Why snow forecasts are sometimes so wrong

[usedtobe]

Below is a the start of a discussion of why forecasts sometimes bust. Some probably think it is a cop out when forecasters equivocate when making a forecast. Hopefully, this article will explain why you shouldn't give much credence to a day 7 GFS and why models sometimes jump around so much with their solutions. It is not because the models stink, it's because of atmospheric processes are complicated and small sampling errors can make big differences to a forecast. Anyway, questions are welcome here or on the CWG web site. Next week I'll post about the effects of the mountains and ocean. They are also big players in why forecasting snow along the east coast is so difficult.

http://www.washingto....html#pagebreak

In the editing process, a discussion of the difficulty that latent heating causes to model forecasts was left out. For those interested, convection and the processes involving water changing states can cause big forecast problems as when gaseous water condenses, heat is released. Essentially, latent heat is the driving force in the development of hurricanes but also plays a role in the development of extratropical storms. Latent heating plays a part in storm development, if you get the heating in the wrong place, it can cause the surface low either be too strong or weak. Get the latent heat in the wrong spot and it can lead to the low ending up a little too far east or west which can change our forecast from snow to rain.

[Huffwx]

Below is a the start of a discussion of why forecasts sometimes bust. Some probably think it is a cop out when forecasters equivocate when making a forecast. Hopefully, this article will explain why you shouldn't give much credence to a day 7 GFS and why models sometimes jump around so much with their solutions. It is not because the models stink, it's because of atmospheric processes are complicated and small sampling errors can make big differences to a forecast. Anyway, questions are welcome here or on the CWG web site. Next week I'll post about the effects of the mountains and ocean. They are also big players in why forecasting snow along the east coast is so difficult.

http://www.washingto....html#pagebreak

In the editing process, a discussion of the difficulty that latent heating cause to model forecasts was left out. For those interested, convection and the processes involving water changing states can cause big forecast problems as when gaseous water condenses, heat is released. Essentially, latent heat is the driving force in the development of hurricanes but also plays a role in the development of extratropical storms. Latest heating plays a part in storm development, if you get the heating in the wrong place, it can cause the surface low either be too strong or weak. Get the latent heat in the wrong spot and it can lead to the low ending up a little too far east or west which can change our forecast from snow to rain.

Good write up Wes, very good. .

[WeatherFox]

Mesoscale banding, thundersnow, and dryslots are difficult to pinpoint until we are in the now forecasting stage.

[Lookout]

Below is a the start of a discussion of why forecasts sometimes bust. Some probably think it is a cop out when forecasters equivocate when making a forecast. Hopefully, this article will explain why you shouldn't give much credence to a day 7 GFS and why models sometimes jump around so much with their solutions. It is not because the models stink, it's because of atmospheric processes are complicated and small sampling errors can make big differences to a forecast. Anyway, questions are welcome here or on the CWG web site. Next week I'll post about the effects of the mountains and ocean. They are also big players in why forecasting snow along the east coast is so difficult.

http://www.washingto....html#pagebreak

In the editing process, a discussion of the difficulty that latent heating causes to model forecasts was left out. For those interested, convection and the processes involving water changing states can cause big forecast problems as when gaseous water condenses, heat is released. Essentially, latent heat is the driving force in the development of hurricanes but also plays a role in the development of extratropical storms. Latent heating plays a part in storm development, if you get the heating in the wrong place, it can cause the surface low either be too strong or weak. Get the latent heat in the wrong spot and it can lead to the low ending up a little too far east or west which can change our forecast from snow to rain.

I heard long ago that it's not the models that are the root problem but lack of real time observations and data that are put into the models. I can't recall the figure since it's been years since I saw someone mention it, but we fall way way short (in the thousands) of the number of observations (especially upper air) that are really needed. Of course making such a network of widespread whole atmospheric observations costs money and despite the fact that billions of dollars would be saved just by having more accurate models, we, and the world, don't do it.

[usedtobe]

I heard long ago that it's not the models that are the root problem but lack of real time observations and data that are put into the models. I can't recall the figure since it's been years since I saw someone mention it, but we fall way way short (in the thousands) of the number of observations (especially upper air) that are really needed. Of course making such a network of widespread whole atmospheric observations costs money and despite the fact that billions of dollars would be saved just by having more accurate models, we, and the world, don't do it.

