Hacker News Reader

Comment by sealaska | Hacker News Reader

Windy is fantastic.

But, I'm curious if there are any apps or weather models that give probability distributions? Telling me that it is going to blow 10 mph is useful, but telling me, for example, that there is a 60% chance that there will be a 10 mph wind and a 40% chance that the wind will blow at 15 mph seems more in line with how I would naively assume weather forecasting works.

Right now, the ECMWF [0] model in Windy shows 6 mph winds gusting to 30 mph for my area (SE Alaska, with lots of mountains and fiords, so a area that is gusty and hard to predict by nature). It almost feels like they're throwing their hands up in the air and admitting that they have no idea what the wind will actually do today. Which, is fine, if that's the case.

I just wish consumer weather forecasts did a better job communicating probabilities and uncertainties instead of spitting out a single value.

[0] https://www.ecmwf.int/

View on HN

ECMWF makes probabilistic forecasts, in the form of an ensemble of 50 IID examples. So this is mostly matter of Windy figuring out how to put that information into their UI.

Thanks for the reply. It sounds like you have a much better grasp of these things than I do.

So, when clients like windy forecast "10 knots east gusting to 20" for ECMWF, do you know if those numbers are directly copied and pasted from the ECMWF model, or do clients take the probabilistic forecast model and make some sort of average prediction that they display?

I'm very interested in how clients like windy go from a probibilistic forecast model to a singular hard number. Unless this is something built-in to ECMWF itself.

GFS takes a simple average of the ensembles for their avg model, I believe. The biggest problem is that you can't use statistics like you normally would in meteorology. Predictions are actually very good for knowing WHAT happens to a weather system, but not WHERE it happens.

I work in the sailing space, so this is what happens there:

A racing offshore sailor would consider a variety of models and ensembles. Then they would consider the various possibilities for the weather systems and fronts (e.g. front crosses your path +/-5% East/West within a +/-3h gap). A sailor would then avoid any possible dangerous situations, and pick a path that on the most number of models gives them an optimal wind.

It is both an art and a science. If you compare the results of sailors who compete in offshore competitions (such as the Vendee globe), where the boats are all constructed in very similar ways, you'll see that the best sailors consistently get good results, and that is based on good meteorology and a good gain to risk ratio for route choice.

That makes a lot of sense. I sail in SE Alaska.

You'd be surprised how many people make their living on the water up here, but their understanding of forecasting seems to amount to "download windy app"!

Anyways, the word ensemble is new vocabulary for me when it comes to forecasting - things are making more sense knowing what an ensemble is.

Ensembles aren't used much outside of competition or serious industry... windy/predictwind/CMAP have usually more than enough weather for the average boater

No argument from me there! But when the forecast completely misses in SE Alaska, or one forecast model displayed in Windy is way different than others, I naturally get curious as to what happened, so have been more interested in the nuts and bolts than is probably practical.

I'm more comfortable on the water knowing terms like ECMFW, myself!

GEFS is the noaa ensemble product. The operational GFS is actually just a single deterministic forecast run.

A lot of private providers roll up different agencies models to provide their own proper ie tart weather model because this is a lot cheaper than running the HPC required for most deterministic models.

The most important thing when checking the weather, especially with apps like windy, is to compare multiple models for your self to see how “stable” the forecasts are that day.

Windy does allow switching between different models, which can be helpful to roughly gauge probability. Definitely not perfect though. And some of the models are consistently inaccurate for specific locations in certain conditions in my experience.

MeteoSwiss does this both for temperature and precipitation. Tends to be the app everyone uses here https://www.meteoswiss.admin.ch

Perhaps something like this https://www.meteoblue.com/en/weather/forecast/multimodelense.... If you check the regular forecasts it also has a predictability figure.

If you click on a specific location, at the bottom you can expand the forecast. Instead of "Basic" select "Wind" which will let you see multiple model forecasts. Not exactly a probability but you can do some mental ensembling to see if models agree or not.

https://imgur.com/a/UezBGqZ

https://en.ilmatieteenlaitos.fi/local-weather/uk/london?fore... shows 50% and 80% confidence interval as well as median for rain and temperature.

Reply on news.ycombinator.com