Figure 3 and Table 2 demonstrate that the LSTM model architecture can achieve good overall simulation performance for the reconstruction of target flows at the three hydrological stations, and can capture extreme flow conditions well. During both the train and test periods, the measured and reconstructed flow series at the three hydrological stations exhibited good fit, with MAPE values ranging from 2.09 to 8.70% and from 5.82 to 10.20%, RMSE values ranging from 1,206.7 to 1,308.8 m³/s and from 1,561.8 to 1,739.3 m³/s, and R² values ranging from 0.9870 to 0.9907 and from 0.9788 to 0.9852, respectively. Furthermore, the small differences in evaluation metrics between the training and testing periods indicate that the partitioning of the dataset did not affect the transmission of model uncertainty. Therefore, the LSTM model can be considered highly reliable and can be used to reconstruct natural flow regimes without the influence of the TGR.