Multi-point settlement deformation prediction method for tailings dams based on GCN-LSTM model

The settlement deformation of tailing dams exhibits complex spatiotemporal coupling characteristics, making it difficult for using traditional methods to effectively capture spatial correlations between monitoring points and the temporal dynamic evolution. For 20 monitoring points of the Hongyangou tailings dam (Initial dam: J1/G1/X1/G2/J2; Accumulation dam: first-stage sub-dam D1/D2/D3/D4/D5, second-stage sub-dam C1/C2/C3/C4; Auxiliary dam: G9/G10/G11/J9/J10/J11), a spatiotemporal hybrid prediction model based on a Graph Convolutional Network-Long Short-Term Memory (GCN-LSTM) architecture is proposed. First, the settlement data are preprocessed using the wavelet threshold denoising method, and a weighted undirected graph of spatial correlations among monitoring points is constructed based on the Pearson correlation coefficient. Subsequently, the GCN module is employed to explore the spatial topological features, which are then coupled with the LSTM module to learn temporal dependencies, thereby establishing a joint spatiotemporal prediction framework. Finally, the Adam optimizer is used to fine-tune the hyperparameters. Experimental results demonstrate that the proposed model achieves significantly higher prediction accuracy than those of traditional methods (RMSE = 0.023 35, R = 0.995 11) and can accurately capture settlement trends even in structurally heterogeneous regions (initial dam/accumulation dam/auxiliary dam). In ablation experiments, the R value of the proposed model increased by 0.189 14 and 0.347 27 when compared with the standalone LSTM and GNN models, respectively. By integrating spatial correlation and temporal dynamic characteristics, the proposed model enables high-precision multi-point settlement prediction of tailing dams, providing a reliable technical basis for safety assessment and risk early warning dam failure.