An interpretable hybrid stacking framework for uncertainty-aware forecasting of structural deformation and damage evolution in concrete dams

Long-term dam deformation is governed by coupled hydrostatic, thermal and aging effects, producing noisy, nonlinear and temporally dependent monitoring series that are difficult to model with conventional statistical or single machine/deep learning approaches. This study proposes an interpretable hybrid stacking framework, RF–DEGWO–LSTM, for uncertainty-aware dam deformation forecasting. Random forest (RF) acts as a nonlinear feature-based learner and provides embedded feature importance; differential evolution grey wolf optimizer (DEGWO) is used to tune key long short-term memory (LSTM) hyperparameters; and a meta-learner fuses RF and DEGWO–LSTM outputs into a unified predictor that explicitly captures time-varying, non-stationary relationships between hydrostatic, thermal and aging factors and displacement. The framework is validated using long-term monitoring data from two dams (a concrete gravity dam and a run-of-the-river dam). Comparative experiments against RF, plain LSTM, recurrent neural networks (RNNs), eXtreme Gradient Boosting (XGBoost) and DEGWO–LSTM show that the RF–DEGWO–LSTM model consistently achieves the highest R ² and the lowest error metrics. In the most challenging scenario, RF–DEGWO–LSTM improves test-set R ² by approximately 5–10% and reduces root mean square error by about 15–20% relative to the best single-model baseline. Residual analyses indicate reduced bias and variance, while interpretability based on RF feature importance and LSTM attention confirms the dominant roles of hydrostatic and seasonal components, with aging effects being secondary. Accurate forecasting of deformation is critical for early identification of damage evolution and integrity loss in large civil infrastructure. The proposed framework thus offers an accurate, robust and interpretable tool for data-driven dam deformation forecasting and early-warning-oriented structural health management.