Multi-site downscaling of heavy daily precipitation occurrence and amounts

Abstract

This paper compares three statistical models for downscaling heavy daily precipitation occurrence and amounts at multiple sites given lagged and contemporaneous large-scale climate predictors (such as atmospheric circulation, thickness, and moisture content at the surface, 850 and 500 hPa). Three models (a Radial Basis Function (RBF) Artificial Neural Network (ANN), Multi Layer Perceptron (MLP) ANN and a Conditional Resampling Method (SDSM)) were applied to area-average and station daily precipitation amounts in northwest (NWE) and southeast (SEE) England. Predictor selection via both stepwise multiple linear regression and compositing confirmed vorticity and humidity as important downscaling variables. Model skill was evaluated using indices of heavy precipitation for area averages, individual sites and inter-site behaviour. When tested against independent data (1979–1993), multi-site ANN models correctly simulated precipitation occurrence 80% of the time. The ANNs tended to over-estimate inter-site correlations for amounts due to their fully deterministic forcing, but performance was marginally better than SDSM for most seasonal-series of heavy precipitation indices. Conversely, SDSM yielded better inter-site correlation and representation of daily precipitation quantiles than the ANNs. All models had greatest skill for indices reflecting persistence of large-scale winter precipitation (such as maximum 5-day totals) or dry-spell duration in summer. Overall, predictability of daily precipitation was greater in NWE than SEE.