[趣闻，惊悚，机器学习] AI天气预报，超过了“大牛 EC”？

欧洲中期天气预报中心的集合模型，ENS： the European Centre for Medium-Range Weather Forecasts’ ensemble model (ENS)

？： GenCast

数值天气预报： numerical weather prediction, NWP

机器学习： machine learning

基于机器学习的天气预报： machine learning-based weather prediction, MLWP

相对经济价值曲线： relative economic value, REV

https://www.nature.com/articles/d41586-024-03957-3

   Price and his colleagues trained the AI on global weather data from 1979 to 2018 and then predicted the weather of 2019. To check its accuracy, they compared GenCast forecasts with actual weather data and ENS forecasts for that year.

   GenCast was more accurate than ENS on 97% of the measures used on a scorecard to evaluate such ‘probabilistic’ forecasts. It was also better at forecasting extreme heat, cold and wind, as well as tropical-cyclone tracks.

   GenCast produces one 15-day forecast within 8 minutes on an AI processing chip. This speed is “quite substantially faster” than the time it takes conventional models, Price says.

   【机器翻译】Price和他的同事们利用1979年至2018年的全球天气数据对人工智能进行了训练，然后预测了2019年的天气。为了检查其准确性，他们将GenCast的预测与当年的实际天气数据和ENS预测进行了比较。

   在记分卡上用于评估此类“概率”预测的97%的指标上，GenCast比ENS更准确。它在预测极端高温、寒冷和大风以及热带气旋路径方面也做得更好。

   GenCast在人工智能处理芯片上在8分钟内生成一个15天的预测。普莱斯说，这种速度比传统型号的时间“快得多”。

https://www.nature.com/articles/d41586-024-03957-3

图1  原文 Fig. 3: The marginal forecast distributions of GenCast are skilful and well-calibrated.

a, CRPS scores for GenCast versus ENS4 in 2019. The scorecard compares CRPS skill between GenCast and ENS across all variables and eight pressure levels. Dark-blue cells on the scorecard indicate a variable, lead time and level combination for which GenCast has 20% better (that is, lower) CRPS than ENS, whereas dark-red cells indicate 20% lower CRPS for ENS (white means they perform equally). The results show that GenCast significantly (P < 0.05) outperforms ENS on 97.2% of all reported variable, lead time and level combinations. Hatched regions indicate neither model is significantly better. b–f, Spread/skill scores for GenCast and ENS for selected variables. Both models are generally well-calibrated with spread/skill close to 1. g,h, REV for predictions of the exceedance of the 99.99th percentile for 2 m temperature and 10 m wind speed, at lead times of 1 day, 5 days and 7 days. GenCast consistently achieves greater REV than ENS whenever either forecast is better than climatology, particularly at small cost/loss ratios.

   【机器翻译】a、 2019年GenCast与ENS4的CRPS得分。记分卡比较了GenCast和ENS在所有变量和八个压力水平下的CRPS技能。记分卡上的深蓝色单元格表示一个变量、交付周期和级别组合，其中GenCast的CRPS比ENS好20%（即更低），而深红色单元格表示ENS的CRPS低20%（白色表示它们的性能相同）。结果表明，在所有报告的变量、交付周期和水平组合中，GenCast在97.2%上显著优于ENS（P<0.05）。阴影区域表示这两个模型都没有明显改善。b–f，所选变量的GenCast和ENS的扩散/技能得分。这两种模型通常都经过很好的校准，点差/技巧接近1。g,h、REV，用于预测在1天的交付周期内，2m温度和10m风速超过99.99%，5天和7天。当任一预测优于气候学时，GenCast的REV始终高于ENS，特别是在成本/损失率较低的情况下。

https://www.nature.com/articles/s41586-024-08252-9/figures/3

https://www.nature.com/articles/s41586-024-08252-9

参考资料：

[1] 许小峰，2018-04-16，大牛“EC” 精选

https://blog.sciencenet.cn/blog-1310230-1109328.html

[2] 光明网，2024-12-12，AI技术引领气象预测领域飞速发展

https://baijiahao.baidu.com/s?id=1818211548496653309&wfr=spider&for=pc

[3] Alix Soliman. DeepMind AI weather forecaster beats world-class system [J]. Nature, 2024

doi:  10.1038/d41586-024-03957-3

https://www.nature.com/articles/d41586-024-03957-3

[4] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds,  Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, Matthew Willson. Probabilistic weather forecasting with machine learning [J]. Nature, 2024

doi:  10.1038/s41586-024-08252-9

https://www.nature.com/articles/s41586-024-08252-9

相关链接：

[1] 2024-05-23，[打听，讨论] ECMWF 人工智能天气预报 AIFS 的核心是什么？

https://blog.sciencenet.cn/blog-107667-1435372.html

[2] 2021-12-10，[资料] 《天气预报的三次跃进》学习笔记

https://blog.sciencenet.cn/blog-107667-1316021.html

感谢您的指教！

感谢您指正以上任何错误！

感谢您提供更多的相关资料！