A two-stage hybrid BPNN prediction model for frost-resistance of rubberized concrete aided by feature selection with information entropy

Abstract The incorporation of waste rubber in concrete is a beneficial solution, it not only can improve the frost-resistance of concrete, reduce environmental pollution, and it also has a role that cannot be ignored in reducing carbon emissions. Benefit from the relative dynamic elastic modulus, the frost-resistance of concrete can be effectively evaluated. Hence, predicting the frost resistance accurately of rubber concrete is urgently needed for its popularization and application in cold areas. In this paper, 153 sets of relative dynamic elastic modulus data of rubber concrete were collected from published literatures, by combining information entropy and back propagation neural network (BPNN), a two-stage hybrid BPNN (5-7-1) prediction model was proposed. In this model, information entropy was used for input features selection, and BPNN was used for predicting frost resistance. The number of neurons in the hidden layer of BPNN was comprehensively determined by empirical formula and the minimum mean square error of the model. The prediction results show that after features selection, the MSE (Mean Square Error), RMSE (Root Mean Square Error), MAE (Mean absolute error), MAPE (average absolute percentage error) and R 2 (determination coefficient) are all superior to the BPNN without feature selection. The results of this research provide a new idea for the frost-resistance prediction for rubberized concrete.