Abstract: To predict the long-term leakage and seepage volume of a concrete-faced rockfill dam, a new method based on the Seq2Seq-Attention model is presented. First, we adopt a Sequence-to-Sequence (Seq2Seq) model to construct a basic prediction framework. To evaluate the importance of each feature and enhance model interpretability, we calculate feature correlation coefficients by combining the Pearson and Spearman correlation analyses, and obtain feature importance indicators by applying the Extreme Gradient Boosting (XGBoost) algorithm, achieving a comprehensive assessment of the impacts of different features on the leakage. Then, a self-attention mechanism from Transformer is adopted to assign differentiated weights to the features at different time steps, so that the interference from irrelevant information is mitigated and the prediction of long-term leakage sequences is improved. A comprehensive evaluation－using the metrics of R2, MAE, RMSE, and MAPE－shows the model￠ calculations of 0.995, 2.23 L/s, 3.40 L/s, and 3.93%, respectively, indicating an average improvement of 0.33-1.31% with superior robustness and applicability, compared to the Seq2Seq and Bidirectional Long Short-Term Memory models. Additionally, we apply Depth-based Adaptive Kernel Density Estimation (DAKDE) to calculate the probability density function and probabilistic prediction results of the model residuals. From the results, the 95% confidence interval achieves the best engineering balance, and the coverage probability agrees with its theoretical calculation, with a moderate interval width and a balanced performance across all the metrics, thereby suitable for practical applications.

Key words: concrete-faced rockfill dam, leakage volume prediction, Seq2Seq-Attention, DAKDE