Abstract: Deep-water gravity flow reservoirs are characterized by strong heterogeneity and complex gas-water distribution, making accurate water saturation prediction critical for reservoir evaluation. However, the complex non-linear relationship between seismic elastic parameters and saturation makes it difficult for traditional methods to accurately characterize subtle fluid features. Deep learning offers a new approach to addressing this challenge. To this end, a seismic water saturation prediction method named TDCN-TransNet, based on the temporal dynamic convolution network (TDCN) and temporal context block (TCBlock), was proposed. This method utilized dynamic convolution kernels and improved coupled dynamic filters to extract highly informative local sequence features. It integrated the TCBlock into the Transformer architecture to capture global contextual information and long-range dependencies. Furthermore, a dual two-way time (TWT) positional encoding strategy was incorporated to enhance the model’s perception of the reservoir’s vertical spatial structure. The model was trained using elastic parameters as inputs and well-log water saturation as labels to establish a precise mapping between elastic parameters and water saturation. Field data testing results demonstrate that compared to DynamicTransformer, ISTNet, and mainstream machine learning algorithms, the TDCN-TransNet model achieves higher prediction accuracy (with an R² of 0.957 2) and stronger generalization ability in the test set. Application to actual 3D seismic data reveals that the predicted water saturation volume aligns well with well logs and geological laws. This study provides a novel approach and method for reservoir water saturation prediction, demonstrating significant practical value.