To get perfect forecasts, you'd have to be able to measure atmospheric variables at every point on the globe without the measurements impacting the variables you're measuring. That's an impossible feat and would cost more than billions of dollars. The non linear nature of the equatiosn governing the atmosphere a large part of the problem. Any slight errors in the initial analysis can grow non linearly leading to a bad forecast. There also is the problem with water changing state which ahs to be parameterized in a model. Convection is a great example. We've all seen times when a tropical system si forecast by a model (often the canadian model) which doesn't verifiy. The model's convective scheme is releasing too much heat over tthe grid box and may not be distributing the heat correctly. The same thing can happen to a developing snow storm. Latent heat works a little differently with an extratropal low than with a tropical one but without latent heat a low will be weaker in a model run. Essentially getting the latent heating released in the wrog place (having the convection in the wrong place) can help pull a low farther off the coast than may actually occur or may if there is too much heating and it is released too low in the atmosphere, may cause a low pressure system to develop too quickly. Then you have to deal with more warm advection to its east which may pump up a ridge. The various feebacks can then alter the forecast.

[usedtobe]

Mesoscale banding, thundersnow, and dryslots are difficult to pinpoint until we are in the now forecasting stage.

I didn't really talk that much about mesoscale banding but I probably should have in part two of the article. The tight thermal gradients associated with snowstorms makes the rain snow line and location of mesoscale banding a tough call. I did not talk about slantwise or elevated upright convection and how that usually occurs where there is strong frontogenesis as I thought it too difficult to explain.

[isohume]

This is one of the reasons we never put precip amounts in the fcst past day 3. A modeled QPF response is one of the most complicated and intricate variables for the models to "figure out". Many things can affect your precip fcst and hence your p/type fcst.

[Rainshadow]

This is one of the reasons we never put precip amounts in the fcst past day 3. A modeled QPF response is one of the most complicated and intricate variables for the models to "figure out". Many things can affect your precip fcst and hence your p/type fcst.

I'd hate to say how wide a spread in qpf occurs even with a run of the mill non convective event in day2 to reach the 67% and 95% confidence intervals in PHI's cwa.

With Boxing Day last winter it was amazing how subtle the differences in phasing made such a tremendous difference along the western periphery of that event.

[usedtobe]

I'd hate to say how wide a spread in qpf occurs even with a run of the mill non convective event in day2 to reach the 67% and 95% confidence intervals in PHI's cwa.

With Boxing Day last winter it was amazing how subtle the differences in phasing made such a tremendous difference along the western periphery of that event.

Yep and that's not that unusual. Seems like around here most events ahve the heavy banded precipitation with the deformation zone end up a little north fo where it is forecast except when it is forecast just to our south. Doesn't take much to screw up a forecast. Even if you look at the o.50" or greater threshold in winter, the threat scores aren't that high. That storm, the Sref esnembles kept insisting we'd get a decent snow, we got next to nothing.

[isohume]

I'd hate to say how wide a spread in qpf occurs even with a run of the mill non convective event in day2 to reach the 67% and 95% confidence intervals in PHI's cwa.

With Boxing Day last winter it was amazing how subtle the differences in phasing made such a tremendous difference along the western periphery of that event.

Yeah many times I'll look at and combine other forcings and moisture fluxes in the models to interpret the mesoscale qpf response. This is a way for me to see where the model may be erroneous in miscalculating the frontogentical development within the storm, over/under responding to mechanical lift, overplaying/underplayng the thermal adv, etc. Many times it relays back to your previous experience and knowing your CWFA wrt a given pattern...ie: complex terrain, avg kinematics, probable adv moisture fluxes, etc.

[Rainshadow]

Yep and that's not that unusual. Seems like around here most events ahve the heavy banded precipitation with the deformation zone end up a little north fo where it is forecast except when it is forecast just to our south. Doesn't take much to screw up a forecast. Even if you look at the o.50" or greater threshold in winter, the threat scores aren't that high. That storm, the Sref esnembles kept insisting we'd get a decent snow, we got next to nothing.

Wes,

Do you know if the culprits were mainly the older eta and rsm members or some of the newer members?

