B27
Well-known member
No 'Hopeless the Weather Girl' was definitely BBC.
Weather Models
- Thread starter lustyd
- Start date
capnsensible
Well-known member
Stemar
Well-known member
Then, of course, there's the Welsh version:
If you can't see across the valley, it's raining.
If you can see across the valley, it's going to rain
If you can't see across the valley, it's raining.
If you can see across the valley, it's going to rain
franksingleton
Well-known member
I agree with much in what you say. There was a time when forecasting was entirely subjective. Forecasters were at most airfields, RAF and civil. These forecasters were good at knowing local conditions. Large scale forecasts, production of forecast charts and guidance came from a central office. Skill was not good.I think it is sad that Lustyd has started a really interesting and what might a really beneficial thread and ruined it by his single mindedness.. His thread, his rules I suppose but I thought this was a forum for the benefit of us all.
As for the weather, I wonder how many of you who use tabular weather data such as wind guru and so actually stop and look at synoptic charts and how they evolve over a period of time. At the minute (or yesterday’s weather - not looked at this mornings weather yet) had a slack area of low pressure over Scotland. The wind forecast was slightly different as we flew north clearly showing the challenges in predicting the pressure gradient in such a system. When I cycled home from work, I started in the dry and 300m later the road was soaking wet from a recent shower. That shower might have been triggered by a farmer ploughing a field miles away triggering a cumulus cloud that then formed a CB of sufficient size to cause local wind effects and so on. Weather is reported in probabilities. Aviation forecasts have multiple layers where different weather has different probabilities. All it takes is a slight change for a weather effect to change its level of probability and whether it falls off the forecast altogether. If models are being run constantly then the forecast will change. I think that might be this issue here. Tabular weather data might be running weather probabilities of say 80% and not reporting weather that probabilities are say 40%. A small change in the model might switch the two weathers in the next forecast.
As for Lustyd’s assertion that weather forecasting will change in the 10 years, I am sure that AI will change everything in life. But the problem I foresee is that because we expect weather forecasts to be cheap, we will rely less and less data derived from organic matter. We see fewer and fewer trained and experienced met observers in aviation, replaced by a laser looking straight up and three visibility machines looking across a metre or so of air. In the old days a human would look across the airfield at the fog bank rolling in from the sea and change the weather report, but the machine is telling Exeter that it’s going to be glorious. You’re in data Lustyd, tell me how a future of poorer data will result in better modelling……
Nowadays, users, civil and military will not pay either for the local, on-site forecaster mor for the luxury of paying human observers Despite their skill. NWP has greatly improved forecasting of the large scale, synoptic pattern. Although synoptic observations contribute to large scale forecasting accuracy, the contribution is small. Major inputs are from satellite based infrared and microwave sounders, high level winds derived fro satellite cloud images. The fourth most useful data input is from measuring the occultation of GNSS signals.
Local forecasting is largely radar/satellite information combined with fine resolution models, grid lengths 1-3 km. However, these are far fro perfect. You have only got to use theMet Office app and compare predictions with actual radar data to see that models do not perform well either in predicting specific shower formation nor the decaying of existing showers.
Like large scale forecasts, these highly detailed local forecasts are constrained by the reality of the highly complex atmosphere. Forecasting has improved greatly since my days but is still far from perfect and will never be perfect.
Going back to your suggestion about synoptic charts, these are useful to get an overall picture. However, if I want to know what winds I will get on my next Channel crossing, I would use GRIB data every time.
capnsensible
Well-known member
Works in Scotland too!
finestgreen
Active member
No, you're all wrong 
The problem isn't the quality of the modelling, it's the presentation. We take complex models of chaotic states and force them to output a set of single point estimates - if we want better forecasts we need to find better ways of consuming the data.
