Abstract: Construction simulation parameters are essential to the accuracy of actual arch dam construction schedule implementation. Of the previous studies on analysis of these parameters, most adopted a univariate model and some resorted to the multivariate analysis method to build a point prediction model, both facing a difficulty in evaluation of their uncertainty under the influence of various factors. To analyze the probability distribution of the parameters effectively, a probability prediction method based on a broad learning system along with the elastic network quantile regression (BLS-ENQR) is presented. BLS is a neural network that, without demanding a deep network structure, has an efficient nonlinear learning capability, so that it can overcome the deficiency of the traditional quantile regression model limited to linear relationship analysis. The method uses the elastic network regularization penalty to reduce the regression coefficient and improve its sparsity so as to avoid model overfitting. Engineering application shows that this new method can effectively analyze the probability distribution of construction simulation parameters, and it has prediction performance better than the support vector machine-ENQR (SVM-ENQR), extreme learning machine-ENQR (ELM-ENQR), or broad learning system-QR (BLS-QR).

Key words: arch dam, construction simulation parameters, probability prediction, broad learning system, quantile regression