Machine learning prediction model of river water level considering riverbed topographic change

In order to establish a machine-learning water level prediction model applicable to river sections with large changes in riverbed topography, the Nanchang River section of the Ganjiang River with large topographic changes from 2002 to 2012 as the research object was selected, and adopting the overflow area of the river at different times of the day to characterize the changes of the riverbed topography, together with the hydrological data of the river as the input features, the water level of the Waizhou station with the forecast period of 1-3 days was predicted, respectively. The results showed that the proposed long short-term memory (LSTM) model had a good performance in predicting the water level at Waizhou, and the NSE and R of predicting the water level from forecast period of 1-3 days reached more than 0.977 and 0.987, respectively. Especially the highest accuracy of predicting the water level from day 1 of the forecast period, and the evaluation indexes of RMSE and MAE were 0.063 m and 0.043 m, respectively, and the NSE and R were both 0.999 with. RMSE and MAE of the predicted water level were reduced by 14.9% and 10.4%, respectively, compared with the results of the model prediction without considering the influence of topographic changes. Through the comparative analysis of different input feature schemes, it shows that for the river channel with large topographic changes, adding the topographic features of the river channel in the model input can help the model learn the mapping relationship between topographic changes and water level changes, and improve the prediction accuracy of the model for water level.