变形预测在地质灾害预警与工程结构健康监测中具有重要意义。针对传统模型难以应对监测数据非线性与动态变化的问题，提出一种基于动态加权集成学习(Dynamic Weighted Ensemble Learning, DWEL)的变形预测方法。该方法首先融合长短期记忆神经网络(Long Short Term Memory, LSTM)、支撑向量回归(Support Vector Regression, SVR)、随机森林(Random Forest, RF)和XGBoost四种模型为基学习器，基于滑动窗口构建误差历史序列，通过动态权重机制实时调整各模型贡献值，增强预测的适应性与稳定性。然后，融合阶段提出“先加权，再回归”的双层融合机制，实现结构上的“堆叠+反馈”双重优化，进一步提升模型的预测精度与泛化能力。基于实际变形监测数据集的实验结果表明，DWEL模型在均方根误差(Root Mean Square Error, RMSE)，平均绝对误差(Mean Absolute Error, MAE)和决定系数(Coefficient of Determination, R²)等指标上均优于单一模型及传统集成方法，预测误差平均降低12%以上，展现出较强的鲁棒性与泛化能力。所提方法为复杂环境下高精度变形预测提供了一种有效的新思路。

Deformation prediction is of significant importance for geological hazard early warning and structural health monitoring of engineering structures. To address the limitations of traditional models in handling the nonlinearity and dynamic variations inherent in monitoring data, this paper proposes a deformation prediction method based on Dynamic Weighted Ensemble Learning (DWEL). The proposed method integrates four base learners: long short-term memory (LSTM), support vector regression (SVR), random forest (RF), and XGBoost. It constructs an error history sequence using a sliding window mechanism and employs a dynamic weighting strategy to adjust the contribution of each model in real-time, thereby enhancing prediction adaptability and stability. During the fusion stage, a two-level fusion mechanism ("weighted average followed by regression") is adopted, achieving dual structural optimization through "stacking + feedback," which further improves prediction accuracy and generalization capability. Experimental results on multiple real-world deformation monitoring datasets demonstrate that the DWEL model outperforms both individual models and traditional ensemble methods across key metrics, including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²). Prediction errors are reduced by an average of over 12%, showcasing strong robustness and generalization ability. This method provides an effective and promising new approach for high-precision deformation prediction in complex environments.