Show HN: Meteosource Weather API: current, forecast and historical data using ML

[Luka78]

As Dark Sky's API will stop working soon I thought I'd suggest a possible replacement.

It is well known that most weather APIs just download data from NOAA and "repackage" them to JSON format. This means you will get data only for the nearest grid point which might be a few kilometers away and at a different altitude than your actual location.

Meteosource combines more models (GFS, ECMW, UKMO, GEM,...). It compares individual models' performance to historical data and uses machine learning to create a single output. This approach minimises inaccuracies of individual models and feeds a proprietary hyper-local model that computes weather exactly for the specified location.

The historical data API is based on actual measurements (not forecasts as usual) and there is a free developer plan available.

I greatly appreciate any feedback and thoughts!

[sosodev]

I find your website kinda sketchy. Your FAQ claims that the company was founded in 2007 but that's not true. meteocentrum.cz was founded in 2007 but this company didn't appear until 2021.

meteosource.com also appears to be a rebrand of forecasts.cloud. Why?

The multiple rebrandings and false claims make me hesitant to trust you.

[Luka78]

It is originally a European-based company founded in 2007 and started as a website, mobile app, and data provider serving mainly corporate clients. Meteosource as a service (low-cost weather API) started at the end of 2020 and this is just a new service operated by the same company, not rebranded company or anything else... I don't think there is anything sketchy in this.

[sosodev]

That's a fine answer. Maybe that should be the answer in the FAQ. The website just felt a bit scam-like to me so I did some more research. Finding out stuff like that, when it isn't clear on the website, only furthers the gut feeling of "this might be a scam".

Your answer ignores why you didn't pursue forecasts.cloud as this new venture instead of meteosource though.

[Luka78]

Thank you for your suggestion - is it mainly the FAQ answer which you find a bit puzzling or anything else we should improve on the website?

Regarding forecasts.cloud - this website will focus on our applied modelling (renewable generation etc.) and will serve as an offer of models for specialised industries, not weather API. It is not fully operational yet (the English version will be launched later this year) - I am impressed you have been able to find it :)

[sosodev]

I think the red flags are:

* vague testimonials

* scrolling list of companies (with no evidence)

* lack of specificity around payments (who's the processor?)

* the random face shown above the "contact us" button

* the lack of details around how the systems actually work (the "how we create our AI weather forecasts" link in your FAQ is just a 404)

* the lack of evidence that your system actually works better the alternatives

* the nearly-pointless blog spam

* the hand-wavy claims of "nearly 100% API uptime"

It's not that I think your API won't work and that you're just trying to steal my money. Instead it seems like this is a marketing effort to sell a low quality service using the magic buzzword "AI".

[Luka78]

Thank you for your comments! We will improve some of the points mentioned.

[probablypower]

Interesting service, this post comes at a convenient time, as I'm looking at different weather data services to bring into my company.

I note that you state 8 years of historical data is available for one of the priced plans, and then on another page there is mention of larger bulk historical data services for custom plans. How much further back can your historical data go? Is it the sort of thing that we could pay once for, for a few thousand GPS points?

The use case is to get historical weather data at each power pole on an electrical transmission system. Obviously the observational and forecast data is of interest to us, but we can't make much value of the observations if we don't have the historical context with which to build real-time models on.

[Luka78]

Thank you for your comment! We are going to extend the historical weather to 20 years soon. At the moment it is possible to prepare a customised dataset based on your needs - please get in touch at support(at)meteosource.com and we will be happy to help with your specific case.

[Slix]

I was disappointed by Accuweather's "local" weather during this spate of Midwest storms. It didn't correctly tell me whether it was raining at my current position.

I miss Dark Sky.

[sosodev]

How does the "hyper-local" computation work if all the sources are very generalized? Does that mean when your service says there's an 89% chance of rain at my location that's an actual, local probability?

[Luka78]

The machine learning model combines more weather models as mentioned and this feeds proprietary local models that compute the weather exactly for your location. If you request a specific GPS, the chance of rain will be computed for the exact point based on the orography and nearby grid points.

[iudqnolq]

Well wouldn't this only be better if you've come up with something superier to all the folks at NOAA. They combine weather models into the grid product every day with local human expert judgement rather than AI.

Have you published a backtest against NOAA's data? Have you looked into why they don't provide more hyperlocal data themselves? There's some interesting papers on the work they've done to get wildland firefighters appropriately local data.

[Gys]

I assumed this was a service were I can upload my own data as actual measurements. But that does not seem to be the case?

[anotheryou]

Is there still weather services not using ML?

Is it all huge modeling still?

[Luka78]

Well, most available weather APIs just download data from NOAA and "repackage" them to JSON format and do not add much value.

[anotheryou]

And this one differs? Is it better than NOAA?

[Luka78]

Meteosource learns from historical data and forecasts from many different models (including those from NOAA), compares the data from different places and meteorological situations, and based on this machine learning creates a single output that minimises individual models' biases and local inaccuracies. This is where it should add value.

[anotheryou]

Interesting the ML corrects and averages the models instead of predicting itself.