Extracting information from an ensemble of GCMs to reliably assess future global runoff change

SummaryFuture runoff projections derived from different global climate models (GCMs) show large differences. Therefore, within this study the, information from multiple GCMs has been combined to better assess hydrological changes. For projections of precipitation and temperature the Reliability ensemble averaging (REA) method was introduced to calculate weighted average ensemble change with an accompanying uncertainty range. In this study the original REA method is compared with three other methods that calculate runoff change by selecting or weighting GCMs on their inter-model similarity for the current and future climate. All methods are applied to distributed runoff fields calculated with the hydrological model PCR-GLOBWB forced with meteorological data from an ensemble of 12 GCMs.Differences between weighted and non-weighted average runoff changes for 2100 are small. Within a validation experiment, where GCM ensemble mean change is derived between the time-slices 1961–1975 and 1976–1990, the non-weighted ensemble change resembled observed change best. Yet, both the weighted and non-weighted average change were too conservative. This underscores the importance of considering an uncertainty range alongside the ensemble average change. In this study the uncertainty range (or 95% confidence interval) is defined by four times the root mean square difference around the ensemble mean change. The uncertainty range derived with the non-weighted method is relatively wide, upper and lower uncertainty bounds show large biases from observed change. However, the uncertainty range was reliably be reduced by using only a selection of GCMs which show high inter-model similarity for the current and future climate.

► Differences between GCM ensemble weighted and non-weighted changes are small.
► GCM ensemble average runoff change is too conservative.
► Ensemble mean runoff changes should be accompanied by uncertainty ranges.
► The Reliability Ensemble Averaging (REA) method reliably reduces uncertainty ranges.
► Weighting based on inter-model similarity for current and future climate shows best performance.