From cyclic degradation to risk mitigation: A deep learning framework for lifecycle reliability management of reservoir landslides

An assessment of the reliability of reservoir slopes is a prerequisite for achieving landslide hazard prevention. Cyclic reservoir operations drive time-varying seepage and pore pressures, and in the hydrological fluctuation zone, they induce wet‒dry cycles that degrade the soils of the sliding zone, causing slope reliability to coevolve with external conditions. Although reliability methods are widely used in geotechnical engineering scenarios, accurately and efficiently evaluating time-dependent reliability remains challenging. Utilizing the Wangjiashan landslide in the Baihetan Reservoir area as a case study, an efficient time-varying reliability framework implemented under long-term reservoir operations is proposed. Safety factors (Fs) computed by coupled, variably saturated seepage and limit-equilibrium analyses serve as supervisory signals. Deep learning surrogates are trained to learn a mapping from the reservoir level and shear strength to the safety factors, while the parameter evolution trend associated with wet‒dry degradation in the fluctuation zone is explicitly incorporated, enabling the prediction of failure probability trajectories under scheduled operations. The results show that the framework reproduces the current failure probability of the Wangjiashan landslide and can be extrapolated to subsequent hydrological years through the degradation model. As the number of wet‒dry cycles increase, the failure probability increases annually, and its growth rate stabilizes from the sixth hydrological year onwards. Compared with full finite-element sequences, the proposed deep learning-based framework improves the computational efficiency level by more than a factor of thirty while maintaining high accuracy, providing practical decision support for reservoir scheduling and risk mitigation tasks involving large reservoir systems. • Novel framework integrating experimental degradation modeling and deep learning. • Efficient Spatiotemporal reliability prediction via surrogate modeling. • Long-term risk projection and mitigation strategy optimization.