融合物理约束与深度学习的防洪调度模型研究及其应用OA

In response to the highly nonlinear and sudden characteristics of flood processes of the section from the Xiaolangdi Reservoir to the Huayuankou Hydrological Stationin the middle reaches of the Yellow River,as well as the computational complexity and lack of timeliness associated with traditional optimization scheduling models,this paper proposed a flood control scheduling model(CNN-BiLSTM-SA)that integrated physical constraints with deep learning.The model employed a one-dimensional convolutional neural network(1D-CNN)to extract local correlation features from hydrological time series,utilized a bidirectional long short-term memory(BiLSTM)network to model global temporal dependencies,and incorporated a self-attention mechanism(SA)to dynamically capture flood lag characteristics.To enhance the model's physical consistency,the principle of water balance was formulated as a physical constraint and incorporated into the loss function.Application results indicate that the model's simulation accuracy on the test set outperforms the baseline model:the Nash-Sutcliffe Efficiency(NSE)for outflow from Xiaolangdi Reservoir and flow at Huayuankou Station reaches 0.9171 and 0.9691,respectively,while the root mean square error(RMSE),mean absolute error(MAE),and mean absolute percentage error(MAPE)are all significantly reduced.It demonstrates the superiority of the hybrid network architecture in handling complex scheduling logic.This study leverages the bidirectional information propagation mechanism of BiLSTM and the dynamic weighting modeling of the self-attention mechanism to more accurately characterize the nonlinearity,long-term delay,and hysteresis of flood wave propagation on long-distance river channels.The proposed approach provides a modeling method for flood control scheduling simulation in basins that balances computational efficiency with physical consistency.