Attention-enhanced spatiotemporal deep learning-based automatic warning model with uncertainty estimation in dam safety

Abstract

Dam safety monitoring and automatic warning have become critically important with the increasing frequency of extreme weather events caused by climate change. This study proposes a spatiotemporal dam joint warning model enhanced by an attention mechanism, incorporating an uncertainty estimation module to quantify the reliability of predictions. The model’s generalization ability under future climate change scenarios is explored for the first time. The model integrates data from multiple dam monitoring points and utilizes the attention mechanism to capture complex spatiotemporal features, building a practical uncertainty-aware prediction framework. The results show that under current climate data, the proposed attention-enhanced spatiotemporal prediction model demonstrates excellent predictive performance across multiple monitoring points, effectively capturing the complex spatial relationships between monitoring points, with lower uncertainty than traditional deep learning methods. In the context of extreme climate change scenarios, the prediction results indicate that the model can effectively capture dam displacement trends under a slight temperature rise (Δt = 2.78°C). However, under moderate and significant temperature rises (Δt = 4.04°C/Δt = 7.50°C), the model’s predictive performance declines, with increased uncertainty.