Slope deformation prediction is an important research field in slope safety analysis of hydraulic projects. The Global Navigation Satellite System (GNSS) technology is one of the main approaches for slope deformation monitoring, its data quality and prediction accuracy are crucial for slope safety evaluation. In response to the issues of noise processing in GNSS data and high-precision prediction of deformation sequences, a denoising and prediction model was proposed based on variational mode decomposition (VMD), convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM) networks and attention mechanism (AM). To improve data quality, VMD is employed to filter and denoise the original deformation sequence; CNN is utilized to extract time-varying features from the processed data, which is then fitted by BiLSTM to obtain the deformation values based on the historical and future informationin of the time series. To optimize the model parameters and thereby improve prediction accuracy, the simulated annealing (SA) optimization algorithm is introduced for the optimal analysis of important parameters in the deep learning network. Subsequently, the output of BiLSTM model is processed in the fully connected layer of AM to get the final optimized prediction results. The prediction results of a hydraulic engineering slope based on GNSS surveillance data demonstrate that the proposed hybrid model achieved an average enhancement of 63.92%, 62.06%, and 89.10% in prediction accuracy in three dimensions(h, x, y), compared with the classic deep learning ensemble models. This model can offer a novel modeling approach for the analysis of GNSS surveillance data in hydraulic engineering slopes.