Trace selection method for a best representative in stochastic downscaling of precipitation

Abstract

Stochastic weather simulation models are used to generate data for hydrologic and agricultural applications. In recent years, stochastic downscaling has been employed in downscaling scenarios-based GCM or RCM output data, especially highly unpredictable hydrometeorological variables (e.g., precipitation). However, downscaling produces a number of traces to apply. In practice, multiple scenarios are difficult to handle. To overcome this difficulty, trace selection methods (TSMs) were proposed and tested in the current study, based on the mean and standard deviation (MS), empirical cumulative distribution function (Ecdf), and density estimate (Dens) methods. Simulation results indicated that key statistics were well preserved by the selected trace with the MS and Ecdf methods, while the Dens method did not show a reasonable performance in selecting a representative trace. The MS and Ecdf methods preserved well the mean and standard deviation of the average of these statistics from all simulated series and the Ecdf method was superior in preserving extrema, such as maximum and minimum. In a real application involving over 62 weather stations from South Korea, results showed that the MS and Ecdf methods were feasible ways to select the best trace to represent all temporally downscaled hourly precipitation from daily data. Overall, the MS and Ecdf can be both good alternatives as a trace selection method (TSM). The MS method can be used in general cases and the Ecdf can be employed in extreme analysis, such as estimating design rainfall with selected traces.