A Displacement Prediction Method for Tailings Dams Based on a Multi−Module Fusion Model Optimized by Grey Wolf Optimization

Displacement is one of the most direct indicators for evaluating the operational state and structural stability of tailings dams. However, the displacement evolution process is jointly affected by rainfall, reservoir water level, phreatic line variation and other coupled factors, which makes the monitoring sequence highly nonlinear, time−dependent and lagged. Traditional prediction models often struggle to balance local feature extraction and long−term temporal dependency modeling when processing such complex monitoring data. To address this problem, this study developed a tailings dam displacement prediction method based on a multi−module fusion model optimized by the Grey Wolf Optimizer (GWO). First, candidate influencing factors were selected from monitoring variables by Pearson correlation analysis and further interpreted from the perspective of displacement evolution mechanism. Then, a Convolutional Neural Network (CNN) was employed to extract local features from multi−source monitoring sequences. A Nidirectional Long Short−term Memory Network (BiLSTM) was subsequently introduced to capture the bidirectional temporal dependency of the displacement sequence. After that, a self−attention mechanism was incorporated to enhance the representation of key time steps and improve the identification of abrupt displacement changes. Finally, GWO was used to optimize the critical hyperparameters of the hybrid model, including the number and size of convolution kernels, the number of hidden units, the dropout rate, the number of attention heads, the attention dimension, and the learning rate. Monitoring data from the Liangzigou tailings pond were used to validate the proposed method. The results show that the proposed GWO−CNN−BiLSTM−Self−Attention model achieves high prediction accuracy, with an R² of 0.9952 and an RMSE of 0.05677 mm on the test set. Compared with several benchmark models, the proposed model presents better performance in trend tracking, peak capture, and overall robustness. In addition, the ablation results indicate that the CNN module effectively improves local feature extraction, while the self−attention mechanism strengthens the model’s ability to focus on critical temporal information. Cross−validation results further demonstrate that the model maintains good stability and generalization under different data subsets. Therefore, the proposed method provides an effective technical approach for displacement prediction, safety monitoring, and early warning of tailings dams.