Predicting the Crest Settlement of Concrete Face Rockfill Dams by Combining Threshold Regression and Support Vector Machine

Abstract

The design and construction of concrete face rockfill dams (CFRDs) usually require a rapid and accurate prediction of deformation behavior to support dam optimal design and safety evaluation. Deformation prediction and control are key issues faced in the construction of CFRDs. This study collects measured data of 75 CFRD case histories. On the basis of the statistical review of the typical dam crest settlement behavior of CFRDs, a prediction model for dam crest settlement combining threshold regression (TR) and support vector machine (SVM) is established. A mixed weight coefficient is introduced to construct an adaptive hybrid kernel function with good learning ability and generalization performance. The particle swarm intelligent optimization algorithm is adopted to optimize model parameters for establishing an improved SVM prediction model. To further improve the generalization ability and accuracy of the improved SVM model, the multivariate TR theory is used to segment the dam crest settlement data according to the dam height. Then, an improved SVM prediction model is established in each dam height interval. The comparative analyses of the prediction results of different models show that the TR–SVM model effectively weakens the nonlinear mutation characteristics of the case data and achieves high prediction accuracy.