A machine learning model for flood forecasting with integrated structural enhancement mechanisms

Key Points

• The aim is to develop a machine learning model for flood forecasting that integrates structural enhancement mechanisms to improve accuracy under complex conditions.
• Developed the ML-P-EF model incorporating structural features like runoff process vectorization, dynamic lag encoding, and event-driven characteristics.
• Validated the model on four typical river basins in the middle Yellow River region.
• Utilized foundational model structures (LSTM, ANN, Transformer) to create process-enhanced and fully enhanced models for 1, 3, and 6-hour forecasting.
• The ML-P-EF model achieved a Nash efficiency coefficient of 0.937 at a 6-hour forecast period, significantly higher than traditional models.
• RMSE decreased by approximately 62.77%, leading to better overall forecast accuracy.
• Peak error reduced by an average of 56.74%, showing a marked improvement in flood peak forecasting.