Short-term river flow forecasting provides a critical time window and decision-making support for disaster prevention, mitigation, and sustainable utilization of water resources. To address the limitations of the single long short-term memory (LSTM) model, such as difficulties in capturing local features, sensitivity to hyperparameters, and poor prediction performance, a novel river discharge prediction model integrating convolutional neural network (CNN), LSTM, and a self-attention mechanism was proposed. The Bayesian optimization (BO) algorithm was employed for hyperparameter tuning, resulting in the construction of a coupled BO-CNN-LSTM-Att model. Taking the Yangtze River Basin in Jingzhou, Hubei Province as a case study, the prediction accuracy of various models was compared and evaluated. The results demonstrate that the proposed integrated model reduces the root mean square error (RMSE) and mean absolute error (MAE) by 51.55% and 53.56%, respectively, compared with the traditional LSTM model, and achieves the highest coefficient of determination (R²) of 0.972 8. The prediction performance is significantly superior to that of incomplete integrated models and other algorithmic models.