Multi-level confidence prediction method for landslide displacement based on BSLO-optimized decomposition and Transformer model

This study proposes a method for predicting step-type landslide displacement using a BSLO-Transformer approach, aiming to overcome limitations such as experience-dependent parameter selection in variational mode decomposition (VMD), insufficient long-sequence processing capability of traditional models, and lack of uncertainty quantification. The method employs the blood-sucking leech optimization (BSLO) algorithm to construct an adaptive optimization framework for optimizing VMD parameters, achieving signal decomposition based on an information entropy minimization criterion. Subsequently, a Transformer model is designed for time series prediction by removing unsuitable components and adding feature enhancement layers. Finally, a multi-level confidence interval prediction framework is constructed to realize uncertainty quantification at multiple time scales. The method was validated using four monitoring points at the Tanjiahe landslide in the Three Gorges Reservoir area. Stable performance was achieved for 1-, 3-, 7-, and 15-day predictions, with R² values exceeding 0.95 across all time scales, RMSE below 5 mm, and PICP values reaching from 0.811～0.986, 0.739～0.975, and 0.617～0.960 under 95%, 90%, and 80% confidence levels, respectively. Coverage rates were close to theoretical expectations, demonstrating high accuracy and reliability. Compared with VMD-SSA-LSTM and CNN-BiLSTM-Attention models, the proposed transformer method demonstrates superior stability and prediction accuracy across all prediction time scales, providing an effective technical solution for landslide monitoring and early warning in reservoir areas.