To address the issue of low prediction accuracy caused by uncertainties in the selection of factors and construction of the seepage safety monitoring model for pumped storage power station dams, this study integrated deep learning models with probabilistic prediction methods. By incorporating the feature extraction capability of convolutional neural network (CNN), the data mining potential of bidirectional gated recurrent units (BiGRU), the parameter optimization advantage of the dung beetle optimization (DBO) algorithm, and the probabilistic prediction capability of quartile regression (QR), a probabilistic dam seepage prediction model based on DBO, CNN, BiGRU, and QR was established. At the same time, to construct an optimal factor set suitable for the seepage safety monitoring model for pumped storage power stations, the lag effect of seepage was fully taken into account, and the kernel principal component analysis (KPCA) was adopted to optimize the influencing factors of the model. Engineering case studies demonstrate that the established probabilistic dam seepage prediction model can not only provide high-accuracy deterministic prediction results of dam seepage pressure, but also yield corresponding probabilistic prediction intervals to reflect the uncertainty of seepage changes, which can provide more comprehensive evaluation information for seepage safety monitoring of pumped storage power station dams.