Study on numerical simulation of sediment transport and prediction model of riverbed evolution based on deep learning

In this paper, a numerical simulation model of sediment transport and riverbed evolution prediction based on deep learning is proposed. The model adopts CNN-LSTM hybrid architecture, combines the advantages of spatial feature extraction and time series modeling, and integrates hydrological, sediment and terrain data through multi-source data fusion technology to realize high-precision prediction of sediment transport and riverbed evolution. The results show that the root mean square error (RMSE) of CNN-LSTM model in sediment transport simulation is reduced by 35%~53% compared with the traditional method, and the determination coefficient (R²) is increased to 0.93. The single simulation takes only 60 seconds, which is significantly better than the traditional hydrodynamic model (HEC-RAS) and empirical formula (Engelund-Hansen formula). In the prediction of riverbed evolution, the mean absolute error (MAE) of riverbed elevation predicted by the model is less than 0.25m, which is 82% consistent with the actual observation, and the average annual error growth rate is only 1.3%. In addition, the introduction of mass conservation constraint significantly inhibits error accumulation and enhances the long-term prediction ability of the model.