A flood forecasting method coupling the CE-QUAL-W2 and PINN models

• The CE-QUAL-W2&PINN coupled model was established to forecast flooding process. • We used the CE-QUAL-W2 model to correct the observed hydrologic data • The PINN model learns deeply about nonlinear variations in flooding The limitation and low accuracy of hydrological data seriously affects the accuracy of flood forecasting. To address the issue, this study proposes a novel flood forecasting method that combines the CE-QUAL-W2 model with the Physical Information Neural Network (PINN) model. Real time observation data correct using the CE-QUAL-W2 model, the corrected data were input into the Xin-An-Jiang (XAJ) model, the Long Short-Term Memory Neural Network (LSTM) model, and the PINN model, respectively. The predictive performance of the CE-QUAL-W2&PINN coupled model was comprehensively evaluated by six evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Nash Sutcliffe Efficiency Coefficient (NSE), Percentage Deviation (PBIAS), Kline Gupta Efficiency Coefficient (KGE), and Wilmot Consistency Index (WI). This study selects Luoma Lake as the study area, selecting 35 representative floods that occurred between 1960 and 2022. The results show that: (1) The CE-QUAL-W2 model simulated the water level and flow of 35 floods, and the R 2 between the simulation results and the observed values was greater than 0.71. (2) Four out of 35 floods were randomly selected for the analysis, and the results showed that the 3-hour predictive lead-time ( R 2 > 0.90) provided better forecasting compared to the 6-hour predictive lead-time ( R 2 < 0.87). The CE-QUAL-W2&PINN coupled model maintained its RE within 18 % for all 35 floods predicted within the 3-hour predictive lead-time, and the R 2 values were all greater than 0.8. (3) Based on a 3-hour predictive lead-time, the optimal values of the CE-QUAL-W2&PINN coupled model in forecasting water levels for six metrics are MAE = 3.894, RMSE = 5.721, NSE = 0.968, PBIAS = − 0.982, KGE = 0.963, and WI = 0.971, respectively. The optimal values for forecasting flow are MAE = 5.943, RMSE = 14.755, NSE = 0.954, PBIAS = − 0.723, KGE = 0.951, and WI = 0.978, respectively, which have higher forecasting accuracy compared to XAJ and LSTM models.