Hybrid CNN-GRU-Attention-XGBoost Model for Rapid and Accurate Prediction of Peak Acceleration and Trigger Time in Dam Safety Monitoring

We have developed a hybrid model combining CNN with GRU, Attention, and XGBoost to improve prediction accuracy of peak acceleration and trigger time in dam monitoring systems. This model overcomes the limitations of traditional methods which rely on nonlinear seismic data patterns. The hybrid architecture features a 1D CNN to extract spatial features, GRU layers to capture temporal dependencies, and self-attention to focus on important time steps, with XGBoost applying residual fitting to enhance the model's predictions. The proposed strategy of adjusting the contribution weights of the CNN-GRU-Attention and XGBoost models based on their ensemble validation errors enhances the overall robustness and reliability of the combined model under varying seismic conditions. Validation results demonstrate the significant efficacy of this approach, compared to baseline models, namely the standalone CNN-GRU-Attention, XGBoost, and Random Forest, its prediction accuracy is markedly improved, with the Mean Absolute Error (MAE) being reduced by 17.14 % relative to the best-performing baseline model. This study presents a new model which combines deep learning and ensemble methods together and provides an extension for serving as an early warning systems for systems in network infrastructures. Its timely and accurate prediction capabilities makes it essential for the protection of important infrastructure from earthquake risks.