Abstract: To predict the long-term leakage seepage volume of concrete-faced rockfill dams, a method based on Seq2Seq-Attention model was proposed. First, a Sequence-to-Sequence (Seq2Seq) model was adopted to construct the basic prediction framework. To evaluate the importance of each feature and enhance model interpretability, Pearson and Spearman correlation analyses were combined to calculate feature correlation coefficients, and the Extreme Gradient Boosting (XGBoost) algorithm was used to obtain feature importance indicators, comprehensively assessing the impact of different features on leakage. Building on this, a self-attention mechanism from Transformer was introduced to assign differentiated weights to features at different time steps, thereby mitigating interference from irrelevant information and improving the prediction performance for long-term leakage sequences. Evaluated comprehensively through metrics such as R2, MAE, RMSE, and MAPE, the model achieves values of 0.9951, 2.2257 L/s, 3.3963 L/s, and 3.9319%, respectively. Compared to Seq2Seq and Bidirectional Long Short-Term Memory models, these metrics show an average improvement of approximately 0.33% to 1.31%, demonstrating superior robustness and applicability. Furthermore, Depth-based Adaptive Kernel Density Estimation (DAKDE) is applied to derive the probability density function and probabilistic prediction results of model residuals, indicating that the 95% confidence interval achieves the best engineering balance, with coverage probability matching the theoretical value, moderate interval width, and balanced performance across all metrics, making it suitable for practical applications. The method proposed in this study enables high-precision point prediction and probabilistic prediction of leakage over long-term sequences, providing scientific support for the safe operation and management of reservoir dams.