Flow simulation based on MISO and LSTM models in the Yangtze River-Dongting Lake System

The study region is the Yangtze River-Dongting Lake system, the largest regulated river-lake system in China. Flow simulation of the Yangtze River–Dongting Lake system has long been recognized as a critical and challenging issue in the Yangtze River basin. Current approaches rely on traditional hydrological and hydrodynamic methods, but they face limitations such as difficulty obtaining required data set, complex modeling processes, long-time consumption, and insufficient accuracy. This study proposed the multi-input and single-output (MISO) model based on system dynamics principles and the long short-term memory (LSTM) neural network model based on deep learning for flow simulation in the Yangtze River-Dongting Lake system. The proposed MISO and LSTM models can achieve accurate flow simulation results using only recorded flow data and rainfall data in the Yangtze River-Dongting Lake system. Both models demonstrate significantly higher computational efficiency than traditional hydrodynamic models, the MISO model can complete simulation within one second, while the LSTM model has the higher Nash-Sutcliffe efficiency coefficients and the smaller relative errors during the training and validation periods, the peak discharge relative errors (PRE) is ranging from −1.40–1.81 % across twelve flood events. This study has provided a novel approach for flow simulation in the complex river-lake system. • The MISO and LSTM models perform well in terms of computational efficiency and simulation accuracy. • The MISO model achieves fast computational efficiency with an average simulation time of 0.003 s. • The LSTM model performs well with the highest simulation accuracy and the smallest relative error. The peak discharge relative errors (PRE) are between −1.40 % and 1.81 % for flood events.