ECMWF 和 GFS 模型

这篇文章是我们的Model Mania系列文章的第三篇，希望为那些没有严格大气科学背景的人简要介绍天气模型。在前两篇文章中，我解释了天气模型究竟是什么/它们如何工作以及区域和全球天气模型之间的区别。这篇文章将深入探讨两个最著名的天气模型，GFS 和 ECMWF。这些都是全球模型，这意味着它们都试图服务于相同的目的（提前 3-10 天预测大规模天气模式），但它们有何不同以及为何不同？

GFS 是美国政府在国家海洋和大气管理局 (NOAA) 及其附属机构的领导下运行的全球模型。GFS 模型由美国纳税人资助，这意味着任何想要它的人都可以免费获得它的预测输出。如果您想从 GFS 模型中获得新鲜的原始输出数据，您可以从 NOAA 的网站上免费下载。虽然原始数据是免费提供的，但使这些数据对最终用户有用需要大量的后处理，其中原始数据被“按摩”成比模型生成的 1 和 0 字符串更容易被人类识别的格式本身。您在weather.us和weathermodels.com 上看到的地图和图表使用我们的气象学家和程序员开发的后处理算法生成。

ECMWF 也是一个全球模式，但它不是由美国政府管理，而是由一个独立的政府间实体管理，并得到 34 个欧洲国家的支持。请注意，这里的命名法有点混乱，但 ECMWF 代表欧洲中期天气预报中心，是组织和模型的名称。ECMWF 该组织的成立方式意味着他们可以并且确实为他们的预测数据收费，尽管国际法规定其中一些（归类为 WMO-Essential）可用于公共利益。如果您访问大多数免费天气模型站点，您将看到的唯一 ECMWF 数据是 WMO-Essential，它仅包含少量分辨率非常低的参数。在撰写本文时，除了我们的网站之外，我们还没有发现任何网站可以让您免费查看全分辨率 ECMWF 数据。然后应用您的后处理算法使其可用于人类。或者，您可以让我们为您处理，并在weather.us或weathermodels.com 上查看最新的 ECMWF 输出。

既然您知道每个模型的负责人以及对其信息的访问有何不同，您可能想知道他们的预测有何不同。上图是2019年12月14日晚美国ECMWF（左）和GFS（右）六天预报的对比，目前还不知道哪个模型预报更准确（我会在一分钟内讨论更多关于 GFS 和 ECMWF 模型的准确性），但立即可以看到一些关键差异。

首先，您会注意到 GFS 模型降水（阴影）场（在此上下文中的场是指网格数据集，或者在网格上的许多不同点有一个值的数据集）比 ECMWF 降水场更加像素化。这是因为 GFS 的运行分辨率低于 ECMWF 模型，这意味着它的网格点相距更远（GFS 模型中每 13 公里放置一个网格点，而 ECMWF 模型中每 9 公里放置一个网格点）。正如您在本系列的前一期文章中所回忆的那样，较低的分辨率通常意味着预测的准确性较低，因为模型不知道有更多的大气和地形特征。

您还会注意到当时美国东部风暴系统预报的位置和结构存在重大差异。这些差异是由于分辨率差异（如上所述）、数据同化差异（告诉模型现在大气中正在发生什么的过程）以及每个模型用来转换其给定的控制方程组的差异的组合将初始条件转化为预测。最后一句话中的术语听起来很陌生吗？查看本系列的第一篇博文，了解初始条件、控制方程以及运行天气模型的更多基础知识。

对每个模型采用的不同控制方程（有时称为物理包）的详细解释需要大量的高级数学，所以我会给感兴趣的读者留下几个链接，这些链接指向最新的详细文档（如of 12/9/19)ECMWF和GFS模型计算每个网格点的预测。详细解释每个模型的数据同化过程需要类似的数学，因此如果您希望将线性代数知识付诸实践，我建议您阅读 GFS 模型用于数据同化的技术（Ensemble卡尔曼滤波）。对于那些有兴趣深入研究数据同化但不想处理线性代数或偏微分方程的人，ECMWF 开发了一个优秀的短期课程（约 1 小时），解释了他们的（更高级的）数据如何同化过程无需深入数学即可工作。

