针对传统大坝渗压预测模型在复杂非线性时序建模与不确定性量化方面存在的不足，本文提出一种融合贝叶斯神经网络与长短期记忆网络的大坝渗压预测模型。该模型利用双层长短期记忆网络提取渗压监测数据中的长期时序依赖特征；引入注意力机制增强模型对重要时序信息与结构特征的表达能力；在输出阶段引入变分贝叶斯神经网络全连接层进行参数概率化建模。本文以新疆提孜那莆河莫莫克水利枢纽工程小时级监测数据集为实验数据，结果表明，所提出模型与基线模型LSTM模型相比，平均绝对误差MAE和均方根误差RMSE分别降低了74.1%和71.9%，R²提高了19.1%，误差显著降低，预测精度与稳定性方面均有所提高；贝叶斯推断输出的后验概率分布与置信区间能够较好覆盖真实观测值，实现预测结果的不确定性量化，为大坝安全风险评估提供了可靠的数据支撑。

[Objective] In order to address the shortcomings of traditional dam seepage pressure prediction models in modeling complex nonlinear time series and quantifying uncertainty, this paper proposes a dam seepage pressure prediction model that integrates Bayesian Neural Network (BNN) and Long Short-Term Memory (LSTM) network. [Methods] The model employs a two-layer Long Short-Term Memory (LSTM) network to extract long-term temporal dependency features from seepage pressure monitoring data. It introduces an Attention mechanism to enhance the model's ability to express important temporal information and structural characteristics. And it incorporates a variational Bayesian neural network fully connected layer at the output stage for probabilistic parameter modeling. Taking the hourly monitoring dataset of the Momoke Water Control Project on the Tizinafu River in Xinjiang as experimental data, model training and comparative experiments are conducted. [Results] The results show that, compared with the baseline LSTM model, the proposed model reduces MAE and RMSE by 74.1% and 71.9%, respectively, while increasing R² by 19.1%. The errors are significantly reduced, and both prediction accuracy and stability are improved. The posterior probability distribution and confidence intervals output by Bayesian inference can effectively cover the true observed values, achieving uncertainty quantification of the prediction results, thereby providing reliable data support for dam safety risk assessment. [Conclusions] The model was evaluated using one project's hourly monitoring data; its applicability to other projects or sampling intervals requires further validation.