Research on tailings dam displacement prediction using an IGWO–Informer–ARIMA model

Abstract Tailings dam deformation prediction plays a critical role in the safety monitoring and early warning of tailings storage facilities. However, displacement monitoring data are typically characterized by strong nonlinearity, temporal dependency, and complex dynamic fluctuations, which make accurate prediction challenging for conventional forecasting models. To address these issues, this study proposes a hybrid prediction framework integrating Improved Grey Wolf Optimization (IGWO), Informer, and Autoregressive Integrated Moving Average (ARIMA) for tailings dam displacement forecasting. First, displacement monitoring data from four monitoring points of the Hetaoqing tailings pond in Yunnan Province, China, were preprocessed through missing-value handling, outlier removal, normalization, and Savitzky–Golay filtering to construct reliable time-series samples. Subsequently, the IGWO algorithm, incorporating Tent chaotic initialization, a nonlinear cosine convergence factor, and an optimal memory retention strategy, was employed to optimize the key hyperparameters of the Informer model. The optimized Informer model was then used to capture the nonlinear temporal characteristics and long-term dependencies of displacement sequences, while ARIMA was introduced to model the residual linear components, forming the IGWO–Informer–ARIMA hybrid forecasting framework. To validate the effectiveness of the proposed optimization strategy, comparative experiments and ablation studies were conducted using multiple benchmark functions. The results demonstrate that IGWO exhibits superior convergence speed, optimization accuracy, and stability compared with GWO, PSO, SSA, and SA algorithms. Furthermore, the proposed hybrid model was comprehensively compared with BPNN, GRU, SVR, RF, CNN, Informer, ARIMA, and IGWO–Informer models under an equal-budget hyperparameter optimization protocol. Experimental results show that the proposed model achieved the best predictive performance for both vertical and horizontal displacement prediction. Specifically, the average test-set R2 values reached 0.9133 and 0.9623, respectively, while MAE and RMSE were significantly reduced compared with baseline models. Additional SMAPE and MASE evaluations confirmed the robustness of the model for near-zero displacement prediction tasks. To further evaluate practical applicability, residual-bootstrap prediction intervals were constructed to quantify forecasting uncertainty, and the proposed model was additionally validated using two other tailings reservoirs with different material types, including a lead–zinc tailings reservoir and a phosphogypsum tailings reservoir. The results indicate that the IGWO–Informer–ARIMA model maintains stable and reliable prediction performance under different deformation mechanisms and environmental conditions, demonstrating strong robustness and cross-scenario generalization capability. Overall, the proposed framework provides an effective technical approach for intelligent monitoring, displacement trend prediction, and early-warning decision-making for tailings dam safety management.