Does Increasing Horizontal Resolution Produce More Skillful Forecasts?

Abstract

This paper examines the impacts of increasing horizontal resolution on the performance of mesoscale numerical weather prediction models. A review of previous studies suggests that decreasing grid spacing to approximately 10 km orless generally produces more realistic mesoscale structures, with particular benefits for orographically and diurnally driven flows. There have been only a few long-term objective verification studies of highresolution forecasts, and these studies suggest, perhaps deceptively, that there are diminishing returns as horizontal grid spacing decreases below approximately 10 km. A multiyear objective verification of the University of Washington MM5 real-time forecastingsystem compares the realism of predicted surface parameters at 36-, 12-, and 4-km grid spacing over western Washington state for periods up to 48 h. Traditional verification statistics (such as mean absolute, bias, and root-meansquareerror) are calculated by interpolating model forecasts to the observation sites. For precipitation, it is shown that model skill over western Washington improves as grid spacing decreases from 36 to 12 km. Verification scores generally degrade as resolution is increased from 12 to 4 km as overprediction develops over the windward slopes and the crestsof terrain. However, for heavy precipitation amounts on windward slopes, the transition from 12 to 4 km does appear to enhance forecast accuracy. Temperature and wind statistics indicate noticeably improved skill as horizontal resolutiondecreases from 36 to 12 km, while only minor increases in skill are evident as grid spacing decreasesto 4 km. In contrast, verification of sea level pressure suggests little improvement as resolution is increased. The benefits of resolution are not uniform over the 4-km domain. For all parameters, the region downwind of a major barrier barrier—the Olympic Mountains—enjoys the largest enhancement of forecast skill as resolution is increased from 36 to 12 km, with noticeable, but lesser, improvements as grid spacing decreases to 4 km. Filtering the verification times to consider only periods with relatively good synoptic-scale forecasts provides only modest forecastimprovement. A case study illustrates that decreasing grid spacing to 4 km produces more realistic mesoscalestructures and amplitudes, a fact that is not revealed by traditional verification approaches because asmall timing error existed. The paper ends by examining the benefits of resolution, some pitfalls associatedwith traditional verification approaches, and suggests future directions for numerical weather prediction and mesoscaleverification. It is argued that higher resolution might be more beneficial than simple skill scores and that new mesoscale verification approaches must be crafted.