Dynamic runoff simulation in a changing environment: A data stream approach

In this study, we introduce a data stream method for dynamic runoff simulation, which allows capturing the evolving relationship between runoff and its impact factors (e.g., temperature, rainfall). The basic idea is to view continuously arriving data of runoff and its impact factors as a data stream, and dynamically learn its relationship. To validate the effectiveness of the proposed method, we compare its performance with that of three data driven models (ANN, SVR, Random Forest) and six representative hydrological models (SWAT, AWBM, SimHyd, SMAR, Sacramento, and Tank) in simulating monthly runoff. The proposed method performs well with the best NSE of 0.88, being superior to comparable models. Furthermore, the data stream model also shows its advantage in the flexibility of combing various impact factors of runoff into the model. The findings demonstrate that the data stream method provides a promising way to dynamically simulate runoff in a changing environment.