Correlation Analysis of Karun-4 Dam Deformation based on Machine-Learning Model Using Combination of Hydrostatic and Micro-Geodetic data

Abstract. This study develops a data-driven framework for modeling and interpreting the deformation behavior of the Karun-4 Dam using Gradient-Boosted Regression Trees (GBRT). The model integrates long-term hydrostatic, meteorological, and geodetic data from April 2011 to May 2025, covering the full operational range of the dam's reservoir. Our analysis reveals distinct seasonal deformation cycles, with displacements increasing during high-water periods and decreasing during drawdown, reflecting the dominant role of hydrostatic loading. The GBRT model accurately captured both the seasonal cycles and multi-year displacement variability, with prediction errors remaining below one centimeter. Sensitivity analysis using Shapley Additive Explanations (SHAP) identified the reservoir head as the primary driver of displacement, followed by air temperature. Interaction analysis revealed a nonlinear coupling between hydrostatic and thermal effects, with the largest displacements occurring when both factors were high. Spatially, upper galleries exhibited larger, more uniform displacements driven by hydrostatic loading, while lower galleries showed smaller displacements with greater sensitivity to temperature. The model was validated through comparisons with displacement data from both internal pendulum sensors and external micro-geodetic measurements, confirming its accuracy. These findings provide valuable insights for real-time monitoring, predictive maintenance, and long-term safety assessments of large concrete dams. This study highlights the effectiveness of machine learning in dam health monitoring, offering a scalable and interpretable tool for infrastructure management.