Floodforecast of Mianhuatan Reservoir based on mechanism anddeep learningmodel

Hydrological forecasting plays a key role in water resource management and flood mitigation. While physically-based models and artificial intelligence models each have their own limitations, combining their strengths to achieve both accuracy and physical interpretability has become a key focus in hydrological forecasting. This paper constructs a coupled hydrological model of the Xin'anjiang (XAJ) model and Long Short-Term Memory Network (LSTM) for the Mianhuatan Reservoir. In this framework, the XAJ model provides the runoff generation module, leveraging its strength in representing physical processes such as evapotranspiration and three-source water separation. The LSTM network, known for its capacity to model nonlinear temporal patterns, forms the routing module. A multi-objective genetic algorithm is introduced for parameter optimization to reduce the uncertainty associated with manual calibration. The model is evaluated using 75 flood events observed at the Mianhuantan Reservoir between 2022 and 2024. Results for representative flood events show Nash-Sutcliffe Efficiency (NSE) values ranging from 0.92 to 0.99, peak flow errors between 1.08% and 6.52%, and peak timing errors within 2-3 hours. Compared to the standalone XAJ model and LSTM model, the coupled XAJ-LSTM improves NSE by 0.24 and 0.03, reduces peak flow error by 5.1% and 1.2%, and shortens peak timing error by 2.75 hours and 0.5 hours, respectively. These findings suggest that the coupled model more effectively captures complex flood dynamics and offers a promising approach for improving flood forecasting in challenging basins.