Weather all’s

[franksingleton]

New Scientist,29 August, has an article on weather apps -not GRIB apps but those that claim, often,to give post code accuracy. It explains, why, discusses at least, why these are often poor. It also casts light on the Met Office app and why, as is my experience, why it is so good.

 

[Lucky Duck]

I’ve noticed the BBC weather forecasts never seems to show a wind speed of 30 something mph, It always seems to jump from the high twenties to forty something

 

[Cerebus]

New Scientist,29 August, has an article on weather apps -not GRIB apps but those that claim, often,to give post code accuracy. It explains, why, discusses at least, why these are often poor. It also casts light on the Met Office app and why, as is my experience, why it is so good.

Yes, I find met office my go to more frequently. Seems to have improved a lot in 20ish years.

 

[johnalison]

My usual resource is just the radar display and forecast on my iPad ‘Weather’, just to see if I need to take my brolly.

 

[Cerebus]

The barometer by my bed (yes by my bed) that shows hourly trends and humidity is also my go to; it seems pretty accurate for imminent weather (as one would expect from such an instrument).

The hourly trends and changes are possibly more important than the immediate pressure reading as an expert ( that I am not) would understand.

 

[Chiara’s slave]

We are ‘Windy’ users, but the most accurate model is Frank’s favourite. If the gust strength looks outrageous, it often does, then hold on to your hat, cos that’s what you’re going to get.

 

[dunedin]

I’ve noticed the BBC weather forecasts never seem show a wind speed of 30 something mph, It always seems to jump from the high twenties to forty something

They suddenly leap from showing average wind speed to gust speed instead

 

[franksingleton]

To avoid thread drift, my OP was about apps, NOT TV or other forecasts such as on Windy, Ventusky, PocketGrib etc.
The NS article tries to understand why apps are different.
A quote near the end is as follows. Apologies for the format, I am not clever or quick enough to improve it!

“Some apps go as far as to extrapolate data that simply isn’t there, says Parker, which


could be a life-and-death matter if you’re trying to gauge the likelihood of flash


floods in Africa, for instance. He’s seen at least four free forecasting products of


questionable utility show rainfall radar data for Kenya. “There is no rainfall radar in


Kenya, so it’s a lie,” he says, adding satellite radars intermittently pass over the


country but don’t give complete information, and his colleagues at the Kenya


Meteorological Department have said they don’t have their own radars running. These


apps are “all producing a product, and you don’t know where that product comes


from. So if you see something severe on that, what do you do with it? You don’t know


where it’s come from, you don’t know how reliable it is”.


On the other hand, the Met Office app will not only use a model that’s fine-tuned to


get UK weather right, but it will also employs all sorts of post-processing to refine the


forecasts and apply the sum total of the organisation’s human expertise to it. Then the


app team goes through a painstaking process to decide how to present that in a simple format.”

 

[Chiara’s slave]

UKV2 was a knot out on the gust strength this afternoon. How do I get my money back?

 

[franksingleton]

UKV2 was a knot out on the gust strength this afternoon. How do I get my money back?

You pay for forecasts?

 

[Chiara’s slave]

You pay for forecasts?

Obviously that was intended to be light hearted, but yes, we all pay for forecasts. You were a public servant, were you not? A government employee, directly or indirectly.
On another note, yes, I pay here and now for forecasts. UKV2 may be free, but the ability to compare is on an hourly basis with other models, with animated maps, swell heights, tidal graphs, all in one place does need to be paid for, and not just our taxes. I regard having the very best weather info as beyond price.

 

[franksingleton]

Obviously that was intended to be light hearted,

So was my response

but yes, we all pay for forecasts. You were a public servant, were you not? A government employee, directly or indirectly.
On another note, yes, I pay here and now for forecasts. UKV2 may be free, but the ability to compare is on an hourly basis with other models, with animated maps, swell heights, tidal graphs, all in one place does need to be paid for, and not just our taxes. I regard having the very best weather info as beyond price.

This really is thread drift. However, I am a realist and a little cynical of the value to us sailors of the value of high resolution models. These are run by the Met Office for good reasons. They are run as ensembles so that, in the event bad weather, they can gibe the best possible advice to the authorities.

