and stability.This article proposes an integrated algorithm model for dynamic lateral pressure of silos,with the main contents as follows:(1)Through literature collection and PFC numerical simulation,the dataset required for machine learning algorithm training and testing was established.After feature selection,the input feature attributes of the silo dynamic lateral pressure prediction model were determined to be: measurement point proportion position,height,diameter,aspect ratio,funnel angle,outlet size,and storage density.The training and testing sets were divided according to an8:2 ratio.In order to compare and analyze the improvement effect of ensemble learning strategies on the prediction performance of a single learning model,a silo dynamic lateral pressure prediction model based on a single learner(decision tree,support vector machine,BP neural network)is constructed as a reference.(2)Based on the three integration strategies(Bagging,Boosting and Stacking),the dynamic lateral pressure prediction model of silos based on the decision tree integration algorithm is established,which are respectively random forest model,Adaboost model and Stacking model.Among them,the Boosting ensemble strategy based on decision trees not only includes Adaboost algorithm,but also gradient boosting algorithms represented by Gradient Boost Tree(GBDT)and its improved algorithms(XGBoost,Light GBM).These six decision tree ensemble algorithms and three single learner algorithms are all trained and predicted on the same training and testing datasets.(3)The prediction effects of nine prediction models on the test set were compared and analyzed through the evaluation index mean relative error(MRE),mean absolute error(MAE),mean square error(MSE),root mean square error(RMSE)and correlation coefficient(R2).Among them,the decision tree in a single learner has the worst prediction performance,and the prediction accuracy is much lower than that of support vector machines and neural network algorithms;The prediction accuracy of decision tree integration algorithm(random forest,Adaboost and Stacking)is significantly higher than that of single decision tree,and higher than that of support vector machine and neural network algorithm.Adaboost algorithm has the highest prediction accuracy;In addition,the prediction accuracy of gradient lifting tree algorithms(GBDT,XGBoost,Light GBM)has been improved compared to Adaboost algorithm,with the highest being XGBoost algorithm.It shows that the Boosting integration strategy is better than Bagging and Stacking,and in the Boosting algorithm,the gradient lifting algorithm that uses the negative gradient value of the loss function to fit the regression tree is better than the lifting algorithm that adjusts the weight of the weak learner according to the error.(4)By comparing the relative errors of 9 prediction models at each sample point in the test set,the XGBoost algorithm model with the best prediction effect for the dynamic lateral pressure of the silo was identified.And conducted silo unloading model tests and numerical simulation tests,compared the dynamic lateral pressure test values,simulation values,and machine learning algorithm prediction values of the silo under the same working conditions,verified the generalization performance of XGBoost algorithm,which can be used to predict the dynamic lateral pressure value of the silo.Based on the tree building basis based on information gain in XGBoost algorithm,the feature importance ranking of dynamic lateral pressure in silos is obtained,providing reference for silo structural design.