那么一般来说哪个模型更准确呢？

从统计上讲，非常明确的答案是 ECMWF 始终比 GFS 表现更好，如上面的模型技能得分图所示。自 2007 年以来（并且可能在此之前的一段时间内），GFS 从未对北半球 20 到 80N 之间的 5 天预测通常比 ECMWF 更准确。话虽如此，在很多情况下，对于特定风暴，GFS 比 ECMWF 更准确。例如，GFS 早在 ECMWF 之前就预测了热带风暴多里安的形成。也许一个更著名的例子是 2015 年 1 月 27 日的暴风雪，ECMWF 预测纽约市将有超过 2 英尺的雪，这将使这座城市完全停顿。另一方面，GFS 预测不到一英尺，这将是破坏性的，但肯定不会造成瘫痪。

尽管 GFS 偶尔会“失利”，但 ECMWF 仍然是中期天气预报领域的一贯领导者，但为什么呢？

这个故事的第一部分是ECMWF组织有一个很大的责任范围较窄比诺阿一样。根据他们各自的网站，这里是每个组织的使命的快速比较。

请注意，ECMWF 的 2/3 目标与生成准确的天气预报或进行研究有关，以生成更准确的天气预报。虽然天气预报是 NOAA 使命中至关重要的一部分，但这只是其整体职责中相对较小的一部分，其中包括更广泛的环境问题。

承认这两个机构都非常重要，并且以相对较小的成本为纳税人提供了非凡的社会价值，承认这一点非常重要。我解释 NOAA 和 ECMWF 之间的组织差异的目的不是争辩说一个比另一个更好，或者 NOAA 应该更像 ECMWF，反之亦然。话虽如此，如果您的目标是准确的天气预报，则有助于将整个组织的大部分精力投入到该特定目标上。这并不是说欧洲不致力于预测气候、海洋和海岸的变化，也不致力于保护和管理其沿海和海洋生态系统，但这些任务并未留给 ECMWF，这使其能够更加专注于提高他们天气预报的准确性。

也许此时您会认为 NOAA 和 ECMWF 之间的组织比较对美国不公平，正是因为 NOAA 负责的职责范围如此广泛。然而，即使在完全致力于天气预报的 NOAA 子机构（国家环境预测中心，NCEP）的子机构（环境建模中心，EMC）内，也出现了类似的模式。虽然 ECMWF 运行一个全球模型和 51 个集合成员（更多关于我们集合系列中的成员），EMC 维护着一套综合模型，范围从 GFS 全局模型到 HRRR（高分辨率快速刷新）区域模型再到 HWRF （飓风天气研究和预测）特定于飓风的区域模型。结果，资源池（时间、金钱、计算能力、

虽然这可能看起来效率低下，但请考虑这些其他模型提供的价值。HRRR 在预测大雪带和雷暴群等小尺度特征何时何地形成以及它们的强度方面表现出相当大的技巧。此外，HRRR 每小时运行一次，这意味着与我们必须每六个小时等待一次全局模型运行相比，预测人员可以更频繁地根据新信息重新评估他们的预测。说到飓风，ECMWF 和 GFS 模型在预测飓风或热带风暴可能去哪里方面做得相当好，但在确定强度方面却出了名的糟糕这将是。HWRF 模型通过提供更准确的强度预测来帮助填补这一空白，这对于做出关于可能需要在风暴前疏散海岸线的哪些部分至关重要的决策。

虽然美国政府天气预报系统支持的区域模型套件限制了我们全球模型 GFS 的准确性，但它可以更全面地了解我们的大气层，而 ECMWF 并非旨在提供。当然，在美国进行天气预报的真正双赢将是为美国的天气建模留出足够的时间、金钱和计算能力，这样我们就可以拥有一个与 ECMWF 技能相匹配的全球模型，并且一整套专业/区域模型，但您必须与您的国会代表讨论。

虽然这篇文章篇幅较长且文字较多，但我希望它有助于您理解 GFS 和 ECMWF 模型是什么、它们有何不同，以及为什么 ECMWF通常比 GFS 更熟练。在下一篇Model Mania 博文中，我将解释为什么需要这两种模型，尽管 GFS 在性能方面始终落后于 ECMWF。

===============================

Hello everyone!

