AI Predicts Dam Failures by Learning from Similar Structures
Researchers combined machine learning with ChatGPT-style AI to forecast safety risks at dams by comparing new structures to historical cases. The system achieved 78% accuracy on real-world Chinese dams, suggesting AI could help utilities and governments prevent costly failures and flooding without expensive on-site inspections.
Originaltitel: Case-Based Safety Risk Prediction for Dams via Hybrid Similarity Modeling and Large Language Models
Dams play a vital role in water management, energy generation, and flood control. However, their structural complexity and long service life make them susceptible to degradation, posing potential safety risks. This study explores a novel approach to dam risk prediction and assessment by combining clustering techniques with large language models (LLMs). We construct a heterogeneous dataset of 515 dams in China, integrating numerical, ordinal, and categorical features, along with textual records of identified risks and mitigation strategies. Using the k-prototypes algorithm, we effectively cluster the dams into meaningful groups, which are validated through t-SNE visualization, classification performance, and SHAP-based interpretability analysis. Building on these clusters and corresponding similarity measurements, we propose a case-based prompting method that retrieves the most similar historical dams and guides an LLM to generate risk predictions and mitigation strategies for a new, unassessed dam. The generated outputs based on the dataset show good alignment in approximately 78% of the real-world cases, demonstrating the potential of this hybrid framework to support dam safety assessment.