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

This study focuses on improving flood forecasting in the tidal reaches of the Tanjiang River Basin in China. Traditional hydrodynamic models, though accurate, are too slow for real-time flood prediction. To address this, the researchers combined a one-dimensional hydrodynamic model with machine learning surrogate models—LSTM, Random Forest (RF), and Support Vector Machine (SVM)—to simulate and predict water levels at key control points. The study evaluates these models’ performance under various flood and tide scenarios, emphasizing accuracy, interpretability, and computational efficiency.

The RF model achieved the highest prediction accuracy and interpretability, while the LSTM model effectively captured temporal dependencies but required precise lag features. The SVM model underperformed in extreme conditions. A stacking ensemble combining LSTM and RF, with XGBoost as the meta-learner, outperformed all individual models, especially under flood-tide superposition. The integration of hydrological principles with machine learning inputs enhanced model reliability, offering a robust tool for real-time flood forecasting and smart water management.