CEEMDAN-FTEA-GCN-Transformer: A Transformer-based model for dam deformation prediction with frequency-spatial feature integration

Dam deformation is a critical indicator for assessing the operational status of dams, making deformation prediction models increasingly valuable in structural health monitoring. However, traditional statistical models and machine learning methods face limitations in information extraction and spatial dependency modeling, resulting in suboptimal predictive accuracy and performance. To address these challenges, this study proposes a deep learning model based on CEEMDAN-FTEA-GCN-Transformer. Initially, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) technique is employed to decompose the nonlinear and non-stationary time series in the raw data into several smoother subsequences, thereby improving signal smoothness and predictability. Subsequently, the Frequency-Time Enhanced Attention (FTEA) Block is designed to enhance feature extraction and fusion across multiple scales by integrating temporal and frequency-domain characteristics. Finally, graph convolutional networks (GCNs) are incorporated into the Transformer framework, enabling the model to effectively capture spatial dependencies among monitoring points. This integration enhances the understanding of overall dam deformation patterns, improving the model's robustness and stability. The proposed model is validated using deformation data from a concrete dam located in Jiangxi Province, China. Performance evaluations across various monitoring points demonstrate that the model achieves R² values exceeding 0.99 for all points, outperforming other comparative models. These results confirm the effectiveness and generalizability of the proposed approach in dam deformation prediction, offering a novel methodology and technical support for dam health monitoring and forecasting.