[usedtobe]

Wes,

Do you know if the culprits were mainly the older eta and rsm members or some of the newer members?

One run about 24 hours before the event, 90% of the members had us getting a decent snow, then it went down to 70% or so, I don't know which once didn't. I suspect it may be been the nam/eta members since the operational nam was less bullish than the GFS. I think the problem was part model physics and part resolution.

[Cheeznado]

Just look at the current situation in the 3-4 day time frame- the Euro and Canadian are closing off a deep 500mb low in the SE, the GFS has a progressive trough hundreds if not a thousand miles farther east- the ensembles are all over the map as you might expect. However, where I work (the media) there are many requests for specific forecasts up to 10 or even more (if a holiday is involved) days in advance and in situations like this where there is a boatload of uncertainty the non-met types just say we still want it. Then when it busts they complain. Such is my world.

[usedtobe]

Just look at the current situation in the 3-4 day time frame- the Euro and Canadian are closing off a deep 500mb low in the SE, the GFS has a progressive trough hundreds if not a thousand miles farther east- the ensembles are all over the map as you might expect. However, where I work (the media) there are many requests for specific forecasts up to 10 or even more (if a holiday is involved) days in advance and in situations like this where there is a boatload of uncertainty the non-met types just say we still want it. Then when it busts they complain. Such is my world.

Yep, right now the models are showing very little consistency between runs. this is an ugly pattern in many ways.

[Lookout]

To get perfect forecasts, you'd have to be able to measure atmospheric variables at every point on the globe without the measurements impacting the variables you're measuring. That's an impossible feat and would cost more than billions of dollars. The non linear nature of the equatiosn governing the atmosphere a large part of the problem. Any slight errors in the initial analysis can grow non linearly leading to a bad forecast. There also is the problem with water changing state which ahs to be parameterized in a model. Convection is a great example. We've all seen times when a tropical system si forecast by a model (often the canadian model) which doesn't verifiy. The model's convective scheme is releasing too much heat over tthe grid box and may not be distributing the heat correctly. The same thing can happen to a developing snow storm. Latent heat works a little differently with an extratropal low than with a tropical one but without latent heat a low will be weaker in a model run. Essentially getting the latent heating released in the wrog place (having the convection in the wrong place) can help pull a low farther off the coast than may actually occur or may if there is too much heating and it is released too low in the atmosphere, may cause a low pressure system to develop too quickly. Then you have to deal with more warm advection to its east which may pump up a ridge. The various feebacks can then alter the forecast.

Fascinating posts as usual from you. However, I realize the bolded part but what I meant was a realistic network of observations...not every single point. The figure I was referring to was also in reference to not making the models perfect, just much more accurate. The largest voids are obviously out in the pacific and atlantic and would cost a lot more but a much more widespread land based network seems doable however vs the billions upon billions of dollars lost every year (by us and virtually civilized nation on earth) by inaccurate forecasts/models.

My point being is we are seriously lacking the much more widespread network than what we have not to make the models more accurate. As you said, it's not possible to have one at every point on the globe but more realistic networks seem possible and would likely be worth the money vs how much is lost every year. And I wouldn't be surprised if that total every year is underestimated when you consider just how many people, businesses, and the government rely on accurate forecasts on a daily basis. And god knows what the real cost is with respect to forecasting hurricanes, large winter storms, etc.

[usedtobe]

Fascinating posts as usual from you. However, I realize the bolded part but what I meant was a realistic network of observations...not every single point. The figure I was referring to was also in reference to not making the models perfect, just much more accurate. The largest voids are obviously out in the pacific and atlantic and would cost a lot more but a much more widespread land based network seems doable however vs the billions upon billions of dollars lost every year (by us and virtually civilized nation on earth) by inaccurate forecasts/models.

My point being is we are seriously lacking the much more widespread network than what we have not to make the models more accurate. As you said, it's not possible to have one at every point on the globe but more realistic networks seem possible and would likely be worth the money vs how much is lost every year. And I wouldn't be surprised if that total every year is underestimated when you consider just how many people, businesses, and the government rely on accurate forecasts on a daily basis. And god knows what the real cost is with respect to forecasting hurricanes, large winter storms, etc.

