Displacement prediction model for seasonally frozen slopes integrating dual signal decomposition and an interpretable deep network

To ensure the safe operation of trains in seasonally frozen regions, achieving accurate and interpretable displacement prediction of tunnel portal slopes is a fundamental requirement. In this paper, we developed a hybrid prediction model that integrates dual signal decomposition with an interpretable deep neural network. The Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) first decomposes the original signals, then performs mode selection and reconstruction based on sample entropy and clustering to suppress redundancy. The high-frequency components are further analyzed using VMD and optimized via DLABC, thereby enhancing multi-scale dynamic feature extraction. On this basis, a Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM)-Attention model, tuned by the Dung Beetle Optimizer (DBO), is constructed to jointly capture local features, temporal dependencies, and key time-specific responses. Results show the proposed model achieves R² > 0.99 and MSE < 0.07 across multiple monitoring points, significantly outperforming single-decomposition models (e.g., VMD-BP, R²=0.896, MSE=1.700). The dual decomposition strategy proves effective in noise suppression and feature enhancement. Additionally, the SHapley Additive exPlanations (SHAP) analysis visualizes the model's decision process, quantifying the contribution of key factors to slope deformation, thus improving transparency and reliability. The model demonstrates specific adaptability to freeze-thaw environments, providing a robust framework for forecasting slope deformation and issuing early warnings in seasonally frozen regions.