Multi-source data-driven unsaturated seepage parameter inversion: Application to a high core rockfill dam

Abstract

The inverse parameter’s estimation of unsaturated seepage models is important to ensure the safety of earth dams, while most existing studies determine the hydraulic characteristics only based on seepage pressure monitoring data, which may lead to inversion results deviating from the intrinsic properties of dam materials. This study proposed an advanced methodology for unsaturated seepage parameter inversion of earth dams driven by multi-source data. The transient unsaturated seepage behavior of earth dams is simulated using the Van Genuchten (VG) model, and the extreme learning machine (ELM) is adopted as the surrogate model to estimate the approximate simulation results. Then the multi-objective function for parameter inversion is constructed by combining the observed transient pressure data and ELM models, and the Pareto-optimal solutions are generated using the elitist non-dominated sorting genetic algorithm (NSGA-II). Multi-source information, including parameter design values, in-situ testing results, and monitoring data, are fused to construct a comprehensive index to drive the parameter decision. The proposed method is applied to the PB core rockfill dam, which has a well-established seepage pressure monitoring system. In addition, the design values and in-situ testing results of the hydraulic conductivity of the dam material are also available. The results showed that the inversion parameters of the proposed method are closer to the physical characteristics of materials and have advantages in prediction accuracy.