Seismic Prediction Method for Reservoir Water Saturation in Deep-Water Gravity Flow Gas Fields Based on TDCN-TransNet

The prediction of water saturation in deep-water gravity flow reservoirs is a critical step in reservoir evaluation and production forecasting, as the highly heterogeneous and complex geological structures lead to uneven gas-water distribution. Traditional logging and seismic methods struggle to accurately capture the microscopic fluid distribution features, resulting in significantly low prediction accuracy for reservoir water saturation, which restricts reservoir evaluation and production forecasting. To address this, a seismic elastic parameters prediction method for reservoir saturation based on the Temporal Sequence Dynamic Convolution Network (TDCN) and Temporal Context Block (TCBlock) is proposed, leveraging the excellent sequence feature extraction and long-range dependency modeling capabilities of these models. The study results show that: (1) The TDCN module, based on dynamic convolution kernels, utilizes the CBAM attention mechanism and an improved coupled dynamic filter to extract high-information sequence features, enhancing the model's ability to represent complex heterogeneous reservoir structures. (2) The TCBlock module extracts global contextual information through bottleneck structures and SENet, achieving more precise long-range dependency modeling. (3) The dual TWT position encoding, by introducing sequence spatial positional information, further enhances the model's understanding of vertical reservoir structures and the modeling of long-distance nonlinear relationships, effectively reducing prediction errors for water saturation. The conclusion is: (1) The TDCN-TransNet model achieved a validation R² score of 0.9572, a root mean square error (RMSE) of 0.0304, a mean absolute error (MAE) of 0.0144, and a mean absolute percentage error (MAPE) of 2.09%, showing improvements of at least 0.0383 in R², a reduction of at least 0.0115 in RMSE, a reduction of at least 0.0088 in MAE, and a reduction of at least 1.04% in MAPE compared to the baseline model and other mainstream machine learning algorithms. (2) The TDCN-TransNet model achieved high consistency and accuracy in water saturation predictions in practical field applications, providing a new approach and method for reservoir water saturation prediction.