Xinhua Finance and Economics, Beijing, November 16 (Reporter Ge Chen) - Google's "Deep Thinking" company based in London, UK, has recently developed the artificial intelligence model GraphCast for predicting mid-term weather. According to current indicators, its calculation speed and accuracy are better than traditional prediction models.
Medium term weather forecast usually refers to the forecast of weather trends within the next 4 to 10 days, and its accuracy is related to the work plans of industries such as agriculture, construction, and tourism during the same period.
The research team published an online paper in the US journal Science on the 14th, stating that GraphCast can use current weather conditions and weather data from 6 hours ago to predict the next 6 hours of weather, and the predicted results every 6 hours are fed back into the model for longer-term predictions.
They first trained GraphCast using global weather data predicted by traditional models from 1979 to 2017. Researchers say that GraphCast utilizes deep learning to skip the tedious equation operation steps in traditional weather prediction, saving a lot of computational power.
Researchers have tested data from the European Center for Medium Range Weather Forecasts since 2018 and found that GraphCast can predict the weather in 10 days within 1 minute; Compared with the hours of operation results of the European Center for Medium Range Weather Forecasting's "High Resolution Forecasting" model, GraphCast's 90% data prediction results are more accurate among 1380 test data points; In certain high-altitude areas, the accuracy of 99.7% of the data prediction results is better than that of "high-resolution prediction".
The researchers stated in the paper that GraphCast can also provide early warning for extreme weather events, by predicting tropical cyclone trajectories, extreme temperatures, and predicting atmospheric water vapor dense transport belts such as "atmospheric rivers" that bring large amounts of rainfall.
The first author of the paper and the head of the research team at Google's "Deep Thinking" company, Remi Ram, said that they trained GraphCast on 32 computers over a period of 4 weeks, resulting in a lightweight algorithm that could run on a desktop computer and produce results in just one minute.
Matthew Chantri, the machine learning coordinator of the European Center for Medium Range Weather Forecasting, said that based on the evaluation of indicators currently used, the GraphCast model is superior to traditional prediction models, but in the future, if other indicators are used for evaluation, the results may be slightly different.
Currently, multiple institutions around the world have developed artificial intelligence weather prediction models. Chantri believes that machine learning is driving the development and changes of weather forecasting, but it is still in the experimental stage and will not completely replace traditional methods. Instead, it can enhance the prediction areas that traditional methods are not good at, such as predicting rainfall within hours.