Predicting landslide surge waves from large-scale physical experimental using machine learning

The impoundment of a hydropower station can cause water levels in reservoir areas to rise, potentially triggering nearby landslides and generating surge waves that pose significant threats to navigation and hydropower infrastructure. Traditional methods for predicting landslide-induced surge waves often struggle to accurately capture peak wave heights and their evolving trends. To address this challenge, this study employs machine learning approaches to enhance the prediction of surge wave characteristics by integrating insight from physical model experimental data. Initially, we utilized multi-peak Gaussian functions to fit the experimental surge wave data, enabling us to characterize surge wave run-up through derived fitting equations. Building on these findings, we developed three machine learning models—Random Forest, Long Short-Term Memory, and Gated Recurrent Unit (GRU)—to predict surge wave behavior. Among these, the GRU model outperformed others, demonstrating exceptional accuracy in capturing the critical first and second wave peaks, which are crucial for disaster mitigation. This study underscores the GRU model's robustness in predicting surge wave dynamics, presenting it as a valuable tool for mitigating risks associated with landslide-induced surge waves. By combining physical modeling, experimental data, and advanced machine learning techniques, this research establishes an innovative framework for enhancing reservoir management and disaster prevention efforts.