Google develops new AI models / Result: ‚almost instant‘ weather forecasts
16. Januar 2020Google develops new AI models / Result: ‚almost instant‘ weather forecasts
New York, 16.1.2020
Weather forecasting is notoriously difficult, but in recent years experts have found that by using machine learning, they can achieve faster and better results. In a blog post published this week, Google presented new research findings that enable „almost instant“ weather forecasts.
While developers are still in the early stages of exploring the possibilities, the initial results look promising. In the non-expert work, Google researchers describe how they were able to predict accurate precipitation forecasts in just a few „minutes“, up to six hours in advance with a resolution of 1 km.
This is a major improvement over existing techniques that would take hours to produce forecasts, although they do so over longer periods of time and use more complex data.
According to the researchers, rapid forecasts will be „an essential tool needed for effective adaptation to climate change, especially in extreme weather conditions“. In a world increasingly dominated by unpredictable weather conditions, short-term forecasts are crucial for „crisis management and the reduction of loss of life and property“.
Google researchers compared their work with two existing methods: optical flow prediction (OF), which looks at the movement of phenomena such as clouds, and simulation prediction, which produces detailed physically based simulations of weather systems.
The problem with these older methods – especially physics-based simulation – is that they are incredibly computationally intensive. For example, US federal weather forecasting simulations would have to process up to 100 terabytes of data from weather stations every day and run for hours on expensive supercomputers.
„If it takes 6 hours to calculate a forecast, that only allows 3-4 runs per day and results in forecasts based on 6+ hours of old data, which limits our knowledge of what is happening,“ Google software engineer Jason Hickey wrote in a blog post.
By comparison, Google’s methods deliver results within minutes because they don’t try to model complex weather systems, but instead make predictions using simple radar data as a proxy for precipitation.
Google’s researchers trained their AI model on historical radar data collected by the National Oceanic and Atmospheric Administration (NOAA) in the neighbouring USA between 2017 and 2019. They say their predictions were as good as or better than three existing methods that made predictions from the same data, although their model was surpassed when trying to make predictions more than six hours in advance.
This seems to be the sweet spot for machine learning in weather forecasting at the moment: making fast, short-term forecasts, while leaving longer forecasts to more powerful models. NOAA’s weather models, for example, provide forecasts up to 10 days in advance.
While we have not yet seen the full impact of AI on weather forecasting, many other companies are also investigating this area, including IBM and Monsanto. And, as Google’s researchers point out, such forecasting techniques will only become more important in our daily lives as we feel the effects of climate change.