Conditioning of Flow Projections under Climate Change on Hydrologic Signatures within the GLUE Framework

Abstract

Climate change impact on water resources is generally quantified in terms of relative changes in characteristicflows (e.g. annual runoff, median annual flows, etc.) over a future period compared to the baseline one. Thesechanges are estimated under the assumed emission scenarios and with one or more modelling chains (combinationsof the Global and Regional Climate Models, and a hydrological model). Since different modelling chains yielddifferent projections, estimates of these relative changes are uncertain. High prediction uncertainty is reflected in awide 90 per cent prediction uncertainty band (90PPU) or in a distribution that resembles the uniform distribution.Therefore, research in robustness of the modelling chains has been conducted. The goal of the research is toappoint higher probabilities to the projections obtained by the more robust chains, and in that way reduce theuncertainty in flow projections under climate change.In this research, the hydrologic projections are conditioned on the hydrologic signatures within the GLUEframework. Namely, a relative change obtained with a modelling chain is assigned a likelihood depending on theperformance of the chain in terms of the hydrologic signatures over the baseline period. High flow projections (2ndpercentile of the daily flows) are conditioned on the high-segment of the flow duration curve (FDC), projectionsof the median flows are conditioned on the FDC mid-segment slope, and the projections of the low flows areconditioned on the FDC low-segment. The projections of total annual runoff are conditioned on the entire FDC.The likelihoods are quantified in terms of Nash-Sutcliffe efficiency coefficient (NSE) evaluated from the FDCs ofthe flows simulated by the modelling chains and the observed FDC.The methodology presented is applied to develop flow projections in the Kolubara River catchment in Ser-bia over the mid 21st century (2041-2070). Hydrologic projections are obtained by the HBV-light hydrologicmodel with input from five climate models (combinations of the Global and Regional Climate Models), all rununder A1B emission scenario. The outputs of the climate models are bias-corrected to reproduce distributions ofthe precipitation depths and temperatures observed in the baseline period (1961-1990). The best twenty parametersets out 25,000 sampled ones are kept in the analysis, resulting in 100 modelling chains.The GLUE conditioning did not significantly affect the median values of the projections, but only thewidth of the prediction bands. All modelling chains perform equally well in terms of the entire FDC and itsmid-segment. Therefore, the GLUE-conditioned projections of annual runoff volume and median flows are similarto the unconditioned ones (i.e. GLUE conditioning yields slightly narrower 90PPUs). Model efficiency in the highflow domain differs between the modelling chains: however, the GLUE conditioning leads to somewhat narrower90PPU. Only few modelling chains performed well in the low flow domain, therefore the width of the 90PPU wasconsiderably reduced by conditioning (from 61.5% to 36.8%)