Reliability and sensitivity analysis of wedge stability in the abutments of an arch dam using artificial neural network

In this study, the seismic stability of arch dam abutments is investigated within the framework of the probabilistic method. A large concrete arch dam is considered with six wedges for each abutment. The seismic safety of the dam abutments is studied with quasi-static analysis for different hazard levels. The Londe limit equilibrium method is utilized to calculate the stability of the wedges in the abutments. Since the finite element method is time-consuming, the neural network is used as an alternative for calculating the wedge safety factor. For training the neural network, 1000 random samples are generated and the dam response is calculated. The direction of applied acceleration is changed within 5-degree intervals to reveal the critical direction corresponding to the minimum safety factor. The Latin hypercube sampling (LHS) is employed for sample generation, and the safety level is determined with reliability analysis. Three sample numbers of 1000, 2000 and 4000 are used to examine the average and standard deviation of the results. The global sensitivity analysis is used to identify the effects of random variables on the abutment stability. It is shown that friction, cohesion and uplift pressure have the most significant effects on the wedge stability variance.