Weather forecast is notoriously difficult, but in recent years experts have suggested that machine learning can help sort the sun out of the ice. Google is the newest company to join, and in one blog post this week shared new research that says it makes “almost immediate” weather forecasts possible.
The work is at an early stage and still needs to be integrated into commercial systems, but early results look promising. In the non-peer review paper, Google researchers describe how they could generate accurate precipitation forecasts up to six hours in advance with a resolution of 1 km from just “minutes” of the calculation.
This is a major improvement on existing technologies. Generating forecasts can take hours, although they do this over longer periods of time and generate more complex data.
Rapid predictions, the researchers say, “will be an essential tool needed for effective adaptation to climate change, in particular for extreme weather.” management, and the reduction of life and property losses. “
The biggest advantage of Google’s approach to traditional forecasting techniques is speed. The company’s researchers compared their work with two existing methods: optical flow (OR) predictions, which look at the movement of phenomena such as clouds, and simulation forecasts, which create detailed physics-based simulations of weather systems.
The problem with these older methods – in particular physics-based simulation – is that they are incredibly computationally intensive. Simulations made by US federal weather forecast agencies, for example, must process up to 100 terabytes of data from weather stations every day and take hours to run on expensive supercomputers.
“If it takes 6 hours to calculate a forecast, it will only allow 3-4 runs per day, resulting in predictions based on 6+ hour old data, which limits our knowledge of what is happening now, “Software Engineer Jason Hickey of Google wrote in a blog post.
Google’s methods, in comparison, produce results in minutes because they do not attempt to model complex weather systems, but instead make predictions about simple radar data as a proxy for rainfall.
The company’s researchers have trained their AI model on historical radar data collected in the neighboring US between 2017 and 2019 by the National Oceanic and Atmospheric Administration (NOAA). They say their predictions were as good or better than three existing methods that make predictions based on the same data, although their model performed better when they tried to make predictions more than six hours in advance.
This seems to be the sweet spot for machine learning in weather forecasts at the moment: fast, short-term forecasts, while longer forecasts are left to more powerful models. NOAA weather models, for example, make forecasts up to 10 days in advance.
Although we have not yet seen the full effects of AI on weather forecast, many other companies are also investigating the same area, including IBM and Monsanto. And, as the Google researchers have noted, such prediction techniques will only become more important in our daily lives because we feel the effects of climate change.