The Utah Snow Ensemble

We are excited to share that the Utah Snow Ensemble is now available on https://weather.utah.edu and has replaced the old NAEFS product.

The Utah Snow Ensemble is an 82-member ensemble for predicting snow over the contiguous western United States based on the ECMWF ensemble (ENS) and the US National Centers for Environmental Prediction Global Ensemble Forecast System (GEFS), with 51 members coming from the ENS and 31 from the GEFS.  The forecasts are provided in 6-h intervals out to 240 hours (10 days).  A summary of the data and methods and graphics is provided below. 

Precipitation Downscaling

Precipitation from all ENS and GEFS members is downloaded on a 0.25°x0.25° grid and downscaled to 800-m grid spacing to provide higher resolution guidance.  The downscaling assumes seasonally varying climatological precipitation-altitude relationships.  See Lewis et al. (2017) for more information.  Such a technique produces improved forecasts, but it does not account for periods during which the strength of orographic precipitation enhancement departs strongly from climatology.  

Snow-to-Liquid Ratio (SLR)

Snow-to-liquid ratio (SLR) is based on a new multiple linear regression algorithm developed using SLR observations from 14 western US snow study sites (thank you to everyone who provided that data!) and temperature and wind profiles from the ERA5 reanalysis.  Although we have developed more sophisticated machine learning algorithms, they were too slow for processing more than 3000 ensemble members and forecast hours, with the small accuracy loss and large speed gain of multiple linear regression a reasonable compromise. 

Snow Level

The multiple linear regression derived SLR is used where the high-resolution terrain used for downscaling is at or above the 0.5°C wet-bulb temperature level.  Below that level, we assume the SLR decreases linearly to 0 (i.e., becomes rain) over a distance of 200 m toward the ground, as illustrated schematically below.  Everything below that level is assumed to be rain.  

This approach is based on snow-level estimation techniques used by the National Weather Service Western Region (click here to access the Technical Report).  It is straightforward and fast.  The technique will struggle, however, when there are warm noses aloft (an issue at times in the Columbia Basin and Cascade Mountains, especially the Columbia Gorge and mountain passes), if the decrease in temperature with height is small or zero (e.g., isothermal), or if there are dense hydrometeors like graupel that can penetrate farther down through the transition zone.  We make no effort to identify such periods or the possible presence of freezing rain or sleet/ice pellets instead of snow when warm noses are present aloft. Improving this method over the west will probably require some sort of machine learning method since the the coarse vertical resolution of temperature data provided for the ENS and GEFS limits our ability to take a direct thermodynamic approach.  This is a subject for future work.  

Four-Panel Plots

We provide loops of four-panel plots of the following variables for several regions: 

• Total precipitation (water equivalent) since the beginning of the forecast period
• Total snow since the beginning of the forecast period
• 24-h precipitation (water equivalent)
• 24-h snow
• 6-h precipitation (water equivalent)
• 6-h snow
• Wet-bulb 0.5°C height above ground level
• SLR

SLR and snowfall are calculated in 6-h intervals, with the resulting 6-h accumulations summed to provide accumulations over longer periods.  This is similar to what you would get if you were measuring snow on a freshly wiped snowboard every six hours and adding it up.  Thus, the 24- and total snowfall should not be confused with the change in snow depth on the ground over long time periods, which would be affected by settlement.  

Each four panel plot includes the the downscaled ENS control forecast at upper left, the downscaled ensemble mean at upper right, the downscaled ensemble minimum at lower left, and the downscaled ensemble maximum at lower right.  Below is an example of the total snowfall through 240-h over northern Utah. 

The choice of the downscaled ECMWF ENS control for upper-left is somewhat arbitrary, but I often find it helpful to be able to look at what one member is doing. The ensemble min and max plots at the bottom are based on the lowest and highest values at each downscaled gridpoint, respectively.  Thus, it is very likely that many ensemble members contribute minima or maxima in these plots so do not use them to illustrate a plausible forecast outcome across a large region.   For instance, if you have one ensemble member that produces heavy snow over far northern Utah but none over central Utah and another that does the opposite, the ensemble max forecast is going to show heavy snow everywhere, even though neither member is calling for such a widespread event.  Additionally, the maximum at any grid point is sometimes a pretty big outlier compared to other forecasts at that point.  Please don't use the ensemble max forecast at lower right for click bait suggesting some sort of apocalyptic storm. 

