Aiming at the problem that existing anomaly detection models for seepage monitoring data of earth-rock dams exhibit weak generalization capability and can only identify specific types of abnormal measurements, leading to low detection accuracy and high false-alarm rates, an anomaly detection model for earth-rock dam seepage monitoring data based on the graph attention network (GAT) mechanism and temporal convolutional network (TCN) is proposed. This model utilizes the GAT mechanism to assign weights to environmental factors, employs the TCN to improve the extraction of temporal features, and identifies abnormal data based on the model’s prediction error. Taking a clay-core rockfill dam in northwest China as a case study, six anomalous scenarios were constructed to validate the model. The results show that the proposed model can accurately identify abnormal seepage measurements of earth-rock dams, with average ROC-AUC and PR-AUC values of 0.948 and 0.968, respectively, meeting the requirements of practical engineering applications. Comparative analyses with state-of-the-art anomaly detection models further confirm that the proposed model exhibits superior anomaly detection performance and robustness.