Abstract: Concrete dams play a crucial role in flood control, water resource allocation, and power generation. Affected by complex service environments, the durability of dam construction materials, extreme events, and operational scheduling, concrete dams still exhibit weak points or uncontrollable factors during their safe service. Deformation is an important response parameter reflecting the service performance of concrete dams, and accurately predicting the deformation response of concrete dams is of great significance for assessing the service status of dams. There are significant nonlinear characteristics between the deformation response of concrete dams and influencing factors such as water pressure, temperature, and time effects. By leveraging the dual-gate mechanism of the Gated Recurrent Unit (GRU), this study characterizes the nonlinear mapping relationship between influencing factors and deformation quantities, and mitigates the problem of gradient vanishing or explosion to improve the model's prediction performance. It introduces the Sine Cosine Algorithm (SCA) to construct a dual-thread iterative strategy for global search and local exploitation, achieving efficient optimization of the GRU model hyperparameter combination. On this basis, to address the potential small-sample problem in dam monitoring data, Generative Adversarial Networks (GAN) are adopted to generate high-quality simulated data, compensating for the limitations of deformation prediction with small datasets and further enhancing the generalization ability and accuracy of the prediction model. Taking the deformation prediction of typical measuring points of a concrete gravity-arch dam as an example, a GAN-SCA-GRU dam deformation prediction model is established and compared with other prediction models such as multi-layer perceptron (MLP) neural network, convolutional neural network (CNN), and bidirectional long short-term memory (BiLSTM) network. Specifically, the radial displacements at the dam crest of dam section 18# (mid-dam) and dam section 26# (dam abutment) are selected as the research objects. The study period includes a total of 117 data points, which constitutes a small dataset. The sizes of the training set, validation set, and test set are 95, 12, and 10, respectively. The results show that the optimization effect of the SCA algorithm varies with different original data and different models. For dam section 18#, the SCA algorithm has a relatively pronounced optimization effect on the CNN and GRU models, a moderate effect on the BiLSTM model, but an insignificant effect on the MLP model, and it is prone to falling into local optima. For dam section 26#, the SCA algorithm exhibits a certain optimization effect on all the adopted prediction models. For both dam sections 18# and 26#, the SCA-GRU model achieves the best overall prediction and fitting performance, followed by the SCA-CNN model. Taking the optimized SCA-GRU model as an example, the GAN method is first used to augment the input sample data before conducting prediction and analysis. The results indicate that the established GAN-SCA-GRU model can further improve the goodness of fit, while reducing indicators such as mean absolute error, mean relative error, and mean squared error, demonstrating that the generalization ability and robustness of the GAN-SCA-GRU model after data augmentation are further enhanced. The prediction model constructed in this study accurately characterizes the nonlinear mapping relationship between the deformation response of concrete dams and influencing factors, effectively improves the accuracy of safety monitoring and early warning of concrete dams, and can provide effective technical support for their safe service. It is suggested that when addressing specific problems, appropriate modeling methods should be selected through comparison in combination with the characteristics of response data. For the case of a small original dataset, optimizing sample augmentation methods and augmentation magnitude is an important direction for future research.