Runoff simulation in Karst regions by integrating deep learning with physically-based hydrological models

The water cycle mechanism of karst regions is complex, posing a significant challenge to the collection of geologic data. Thus, further improvement of the runoff simulation performance of hydrological models in these regions is a research challenge that requires attention. By coupling an improved EXP-HYDRO model, a two-layer NRIHM model, and a recurrent neural network algorithm, a hydrological model was developed that integrates the physical mechanisms and artificial intelligence (AI). Simulation and validation of the model were performed for two karst regions: the Nanyang River basin in the middle reaches of the Yangtze River and the Yeji River basin in the middle reaches of the Wujiang River. The findings suggest that the coupled hydrological model can significantly improve the simulation accuracy of the daily runoff process in karst areas. Compared with conventional hydrological models, the coupled model can improve the Nash-Sutcliffe efficiency by approximately 19.7% in the Nanyang River basin and 37.1% in the Yeji River basin. Moreover, owing to the learning of physical mechanisms, the coupled model exhibits better simulation accuracy than purely data-driven approaches, suggesting that the comprehensive prediction performance of AI models can be effectively improved by properly selecting and learning different physical mechanisms.