A year of reservoir-slope reliability analysis drops from 6.7 hours to 15 seconds

Time-dependent probabilistic stability analysis of reservoir slopes no longer has to rerun a numerical model for every case. Traditional numerical methods for this analysis are computationally demanding, while existing surrogate models often overlook temporal correlations or spatial variability. On the Baijiabao slope in the Three Gorges Reservoir Area, the network reached a coefficient of determination of 0.975, against 0.922 and 0.946 for two baseline models.

The network learns the spatial random field of soil parameters and the time series of loads separately, then fuses them to predict the factor of safety and its probability distribution. A residual convolutional network extracts spatial features from random-field representations of heterogeneous soil parameters, a convolutional block attention module adaptively recalibrates those features, and a bidirectional long short-term memory network learns the temporal evolution of environmental loads in parallel. The evaluation uses the Baijiabao slope in the Three Gorges Reservoir Area, with precipitation and reservoir water-level fluctuations as the time-varying environmental loads.

The report covers time-dependent probabilistic stability analysis on the single case of the Baijiabao slope, with soil-parameter random fields and precipitation and reservoir water-level fluctuations as inputs, and a time-dependent factor of safety and its probability distribution as output. It is aimed at researchers and engineers working on slope reliability and reservoir geohazard assessment who want to judge whether such a surrogate belongs in their own analysis workflow.

The results rest on a single reservoir slope case, and the material does not state how the network performs on other slopes or under different load conditions. The computation-time comparison is inference time after training; the cost of training itself is not given in the material.