A surrogate model for the Variable Infiltration Capacity model using deep learning artificial neural network

•A surrogate model for VIC is builded with SKC and LSTM.•The surrogate model alleviates the computing burden of VIC with only slight losses of model fidelity.•The surrogate model keeps partial spatial information. The Variable Infiltration Capacity (VIC) model is a widely used distributed hydrological model. However, VIC is computationally expensive in hydrologic prediction or forecast, which needs tens of thousands of runs of the model. To alleviate the burden of computation and reduce the losses of model fidelity, a new surrogate model (SM) coupling the self-organizing map and K-means clustering algorithm (SKC) with long short-term memory network (LSTM) is proposed. SKC is utilized to divide the subwatershed and select the representative cells in each subwatershed. LSTM is used to simulate the streamflow with the runoff of the representative cells. The new model is successfully applied in the Upper Brahmaputra River (UBR) basin, Southeast China. The results show that the SM-simulated streamflow has little difference from the VIC-simulated streamflow in terms of the Nash-Sutcliffe coefficient efficiency, the main metric of SM performance, 0.9677 at Yangcun Station and 0.9696 at Nuxia Station. For the representative cells, SM retains partial spatial information of the runoff in the study area. The computational savings achieved through the use of SM are over 97% with only slight losses of accuracy in the application to the UBR basin.