Google creates superfast weather forecast
Scientists at Google have used artificial intelligence to make faster, more accurate local weather predictions.
They are working on a ‘‘deep learning’’ system to make forecasts within minutes by looking at radar images of a given location and working out if it will rain within the hour. The technique is known as ‘‘nowcasting’’.
Google said that the tool would take 10 minutes to create a six-hour forecast for a square kilometre. Conventional technology can take three hours to produce a forecast, ruling out short-term predictions. Traditional systems also need massive computing power: the world’s biggest forecasting centres process up to 100 terabytes daily and then take hours to run simulations on supercomputers.
Developers hope that the tool could one day be used for everyday activities, such as planning cycling routes that avoid rain, or pinpointing the ideal spot on a beach for a wedding. An instant system could be used for farming, disaster prevention and even by pilots wanting to avoid turbulence.
‘‘Weather predictions can inform people about whether they should take a different route to work, if they should reschedule the picnic, or even if they need to evacuate due to an approaching storm,’’ Jason Hickey, a senior software engineer at Google Research, said.
Engineers at the US software giant taught the system to make fast predictions by making it analyse vast numbers of precipitation patterns around America and noting how they changed. The network became good at predicting precipitation by looking at an image and analysing the information within.
This is different to forecasting that relies on data such as cloud motion, pressure, temperature and humidity. However, Google has admitted that its system struggles beyond a six-hour forecast and is less good when it comes to forecasting for far larger areas. The work is also in the early stages and has yet to be peer-reviewed or used in any commercial systems.
Last year IBM said it had created an app that could provide hourly updates for any location on Earth to the nearest three kilometres.