又到了每周一次的 Nature Podcast 时间了！欢迎收听本周由Ali Jennings和Benjamin Thompson带来的一周科学故事，本期播客片段里讨论了基于AI模型的即时预测降雨方法。欢迎前往iTunes或你喜欢的其他播客平台下载完整版，随时随地收听一周科研新鲜事.
音频文本：
It’s the age-old question: should I have brought my umbrella today? Predicting precipitation in the very near future is important in many situations – managing emergency services, giving early flood warnings and controlling air traffic, as well, of course, as optimising your daily commute. This kind of short-term forecasting is known as nowcasting – predicting precipitation patterns up to just two hours ahead. One popular way to do this is to use radar. Microwave pulses bounce back off airborne particles and can be measured to provide a picture of the surrounding precipitation. By combining radar data with an atmospheric physics equation that calculates how air moves, nowcasters can build up short-term precipitation predictions. But there’s a problem.
Yeah, so that method is not particularly great at predicting thunderstorms and other heavy precipitation events.
This is Suman Ravuri from the company DeepMind in the UK. Suman is part of a team exploring how deep learning – a particular kind of artificial intelligence – could help solve nowcasting problems like these. The physics equations used in nowcasting rely on assumptions that don’t always map onto reality, but deep learning models don’t rely on these equations. Instead, they’re trained entirely on radar data, and lots of it. For example, Suman’s model is trained on 1.5 million examples of previous rainfall in the UK. Suman and the DeepMind team aren’t the first to use deep learning for nowcasting, but their model is a little different from what has come before. It’s called a deep generative model.
So, a deep generative model is a model that gives you alternate futures, in this case, of precipitation.
Suman and his team trained their new deep generative model to imagine what future rain might look like, based on radar data taken over the previous 20 minutes.
We spent a lot of time making sure that the predictions themselves didn't only look like pretty videos of rain but were ones that were actually consistent with what happened in reality. It's really exciting kind of seeing this model as it’s training over time, going from making predictions that look nothing like reality to sort of slowly getting to things that look like possible evolution of precipitation.
The predictions made by the DeepMind team’s model are judged by another program called a discriminator. The discriminator tries to discriminate between the precipitation predictions the nowcasting model has generated and what actually happened in reality. When their discriminator couldn’t tell the difference between their model and reality, the team were happy with their model’s accuracy, and they asked 56 expert meteorologists from the UK Met Office to judge the predictions of their model against predictions from other existing models.
Interviewee: Suman Ravuri
They preferred our predictions 89% of the time. And they also told us that this is something that is a step change in terms of what they're used to working with.
Interviewer: Ali Jennings
The model had much higher detail than what meteorologists are used to and was the best at representing risk, according to some of the meteorologists who took part. And Suman wants his model to have real, practical applications.
Interviewee: Suman Ravuri
I hope either this or some future work of this does form the base of predictions for issuing flood forecasting or other warnings.
Interviewee: Suzanna Maria Bonnet
In Rio de Janeiro, some streets have flood problems and landslides, and it's very important to take decisions to mobilise security sectors.
Interviewer: Ali Jennings
This is Suzanna Maria Bonnet, a forecaster and researcher at the Federal University of Rio de Janeiro.
Interviewee: Suzanna Maria Bonnet
This type of model is very important to give this type of information to the meteorologists who will call the authorities.
Interviewer: Ali Jennings
But although Suzanna thinks that DeepMind’s model could be extremely useful for disaster prediction, she points out a significant problem for making this model accessible to a wide group.
Interviewee: Suzanna Maria Bonnet
We can have a very big, complex model, but if we don’t have computational resource to run this model, we will have to cut the model.
Interviewer: Ali Jennings
And there’s a more specific issue that could affect whether forecasters like Suzanna could use this prediction tool, this time with the parameters within the model shaping its predictions.
Interviewee: Suzanna Maria Bonnet
If I want to use this model here in Brazil, some parameters are not specific for my region. it's a problem we have with numerical weather prediction models. Many studies that change parameterisations inside the model, they are made in the higher latitudes. It's not applied for our tropical region. It changes a lot the results.
Interviewer: Ali Jennings
Even back in the UK, Suman appreciates that they’ll need to work on the model further with meteorologists and experts in the field to get it to a stage where it can be deployed in practice. But he and Suzanna agree that they’d also like the model to take account of more different kinds of data.
Interviewee: Suman Ravuri
I think maybe we’re, at the moment, 20 to 30% of the way there. I think there’s a lot more work that we can do about incorporating other types of weather data. However, what’s really interesting to me is incorporating more physics into our deep learning models. So, that’s something that we haven’t scratched the surface of yet and something that I really hope that a lot of people think about and work on.
Interviewer: Ali Jennings
And predicting rain isn’t the only thing on DeepMind’s radar. Wind patterns and temperature change are both things DeepMind are considering tackling in the future. So, perhaps one day I’ll finally know what to wear on the way to work.
That was A
li Jennings. For this story he spoke to Suman Ravuri from DeepMind in the UK and Suzanna Maria Bonnet from the Federal University of Rio de Janeiro in Brazil. To find out more, look for a link to the paper in the show notes
联系客服
