Here Come the Raw Deterministic Snow Forecasts!
It is the time of year where a quick glance at deterministic medium-range models can show you whatever you are looking for: snow, severe, and anything in between. There is never a shortage of forecast hour 240 model output of snow circulating on social media (like the GFS plot from pivotal weather above).. While we can’t prevent these from being irresponsibly posted, we can take back control of the message. Instead of pointing at them and saying “it’s too soon to talk about anything”, or flat out ignoring them, we can tell the truth, and put those wild doomsday images into context. Doing this will actually increase trust in later forecasts!
The Week Ahead: Ridge-y West, Trough-y East
If we are going to look at the medium range, that can only end well by looking at ensembles. So, taking a look at the back half of the coming week, we can see a pretty familiar cool-season pattern developing over North America in the above plot of the EPS 500mb height and anomaly. Something akin to a rex block on the west coast, with downstream troughing over much of the eastern US. Such high amplitude patterns are usually always interesting, so it’s no surprise we are starting to see some interesting deterministic forecasts. But how confident are we that the ensemble mean isn’t masking some uncertainty? For that, we’ll turn to clusters.
In the above gallery are two different views of the WPC Clusters from the EPS, GEPS, and GEFS combined grand ensemble for the 24 hours ending 00Z Friday, October 14th. In the first image, the shading is the 500mb anomaly from climatology, and in the second, the difference from the grand mean (in case you had any trouble discerning the differences between the clusters in the first image). If you find the scenarios as depicted by the clusters to all be similar, we are in agreement. We can see slight differences in the depth or exact placement of the trough axis in the different clusters, but each and every member has a deep trough in the eastern US. So, we can be highly confident in this synoptic feature, but the devil is now in the details. First we can check to see if the slight differences noted above have any impact in sensible weather like temperatures or precipitation. The WPC site does let us get a little look at just that.
If we look at the corresponding resulting max temperature and QPF from the clusters we previously looked at, we can see some slight differences in maximum temperatures (but all telling the same story), and much bigger differences in QPF. Remember the shading in these is difference from the all-member mean (not climo). We can see that in cluster 2, which was a bit slower than the others, our band of QPF across the east is slower, but even the all-member mean has widespread measurable QPF for much of the eastern US. If we wanted a gauge timing uncertainty a bit more, we can swap out the plan view maps, for something more one dimensional, such as an ensemble meteogram. For this task, we can go to weatherbell.com (yes, it does cost money, but if you aren’t quite sold on investing yet, you can see similar plots of older runs for free if you head to weathermodels.com).
In the slider above, we can compare 80% of the 100-member grand ensemble (EPS + GEFS) we were looking at with clusters. The top panel of both plots shows each ensemble member as a row, with time increasing to the right. Each cell has a color/value of that member’s QPF for the 24-hrs ending at that time. We can see that later in the week, we have a pretty decent column of certainty in both ensembles, especially compared to some of the activity later in the forecast. Even though at first glance, the columns line up nicely, we can see some subtle differences in the peak timing, with some members peaking early on (06Z Wed 10/12), while some members do not peak until 12 or 18Z on the 13th. Again, the devil is in the details here, not necessarily the synoptic pattern happening or not, which is not something we can say every time.
The Power of Ensembles
So, let’s take another look back at deterministic models, specifically the GFS. We’ll use the model trend loop feature on pivotalweather.com to look at the last handful of runs of the GFS.
Just from this one deterministic model, we can see wildly varying solutions, especially if you are interested in just one area. You might even notice one or two runs could be concerning if you live in the northeast, or parts of the Great Lakes region. We could give up here, and say its bouncing around because it’s just that uncertain, so we do not know what is going to happen. Or, we could start leveling up. What happens when we change that same plot to the GEFS mean? (spoiler, there is a lot less run-to-tun change). But, we could level it up even more by revisiting those meteograms to put that deterministic run into context; so let’s do that!
Now, we have some context for that one deterministic run. That 6″+ snow represents 1 of 30 possible outcomes in it’s own ensemble (which is a little over a 3% chance by my math), and it is not associated with this initial round of precipitation (it happens much later). Responsible messaging shouldn’t shy around providing that context vs just ignoring that scary map that will generate clicks. But we can go further. This is raw ensemble data using a 10:1 ratio. Let’s look at something that is calibrated that ingests a much larger grand ensemble: the National Blend of Models (NBM).
In the plot above (the code of which, I will release as part of an AMS presentation if the abstract is accepted), I had to extend the forecast timeframe to 72 hours and go well into midweek next week to get actionable probabilities. This isn’t to say there won’t be precipitation, or that it won’t be cold, as I’ll show next.
It’s looking wet
As mentioned before, and as we saw with clusters, although snow potential is pretty much nil initially, it will still be wet. Shown in the gallery above is the probability of at least a quarter-inch of rain for Tuesday, Wednesday, and Thursday (24-hrs valid 12Z the following day). We can glean a pretty decent stab at timing from these, and even notice that probabilities increase as we step forward in time. So let’s take a closer look at the northeast for Thursday.
If we up our threshold to one inch, we still get probabilities well in excess of 50% to receive at least one inch of rain. That’s starting to sound like some heavy rain potential. So let’s turn to a tool designed to identify excessive rainfall, CSU Machine Learning Probabilities.
Now, with a short caveat first, the CSU MLP is trained only on the GEFS, so it should look similar to and inherit any GEFS trends. With that, we see that CSU MLP forecasts a 5-15% chance of excessive rainfall (long story as to what that means) for that same day. Notice it is a bit westward of the highest NBM probabilities, but we also already identified some 12+ hour timing uncertainty earlier on.
It’s looking cold
As we mentioned before, with such a high amplitude trough, there’s bound to be cold air, and it certainly looks that way. Much of this area has frozen this weekend, but probabilities of hard freezes extend southward once again by next week. Notice this is quite a few days AFTER the precipitation. The northerly flow on the back side of the upper-level trough brings in the cold air, well separated from the moisture and dynamics. What will bear watching is subsequent shortwaves that can bring additional moisture and forcing over the top of this cold air once it gets established- especially if the upper-level pattern doesn’t change much.