The comparative study of machine learning agent models in flood forecasting for tidal river reaches

The tidal reach of the Tanjiang River Basin, influenced by both runoff and tidal forces, exhibits complex and variable water levels and river flow conditions. Traditional hydrodynamic models struggle to meet the timeliness requirements for flood forecasting and early warning. Moreover, existing machine learning studies have limitations in integrating hydrological principles and explaining model applicability. This study focuses on this region, constructing a one-dimensional hydrodynamic model based on Saint-Venant's equations to simulate flood-tide evolution. Using the model's results, three typical machine learning surrogate models-Long Short-Term Memory (LSTM), Random Forest (RF), and Support Vector Machine (SVM)-are developed to predict the water level at the key control section of Changsha Station. The results show that the surrogate models offer advantages in both timeliness and convenience. Among them, the RF model has the highest prediction accuracy and can accurately identify key hydrological influencing factors. The LSTM model effectively captures the temporal dependencies of hydrological factors, considering the influence of historical moments on current water levels. However, the SVM model performs poorly under extreme flood-tide conditions. SHAP analysis reveals that the feature contribution values in both LSTM and RF models better align with the actual hydrology of the basin. Incorporating hydrological principles into the selection of input features helps improve model accuracy, while employing Stacking model to combine the strengths of different models further enhances prediction precision. This research provides a scientific basis and effective reference for the application of machine learning models in flood forecasting for tidal reaches, with important implications for advancing smart water conservancy construction.