Wellington bad local interp
-
On the east side of Wellington (New Zealand) harbour, predictions had been good but are now consistently 3-4 °C cold and winds are wrong for all models. Has the ML been reset? It is now unusable and compares badly to all other sources. There are now a number of wunderground stations - you have one that shows the errors near me (Eastbourne). Admittedly, we've been having a patch of unusually good and calm weather but it really makes the errors stand out.
-
@tbaisden
Hi, what do you mean by ‘Has the ML been reset?’
What is ML?Concerning weather models I understand you have tried all of them. During the hot season, at a location very close to the sea, they usually predict lower temperatures in the middle of the day. This is because some of the model grid points used to interpolate parameters at this location are over the sea, resulting in lower temperature.
See : https://community.windy.com/topic/23111/temperature-forecasts-appear-underestimated/8?_=1672128990499Regarding weather stations, those labelled pws or madis are personal weather stations and some of them are not accurate due to bad position of the station.
Better to refer to an official one, like that of Wellington airport, labelled ad.
The Observation vs forecast feature shows clearly that temperature forecast (dotted curve) is too low during the day compared to the reported temperature (colour shaded curve). Here shown with ACCESS forecast.
-
@idefix37 That's a very disappointing response. ML is machine learning (sorry for not explaining but I hoped it was obvious at this level). Call it what you want, but most weather forecasters including the NZ met service use it to correlate local conditions with the high quality stations and models below model grid resolution.
It could also be that the model lost the NZ topography, but it still seems to adjust for the mountain hut locations.
The bottom line is that something is wrong that used to work well. It is wrong at the airport (winds like you've picture become southerlies or calm around Wellington due to topography). It is more wrong in the east harbour. I was being generous with regard to the magnitude of the error - really it is unusable for the east side of the harbour now and data confirms that. It used to work - the area is a is important for yacht clubs and board sports. So it should the heart of windy's business.
Briefly, the airport and the top of the Metservice building are great stations, but are uniquely unprotected locations still influenced by local topography around them. Certainly use them but be aware they need to be augmented. I don't think it helps at all to demean local users who have put up and made high quality local stations available via the weather underground model. They do a good job of showing weather people experience and it is also possible to access and presumably multiple sites run by the port authority.
Based on this performance and response I gather the Windy approach is broken - just a pretty interface for the models and not interested in local conditions anymore? I've discontinued my premium Subscription but will continue to bang on about this since it renewed recently. And because Windy should work better and shouldn't break without warning in the windiest capital city in the world?
I don't think I should have to post screen shots for a service I paid for, but here's an example. (The peg shows the model, and is 15-18 SE throughout the area)
/var/folders/d4/cnxk8k9c8xj1qk001s7hr9br0000gr/T/com.apple.useractivityd/shared-pasteboard/items/7B2506D8-7CC1-4928-B984-E89D83175B2E/IMG_1630.jpeg -
Sorry to have disappointed you in trying to give you an answer. But Windy display raw forecasts of weather models like ECMWF, GFS, ICON, ACCESS…at their grid resolution with interpolation for local forecast. That’s all and nothing has changed.
It is obvious that the forecasts in a mountainous island and by the sea could be improved by the methods you explain. As far as I know, being a user not a developer, Windy does not make any local adaptation of the raw forecasts according to the terrain topography. What is achieved by the Services you mention in New Zealand would be difficult to achieve at whole world scale.
Apart from the models that I have mentioned, Windy offer the forecasts of Meteoblue, a Swiss company which performs the adaptation to the terrain and which obtains good results in the mountains. Their forecast for New Zealand is only accessible in the diagrams at a location but not as map form. -
@idefix37 Thanks for the replies.
Alas, I guess I'll accept the problems are baked into the models - at least into the relatively unusual conditions we've been having for the last couple weeks. I'm fascinated that Windy seemed to me to perform better in the past.
We did have a more normal northerly yesterday, and my local queries turned up an app that now provides two 1km model runs for the area. These really show the sort of difference that is so interesting and important on on the lee side of the harbour. I do think machine learning can mostly get this sort of stuff right, but the difference between the two models is also interesting. The mist above the harbour matched up pretty well with the areas shown in red, and unsurprisingly the two coarser scale models captured none of this. https://share.icloud.com/photos/042mTFFT9xfxwyQ9uGWWv4Y8Q
I hope you find the video interesting. I'll certainly be interested to know if there are other areas or apps with strategies for local interpretation matched to stations people have put up or found of interest. Even correcting the bias at the official Wellington stations might be of interest if it continues. I'll be interested to look at the reanalysis product at some point.
-
Thank you for the video, but I haven’t been able to see it totally and in detail because of a poor internet connection, being at the moment in mountain.
I see you are using PredictWind with their homemade models PWG and PWE, probably with 1km resolution?
I tried them but with the free app and the 8km resolution version only. So they don’t bring me any real advantages since Windy offers AROME with 1km resolution in France and neighbouring countries.
https://www.windy.com/menu?arome,45.921,6.372,5I don’t know how PredictWind get these 1km versions. They extrapolate them from lower resolution weather models using may be AI and ML means… But I don’t find them as good as ECMWF at 4 or 5 days ahead.
