AIGFS

[franksingleton]

I think in another thread you also indicated a distrust of these models.

Why is this?

As someone who has to make decisions, who has to point the boat one way or another, I find them _extremely_ useful.

For coastal and inshore races that last say up to two days, I typically only occasionally look at global models once the race has begun.

I have used HRRR for racing more than Arome, and much more than UKV, so I can mostly talk about that. The 1hr update frequency is really valuable. I pay close attention to whether f0, f1... match what I see - my instruments, and local weather stations - and when it does, i often go with it.

If the model initializes well, and it says if i go left, i can tack on a header in 3 or 4 hours - there is a good chance I am going to do that. And, I trust the high res model much more for the timing of the header than a global model. The global models take longer to assimilate data, longer to run and longer to release, so they are already old by the time i get them.

Another thing they can be really good at is local variation in wind speed and direction - which side of the course is going to have more pressure? Is one side going to have a slightly more favourable angle?

Have I ever been misled by these models - yes. But I think in part, it is my fault for not using the model correctly. For example, if a new run comes out suggesting a big change in the race strategy - maybe wait for the next run to see if it is confirmed.

Racing puts you in a situation where no decision is not an option, and you also can't wait forever to make a decision. So, the question is; what information are you going to use to help you make those decisions? Hi res models are not perfect, but they are probably the best thing available.

I am really looking forward to the new AI hi res models.

I tried to explain earlier. I will try again.

See Numerical weather prediction models. The area covered by UKV is 622 km from east to west, 810 km north to south. The starting point has no fine scale data outside that area. At a fairly normal speed a weather feature can, typically, move at 20 kts, 40 km/r roughly. So, if you are near the west of the British Isles, in a very few hours the weather that you experience will have originated in an area where the computation grid is 4km spacing ie from an area where the effective resolution is about 20 km.

The outer domain for UKV is 950 x 1025 km, beyond that the UK UM has a 10 km grid, 50 km effective grid.

As I said earlier, the main input to UKV is the gloral model. It can, or should, improve on the global model but, depending on where you are, maybe only for a short time. My word “distrust” was mot well chosen in isolation. I should have said something like, “distrust claims to be better or more useful.” Small weather details have short lifetimes. Beyond very few hours, UKV and any other detailed model will only show what could happen. Such models cannot, for example predict that a shower will form over a specific area.

Doing a Channel crossing, my main input would be the global model. In the event of the weather being marginal, as happened 0n our return this year, I might, but rarely have done, take a look at UKV and/or AROME.

Detailed models have their short term uses especially in severe weather where they are a valuable asset for specific locations. For the reasons above, their use is limited. Great for telling me if I can walk down to the village and back in the dry this morning.

 

[franksingleton]

OT
FYI, there is another version of Arome named AROME PI (for Prevision Immediate), it is run every hour, up to t+6 hours with time step 15min. Initial state is from Arome HD, it basically assimilates Radar and aviation data and most outputs are related to precipitation/convection etc. Wind speed components (U-V) are output at .025deg grid, whereas Wind Gusts and Max Gust are output at .01deg grid.
Data can be accessed here (good luck )
MeteofranceWeb
or there is a viewer here
Meteociel - Modèle Numérique AROME PI Meteo-France pour le Nord-Ouest de la France

Description
https://donneespubliques.meteofranc...0201109_donnees-publiques-arome-pi-vf_259.pdf

Edit
The GIF is too big for YBW forum, I posted it here
https://blogger.googleusercontent.c...dKvwBx0wHvfaeVLyd5-TieieuZy/s768/animgwz1.gif
AROME PI

The 6 hour period is sensible although some small detail will have shorter lifetimes.

 

[oldbloke]

I may be (probably am) totally wrong here but my analysis of these exchanges between 2 intelligent and usually worthwhile contributors is.....
Scientific meteorology has spent 2 centuries working on the basis that if we know enough about the weather and how it works we will be able to say that if it rains on Tuesday morning it will probably be raining on Thursday.
Whereas Lusty says that AI will be able to say that from past experience that if it rains on Tuesday morning it will rain on Thursday but that it doesn't need to know the mechanisms that make it so just that that is what happens

 

[franksingleton]

I may be (probably am) totally wrong here but my analysis of these exchanges between 2 intelligent and usually worthwhile contributors is.....
Scientific meteorology has spent 2 centuries working on the basis that if we know enough about the weather and how it works we will be able to say that if it rains on Tuesday morning it will probably be raining on Thursday.
Whereas Lusty says that AI will be able to say that from past experience that if it rains on Tuesday morning it will rain on Thursday but that it doesn't need to know the mechanisms that make it so just that that is what happens

I do not think that there is a great difference between lusty and myself. The ECNWF AIFS works well. There is some concern about the length of the period that can be used for training. However, it is a halfway house as the AI training uses a high quality, well ordered data set of carefully re-analysed data. This necessitates the use of a physical model based analysis. I am more cautious in that I think that an AI system using raw data will be difficult to develop. We may have to settle for the halfway house. That would have several benefits. There are already suggestions of some combination of AI and physical modelling. Research into weather, not least climate wil require understanding of the physics of the atmosphere.

One area where we do differ is the problem of chaos. However you predict weather, chaos will always be a determining factor.