Efficient and accurate performance of river runoff calculation is of paramount importance for water resources optimal operation. To effectively extract the characteristics of river runoff time series information and enhance the nonlinear fitting capability of river confluence process simulation and prediction, a river confluence prediction model (MABLFN) integrating Bi-directional Long Short-Term Memory network (Bi-LSTM), Multi-Head Attention mechanism, and Feed-Forward Neural Network (FFNN) is proposed. To validate the effectiveness of the MABLFN model, a case study was conducted using measured data from typical stations in the mountainous section of the Yongding River. And the prediction results of MABLFN model were compared and analyzed with those of standalone LSTM, Bi-LSTM models, and the MIKE11 model, assessing the predictive performance of the model for runoff processes across different forecast period. The results indicate that the MABLFN model can predict river runoff processes across different forecast lead times with improved accuracy. Compared to the standalone Bi-LSTM, LSTM, and MIKE11 models, the Mean Squared Error (MSE) is reduced by 2% to 76%, the Root Mean Squared Error (RMSE) is decreased by 1% to 51%, the Mean Absolute Error (MAE) is lowered by 34% to 42%, and the Nash-Sutcliffe Efficiency (NSE) is increased by 9% to 10%. In terms of computational time, the MABLFN model requires 1.2 seconds, which is slightly longer than the 0.26 seconds needed by LSTM and Bi-LSTM models but significantly shorter than the 360 seconds needed by the MIKE11 model.