Quantifying Predictive Uncertainty in Deep-Sequence Modeling for Early Warnings of Embankment Dam Failure

Reliable early warning of embankment dam instability remains a critical challenge due to nonlinear hydrological forcing, soil–structure interaction, and sparse monitoring data. This study introduces a probabilistic deep-sequence modeling framework that integrates a Mixture Density Network (MDN), Monte Carlo dropout, and hybrid ANN–LSTM architecture to predict settlement evolution while quantifying epistemic and aleatory uncertainties. FEM-derived stress–strain features and PCA-compressed hydro-climatic indicators enforce physical consistency and enhance generalization. The framework is rigorously validated using a 104-week temporal hold-out test, k-fold cross-validation, and distributional calibration diagnostics, achieving near-ideal uncertainty calibration (PICP = 95%). Applied to the Megech Dam failure case, the model exhibits distinct uncertainty signatures across stable, accelerating, and pre-failure phases. A new exponential instability metric, γ, derived from the MDN predictive distribution, provides a quantitative early-warning threshold; the γ-based alert system detects instability at γ = 0.08, approximately 11 weeks before the December 2021 failure, matching crack initiation observed in October 2021. These findings demonstrate that uncertainty evolution provides early, physics-consistent precursors that deterministic models cannot capture, offering a validated pathway toward next-generation early-warning systems for geotechnical infrastructure.