The problem isn't the quality of the modelling, it's the presentation. We take complex models of chaotic states and force them to output a set of single point estimates - if we want better forecasts we need to find better ways of consuming the data.
westernman
Well-known member
It would help if we had some kind of representation of confidence level and what the alternate forecasts and their confidence levels could be. No idea how this could be represented in a meaningful way though.No, you're all wrong
The problem isn't the quality of the modelling, it's the presentation. We take complex models of chaotic states and force them to output a set of single point estimates - if we want better forecasts we need to find better ways of consuming the data.
At one level it would be most likely wind strength, 90% confidence level it won't exceed x, 90% confidence level it won't be less than y, and direction between a-b, with a certain confidence etc.
Could that be done??
franksingleton
Well-known member
It is not really quite like that. The models smooth their outputs. The GFS calculates on grid of about 13 km but smoothing means that the data are averaged over an area of around 60+ km. Variations in weather and wind occur on much finer scales leading to complaints that the models are not accurate. If models were judged on average vale’s, they would be much better than is apparently the case. Even the Meteo France AROME can only, at best, provide a forecast ove areas around 5 or 7 km, despite what providers say or imply about revolution.No, you're all wrong
The problem isn't the quality of the modelling, it's the presentation. We take complex models of chaotic states and force them to output a set of single point estimates - if we want better forecasts we need to find better ways of consuming the data.
Added to all this is the model uncertainty as seen when you look at model ensemble grid point values. It is not so much the models that are at fault, nor the presentation. It is really the way in which users interpret the model output and, in some cases, claims made by third party providers who might not be too clear about NWP.
franksingleton
Well-known member
This is possible. An algorithm could be written or AI used. In a crude way, it is already done in marine forecasts using undefined terminology. A counter argument would be that putting numbers to the occurrence of a gale, for example, would imply precision where it does not exist.
finestgreen
Active member
That's not really what I mean. Absolutely they smooth and average, but then what comes out is a point estimate for each blob of space and time - ie, in this place at 3pm the wind will be 12 knots from 272 degrees, it'll be 18C and it'll be raining 3mm/hour.
At best you might get some uncertainty expressed in "percentage chance of rain" but even then you're throwing away a ton of modelled information.
The flip flopping forecast that started the thread is probably an indication that the modelled probability distribution is multimodal - ie, there's a very good chance we'll be in one of two completely different regimes and they're roughly equally likely. On any given run noise is pushing one or the other to be *most* likely.
We'd have a much better result if the model was allowed to convey not just the level of confidence but also dependencies ("if the wind direction changes by 3pm there'll be heavy weather, if it doesn't then it'll be clear")
westernman
Well-known member
I would absolutely love to see this kind of information.
westernman
Well-known member
If the weather forecast says it will calm and sunny, and it turns out to be blowing a gale and raining, then I am annoyed.
If the weather forecast had said 55% sure it will be calm and sunny. 45% change of a gale and rain, then I would have stayed in the bar.
Not perfect, but better than the current situation where I often end up in a gale at Cap Bear although the forecast said calm.
boomerangben
Well-known member
The met office app already give a percentage chance of rain for pretty much anywhere in the uk every hour for that day and the next with up the next 5 days at 3 hour intervals. And in my experience it’s pretty accurate. If you live/sail near an airport that issues forecasts and a bit of goggling for a decode, you can use aviation forecasts which offer probabilities for wind gusts and changes
franksingleton
Well-known member
You make some valid points. You, I and some others in the forum would like to see some indication of reliability of the forecast. I would like to see ensemble grid point values. However, to the majority, particularly among the general public such data would cause confusion. I still think that the smoothing effect would defeat the object.
The uniformity is unrealistic.
Each centre, NOAA, ECMWF, UK, DWD etc optimises its model and data analysis software according to its own criteria. The forecast issued is the outcome. As far as I understand, it is not necessarily the most likely outcome that you might expect from the ensemble process. Further, the ensembles are run using double the grid length as the forecast run of the model. Ensembles do not generate alternative forecast output. They show the kind of spread that is likely in model output at the time. As far as I am aware, there is not enough computer power to run the full model 50 or however many times the ensemble is run.
