A statistical downscaling procedure for improving multi-model ensemble probabilistic precipitation forecasts

The National Weather Service, via its Hurricane Sandy Supplemental-funded “Blender” project, is developing new methods for producing automated post-processed weather forecast guidance at high spatial resolutions using global multi-model ensembles and deterministic forecasts. Previously, the author has shown that for probability of precipitation forecasts, multi-model ensemble guidance is relatively skillful and reliable when verified against precipitation over the US on a 1-degree grid. However, similar guidance verified against much higher resolution precipitation analyses is not as skillful nor as reliable, suggesting that a primary problem with multi-model ensembles for such applications is the lack of downscaling to the scale of the verification data.

A statistical downscaling solution is proposed and tested here to improve skill and reliability of multi-model ensembles verified when against high-resolution analyses. Point by point, the multi-model ensemble members (and/or deterministic forecasts) are compared against precipitation analyses upscaled to the same grid. A list of dates is generated which represents the dates with the most similar past weather. An ensemble of the deviations of the fine-scale precipitation analyses from their upscaled mean is then multiplied by the ensemble of forecast precipitation amounts, thereby producing a much-higher resolution precipitation forecast with implicit statistical downscaling. At the conference, the impact of this method on precipitation forecast skill and reliability over the US will be discussed.