Abstract: To accurately evaluate the time-varying reliability of reservoir slope, it is necessary to consider both the spatial variability of soil strength and hydraulic parameters and the temporal effects of environmental loads, such as rainfall and reservoir water-level fluctuations. In time-varying reliability analysis, traditional numerical methods often require static reliability to be evaluated at each time step, making the process computationally expensive. Therefore, a hybrid deep-learning-based method is proposed for the efficient analysis of the time-varying reliability of reservoir slopes. The proposed method incorporates a convolutional block attention mechanism (CBAM)-enhanced residual network (ResNet) to extract spatial features from random-field images of soil parameters. It also integrates a bidirectional long short-term memory network (BiLSTM) to capture the temporal evolution of environmental loads. The outputs of the two models are fused to achieve time-series prediction of the safety factor of reservoir slopes under the uncertainties of soil parameters and environmental loads, thereby realizing efficient reliability evaluation. The proposed method is validated using the Shiliushubao landslide in the Three Gorges Reservoir area as a case study. The results show that the proposed method can simultaneously learn from multiple random-field images of soil parameters and multimodal temporal data describing environmental loads. The predicted safety factors are consistent with those obtained using traditional numerical methods, while computational efficiency is improved by approximately 305 times. Using monitoring data spanning 0 to 60 months, the proposed method achieved a coefficient of determination (R2) of .932 for failure probability prediction, outperforming conventional deep learning approaches and demonstrating superior predictive capability.

Key words: reservoir slope, time-varying reliability, surrogate model, digital image, spatial variability