Short-term river flow prediction provides a critical time window and decision support for preventing extreme hydrological events and water resource planning management. Due to the problems such as difficulty in capturing local features, sensitivity to hyperparameters, and poor prediction performance of the single long short-term memory (LSTM) model, a water flow prediction model integrating Transformer, LSTM, and pointwise attention mechanism is proposed. The data set is preprocessed using variational mode decomposition (VMD), and the hyperparameters are tuned using the Optuna optimization (OP) algorithm. A coupled model OP-Transformer-LSTM-Att is constructed. Taking the Chester site on the Mississippi River in Illinois, USA, as an example, the flow prediction accuracy of various models is compared and evaluated. The results show that the integrated model proposed in this paper reduces 53.7% and 46.55% in the root mean square error (RMSE) and mean absolute error (MAE) respectively compared with the traditional LSTM, and has the highest coefficient of determination (R2) of 0.9892. The prediction effect is significantly better than the non-integrated models and other algorithm models.