A Data-Driven Approach to Dam Infrastructure Monitoring: Enhancing Prediction Accuracy by Systematic Rainfall Event Classification

Abstract

Aging dam infrastructures are increasingly susceptible to structural deterioration influenced by environmental conditions. Ensuring safety under such circumstances requires not only reliable monitoring data but also robust handling of data irregularities and missing values, particularly during rainfall. However, existing predictive models often overlook the distinction between rainfall and non-rainfall conditions, limiting their effectiveness. This study presents an integrated predictive framework that incorporates the Extreme Gradient Boosting (XGBoost) model with tailored preprocessing for turbidity and leakage data. The Pruned Exact Linear Time (PELT) algorithm is employed to segment data based on rainfall events, enabling the model to reflect environmental impacts on sensor behavior. Short duration missing data are interpolated, and anomalies are corrected using structural constraints. The preprocessed dataset is then used to train an XGBoost model optimized via automated hyper-parameter tuning. Comparative analysis demonstrates improved prediction accuracy and robustness, supporting proactive dam safety monitoring and enhancing infrastructure resilience under diverse environmental scenarios.