Lay users too easily forget that all NWP models have smoothing built in so that the effective resolution is about 5 grid lengths. Post code accuracy is a nonsense. Further, the area of detailed data input is small. Weather systems move so, in the upwind part of the UKV area, very quickly, both forecast and actual weather will be determined by larger scale of weather. Depending on the movement of the large scale pattern, the accuracy of the limited area forecast quickly becomes determined by the coarse scale regional and global models within which they are nested. Unless you are wanting information on fine detail for a short period, there is little to be gained using high resolution models. Recently, 7th September, we did a Channel crossing, St Peter Port to Dartmouth in a, mainly, 5/6. We encountered detail that was unpredictable before departure and about which we could have done anything differently.

Having said that, UKV is probably the best detailed model in use today. AROME does not run at every hour. HRRR is based on the US GFS of a variant. These are considerably inferior to UK UM.

 

[finestgreen]

Do you have a view on windy.app's "EXP3" model? It claims to be "terrain aware" to take account of coastal effects. I haven't done or seen any comparisons on it's accuracy, it might just be a gimmick

 

[Chiara’s slave]

Do you have a view on windy.app's EXP3 model? It claims to be "terrain aware" to take account of coastal effects. I haven't done or seen any comparisons on it's accuracy, it might just be a gimmick

I’ve found it to be not worth paying attention to, in our area.
Grid size, I guess we all need to remember that it’s a chaotic system. The prediction is a ‘most statistically likely’ and needs to be treated as such. Small terrain variables can make a big difference, and would be hard, or even impossible, to take into account. But that’s why racers at their home waters have an advantage. Overall, UKV2 works well for us, and we look at the others to see if there’s anything anomalous.

 

[lustyd]

I don’t think terrain is hard to take into account. Right now it’s costly because of the physics based approach needing to calculate everything, so current and probably future supercomputers are not sufficient to do so. That doesn’t mean other approaches wouldn’t work to extrapolate quite well the effects of terrain without all of the processing happening every day.

We need good labelling though so it’s clear what sources and techniques an app uses. Unfortunately the sources are often poor at providing this too. I worked with some government agencies on data metadata projects and to say it’s a lost cause is an understatement in public sector!

 

[Chiara’s slave]

Terrain might only affect a sq km or less though. That’s a tough call. I can think of 2 well known Solent harbours where a few degrees difference in wind direction can completely change the conditions in the offing. I can’t imagine they’re the only 2 places on the south coast, let alone the wider world. Sea breezes too are very local. We’ll need another quantum jump in computer power to deal with that, pun intended.

 

[lustyd]

Terrain might only affect a sq km or less though. That’s a tough call. I can think of 2 well known Solent harbours where a few degrees difference in wind direction can completely change the conditions in the offing. I can’t imagine they’re the only 2 places on the south coast, let alone the wider world. Sea breezes too are very local. We’ll need another quantum jump in computer power to deal with that, pun intended.

No, we need a change of approach as I said. The very thought that increasing resolution could ever be the answer is absurd. It’s not a linear scaling problem so as resolution increases compute requirements go up exponentially. This is why physics modelling has not continued to improve, we just can’t add enough parameters and keep the grid size small enough to be useful. This is why AI is being used since you don’t need to calculate every point on a grid, the model learns as local sailors do and produce the required results without all the compute requirements.

 

[Chiara’s slave]

No, we need a change of approach as I said. The very thought that increasing resolution could ever be the answer is absurd. It’s not a linear scaling problem so as resolution increases compute requirements go up exponentially. This is why physics modelling has not continued to improve, we just can’t add enough parameters and keep the grid size small enough to be useful. This is why AI is being used since you don’t need to calculate every point on a grid, the model learns as local sailors do and produce the required results without all the compute requirements.

The computing power is needed somewhere. The weather forecasters will just be outsourcing it. AI is pretty processor intensive.

 

[lustyd]

The computing power is needed somewhere. The weather forecasters will just be outsourcing it. AI is pretty processor intensive.

AI training is intensive but needs doing only once. Running an AI model is not intensive at all, it’s effectively a lookup in a very large table based on various parameters. As such, we can drastically ramp up the number of parameters and decrease the size of the grid, making localised forecasting feasible.
Initially the physics model will be required and the output modified. This is happening now. Eventually, the physics model will be phased out. The AI model will likely “understand” the physics as it’s done now but will be able to take many more parameters into account.

 

[Chiara’s slave]

AI training is intensive but needs doing only once. Running an AI model is not intensive at all, it’s effectively a lookup in a very large table based on various parameters. As such, we can drastically ramp up the number of parameters and decrease the size of the grid, making localised forecasting feasible.
Initially the physics model will be required and the output modified. This is happening now. Eventually, the physics model will be phased out. The AI model will likely “understand” the physics as it’s done now but will be able to take many more parameters into account.

Is that response composed by AI?