渗流分析是大坝安全与稳定的重点研究内容，通过构建高精度的土石坝渗流量预测模型对于大坝灾害风险管控具有重要意义。为了进一步提高土石坝渗流预测能力，提出了一种结合长短期记忆神经网络（long short-term memory, LSTM）、卷积神经网络（convolutional neural networks, CNN）和注意力机制（attention mechanism, Attention）的预测模型。该模型首先利用CNN挖掘数据的深层特征，然后通过LSTM提取渗流量监测数据的时间序列特征，最后将注意力机制添加到池化层和全连接层中，确定不同时间特征的重要性并分配权重。通过工程实例应用分析，与CNN、LSTM、CNN-LSTM模型相比，CNN-LSTM-Attention模型预测效果更好，其可决系数R²高达0.98以上，并且能够同时捕捉到渗流量数据的空间特征和时序依赖性，在土石坝渗流量预测中表现出了较强的可靠性与稳定性。

Seepage analysis is the key research content of dam safety and stability, and it is of great significance for dam disaster risk control by constructing a high-precision prediction model of seepage quantity for earth-rock dam. In order to further improve the seepage prediction capability of earth-rock dam, a prediction model combining long short-term memory neural(LSTM) networks, convolutional neural(CNN) networks, and attention mechanism (Attention) was proposed. Firstly, CNN was used to mine the deep features of the data, then the time series features of the seepage flow monitoring data was extracted through LSTM, and finally the attention mechanism to the pooling layer and the fully connected layer was added to determine the importance of different time features and assign weights. Through the application analysis of engineering examples, compared with CNN, LSTM and CNN-LSTM models, the CNN-LSTM-Attention model has better prediction effect, and its coefficient of determination R² is as high as more than 0.98, and it can capture the spatial characteristics and temporal dependence of seepage data at the same time, which shows strong reliability and stability in the prediction of seepage flow of earth-rock dam.