A novel prediction method for slope deformation based on adaptive mode decomposition and interpretable deep learning

To address the issues of non‑stationary signal processing and the black‑box nature of models in slope deformation prediction, this study proposes a prediction framework based on the Dung Beetle Optimizer (DBO), Variational Mode Decomposition (VMD), and Long Short‑Term Memory (LSTM), and introduces SHAP for interpretability analysis. DBO simultaneously optimizes the VMD decomposition parameters (number of modes and penalty factor) and LSTM hyperparameters (number of hidden layer neurons, learning rate, and number of iterations), achieving end-to-end adaptive optimization. A case study on a red clay high slope in Jiangxi Province showed that the proposed DBO‑VMD‑LSTM model achieved an RMSE of 0.155 mm and an R² of 0.993, significantly outperforming comparison models such as SSA‑VMD‑LSTM ( p < 0.001). SHAP interaction analysis revealed an alternating coupling pattern between displacement rate and multi‑scale rainfall—positive for short‑term, negative for medium‑term, and positive for long‑term—and the magnitude of interactions differed markedly between periodic and trend components (the maximum for the trend component was only 15% of that for the periodic component), validating the physical effectiveness of the VMD decomposition. This study provides a high-precision and interpretable method for predicting high-slope deformation.