<p>Accurate prediction of production dynamic parameters is important for optimising reservoir development and improving recovery. In this study, a composite deep learning model combining convolutional neural network (CNN), bi-directional long and short-term memory network (BiLSTM), and self-attention mechanism is proposed for predicting the dynamic parameters of reservoir production, especially daily oil production and water cut rate. CNN is used to extract the local temporal features, BiLSTM captures the long term dependencies, and the Attention mechanism enhances the model to pay attention to the key temporal patterns. Combined with multiple well studies, it is proved that the CNN-BiLSTM-Att model has a very high accuracy in the prediction of production dynamic parameters. Through the comparative analysis with traditional models and the prediction of single-well effects based on the evaluation indexes of MAE, MAPE, MSE, RMSE, and R², the results show that the composite model is excellent in both prediction accuracy and robustness are better than the existing models. This study provides valuable insights for intelligent reservoir management and a novel prediction method for solving complex reservoir production nonlinear dynamic problems.</p>