The use of an FV3 core should allow for a more computationally efficient model.
Read the whole story
Read the whole story
NOAA is about to make some big changes to its global weather model
- Thread starter JournalBot
- Start date
Ominous wall cloud portends possible violent weather in Nebraska early in 2018.
I already have my own weather model that tells me what's likely to happen and what I should do (seek cover!) if that sort of cloud developed overhead.
Upvote
-2
(24
/
-26)
In general, I find NWS' forcasts to be really reliable, but I'd love to see an article unpacking the differences between this model and the (generally regarded to be suprior) Euro model.
Upvote
73
(75
/
-2)
The takeaway for me, as a CFD person, is here (from the linked website):
Whether or not FV3 is more accurate today, the numerical methods are more consistent with the underlying physics. "Consistent" in a very technical sense -- that is, as you reduce (e.g.) the grid size and time-step size, you can rigorously show that the errors introduced by the discritization limit to zero.
You can tune and massage an inconsistent model to hell and gone in order to get the right answer, and that's the approach taken in a lot of applications. If you don't have the horsepower to do things rigorously, sometimes this is the better approach: especially in weather modeling, it's useless to have a more accurate model that can't run fast enough to give a forecast. But there's a limit to how much you can improve such a model. A heavily tuned solver is intrinsically sort of limited in how predictive it can be -- tuning implies that your model is going to want to "predict" only things that have already happened.
Better to respect the underlying physics with your choice of numerics, if you can afford it. Most engineering applications are a decade or two ahead of the geophysical community in this sense, since our computational domain usually isn't the whole damn planet.
This is a good and necessary move for future improvements.
Whether or not FV3 is more accurate today, the numerical methods are more consistent with the underlying physics. "Consistent" in a very technical sense -- that is, as you reduce (e.g.) the grid size and time-step size, you can rigorously show that the errors introduced by the discritization limit to zero.
You can tune and massage an inconsistent model to hell and gone in order to get the right answer, and that's the approach taken in a lot of applications. If you don't have the horsepower to do things rigorously, sometimes this is the better approach: especially in weather modeling, it's useless to have a more accurate model that can't run fast enough to give a forecast. But there's a limit to how much you can improve such a model. A heavily tuned solver is intrinsically sort of limited in how predictive it can be -- tuning implies that your model is going to want to "predict" only things that have already happened.
Better to respect the underlying physics with your choice of numerics, if you can afford it. Most engineering applications are a decade or two ahead of the geophysical community in this sense, since our computational domain usually isn't the whole damn planet.
This is a good and necessary move for future improvements.
Upvote
121
(121
/
0)
With the practical differences apparently being so small, being 3rd or 4th best seems fine to me.
Upvote
-3
(4
/
-7)
>>The nation's weather and climate organization, NOAA, has appointed a new director of its Environmental Modeling Center.<<
Given how the current gov't seems to put people actively hostile to the Agency in charge of said agency, I was half expecting a weather-denier to be put in place...
Glad to see I was (hopefully!) wrong.
Given how the current gov't seems to put people actively hostile to the Agency in charge of said agency, I was half expecting a weather-denier to be put in place...
Glad to see I was (hopefully!) wrong.
Upvote
21
(35
/
-14)
If it's more efficient, will they get results faster or will they bump the resolution to try to get better accuracy in the same amount of time?
Upvote
22
(22
/
0)
Read the Area Forecast Discussions put out by the National Weather service. The meteorologist will talk about leaning toward one model or the other, or splitting the difference based on their own experience.
NWS Feed hosted by Iowa State University. On the left, locate your local weather office "chat room". One of the many hats I wear at work is supporting Crisis Management and given that I live in Huntsville, AL - tornados are a real possibility. I often get alerts before phones or radio (just seconds or minutes - but still).
And I lost the point I was trying to make. Even in short term forecasts there are enough differences that the meteorologist has to make his own determination about what to put as a forecast.
Upvote
34
(34
/
0)
Typo:
(May even want to make it "third- or fourth-best" to match hyphenation in "second-best".)
(May even want to make it "third- or fourth-best" to match hyphenation in "second-best".)
