Machine-learning ensembled probabilistic methods for time-dependent reliability analysis of reservoir slopes under...

Probabilistic reservoir slope stability analysis usually suffers from computational inefficiency of complicated implicit performance functions. To address this problem, numerous machine learning (ML) algorithms have been successfully applied to calculate the index of factor of safety (FS) to facilitate slope reliability analysis. However, current ML surrogate models generally select an optimal algorithm with good performance while discarding all the others. Meanwhile, ML models mainly provide a point-value deterministic evaluation of failure probability and ignores the predictive uncertainties associated with the models. How to ensemble the strengths of different ML algorithms to improve the estimation accuracy of slope failure probability and rationally quantify the predictive uncertainties still remains an open issue. This paper proposes ML ensembled probabilistic methods for time-dependent reliability analysis of reservoir slopes under rapid water level drawdown using Bayesian model averaging (BMA). A practical slope in the Three Gorges Reservoir Area is selected to illustrate the proposed approach. Firstly, two machine-learning models, namely support vector machine (SVM) and back propagation-based neural network (BPNN), are used to evaluate the safety factor of reservoir slopes. Then, the deterministic evaluations of FS are ensembled together by the weighting of each member using BMA to make probabilistic evaluation of time-dependent failure probability. Finally, the predictive performance of individual ML models and the ensemble BMA model are systematically investigated. Results show that the ensemble surrogate model has better predictive performance on slope failure probability than individual ML model. Ensemble models can not only combine the strengths of different models to improve the prediction accuracy of time-dependent failure probability of reservoir slope, but also provide probabilistic forecasts to reasonably quantify the predictive uncertainty.