Hybrid intelligence approaches for reservoir landslide displacement prediction driven by data-mechanism coupling

FC Zhao and FS Miao and YP Wu and SQ Gong and TY Li and JH Kou, BULLETIN OF ENGINEERING GEOLOGY AND THE ENVIRONMENT, 84, 601 (2025).

DOI: 10.1007/s10064-025-04632-1

Landslide displacement prediction (LDP) is instrumental in mitigating the escalating global risk associated with reservoir induced slow-moving landslide reactivation. However, the pronounced temporal lag of triggering factors significantly impedes the practical application of data-driven LDP. This study proposes a quantitative LDP framework that couples data and mechanism by explicitly considering temporal lag effects, with application to the Baishuihe landslide in the Three Gorges Reservoir area. Herein, variational modal decomposition (VMD) was employed to decompose cumulative displacements and hydrometeorological factors into periodic and random components. The cross-correlation function (CCF) and Pearson correlation coefficient (PCC) identified lag periods of periodic terms before dual-factor diagnosis. Grey Wolf Optimized Support Vector Regression (GWO-SVR) was employed to select the optimal feature combinations for prediction, while SHapley Additive exPlanations (SHAP) provided interpretability of factor contributions at different slope positions. Three major challenges were addressed: (1) Integrating metaheuristic optimization to improve predictive accuracy. (2) Incorporating lag analysis of periodic terms to capture the coupled effects of rainfall and reservoir level. (3) Revealing differential slope responses to triggering factors. Results indicate that the time- lag effect of reservoir water level dominates periodic displacement. The hybrid model that accounts for lag achieves superior performance compared with conventional approaches. Moreover, spatial heterogeneity of triggering factors is evident: the front edge responds mainly to drawdown of reservoir levels, while the rear edge is more sensitive to intense rainfall. This hybrid data-mechanism framework demonstrates strong potential for advancing early-warning and risk-mitigation strategies for reservoir-related landslides worldwide.

Return to Publications page