This post is the third in ourModel Maniaseries which hopes to give a brief introduction to weather models for those without a rigorous atmospheric science background. In the previous two posts, I explainedwhat exactly weather models are/how they work andthe difference between regional and global weather models. This post will take a deeper dive into the two most famous weather models, the GFS and ECMWF. These are both global models, which means they’re both trying to serve the same purpose (predict large scale weather patterns 3-10 days in advance), but how and why are they different?

The GFS is the global model run by the US Government under the leadership of the National Oceanic and Atmospheric Administration (NOAA) and its subsidiary agencies. The GFS model is funded by American taxpayers, which means its forecast output is freely available to anyone who wants it. If you want the raw output data fresh from the GFS model, you candownload it for free off NOAA’s website. While the raw data is available for free, making that data useful to end users requires a substantial amount ofpost-processingwhere the raw data is ‘massaged’ into a format more recognizable to humans than the strings of 1’s and 0’s produced by the model itself. The maps and graphs you see atweather.usandweathermodels.comare produced using post-processing algorithms our meteorologists and programmers have developed.

The ECMWF is also a global model, but instead of being run by the US Government, it is run by an independent intergovernmental entity supported by 34 European nations. Note that the nomenclature here isa bit confusing, but ECMWF stands for the European Center for Medium-Range Weather Forecasts and is the name of both the organization and the model. The way ECMWF the organization is set up means that they can and do charge for their forecast data, though international law dictates that some of it (classified as WMO-Essential) is available for the public good. If you go to most free weather model sites, the only ECMWF data you’ll see is WMO-Essential, which only includes a small handful of parameters at a very low resolution. As of this writing, we are not aware of any site besides ours that lets you view full-resolution ECMWF data for free. If you want the raw forecast output from the ECMWF, you’ll have to pay for it andthenapply your post-processing algorithms to make it usable for humans. Alternatively, you could let us take care of that for you and view the latest ECMWF output atweather.usorweathermodels.com.

Now that you know who’s in charge of each model and how access to their information differs, you’re probably wondering how their forecasts are different. The above graphic shows a comparison of the ECMWF (left) and GFS (right) six-day forecasts for the US on the evening of December 14th, 2019. I don’t yet know which model forecast will be more accurate (I’ll discuss more about the accuracy of the GFS and ECMWF models in a minute), but immediately a few key differences are visible.

First, you’ll notice the GFS model precipitation (shaded) field (fieldin this context means gridded dataset, or a dataset for which there is a value at many different points on a grid) is much more pixelated than the ECMWF precipitation field. This is because the GFS is run at a lowerresolutionthan the ECMWF model, meaning it has grid points located farther apart (one is placed every 13km in the GFS model compared to every 9km in the ECMWF model). As you’ll recall from the previous installment of this series, a lower resolution generally means a less accurate forecast, as there are more atmospheric and topographic features the model is unaware of.

You’ll also notice substantial differences in the placement and structure of the storm system forecast over the Eastern US at the time in question. These differences are due to a combination of resolution differences (mentioned above), differences indata assimilation(the process of telling the model what’s happening in the atmosphere right now), and differences in the sets of governing equations each model uses to turn its given initial conditions into a forecast. Did the terms in that last sentence sound unfamiliar? Check out the first post of this series to learn about initial conditions, governing equations, and more of the basics of running a weather model.

A detailed explanation of the different governing equations (sometimes referred to asphysics packages) employed by each model would require a substantial amount of advanced math, so I’ll leave the interested reader with a couple links to detailed documentation of how the latest (as of 12/9/19)ECMWFandGFSmodels calculate their forecasts at each gridpoint. A detailed effort to explain each model’s data assimilation process would require similar mathematics, so if you’re looking to put your knowledge of linear algebra to practical use, I’ll suggest reading up on the technique the GFS model uses for data assimilation (Ensemble Kalman Filtering). For those who are interested in a deeper dive into data assimilation but don’t want to deal with linear algebra or partial differential equations, theECMWF has developed an excellent short course(~1 hour) that explains how their (much more advanced) data assimilation process works without getting too deep into the math.

So which model isgenerally speakingmore accurate?

