Interpretable Machine Learning Stacking Ensembles for Dam Displacement Prediction and Safety Assessment

Abstract

Monitoring dam movements is essential to ensure structural safety and to plan maintenance interventions. However, dam displacement is influenced by multiple interacting factors, including reservoir water-level fluctuations, air–water–concrete temperatures, wind speed, and relative humidity, making its behavior highly complex and nonlinear. To address this challenge and improve short-term displacement forecasting for the Beni Haroun Dam in Algeria, this study develops a machine learning–based stacking ensemble model. Five regression algorithms, Random Forest (RF), Extra Trees (ET), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Multiple Linear Regression (MLR), were first trained using a fivefold cross-validation procedure. Their out-of-fold predictions were then combined as metafeatures to build a higher-level stacking ensemble model, with CatBoost employed as the metalearner. Model performance was evaluated using NSE, RMSE, R², MAE, PBIAS, and MAPE. The results demonstrate the clear superiority of the stacking approach, which achieved R² = 0.971, RMSE = 0.292, MAE = 0.236, NSE = 0.971, and near-zero bias (PBIAS = 0.000). This represents a substantial improvement over the best-performing individual model, ET (R² = 0.843, RMSE = 0.686), as well as over other models such as CatBoost (R² = 0.831), RF (R² = 0.811), MLR (R² = 0.803), and XGBoost (R² = 0.797). The MLR model showed comparatively lower performance (R² = 0.803, RMSE = 0.768), confirming the advantage of nonlinear approaches. SHAP analysis identified reservoir water level as the most influential factor in predicting dam displacement. The results confirm that the stacking ensemble model effectively captures complex dam behavior, providing accurate forecasts that support early anomaly detection and improved dam safety management.