Accurate deformation forecasting is essential for concrete dam safety monitoring, yet real‐world deformation series are often nonlinear, nonstationary, and contaminated by noise, missing values, and outliers, which limits the performance of traditional prediction models. This paper proposes a high‐precision forecasting framework that integrates variational mode decomposition (VMD), adaptive wavelet‐threshold denoising, and the informer model. First, VMD decomposes complex deformation signals into intrinsic mode functions to mitigate mode mixing and enhance frequency‐specific interpretability. Then, an adaptive wavelet‐threshold strategy is introduced, where the threshold is dynamically adjusted using signal‐to‐noise ratio (SNR) guidance and Stein′s unbiased risk estimate (SURE), enabling effective noise suppression while preserving informative deformation patterns under nonstationary conditions. Finally, a multichannel informer architecture is employed to fuse multiscale components and capture long‐range dependencies efficiently via ProbSparse self‐attention. Experiments on deformation monitoring datasets from multiple concrete dams demonstrate that the proposed VMD‐AWT‐informer consistently outperforms mainstream baselines across forecasting horizons. For instance, at a 30‐day horizon, the proposed method achieves an RMSE of 0.55 mm, reducing errors by 23.6% compared with VMD‐LSTM and by 55.6% compared with SVR, while maintaining strong goodness of fit ( R 2 = 0.887). Robustness tests further confirm improved stability under noisy, incomplete, and outlier‐corrupted inputs. These results indicate that the proposed framework provides an effective and practical tool for long‐horizon dam deformation prediction and early warning.
Qi et al. (Thu,) studied this question.