Wind forecasts: Can AI and weather stations provide the perfect forecast for Lake Garda and Walchensee?

Addicted Sports runs a whole range of Webcams and wind gauges at surf spots. The founders, Andy Mattausch and Rainer Motloch, have now linked their database of measurements from these stations, using neural networks (AI), with up to 80 forecast parameters from the OpenMeteo database dating back to 2022, and have used this to create a a brand-new, in-house wind forecast programmed. Various ‘drivers’ for wind – including solar radiation, the exact wind direction, various pressure gradients in the Alps, high-altitude winds, atmospheric stratification and the Föhn forecast – are fed into the model. This produces a forecast for the prevailing wind and gusts – along with an assessment of whether it will be a “windy day” or not. We wanted to find out who is putting so much energy and dedication into this project and what the “drivers” behind it are.

​Andy, to start with, why don’t you tell us: Who are you actually, how did it all begin, and where are you at today?

Rainer Motloch and I founded Addicted Sports about 15 years ago. The reason was quite simple: we used to go surfing a lot and would travel to the lake time and time again, driven by our passion for it. Back then, we’d often drive from Munich to Lake Walchen or Lake Kochel – you’d set off in the dark, only to arrive at the spot and realise there wasn’t any wind at all. We didn’t want that anymore. The solutions available at the time were sometimes good, sometimes not so good. So we said: ‘We can do that ourselves.’ That was also when Foto-Webcam.eu came along. We thought the idea of using high-quality SLR cameras as webcams was brilliant. It meant you could still tell whether there was any wind, even at night. That was essentially the birth of Addicted Sports. Our first location was the Trimini on Lake Kochel – and we were lucky enough to be allowed to launch the project there, despite having no track record.

So your business originally started out as a service for streaming webcam footage for windsurfers?

Exactly. We wanted to know if there was any wind. That’s why we set up webcams and installed weather stations at the same time. During the day you can see the image; at night, the wind data comes in handy. That’s how the whole portal came about.

Our own webcams and weather stations

Many people probably think they’re using existing webcams or weather stations. Is that the case?

No. The webcams all belong to us, just like the weather stations. At Walchensee, for example, we have a very special measuring station: the anemometer is mounted on a buoy right in the middle of the lake. We set that up ourselves. That way, we get exactly the data we need.

And how does that work in practice?

We work with partners who provide us with the venues. However, all the technical equipment is supplied by us.

What sort of cameras do you use?

Digital single-lens reflex cameras, mostly Canon EOS 2000Ds, and occasionally a 600D. This enables us to achieve a very high standard of image quality. At night, the cameras have exposure times of up to 30 seconds. This means you can even see thunderstorms, lightning, stars and, sometimes, the Northern Lights. If you take a look at the gallery of our best photos, you’ll find lots of impressive night-time shots there.

Funding and personal background

How are you funding the whole thing?

About advertising and partners on the platform.

Is Addicted Sports your main job?

No.

You studied computer science and took meteorology as a minor. Was that with a view to Addicted Sports?

Addicted Sports had been around for ages by the time I was at university. I actually chose to take a minor in meteorology because of this interest – I wanted to understand the weather better and be able to gauge the wind at Walchensee more accurately.

The weather forecasts on your website are still quite new. How long have they been available?

It’s only been a few weeks. The actual development, however, has been going on for much longer. I made my first attempt in 2018 as part of my Master’s thesis. I was already working with neural networks back then. You could call it neural networks, or machine learning, or AI. In my view, they’re all synonyms. The basic principle worked, but there was a lack of historical forecast data to use as a training basis. Open-Meteo has only been recording such historical forecasts since 2022. It is precisely this data that has made it possible to implement the model now.

You’ve collected a lot of data yourselves over the years, haven’t you?

Exactly. On our website, there’s a history section for every location. There, you can see when a webcam was first installed, how many images have been captured, what weather data is available, and how many windy days there have been. Users can also set their own threshold for what constitutes a ‘wind day’. For the forecasts themselves, however, we use a fixed definition: a so-called ‘wind day’ begins when the base wind reaches eleven knots for at least one hour. This is how the forecasts are generated: the neural network learns.

Could you explain, as simply as possible, how your prediction works?

We are training a neural network. Historical weather forecasts from Open-Meteo serve as the input. As the output, we use the measured wind data from our own weather stations. The neural network therefore learns which weather conditions actually generated which winds at this exact spot at a later date.

​The neural network therefore learns which weather conditions actually produced which wind at that exact spot at a later date

What data, specifically, does the model process?

Depending on the location, there are around 80 different parameters. At Walchensee, for example, these include sunshine, cloud cover, temperature, wind direction at altitude and the night-time temperature. There are also a number of particularly important factors that we display alongside each forecast. This allows experienced users to assess the weather conditions for themselves as well. Most people know, for example, that without sunshine, the chances of good conditions at Walchensee are rather slim. The model itself, however, processes significantly more data than this overview shows.

How often are the models recalculated?

Three times a day – each time following the release of new forecast data from Open-Meteo at 6 am, 12 noon and 6 pm. Our models are then automatically recalculated.

Why are your forecasts for places like Walchensee or Lake Garda better than traditional weather models?

The large-scale weather models make predictions for grid cells several kilometres in size. This is too coarse for thermal spots. Our model, on the other hand, is trained directly on the specific spot in question and learns its local characteristics. Of course, the uncertainty remains the same as with any weather model: if the large-scale weather forecast is incorrect, our local forecast cannot be accurate either. After all, the forecast is based on that data.

Which advert was particularly difficult?

It took me several attempts to get the Lake Garda model right. Gradually, I added further influencing factors, such as weather developments north of the Alps. Rare weather conditions – which might only occur a few times a year – are particularly difficult to model. There hasn’t been enough training data available for this since 2022. The more years we add, the better our models will become. We also want to involve users more closely. Anyone who was there can best judge whether the forecast was accurate. Users can rate the forecast via our reports. This feedback is incorporated into the training of future models. So everyone can help to make the forecast even better in future, and I’d like to encourage everyone to do just that.

This is how accurate the wind forecasts are

Thermals are much harder to predict than a day of strong winds in Denmark. How well does it work these days?

For the Ora winds on Lake Garda, we currently achieve an accuracy rate of around 76 per cent. If we allow for a margin of error of one knot, this rises to around 80 per cent. It was also important to us to make the forecasts transparent. That’s why you can view older forecasts alongside the actual wind data and compare for yourself how accurate the model was.

You’re showing a four-day forecast. When will it actually be reliable?

The closer we get to the date, the more accurate the forecast becomes – just like with traditional weather models. A four-day forecast tends to indicate a trend. One day in advance, the prediction is, of course, much more reliable. That’s why we also show probabilities. If a day is classified as highly likely to be windy, then I’m now very confident that there will indeed be wind on that day.

​For example, if you forecast 12 to 17 knots – what are these figures based on?

Always for our specific measuring station. The model has been specifically trained for this station. Anyone familiar with the spot often knows how these figures can be applied to the actual surfing area. It is up to the user to make this connection. The stations and their readings have been in place for quite some time now, so many surfers are well acquainted with the specific characteristics of their local area.

Do you plan to include any more areas in future?

Yes. In principle, we aim to cover all locations where we have our own measuring stations. This applies not only to thermal flying areas, but also to classic wind spots such as Ammersee, Starnberger See and Chiemsee. In all these places, there are also local effects that aren’t fully accounted for in large-scale weather models, and which we can capture using our own models.

Andy, thank you very much for the interview, and I wish you continued success with your projects!