基于TOPMODEL-LSTM耦合模型的安墩水流域洪水模拟研究OA

This paper proposes a TOPMODEL-LSTM coupled model to address the difficulty in integrating physical characteristics of watersheds into Long Short-Term Memory(LSTM)neural network models.Based on hydrological data in the Andun River Basin from 2008 to 2018,the performance of runoff simulation over the 1～6 hour forecast periods under three input schemes was compared and analyzed.The results indicate that:① the coupled model achieved better forecast accuracy than the LSTM model in all forecasting periods,with average reductions of 15%in RMSE values and 13%in MAE values,and an increase in R2 from 0.78 to 0.83,which improved the accuracy of flood simulation to some extent;② in the simulation of hydrological processes in a watershed,the scheme of using hourly precipitation data and simultaneous runoff observations as the basic input data for the model was validated through multiple comparative experiments,and its simulation accuracy was significantly superior to that of a single runoff input or other combined input schemes;③ as the forecast period extended,the simulation accuracy of all models and their schemes decreased,but the coupled model effectively alleviated the memory decay issue of LSTM and exhibited more stable simulation performance.The research results can provide a technical reference for the coupling of traditional flood simulation models and artificial intelligence models.