Evaluation of large-scale precipitation data sets for water resources modelling in Central Asia

Central Asia features an extreme continental climate with mostly arid to semi-arid conditions. Due to low precipitation and therefore low water availability, water is a scarce resource and often the limiting factor for socio-economic development. The aim of this model study was to compare the uncertainties of hydrological modelling induced by global and regional climate data sets and to calculate the impacts on estimates of renewable water resources. Within this integrated model study the hydrological and water use model Water Global Assessment and Prognosis 3 (WaterGAP 3) is being applied to all river basins located in Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan, Uzbekistan, and Mongolia in five arc minutes spatial resolution (similar to 6 9 9 km/grid cell). The model was driven by different global and regional climate data sets to estimate their impact on modelled water resources in Central Asia. In detail, these are the global TS data set of the Climate Research Unit (CRU), the WATCH forcing data (WFD) developed within the EU-FP6 Project "WATer and global Change", the Global Precipitation Climatology Centre (GPCC) Reanalysis product v6, and the regional Aphrodite's Water Resources data set (Aphrodite). The performance of the model is then being evaluated by comparing modelled and observed river discharge for the time period 1971-2000. Finally, the uncertainties in modelled water availability induced by the different data sets are quantified to point out the consequences for water management. Over the entire region, mean and maximum annual precipitations given by the various data sets differ by 13 % and up to 42 %, respectively. In addition, considerable deviation of temporal dynamics is found in some locations, where a pairwise comparison showed poor agreement between CRU and GPCC/WFD (Nash-Sutcliffe efficiency 0.27/0.23). Thus, also modelled discharge shows high temporal and spatial variations which leads to differences in median model efficiency of 0.11 between the data sets.