Statistically speaking, the very clear answer is that the ECMWF consistently performs better than the GFS, as the model skill score graph above shows. At no point since 2007 (and likely for a while before then) has the GFS produced an generally more accurate 5-day forecast for the Northern Hemisphere between 20 and 80N than the ECMWF. That being said, there have been many cases where the GFS has been more accurate than the ECMWF for specific storms. For example, the GFS predicted the formation of Tropical Storm Dorian long before the ECMWF did. Perhaps a more famous example was the snowstorm of January 27th, 2015 where the ECMWF forecasted over two feet of snow for New York City which would have brought the city to an absolute standstill. The GFS on the other hand predicted less than a foot, which would be disruptive but certainly not crippling. Central Park ended up recording 9.8″ of snow.

Despite the occasional ‘loss’ to the GFS, the ECMWF remains the consistent leader in medium-range weather prediction, but why?

The first part of this story is that ECMWF the organization has amuchnarrower range of responsibilities than NOAA does. According to their respective websites, here is a quick comparison of the mission each organization has.

Notice that 2/3 of ECMWF’s objectives relate to either producing accurate weather forecasts or conducting research for the purpose of producing more accurate weather forecasts. While weather forecasts are a crucially important part of NOAA’s mission, it is only one relatively small part of their overall responsibilities which includes a much broader range of environmental concerns.

It is extremely important to acknowledge thatboththese institutions are tremendously important, and provide extraordinary societal value for the relatively small costs they impart on taxpayers. My goal in explaining the organizational differences between NOAA and ECMWF is not to argue that one is better than the other, or that NOAA should be more like ECMWF or vice versa. That being said, if your goal is accurate weather forecasts, it helps to devote most of your entire organization to that specific aim. That’s not to say that Europe isn’t working on predicting changes in climate, oceans, and coasts, or working to conserve and manage its coastal and marine ecosystems, but those tasks are not left to ECMWF, which enables it to focus more intently on improving the accuracy of their weather forecasts.

Perhaps at this point you’re thinking that the organizational comparison between NOAA and ECMWF is unfair to the US precisely because NOAA is charged with such a broad range of responsibilities. However, even within the sub-agency (the Environmental Modeling Center, EMC) of the sub-agency (the National Centers for Environmental Prediction, NCEP) of NOAA devoted entirely to weather prediction, a similar pattern emerges. While the ECMWF runs one global model and 51 ensemble members (more on those to come in our ensemble series), EMC maintains a comprehensive suite of models ranging from the GFS global model to the HRRR (High Resolution Rapid Refresh) regional model to the HWRF (Hurricane Weather Research and Forecasting) hurricane-specific regional model. As a result, the pool of resources (time, money, computational power, and personnel) the US has dedicated to numerical weather prediction is split among a wide range of models, each with a specific purpose.

While this might seem inefficient, consider the value provided by these other models. The HRRR has demonstrated considerable skill in predicting when and where small-scale features like heavy snow bands and thunderstorm clusters will form, and how intense they’ll be. Additionally, the HRRR is run every hour which means that forecasters can re-evaluate their predictions based on new information much more frequently than if we had to wait for a run of the global models every six hours. When it comes to hurricanes, the ECMWF and GFS models do a fairly good job predictingwherea hurricane or tropical storm might go, but are notoriously bad at figuring outhow strongit will be. The HWRF model helps to fill in that gap by providing more accurate intensity forecasts that are critically important to making decisions about which parts of the coastline might need to be evacuated ahead of a storm.

While the suite of regional models supported by the US Government’s weather prediction system has limited the accuracy of our global model, the GFS, it enables a much more comprehensive view of our atmosphere that ECMWF isn’t designed to provide. Of course, the real win-win-win for weather prediction here in the US would be to have enough time, money, and computational power set aside to weather modelling in the US so we can have both a global model that matches ECMWF’s skillanda full suite of specialized/regional models, but you’ll have to take that up with your congressional representatives.

While this post was on the longer and text-heavier side, I hope it has been helpful to your understanding of what the GFS and ECMWF models are, how they’re different, and why the ECMWF isgenerallymore skillful than the GFS. In the nextModel Maniapost, I’ll explain why both models are needed despite the fact that the GFS consistently lags behind the ECMWF in terms of performance.