Plume Plots

We have completely upgraded the popular forecast plumes to provide more information about snow level and SLR uncertainty.  As shown below, the plumes now include six panels, with accumulated precipitation and snow on top (including GEFS and ENS means), 6-h precipitation and snow in the middle, wet-bulb 0.5°C level at bottom left, and snow-to-liquid ratio at bottom right. 

The bottom four panels are violin plots, which are designed to provide detailed information about the distribution of forecasts produced by the 82 ensemble members. The violin itself (filled with blue) summarizes the distribution of all 82 forecasts.  The width is standardized, so the 82 forecasts are all similar when the violin is short.  More forecasts fall into areas where the violin is wide.   In contrast, the 82 forecasts produce a wider range of possibilities when the violin is tall and there are few forecasts in areas where there is just a line. 

An example for snow-to-liquid ratio is below, which is extracted from the plot above.  I highlight four violins.  The violin at 12Z 2 October is short and extends from an SLR of 7 to 10.  That's a small forecast range with all of the forecasts falling between 7 and 10.  At 6Z 4 October, the range is larger, going from 0 to 9 and the violin has two bulges, indicating that there is clustering of the forecasts near zero and then at about 7.  Such a distribution is called bimodal and suggests two possible outcomes, one with rain (i.e., SLRs at 0) and others with dense snow (SLRs near 7).  At 18Z 5 October, almost all of the forecasts are near zero, but there is a long whisker going to 9 suggesting there may be a one or two forecasts that call for snow.  Finally, at 0Z 7 Oct, all of the forecasts are zero or no snow.  

Some context for the above SLR forecasts is provided by the wet-bulb 0.5°C violins below.  The wet-bulb 0.5°C level is an approximation for the upper-part of the transition zone where we would expect wet snow.  At 12Z 2 October, when the SLR range is small, most of the forecasts show that the site is above the 0.5°C level.  That means the forecasts put the site near the top of the transition zone or above it.  Hence the SLR range is small.  At 6Z 4 October, most of the forecasts put the wet-bulb 0.5°C level above the site, some well above it, but there are a few that put it below the site.  This leads to the bimodal distribution, a clustering of SLR at 0, but some forecasts with an SLR near 7 produced by the colder members.  At 18Z 5 October, nearly all of the forecasts put the wet-bulb 0.5°C level well above the site, but there is one outlier that puts it below the site and possibly one or two others that put the site in the transition zone.  As a result most of the forecast SLRs are near zero, but there are a couple of outliers that have dense snow with SLRs in the high single digits. Finally, at 6Z 7 October, all of the forecasts put the wet-bulb 0.5°C more than 1000 feet above the site, and the SLRs are all zero.  

Note that we provide SLR distributions whether or not precipitation is forecast by the model.  Thus, an SLR of 0 doesn't mean it's raining, but it means if there is precipitation, we are diagnosing that it would be rain.  

The violins contain additional statistical information, summarized below. The thick horizontal red line indicates the median of the forecasts.  The thick vertical black line indicates what is known as the interquartile range (IQR) or the middle 50% of the forecasts.  The vertical red line indicates the middle 90% of the forecasts.  Finally, the blue horizontal lines or whiskers denote the minimum and maximum extrema. 

The top of the brown region is the station elevation.  Note that due to the limits of resolution, the model elevation can differ from this elevation.  We provide the model elevation in plume plot header so you can compare. Sometimes this is important.  For the plots above, the site is Washington Pass.  The site elevation is 5450 feet, but the model elevation is a bit higher, at 5713 feet.  

Caveats and Disclaimers

This is an experimental product.  Feedback is helpful to us as we are trying to find ways to better forecast snow and its characteristics and squeeze everything we can out of the operational model suite.  Tell us what works and what doesn't.  

The Utah Snow Ensemble is based on data and products of the European Center for Medium Range Weather Forecasting (ECMWF), National Centers for Environmental Prediction (NCEP), University of Utah, and other groups.  These groups do not accept any liability whatsoever for any error or omission in the data and their availability, or for any loss or damage arising from their use. 

This blog post may be updated as needed.