Could artificial intelligence have shown the power of storm Poly earlier in the weather forecasts? A weather forecast is made with large calculation models. They calculate how the state of the atmosphere will change in the coming hours or days, based on physical principles relating to radiation, clouds, wind and precipitation, among other things. Such an expectation is never perfect, the always chaotic atmosphere is difficult to calculate. It is especially difficult to predict the exact location of a shower.
Artificial intelligence may be able to significantly improve forecasting in the coming years. This has been shown in two different studies global expectations and local precipitationwhich is the scientific journal Nature published Wednesday.
Recognize patterns
Artificial intelligence (AI) is extremely suitable for learning to recognize patterns in large amounts of data, which has already been proven in numerous applications. Plenty of experiments are therefore being done to see whether AI can also offer advantages for making weather forecasts. For a long time, however, artificial intelligence underperformed.
The first research that Nature now publishes, done by researchers from the Chinese company Huawei, shows an AI weather model that can make global forecasts up to a week ahead. It has been trained on a whopping 39 years of historical weather data. The performance of the model is comparable to that of existing weather models that make forecasts on this large scale.
What is special is that the AI model gives results 10,000 times faster. That speed is an important advantage, because running existing weather models requires a lot of computer power. Many weather bureaus run the model a number of times for a forecast, each time with slightly different starting values. The final expectation is a probability forecast based on those different outcomes. If this process can be so much faster, it will save computer capacity, time and energy.
I found this methodologically quite innovative
Maurice Schmeits KNMI
“This model is causing quite a stir in the meteorological world. It’s been public for a while so people already have it can test with”, says Maurice Schmeits, who is involved in R&D in the field of weather models at KNMI and who also looks at what AI can mean for KNMI. “The speed is impressive. The models we use now divide the world into compartments and solve all sorts of mathematical equations for each compartment. That takes a lot of calculation time, it goes in steps of a few minutes, up to 15 days ahead. The AI model was trained on a lot of data, which also took a lot of time, but now that it has been developed, very little computing power is required. Incidentally, this model still relies heavily on the data from the European Weather Center, for both the training data and the initial values that the AI model uses for calculations.”
“The model performs very well on a number of variables, sometimes better than the best global weather models, and that is really an achievement,” says Schmeits. “But the study does not show anything about precipitation, while that is of course very important. I can imagine that in a few years a kind of hybrid form will be used for these kinds of expectations. Then AI models are used, for example, to post-process the results from the existing weather models in order to correct them, something that we are also researching at KNMI.”
Heavy local showers
The second investigation Nature is entirely focused on forecasting precipitation in the short term. Heavy local showers in particular cause a lot of danger and nuisance, so improvements in this area are valuable. The researchers (from the University of California, Berkeley in the US and Tsingua University in Beijing, China) made a so-called ‘nowcasting’ model. In such models, an area is divided into small squares and a short look ahead – in this case up to three hours ahead in squares of 2 by 2 kilometers. What is striking about this AI nowcasting model is that it not only does pattern recognition, but also includes a bit of the physics behind the processes in the atmosphere. The AI model outperforms existing nowcasting models in 71 percent of cases.
“I found this methodologically quite innovative,” says Schmeits. “Because a little physics is also included, you can get better forecasts than AI that only looks at precipitation images from the radar. Our current numerical models are also not very good at nowcasting and forecasting showers. I also envision a hybrid form of the existing approach and AI models with nowcasting.”
“The disadvantage of AI is that it is only as good as the data it was trained with. The question is how it deals with new extremes,” says Schmeits. “Due to climate change, we are more likely to experience extreme weather. You can’t train the model on that so easily, something clever has to be devised for that.”