We actually lost soundings when the Soviet Union broke apart. With the weak economy here and abroad, I don't see any push to get more sounding in the near future. The bright side of chao is the weather surprises that still arise. It's fun when the models go from an offshore snowstorm to a major one inside of 72 hours. It would be boring if 48 hour forecasts were almost always right.

[dtk]

We actually lost soundings when the Soviet Union broke apart. With the weak economy here and abroad, I don't see any push to get more sounding in the near future. The bright side of chao is the weather surprises that still arise. It's fun when the models go from an offshore snowstorm to a major one inside of 72 hours. It would be boring if 48 hour forecasts were almost always right.

Given the economic situation/environment, the reality is that there has been quite a push to reduce the global observing system (this includes a threat to remove funding from JPSS, planning for "radiosonde replacement" networks, etc.). Along these lines but somewhat of a tangent....there is much more to the observing system than 'sondes/soundings. In fact, in terms of reducing forecast errors, we actually get a bigger (or as big of an) impact from remotely sensed microwave and hyperspectral infrared instruments (AMSUA, AIRS, IASI). Don't get me wrong, the radiosonde network is a vital component to the global observing system and we'd be in big trouble without it....but we do have tons of other (good) stuff available.

So, given the reality that we cannot afford to measure every variable (this is really important, since a huge percentage of our observations are mass [temperature/pressure]) at some resolution finer than what we run the models at....we need to continue to devise better ways to extract even more information out of the observations we do have. One example: I believe we do a really poor job in terms of analyzing and modeling moisture/clouds (this fits right into Wes's post and the larger discussion).

Having said all that, there is hope. If you look at trends in terms of skill, the slope is still positive and we're still getting better (with much more room to improve). There are algorithm advances, such as hybrid variational/ensemble (recently implemented at the UKMet Office, and will be implemented at NCEP for the GFS in spring) and ensemble-based data assimilation algorithms have shown a lot of promise, but there is still so much work left to do. There is a lot of work to do in terms of how to better use remotely sensed data (i.e. satellite radiances in particular). There is still a lot of work to do on the modeling side. Advances in computing will allow us to continue to increase resolution and become less reliant on parameterizations for certain processes (hmm, global scale, cloud resolving model anyone?)

[usedtobe]

Given the economic situation/environment, the reality is that there has been quite a push to reduce the global observing system (this includes a threat to remove funding from JPSS, planning for "radiosonde replacement" networks, etc.). Along these lines but somewhat of a tangent....there is much more to the observing system than 'sondes/soundings. In fact, in terms of reducing forecast errors, we actually get a bigger (or as big of an) impact from remotely sensed microwave and hyperspectral infrared instruments (AMSUA, AIRS, IASI). Don't get me wrong, the radiosonde network is a vital component to the global observing system and we'd be in big trouble without it....but we do have tons of other (good) stuff available.

So, given the reality that we cannot afford to measure every variable (this is really important, since a huge percentage of our observations are mass [temperature/pressure]) at some resolution finer than what we run the models at....we need to continue to devise better ways to extract even more information out of the observations we do have. One example: I believe we do a really poor job in terms of analyzing and modeling moisture/clouds (this fits right into Wes's post and the larger discussion).

Having said all that, there is hope. If you look at trends in terms of skill, the slope is still positive and we're still getting better (with much more room to improve). There are algorithm advances, such as hybrid variational/ensemble (recently implemented at the UKMet Office, and will be implemented at NCEP for the GFS in spring) and ensemble-based data assimilation algorithms have shown a lot of promise, but there is still so much work left to do. There is a lot of work to do in terms of how to better use remotely sensed data (i.e. satellite radiances in particular). There is still a lot of work to do on the modeling side. Advances in computing will allow us to continue to increase resolution and become less reliant on parameterizations for certain processes (hmm, global scale, cloud resolving model anyone?)

Daryl, Thanks for posting. Remote sensing has been the saving grace allowing for the steady improvement in model forecasts despite the loss of soundings. To me it's always been interesting that the improvement in threat scores (not necessarily the greatest way of measuring precipitation accuracy) has been steady and the improvement has been pretty linear. That said, the day 48-72 hour forecasts for the 0.50" or greater threshold are now higher than the day 1 forecasts were back in the mid 80 so there has been lots of improvement.