Personally, I prefer to take a simple approach that does not try to extract more information than is really there. If a F5, say, is predicted, I assume that there will be some F6 with the chance of a F7 and similarly F4 and F3. I think that forecast centres use ensemble data in a subjective manne as guidance to human forecasters so that they can express levels of confidence.
When planning ahead, for example our return to Dartmouth from St Peter Port last Monday, I watch forecasts continually and look for consistency between successive model runs 24 0r 12 hours apart. About 7 days ahead a crossing on our preferred date looked possible. 6 days ahead, much less so. By 3 days ahead forecasts came back to being OK and remained so. The date was critical as one of the crew had a pressing need to return by the day after. I was under pressure to get it right. Anhinga does not motor well with an adverse wind so we were looking for a wind that might be sailable or possible to motor sail comfortably. It was that crew member’s second Channel crossing and she was nervous.
franksingleton
Well-known member
My late colleague David Houghton did a survey of “experiences” quoted by local sailors. He said that when someone claimed that sp-and-so always or often occurred, it really meant less than half the time. Memories are selective, so I take some of the claims made in this thread with a rock of salt. Having said that, to provide the kind of information suggested is probably a good, maybe ideal, application for AI. We will all have fairly specific requirements.
However, despite the continual improvements in modelling a the increasing availability of detailed global data, the atmosphere will have the last word. There are limits to predictability that, I believe, are not well understood.
However, despite the continual improvements in modelling a the increasing availability of detailed global data, the atmosphere will have the last word. There are limits to predictability that, I believe, are not well understood.
Stemar
Well-known member
We all remember the times the forecasters got it wrong, but forget the times they get it right. Also, "Yeah, they promised us a nice F4-5, but we got a 7 off that headland." Well, yes, that happens round headlands. What about the nice F4-5 for the rest of the journey?
Supertramp
Well-known member
It has made it hard this year, especially when cruising with family and not wanting to scare their socks off.
As others have said, I have found checking several weather models and keeping a close eye and ear to the coastguard forecasts during the day has given some clarity of what is 24 to 48 hrs ahead. But I have twice set off in forecast 4-5, then small craft warnings for 6+ issued and fully delivered. A good plan is to also monitor the actual weather and history at any nearby weather buoys (rather than land based).
Never seemed to have this trouble before the Internet....
As others have said, I have found checking several weather models and keeping a close eye and ear to the coastguard forecasts during the day has given some clarity of what is 24 to 48 hrs ahead. But I have twice set off in forecast 4-5, then small craft warnings for 6+ issued and fully delivered. A good plan is to also monitor the actual weather and history at any nearby weather buoys (rather than land based).
Never seemed to have this trouble before the Internet....
franksingleton
Well-known member
We all have our pry memories. Not long retired, we had a F9 off Start Point. The forecast “Perhaps gale 8 later” was still 12+ hours away. I complained to my former colleagues. The strongest actual reported win around that time over SW England was F6.
One of my truisms is that the atmosphere does not know itself to within one Beaufort force. Another is that it is often difficult to guarantee that there will not be winds up to F6 at some time, especially when there is a sea breeze effect. I never take the forecast in knots very seriously. I think in Beaufort force terms.
One of my truisms is that the atmosphere does not know itself to within one Beaufort force. Another is that it is often difficult to guarantee that there will not be winds up to F6 at some time, especially when there is a sea breeze effect. I never take the forecast in knots very seriously. I think in Beaufort force terms.
franksingleton
Well-known member
Some good advice there. Whilst we still have human input in marine forecasts, they should be heeded. It is significant that there is still human input into GMDSSh forecasts.
franksingleton
Well-known member
I have just seen https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj.3410, the abstract of a paper in the R Mer Society Quarterly Journal, a leading scientific publication. The full paper is behind a pay wall. It discusses machine learning, AI? as a technique for assessing confidence in forecasts. This was published 5 years ago.I