AI developed by and the in the UK can predict rain more accurately than current forecasting models over the very short term.
UK-based DeepMind 鈥� an AI subsidiary of Google parent company Alphabet 鈥� has previously achieved high-profile success with neural networks trained to play the game Go and to investigate protein folding. It has now applied its deep-learning approach to short-term rain 鈥渘owcasts鈥�. These usually focus on the next 2 hours of weather and are used by outdoor event organisers, flood warning systems and the aviation industry.
The research team trained a neural network on weather radar data collected from 2016 to 2018 and tested it using data from 2019. Prior to 2016, radar data was less detailed. The resulting model can make predictions over areas measuring up to 1536 kilometres by 1280 kilometres and predict the chance of rain in a given 1km by 1km area from 5 to 90 minutes ahead using 5 minutes of historical radar data.
Advertisement
In experiments the team compared predictions from the DeepMind tool with those from two existing tools, one which attempts to model the atmosphere using mathematics and another which also uses a deep-learning approach. DeepMind claims that in a blind study of 50 meteorologists from the Met Office, its AI model was ranked first for accuracy in 89 per cent of experiments.
聽at DeepMind says current weather prediction models for long-term forecasts tend to be based on computationally intensive mathematical simulation and perform poorly in the short-term, and that specialist nowcasting tools used to fill the gap leave room for improvement.
鈥淸The existing technology is] really amazing from sort of 6 hours up to two weeks in terms of weather predictions, but there is this area, especially around 0 to 2 hours, in which the models perform particularly poorly,鈥� he says.
Ravuri says humans are likely to remain involved in forecasting even if AI tools are adopted. 鈥淎t some point you are going to need experts and humans in the loop in order for them to make sure what they鈥檙e understanding in terms of the predictions are sensible and then how to communicate that to the public,鈥� he says. 鈥淚t鈥檚 hard to see an automated system doing all of that.鈥�
at the Met Office said: 鈥淭his research demonstrates the potential AI may offer as a powerful tool for improving our short-term forecasts and our understanding of how our weather patterns are evolving.鈥�
Nature
