Abstract: Accurate flood forecasting and real-time correction are essential for modern flood prevention and mitigation. Taking the Sandaohezi hydrological station in the Luan River Basin as an example, this study develops a hybrid PHY-FTMA-LSTM flood forecasting model by incorporating attention mechanisms and physical constraints into a Long Short-Term Memory neural network. The K-Nearest Neighbors algorithm is then applied for real-time dynamic correction of forecast sequences, and the performance changes of different models after KNN correction are evaluated. The results show that KNN correction effectively improves model accuracy, reduces errors in peak flow and peak timing, and suppresses anomalous oscillations during flood process simulation. This method provides a reference for understanding flood process characteristics in the study area and has practical value for improving flood forecasting in small and medium-sized rivers, enhancing warning timeliness, and supporting disaster prevention decisions.