Simulation of Discharge Evolution in the Main Streams of Large Rivers Using a Deep Learning Model: A Case Study of the Yellow River, China

Abstract

Abstract Simulating river discharge evolution is fundamental to water resources management in large river basins. Nevertheless, human activities such as intensive water abstraction, reservoir operation, and other anthropogenic interventions have rendered the patterns of river discharge evolution highly complex, marked by significant spatiotemporal variability and nonlinear hydrological–anthropogenic interactions. Therefore, to improve the accuracy of large river discharge evolution, taking the Yellow River as an example, we developed a discharge evolution model by employing the convolutional neural network–long short-term memory (CNN-LSTM) model. Subsequently, through parameter optimization using the shuffled complex evolution–University of Arizona (SCE-UA) algorithm, the convolutional neural network-long short-term memory coupled with the shuffled complex evolution algorithm developed at the University of Arizona (CNN-LSTM-SCE-UA) model effectively mitigates the error propagation phenomenon between river reaches. Additionally, a comparison was conducted between the deep learning model and the physical mechanism model for river discharge evolution, aiming to explore the causes of simulation discrepancies from the perspective of physical mechanisms. Results showed that (1) the hydrological method, CNN-LSTM model, and CNN-LSTM-SCE-UA model achieved accuracies of 0.627, 0.613, and 0.774, respectively (evaluated via coefficient of determination, root-mean squared error, and mean absolute error; values closer to one indicate better agreement with observed data); (2) the spatial distribution pattern of unbalanced water volume—an external manifestation of the physical mechanism governing discharge evolution in the Yellow River main stream—followed an overall trend: middle reaches > lower reaches > upper reaches; and (3) for both the hydrological method and CNN-LSTM-SCE-UA model, simulation accuracy in the upper and lower reaches exceeded that in the middle reaches, consistent with the aforementioned spatial distribution pattern of unbalanced water volume.