Artificial Neural Network Model to Predict the Factor of Safety in Earth Dams Subjected to Rapid Drawdown

Abstract

Rapid drawdown has been identified as one of the most frequent causes of slope failures due to the effects associated with drought and operational changes when incorporating hydroelectric plants, which influence the filling level of earth dams. The main goal of this research is to obtain predictive models based on Artificial Neural Networks that return the factor of safety of the upstream slope in homogeneous earth dams in the face of the effect of rapid drawdown. Three geometries and 40 soils were defined to form the embankment, from which hybrid numerical models of transient water flow with unsaturated soils were built, considering three discharge speeds. From these results, a database was built to develop the predictive models, by means of the KNIME program and an algorithm based on Artificial Neural Networks. The behavior of the factor of safety as a function of time is also analyzed to establish its recovery intervals. Main results show that the minimum factor of safety is obtained between 52 % and 88 % of the total drawdown time. Regarding the predictive models, the adjusted R2 determination coefficients were greater than 95 % in all cases and the errors remained below 10 %. This demonstrates a high effectiveness of this algorithm and the validity of its application to geotechnical problems.