Quantifying uncertainty sources in hydrological climate-impact projections with Hydrological Ensemble Prediction System and ANOVA

Collections of hydrological model output, called ensembles forecast, can be used to explore the uncertainties associated with projections of climate change impacts on hydrology. Uncertainty in hydrological climate-impact projections stems from multiple sources, including climate model (CM), the bias correction method (BM), the hydrological model (HM) ,the parameter value (PM),and so on. To disentangle each uncertainty component and assess how each of them contributes to the total uncertainty helps to better understand and quantify the sources of variability in hydrological forecasting. This study is intended to quantify the uncertainty due to four different components (CM, BM, HM and PM) and their interactions via a case study involving the upper Pearl River basin in South China. A hydrological ensemble prediction system (HEPS) was developed with the purpose of combining information from an ensemble of three precipitation datasets, three bias correction methods, three distributed hydrological mode structures and three sets of model parameters. We construct an Analysis of Variance model (ANOVA) based on the hypothesis that the CM, BM, HM and PM have an influence on the correlation coefficient, to allow us to split the total sum of the squares (SST) into sums of squares due to the individual effects and their interactions. The result of different uncertainty sources contributing to the total uncertainty depicts in Figure 1. The CM and PM generally are the dominant source all year round. During wet days, the biggest uncertainty is observed due to the PM, whereas during dry period, the uncertainty due to the CM gains importance and partly dominates. In addition, in distinction to some previous researches, the results indicate the individual uncertainty from the BM is not attributable to the total ensemble uncertainty, but the interaction effect among the CM and BM contributes a considerable percentage, account for about 5%-20%. None of the targeted uncertainty sources are negligible and more attention should be focused on interaction study.