| To address the structural rigidity and peak flow forecast deviations in existing machine learning models for flood forecasting under complex rainfall patterns and delayed response conditions, this study proposes a machine learning model for flood forecasting integrated with a process-enhanced mechanism (ML-P-EF). This method introduces three types of structural features: Runoff Process Vectorization, Dynamic Lag Encoding, and Event-Driven Features, which process the input data for flood forecasting from the aspects of process structure, time-lag response, and event attributes, respectively. Using four typical watersheds in the middle reaches of the Yellow River as validation cases, and based on three fundamental model structures (Long Short-Term Memory (LSTM), Artificial Neural Network (ANN), and Transformer), we constructed three process-enhanced models and three full-structure-enhanced models to conduct flood forecasting with lead times of 1h, 3h, and 6h. The results demonstrate that under the 6h lead time condition, the ML-P-EF model improved the average Nash-Sutcliffe Efficiency (NSE) by 0.725 and 0.188 (increasing from the basic model ML's 0.212 and the process-enhanced model ML-P's 0.749 to 0.937), reduced the Root Mean Square Error (RMSE) by approximately 62.77%, and decreased the peak flow error by an average of 56.74%. Taking the Daning Station as an example, the NSE of the LSTM-structured model increased from 0.165 (basic model) to 0.775 (process-enhanced model) and further to 0.982 (full-structure-enhanced model), while the RMSE decreased from 46.11 m3/s to 23.96 m3/s and then to 6.75 m3/s, and the peak flow error changed from -68.41% to -41.48% and finally to +3.32%. The ML-P-EF full-structure-enhanced model developed in this study significantly outperforms the basic model in peak response, temporal fitting, and error control, particularly demonstrating stronger generalization capability and stability under the 6h lead time condition. The research findings provide a new structural awareness modeling approach for watershed flood forecasting. |