Multi-site simulation of precipitation by conditional resampling

Abstract

A single-site, regression-based downscaling method is extended to multi-site synthesis of daily precipitation at stations in Eastern England and the Scottish Borders. Area-averaged precip- itation series for each region are downscaled using gridded predictor variables selected from a can- didate suite representing atmospheric circulation, thickness and moisture content at length scales of 300 km. Three simulation methods are compared: (1) a deterministic model in which daily precipita- tion occurrence and amounts are conditioned by atmospheric predictor variables (DET); (2) a hybrid model in which the unexplained variance of the deterministic model is represented stochastically (VAR); and (3) an unconditional resampling procedure (RND). Downscaled daily area averages are, in turn, used to resample daily precipitation amounts at multiple sites contributing to the unweighted areal mean. The temporal dependence of precipitation amounts at individual sites was explored using the standard deviation and autocorrelation of daily amounts, and N-day winter maxima. Spa- tial dependency was examined using inter-site correlations, correlation decay lengths, and Kendall's τb statistic for joint exceedance of N-day precipitation totals. The DET procedure underestimated the variance of daily amounts and N-day totals and over-estimated observed autocorrelation, leading to generally poor representations of spatial dependency. The RND procedure reproduced the distribu- tion of daily amounts and inter-site correlations, but yielded poor representations of N-day exceedances for large N. Overall, the VAR procedure was the most successful downscaling approach. Even so, suggestions are made for refinements to VAR in order to better capture seasonal variations in decay distances and pairwise correlations of multi-site precipitation amounts.