Holistic uncertainty quantification and attribution for real-time seasonal streamflow predictions: Insights from input, parameter and initial condition

This is the authors' abstract. We don't add key points for this paper.

Two mesoscale watersheds in the East China, where increased occurrences of hydrometeorological extreme events (drought-flood abrupt alternation) were observed over the last few years. To investigate the impact of different uncertainty sources in seasonal streamflow predictability with lead times from + 1 M to + 9 M, the study proposed an uncertainty quantification and attribution framework leveraged by a three-way ANOVA by considering uncertainties from forcing inputs (CFSs), hydrological model parameters (PARs) and basin initial conditions (ICs). The annual and sub-annual variations in predictability and uncertainty contribution were analyzed and elaborated through four well-designed experiments, by using both deterministic and uncertainty metrics. The results show that CFSs-driven monthly streamflow predictions are effective up to + 3 M. Additionally considering PARs uncertainty enhances the predictability for all lead times, whereas additional consideration of ICs uncertainty exerts a positive impact on streamflow prediction only in shorter lead times (<+2 M). Uncertainty contributions vary in lead times and seasons. For the + 1 M predictions, the uncertainty is influenced by complicated contributors, with 47 % from CFSs, 18 % from ICs, and 17 % from PARs, respectively. Regarding lead times beyond + 1 M, CFSs and PARs are the dominant contributors, with overall contributions exceeding 90 %. Seasonally, ICs and PARs play a dominated role during low-flow period, nevertheless CFSs more pronounced during high-flow period. Our finding provides insights for modelers into the relative impact of individual uncertainty sources in the hydrological forecasting process, and supporting valuable suggestions for related applications. • A holistic uncertainty quantification and attribution is developed for real-time seasonal streamflow prediction. • Uncertainties in CFSs, PARs and ICs and their synergy are analyzed and quantified for sub-annual variations. • Effects of ICs are substantial for lead times <= +3M, for longer lead times the CFSs uncertainty dominated. • Valuable suggestion has been made for regional water resources management.