Machine learning-based data mining of reservoir landslide triggering mechanisms and failure time prediction for displacement sudden state

Reservoir landslides, driven by rainfall and water level fluctuations, exhibit step-like displacements, posing significant geotechnical risks. This study employs machine learning and data mining to elucidate deformation mechanisms and predict failure times during sudden displacement states, enhancing disaster mitigation. Using GPS displacement data from three profiles, we quantify factors influencing front, middle, and rear-edge displacements in reservoir landslides. An interpretable interval prediction model, refined by Markov Chain Monte Carlo (MCMC) Bayesian updating, ensures robust failure time estimates. Results indicate that front-edge displacements are primarily triggered by water level fluctuations, with rapid drawdowns exceeding 8 m amplifying deformation under low-intensity rainfall. Middle and rear-edge displacements arise from combined rainfall and water level effects. SHAP analysis reveals rainfall's direct influence and water level's indirect role via rainfall interactions, with current-month (a5) and two-month (a7) water level changes driving short- and long-term displacement patterns, respectively. The CEEMDAN-TTAO-BiGRU model delivers high-accuracy predictions for periodic and total displacements, yielding narrow 95 % confidence intervals. Seven onset of acceleration (OOA) points, identified via the MACD indicator, show six with prediction errors ≤1 month. MCMC-based Bayesian updating estimates a mean failure time of 29.95 months (95 % CI: [28.38, 31.52] months), advancing landslide monitoring and early warning systems. This study offers scientific insights into reservoir landslide deformation by combining interpretable AI with physical-process understanding, and provides an engineering tool for accurate failure prediction to support intelligent monitoring and early warning.