A hybrid approach for interval prediction of concrete dam displacements under uncertain conditions

Accurate and credible displacement prediction is essential to dam safety monitoring. However, due to the inherent uncertainties involved in dam systems, errors of conventional deterministic point predictions are inevitable and sometimes large. In this paper, prediction intervals (PIs) are used instead of deterministic values to quantify the associated uncertainties and improve the reliability of dam displacement prediction. A hybrid modeling approach is proposed to synthetically evaluate the aleatoric and epistemic uncertainties through PI construction, which integrates the non-parametric bootstrap, least squares support vector machine (LSSVM), and artificial neural network (ANN) algorithms. Specifically, the PIs of dam displacement are constructed in two stages. In the first stage, multiple bootstrap-based LSSVMs are utilized to estimate the true regression means of future displacements and the variance of model uncertainty. In the second stage, a modified ANN (MANN) is developed and applied to estimate the variance of data noise. The final PIs are calculated by combining the true regression means and the variances of model uncertainty and data noise. The performance of the bootstrap-LSSVM-MANN model is verified using monitoring data from a real concrete dam. The results show that the proposed method can generate computationally efficient high-quality PIs and can effectively deal with multiple uncertainties in data-driven modeling and prediction. The novel approach has great potential to support the decision-making activities in an environment characterized by uncertainties and risks.