A Digital Twin Model for Point-Interval-Probability Prediction of Reservoir Landslide Displacements

Accurately predicting reservoir landslide displacements remains a significant challenge due to raw data noise, model overfitting, and high uncertainty. This study introduces a "digital twin model" for point-interval-probability prediction of reservoir landslide displacements, addressing these limitations through a hybrid framework. The model employs complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose displacement sequences into periodic and trend components, effectively reducing noise. For point prediction, a bidirectional gated recurrent unit (BiGRU) is used, with hyperparameters optimized by the triangular topology aggregation optimizer (TTAO) to mitigate overfitting. Adaptive bandwidth kernel density estimation (ABKDE) then provides interval and probability predictions, improving uncertainty quantification. The model's performance was validated on the Jiuxianping landslide dataset. In point prediction, it achieved a coefficient of determination (R2) of 0.992, a mean absolute error (MAE) of 3.617 mm, a root mean square error (RMSE) of 4.418 mm, and a mean absolute percentage error (MAPE) of 0.005. The model delivered optimal coverage metrics for 95%, 80%, and 70% confidence intervals for interval prediction. In probabilistic prediction, it attained a continuous ranked probability score (CRPS) of 0.135 at the 95% confidence level. The proposed "CEEMDAN-TTAO-BiGRU model" surpasses existing methods in accuracyand robustness, providing valuable insights for early warning and disaster mitigation in reservoir landslide management.

Keywords: Empirical mode decomposition, BiGRU, Meta-heuristic algorithm, ABKDE.