Upvote
8
(9
/
-1)
snowman<ca>
Ars Praefectus
I am pretty sure Eric did that in an article here about a year or two ago, though I think it's main focus was on a couple of the newer alternatives which would be a much bigger shift by NOAA to it's current model.
Upvote
14
(14
/
0)
donovan1983
Wise, Aged Ars Veteran
I really do think the openness of NOAA data is under-appreciated. Being able to get various model output, and so much more data free-of-charge is just plain awesome for those at all interested in dabbling in citizen science. They even publish source code and software for handling a lot of that data. Partnerships like UCAR also have a lot of useful software, such as the NetCDF-Java library which makes handling the published model output files a bit easier.
Upvote
75
(75
/
0)
Upvote
96
(99
/
-3)
Not that I don't care about our Southern Hemispherical friends, but I noticed that the FV3 Northern numbers are just short of the European best-in-class and well past the Met scores for their global result.
How do the European and the Met models perform in the Northern Hemisphere? I'm wondering if the FV3 has biased performance relative to them are do they all have that significant a swing.
How do the European and the Met models perform in the Northern Hemisphere? I'm wondering if the FV3 has biased performance relative to them are do they all have that significant a swing.
Upvote
5
(5
/
0)
Maybe. It's unclear what the 'score' really means in detail. And weather prediction is judged by details, e.g., will it rain -here- tomorrow.
Upvote
4
(4
/
0)
It could be a climate change denier, that doesn't matters for the execution of the weather prediction model, because weather != climate.
Upvote
-6
(10
/
-16)
As an accomplished and competitive sailboat racer (hobby, not professional), who is frequently in charge of tactics and navigation for distance races (usually in the ~50-250 nautical mile range), I can personally attest to the low accuracy of ALL the weather models.
My strategy for success with weather is to get all the latest models for my given location, compare that to the weather conditions we are currently experiencing at the moment, then after finding the one that most closely matches what we are experiencing, use that model, compensated in my head for any variances it is failing to predict, for my tactical decisions for what I plan to do for the next 3-6 hours. In another hour or two when I re-evaluate my tactical situation relative to the weather and the strategy, I repeat the process from scratch. Once a particular weather model has shown to be consistently the most reliable for a given day over a few of these cases (which doesn't always happen), I will defer primarily to that model for the next ~24 hours unless something significant changes with the weather (which often does).
Note, I'm not measuring gross conditions like the TV weatherman report covers. I am paying attention to specific detailed information relating to wind speed and direction. Changes in as little as 5 or 10 degrees of true wind direction and 2-3 kts of wind speed are very important when you are trying to cover ~250 miles of ocean using only the power of the wind in your sails to do so, while competing with 50-100 other teams trying to do it faster than you.
My strategy for success with weather is to get all the latest models for my given location, compare that to the weather conditions we are currently experiencing at the moment, then after finding the one that most closely matches what we are experiencing, use that model, compensated in my head for any variances it is failing to predict, for my tactical decisions for what I plan to do for the next 3-6 hours. In another hour or two when I re-evaluate my tactical situation relative to the weather and the strategy, I repeat the process from scratch. Once a particular weather model has shown to be consistently the most reliable for a given day over a few of these cases (which doesn't always happen), I will defer primarily to that model for the next ~24 hours unless something significant changes with the weather (which often does).
Note, I'm not measuring gross conditions like the TV weatherman report covers. I am paying attention to specific detailed information relating to wind speed and direction. Changes in as little as 5 or 10 degrees of true wind direction and 2-3 kts of wind speed are very important when you are trying to cover ~250 miles of ocean using only the power of the wind in your sails to do so, while competing with 50-100 other teams trying to do it faster than you.
Upvote
26
(28
/
-2)
If the European Model is so great, why don't the US use that? (Apparently, the European model was better at predicting Hurricane Sandy than GFS.)
Upvote
-1
(6
/
-7)
At some point I'd like to do that. In general, it comes down to two factors. ECMWF has more resources available to dedicate solely to accurate 10-day weather forecasts. Secondly, their system for assimilating real-world weather observations from a very broad array of sources into their global model is vastly superior.
Upvote
62
(62
/
0)
There's value in having different weather models, written by different teams and using different assumptions and approximations, even if one of the models seems to usually issue the more accurate forecast.
