Seasonal Runoff Forecasting Using Precipitation from Meteorological Data Assimilation Systems

Abstract

In semiarid mountainous regions such as central Asia, runoff from snowmelt often represents the dominant contribution to river flow and freshwater supply during the dry season. The estimation of snow accumulation during the preceding seasons then provides a key to seasonal runoff forecasting with lead times of a few months, and it requires appropriate coverage with surface precipitation and/or snow water equivalent observations. This study tests whether the lack of conventional precipitation and snow observations can be overcome by using model-based precipitation estimates from meteorological data assimilation systems. To this end, a detailed examination is undertaken of the ability of model-assimilated precipitation data to represent the interannual (year-to-year) variations of observed runoff in the Aral Sea basin in central Asia. Precipitation from the 15-yr Re-Analysis (ERA-15) of the European Centre for Medium-Range Weather Forecasts (ECMWF) for the period 1979?93 is compared against precipitation estimates derived from rain gauge networks, and against the observed natural runoff in the Syrdarya (166 400 km2) and Amudarya (320 520 km2) basins. It is demonstrated that the ERA-15 dataset is able?despite its low spatial resolution?to describe the seasonal cycle and the larger-scale geographical distribution of precipitation in central Asia. For the Syrdarya basin, it is found that December?April ERA-15 precipitation correlates well with observed May?September natural discharge. The correlation coefficient between the two time series amounts to r = 0.92. It is also demonstrated that ERA-15 precipitation is a better predictor for subsequent runoff than rain gauge?based precipitation analyses, presumably because of the poor coverage with rain gauge stations. The high correlations suggest that a reliable seasonal runoff forecasting system can be constructed from the statistical relationship between model-assimilated precipitation and subsequent runoff. Cross-validation hindcasting techniques are used to confirm this conclusion. A real runoff forecasting system would, however, require using a real-time precipitation product from an operational data assimilation system. For the Amudarya basin, the correlation between precipitation and subsequent runoff is substantially lower, presumably because of a lower quality of ERA-15 precipitation estimates within the tropical weather system, and/or due to a lower quality of the withdrawal-corrected runoff figures.