Early warning of dam failure risk via TCN-attention modeling of hydrometeorological time series

ABSTRACT

Catastrophic dam failures are rare but high-impact events, making reliable early warning systems essential for effective risk mitigation, particularly under severe data imbalance and limited failure observations. This study proposes a deep learning-based early warning framework for dam safety monitoring that integrates Temporal Convolutional Networks (TCN) with a temporal attention mechanism to model complex hydrometeorological time-series dynamics. The proposed approach learns normal operational behavior from historical data and identifies abnormal deviations using reconstruction-based risk scoring, enabling continuous monitoring without requiring labeled failure events. The framework is evaluated using multivariate hydrometeorological observations associated with the Sardoba dam system, which experienced a catastrophic failure event and provides a realistic safety-critical validation scenario. Experimental results demonstrate strong detection performance, achieving an accuracy of 0.968, a precision of 0.906, an F1-score of 0.951, a perfect recall (1.000), and a ROC–AUC of 1.000. Importantly, no hazardous states were missed, confirming the suitability of the proposed framework for early warning applications. Overall, the proposed TCN-attention model effectively captures both short-term fluctuations and long-range temporal dependencies, providing a robust data-driven solution for proactive dam safety monitoring.