Ensuring the operational safety of dam structures under complex environmental conditions remains a formidable challenge in the fields of hydraulic and civil engineering. This study proposes a method based on the Temporal Fusion Transformer (TFT) for evaluating the quality of dam monitoring data, which provide a reliable foundation for subsequent dam safety assessment. By integrating structural feature decomposition, deep predictive modeling, and a multi-indicator scoring mechanism, the method enables intelligent identification of anomalies in dam monitoring data. First, the original deformation data are standardized and decomposed into three feature components: trend, oscillation, and periodicity (seasonality). Based on feature similarity metrics, monitoring points are classified into seven data categories. A multi-indicator data quality evaluation system is then applied to select high-quality datasets for model training. Second, a TFT prediction model is constructed by combining Prophet decomposition and multiple environmental factors, including water level, temperature, and rainfall. The Sea Horse Optimization (SHO) algorithm is employed to adjust hyperparameters and improve the generalization capability of the model. Finally, nine evaluation metrics are calculated, including the coefficient of determination, root mean square error, and normalized mean absolute percentage error. Three additional indicators—global fit, local precision, and prediction stability—are introduced to comprehensively assess model performance. A multi-indicator weighted scoring mechanism is then established to classify the safety levels of the dam monitoring points. Case studies demonstrate that the proposed method not only enhances prediction accuracy but also provides excellent interpretability and stability, offering strong support for practical dam safety evaluation.
Zhuo et al. (Mon,) studied this question.