Multi-level factorial analysis for ensemble data-driven hydrological prediction

Assessing the impacts of climate change on hydrologic regimes through hydrologic modeling is challenged by data uncertainty, predictor-selection uncertainty, and model uncertainty as well as their interrelationships. In this study, a Multi-level factorial ensemble data-driven hydrological model (MFEDHM) is developed to quantify the interactive, individual, and integrative impacts of multiple boundary conditions (e.g., climate conditions) on hydrological processes, and revealing the spatial heterogeneity of these impacts under various uncertainties and non-predictor impacts. In the MFEDHM, multi-level factorial analysis is integrated with ensemble prediction (i.e., Bayesian Model Averaging) and data-driven hydrological model. The MFEDHM is applied to quantitatively analyze the rainfall-runoff relationships of 16 catchments over China. Results reveal that the multilevel factorial analysis can accurately reveal both individual and interactive impacts of climate variables on hydrologic processes, and the impacts of non-climatic factors. As the most important finding of this study, climate-change impacts on hydrology show significant spatial heterogeneities over China. For instance, contemporaneous climatic conditions dominate (57%-64%) runoff changes and variations in Southern China, while precedent climate conditions pose significant impacts (20%-67%) on runoffs in Northern China; the overall influence of non-predictor factors (anthropogenic) on runoffs may decrease by 0.07% for the catchment-area increment of 10000 km2 and ranges from 4% to 27% over China. The development of the MFEDHM can enhance the reliability of ensemble hydrologic prediction, and provide scientific support for climate-change impacts assessment and adaptation under complexities.