Abstract: Traditional prediction methods are hardly applicable to the dam deformation featured with high dimensions and nonlinearity; they can predict the deformation at location points of a dam body, but fail to effectively quantify the uncertainties from data with random noise, subjectivity in input samples, and randomness in parameter selection. To solve this problem, we develop a new dam deformation prediction model based on the Bootstrap algorithm and an improved cuckoo search–multiple kernel extreme learning machine (ICS-MKELM) algorithm. The model quantifies the uncertainty through interval prediction and can realize accurate point prediction of dam deformation. First, based on high-precision MKELM, we construct a dam deformation prediction model that integrates the advantage of KELM efficiently handling strong nonlinear regression problems with the superiority of a hybrid kernel of high generalization and strong learning capability. And an ICS algorithm, based on the inertia weight and chaos theory, is adopted to optimize MKELM’s kernel parameters and regular coefficients, offsetting its disadvantage of easy falling into local optimization. Then, a Bootstrap interval prediction method is used to quantify the uncertainty from the model and data. Our model is applied to the deformation prediction of a real dam, and its consistency and superiority are demonstrated through an analysis on the influence of different sizes of the training sets on prediction accuracy and a comparison with other five commonly-used prediction algorithms.

Key words: dam deformation, interval prediction, multiple-kernel extreme learning machine, improved cuckoo search algorithm, uncertainty