Landslide displacement prediction using a hybrid deep learning model combined with variational mode decomposition

Landslide displacement prediction is critical for disaster prevention and mitigation. Addressing the distinct nonlinear and "step-like" evolution characteristics of reservoir bank landslides, as well as the limitation that single models struggle to simultaneously capture local features and long-term dependencies, this paper proposes a hybrid prediction framework integrating Variational Mode Decomposition (VMD) and a channel attention mechanism. First, the preprocessed cumulative displacement is decomposed into trend, periodic, and random components via VMD. A GRA-MIC-Pearson hybrid strategy is employed to screen environmental triggering factors with the strongest correlation to displacement components. Subsequently, a TCN-Attention-BiLSTM hybrid model is constructed to predict the highly nonlinear periodic component. Its performance is benchmarked against BiLSTM, BiGRU, and TCN-BiLSTM models. Additionally, a cubic polynomial and a lightweight LSTM are combined to predict the trend and random components, respectively. Experimental results on the Baijiabao landslide in the Three Gorges Reservoir Area demonstrate that the proposed hybrid model achieves superior prediction accuracy, providing a novel approach for landslide displacement monitoring.