[Objective] The primary objective of this study is to develop a high-precision deformation prediction method capable of effectively handling the inherent nonlinearity and dynamic variations in monitoring data. Traditional single models and static ensemble methods often struggle to adapt to the shifting data distributions and sudden changes characteristic of real-world deformation scenarios, such as landslides and structural settlements. This research aims to overcome these limitations by proposing a Dynamic Weighted Ensemble Learning (DWEL) model that intelligently integrates multiple base learners and dynamically adjusts their contributions based on recent performance, thereby enhancing prediction accuracy, robustness, and generalization ability in complex environments. [Methods] The proposed DWEL model integrates four diverse and complementary base learners: Long Short-Term Memory (LSTM) networks, Support Vector Regression (SVR), Random Forest (RF), and XGBoost. This selection ensures comprehensive feature extraction from different perspectives. The core innovation of DWEL lies in its two-stage optimization mechanism: (1) Dynamic Weighting Mechanism. A sliding window is employed to continuously track the prediction error (absolute error) of each base learner over a recent period. (2) Two-Level Fusion Strategy. A novel “weight-then-regress” hybrid fusion strategy is implemented. Experiments were conducted on two real-world deformation monitoring datasets: a landslide monitoring dataset (Dataset A) characterized by high non-stationarity and abrupt changes, and a tunnel settlement dataset (Dataset B) with relatively stable long-term trends but periodic fluctuations. The model’s performance was evaluated against single models (LSTM, SVR, RF,XGBoost) and traditional ensemble methods (Averaging, Bagging, Static Stacking) using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²). Ablation studies were performed to validate the contribution of each core component. [Results] (1) Experimental results on both datasets demonstrated the superior performance of the proposed DWEL model.On the complex Landslide Dataset A, the complete DWEL model (DWEL-2) achieved the best results with an RMSE of 2.72 mm, MAE of 2.08 mm, and R² of 0.951. This represented a significant improvement over the best single model (LSTM: RMSE=3.92 mm, R²=0.897), reducing the prediction error by approximately 30.6%. It also outperformed traditional ensemble methods like Static Stacking (RMSE=3.52 mm, R²=0.918). Similarly, on the Tunnel Settlement Dataset B, DWEL-2 attained an RMSE of 0.63 mm, MAE of 0.52 mm, and R² of 0.951, again surpassing all competitors and reducing the error of the best single model (LSTM) by 32.3%.(2) Analysis of the dynamic weight changes revealed the model’s adaptive capability. For instance, during the accelerated deformation phase (hours 150-160) in Dataset A, the weights of LSTM andXGBoost increased significantly, indicating their stronger ability to capture nonlinear mutations, while RF and SVR maintained higher weights during stable periods. This visual analysis confirmed that the dynamic weighting mechanism effectively allocated influence based on temporal data characteristics.(3) The impact of the sliding window length (w) was systematically analyzed. For the highly dynamic Landslide Dataset A, a medium window length (w=16) yielded the optimal balance between responsiveness and stability (RMSE=2.72 mm), whereas shorter (w=8) or longer (w=24) windows led to increased errors due to noise sensitivity or delayed response, respectively. For the more stable Tunnel Dataset B, a longer window (w=24) performed best, demonstrating the method’s adaptability to different data dynamics. (4) Ablation studies conclusively proved the necessity of both innovative components. Using static weights instead of dynamic weighting increased the RMSE on Dataset A by 15.8% compared to the full DWEL-2 model. Using only the meta-learner without dynamic weights performed similarly to Static Stacking. Furthermore, the two-level fusion (DWEL-2) provided an 11.7% reduction in RMSE compared to using only dynamic weighted averaging (DWEL-1), highlighting the significant contribution of the regression-based residual correction in the second fusion layer. [Conclusions] This study successfully developed and validated a novel Dynamic Weighted Ensemble Learning (DWEL) model for deformation prediction. The key innovations include: 1) A dynamic weighting mechanism based on sliding window error feedback, which effectively addresses the lag issue of static ensemble methods during data distribution shifts, improving response speed at mutation points by over 60%; and 2) A two-level “weighted average followed by regression” fusion strategy that synergizes rapid dynamic adaptation with enhanced non-linear fitting capability, reducing prediction errors by an additional 11.7% compared to single-level fusion.The experimental results robustly demonstrate that the DWEL model significantly outperforms existing single and traditional ensemble models across different deformation scenarios (landslide and tunnel settlement). It exhibits remarkable accuracy, robustness, and generalization ability, particularly during critical periods of trend mutation. The method provides an effective, reliable, and high-precision prediction tool for geological hazard early warning and structural health monitoring. Future work will focus on the adaptive optimization of the weighting function and the integration of multi-physics field coupling modeling to further enhance prediction reliability in even more complex environments.