Deepmind presents an AI model for weather forecasting
When it comes to searching the world wide web, Google is the king. The tech giant is of course not sitting still and thinks it can do many more things better. Take predicting the weather. There’s Google’s AI division Deepmind has now developed an AI-enriched model for this. And because nothing is as changeable as the weather, this is a development that, if the accuracy is indeed as high as Google wants us to believe, could make quite a few people and organizers of (sports) events happy.
Better long-term weather forecasts
How often do we have to conclude that the forecasts of the KNMI, MeteoGroup (formerly Meteoconsult) and other weathermen and weather forecasters do not entirely approximate the truth. And there you are, with your shorts behind the BBQ in the pouring rain. Now, predicting the weather for the next day, 24 to 48 hours, is normally possible, but if you have to plan a party or other outdoor event for one or two weeks, it rarely happens that the so-called medium term expectation (10 to 14 days ahead) is also correct.
Of course, this all has to do with weather systems, the jet stream and many other factors. (Super)computers then calculate the most likely weather forecast, based on current weather maps and information from the past. Meteorologists then shed their light, knowledge and experience on it and a multi-day forecast is put together. However, this is rarely the same for all weather sites and services and is also subject to change.
Follow the ten or fourteen day weather forecasts on various weather sites or apps for a week. Then you not only see that they almost never all show the same expectations and that those multi-day expectations change every day, sometimes even several times a day. For us as consumers we just have to hope that we get the weather we hope for.
Google will be the best AI weatherman
Anyway, Google Deepmind thought that predicting the weather with the help of AI could be a lot better. They have worked on this, tested it and now also have a study on it published. The model is called GraphCast. And it turns out that it can make weather forecasts for ten days in advance better than any other weather model or computer.
Like all AI tools, data was used to train GraphCast, a lot of historical weather data in this case. Using that data, the AI weather model has learned how the weather develops, how weather systems such as high and low pressure areas arise, but also storms and other extreme weather and which factors influence this. In addition, GraphCast also looks at the weather at the time of calculating the forecast and the weather in the hours preceding it. With all that data and historical knowledge, GraphCast then makes a weather forecast.
GraphCast can predict the weather for the entire world. For this purpose, the AI model has divided the globe into areas of (calculated at the equator) 28 by 28 kilometers (0.25 square latitude/longitude). For each point in that grid, more than a million, a number of variables are calculated, at no fewer than 37 different height levels. These variables include temperature and wind direction, wind force and air pressure (corrected to sea level). In addition, a number of atmospheric variables are also calculated at each of the 37 altitude levels, such as relative humidity, temperature, wind direction and speed.
Better than the forecaster so far
It will come as no surprise that GraphCast training is quite a computationally intensive exercise. However, this now means that creating 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. By comparison, a 10-day forecast using a conventional approach such as HRES can take hours of computation in a supercomputer with hundreds of machines.
Great, GraphCast is fast, but of course we all want to know how good and accurate the predictions are. Google Deepmind has already tested this and the results are hopeful to say the least. HRES is currently the so-called gold standard for weather forecasting.
Compared to HRES, GraphCast’s predictions were better more than 90 percent of the time. Looking at the predictions for the troposphere – the 6-20 kilometer high region of the atmosphere closest to the Earth’s surface – GraphCast actually outperformed HRES 99.7 percent of the time.
Let’s hope that we won’t have to pay tens of euros for a subscription to Google GraphCast weather forecasts. Otherwise I’ll just look at the weather maps myself and ask the farmer on the corner if it’s going to rain.