Evaluating the downscaling uncertainty of hydrometeorological data in snowmelt runoff simulation

Abstract

Snowmelt runoff is the main water source in cold and arid areas, and hydrological models provide a useful tool for water resource management in such areas. The downscaling of hydrometeorological data is an important method of obtaining input data for hydrological models. In this study, the snowmelt runoff of the Cele River basin was simulated using the snowmelt runoff model (SRM), and large-scale precipitation and temperature data were downscaled and used as the input of the SRM. To evaluate the downscaling uncertainty in snowmelt runoff modeling, four commonly used downscaling methods were selected, and Bayesian stacking was used to reduce the downscaling uncertainty by combining the predictions from these downscaling methods. Additionally, Markov chain Monte Carlo simulation was conducted to calibrate the model parameters and generate predictions for the four SRMs, corresponding to the four downscaling methods. Four metrics were used to evaluate the performances of the four SRMs and Bayesian stacking in daily and monthly runoff predictions. The results demonstrated that the performances of the four downscaling methods in runoff prediction differed, and none of the downscaling methods were superior to the others in runoff prediction for all evaluation metrics. Model parameter and downscaling method contributed 45.62% and 54.38% of the predictive uncertainty in runoff prediction, respectively. Thus, downscaling could lead to non-negligible uncertainty in snowmelt runoff modeling. Bayesian stacking achieved good and reliable performance in daily and monthly runoff predictions and effectively reduced the uncertainty of snowmelt runoff modeling.