Research on hydrological simulation in karst areas coupled with VIC-machine learning model

In view of the difficulties in simulating hydrological processes under special hydrogeological conditions in karst areas, taking the sub-basin above Yangchang Station of Sanchahe in the karst landform area of the Wujiang River Basin as the study area, based on the construction of a VIC-FLASH hydrological model with a resolution of 1 km * 1 km, By coupling VIC-FLASH with the deep learning methods CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) models, feasible schemes for improving the simulation of hydrological processes in karst areas were explored. The results show that: (1) The coupling scheme of the VIC-FLASH model and the LSTM model improves the flow simulation accuracy during the dry season by optimizing the output of the hydrological model; Compared with the VIC-FLASH hydrological model that does not consider the coupling scheme, the relative error of the flow simulation is reduced by 10%. (2) The coupling scheme of VIC and CNN-LSTM takes advantage of the physical mechanism model and the hybrid architecture of CNN-LSTM, significantly improving the accuracy of flow simulation. The relative error of flow in the test set during the dry season is reduced by approximately 35%, and the error fluctuation is significantly improved. The research results provide an important reference for improving the hydrological simulation effect in karst areas.