Multi-site precipitation downscaling via an expanded conditional density network

Abstract

Statistical downscaling models are used to estimate weather data at a station or stations based on atmospheric circulation data defined at a coarser resolution, for example gridded outputs from a Global Climate Model (GCM). Downscaled data can be used as inputs to environmental models that require finer-scale climate fields than are currently available from GCMs. Maintaining realistic downscaling relationships between sites and variables is particularly important in hydrological models, as streamflow depends strongly on the spatial distribution of precipitation in a watershed and on interactions with temperature that determine whether precipitation falls as rain or snow. More generally, precipitation is a difficult variable to downscale because of its non-normal distribution and its spatial and temporal patchiness.A downscaling algorithm for daily precipitation series at multiple stations is presented. The expanded conditional density network (ECDN) models the conditional density of the Poisson-gamma distribution via an artificial neural network. ECDN is capable of (1) specifying the conditional distribution of precipitation at each site; (2) modeling occurrence and amount of precipitation simultaneously; (3) reproducing observed spatial relationships between sites; (4) randomly generating synthetic precipitation series; and (5) predicting precipitation amounts in excess of those in the observational record. The ECDN model is applied to two downscaling problems: the first is a benchmark precipitation downscaling task previously evaluated by other modeling groups; the second is a multi-site precipitation dataset from British Columbia, Canada. Results suggest that the ECDN approach is capable of generating spatially and temporally realistic precipitation series that are suitable for use as inputs to hydrological models or to spatial interpolation schemes.