I could write all day on why this is so, but here's a simple example: if three weather models which don't share a single line of code issue three forecasts which are relatively in line with each other, you can have considerably more confidence in the forecast than if the three models issue wildly different forecasts.
Upvote
67
(67
/
0)
That depends on the agency. Sometime in the last year or two I read about one national entity that stopped allowing their meteorologists to tweak presented results after finding the model was now good enough that the notional expert adjustments were now wrong more often than they were right.
Upvote
10
(10
/
0)
The article touches on this a bit but the main goal with this FV3 upgrade is to get on the new core while keeping the results similar. Once that's completed and checks out, they'll work on the physics packages, etc. to, hopefully, improve the forecasts. I would compare it to replacing Win95 with WinNT on your computer -- it looks similar from a user and API standpoint but underneath it's completely different: a new and better core to build on.
IIRC, the FV3 will be used for more than just the GFS in the future.
IIRC, the FV3 will be used for more than just the GFS in the future.
Upvote
34
(34
/
0)
As Michael hit the panhandle, we heard a lot about our inability to predict the rapid intensification that Michael and other hurricanes this year demonstrated. Is there anything in the new model that helps with that?
Upvote
7
(7
/
0)
Upvote
43
(43
/
0)
Upvote
13
(13
/
0)
This is correct. FV3 will also be implemented for the NAM, HRRR, and other regional models that NOAA and NWS use.
Upvote
20
(20
/
0)
Upvote
0
(2
/
-2)
In this case it appears to've hit a basket of eggs, bounced back, hitting the thrower in the head and covering her face with egg.
Upvote
21
(21
/
0)
Speaking of weather, I expect that burn to cause a major heat wave. Break out the shorts.
Upvote
27
(29
/
-2)
The models produce numerous Cat 6 winter storms in their 240+ hour forecasts. It would be fun to add up all the model snow we get a year here in Atlanta and compare that to actual accumulation!
Upvote
-3
(1
/
-4)
Each model has its own quirks and biases. Typically meteorologists use a blend of different models to form a forecast. Even though the Euro tends to be more accurate on average, it still can be wrong. One prominent example was the northeast blizzard of 2015. I am nowhere near a professional meteorologist but I like following the local NWS forecast discussions.
As the storm was nearing, the Euro was consistent for days in predicting a massive blizzard with over 24 inches of snow for NYC and the surrounding areas but the GFS model disagreed and was varying a bit between each new run. Usually consistency from a model implies that it is more certain than others that might be varying more from run to run. Plus, the GFS was recently upgraded and meteorologists noticed a few new quirks (like in the article) and might have been hesitant to rely on it too much. Still, there were other models to work with like the Canadian, NAM, UKmet, etc.
Since Superstorm Sandy was still in everyone's memory at the time and "everybody knows" the Euro tended to be more accurate with this kind of thing, somewhere along the line the local NWS mets basically decided to discount the other models almost completely. They basically threw all their eggs in one basket. This turned out to be a Bad Thing™. Less than 10 inches fell in NYC, and even less in New Jersey. It turned out that in this case the GFS had a much better handle on things.
A few links:
https://dotearth.blogs.nytimes.com/2015 ... forecasts/
https://fivethirtyeight.com/features/ho ... d-of-2015/
Upvote
16
(16
/
0)
It's interesting to look at the Forecast Discussion for NOAA weather and see how local forecasters deal with the models. It hasn't been unusual lately, in my area, for the GFS and Euro models to be more or less orthogonal, leaving the forecaster to guess, based on local history, which one might be closer to reality in a day or 2, or maybe to just average them. If all else fails, I guess, there's always John's Weather Stone (above).
Upvote
5
(5
/
0)
Hopefully he or she ends up saying "Based on [reasons] I'm forecasting X, but because there's a lot of disagreement in the models, there's more uncertainty than usual in this forecast"
Upvote
11
(11
/
0)
Very common, yes!Hopefully he or she ends up saying "Based on [reasons] I'm forecasting X, but because there's a lot of disagreement in the models, there's more uncertainty than usual in this forecast"
Upvote
7
(7
/
0)