2026年11月12日(木)マイページ
← 記事にもどる 中学生版高校生版 Word

「印刷する」→ 送信先で「PDFに保存」を選ぶと PDF になります。解答は最後のページです。

原田英語の時事英語ジュニア|英語ニュース ワークシート(高校生レベル)

Class ____ No. ____ Name ____________________

天気予報は2週間が限界?昔の予報のまちがいを学ぶAIで、3〜4週間先がよくなった

AI Tool That Learns From Old Errors Improves Forecasts Weeks Ahead

音声・ゲームはこちらから

1An AI tool that learns from old forecast errors has improved weather predictions three to four weeks ahead, a study has found. The team included scientists from Microsoft Research, MIT and the European Centre for Medium-Range Weather Forecasts (ECMWF). The study has been published in the journal Nature.

2Forecasts for the next two weeks have improved steadily, but their skill falls sharply after that. Small errors build up at each step, and computer models have biases, which are errors that appear again and again. The atmosphere is also chaotic. The idea of a two-week limit dates back more than 50 years, to work by MIT scientist Edward Lorenz.

3The new method, called probabilistic bias correction, does not make forecasts on its own. Instead, part of it compares about 20 years of past forecasts with the observed weather, place by place and for each time of year. It then subtracts the errors that keep coming back. The tool needs little computing power, so it can be retrained every day.

4The team tested it on forecasts of temperature, air pressure and rainfall for weeks three and four, which in the study meant days 19 to 32. Skill here means how much better a forecast does than just using the usual weather for the season. In tests covering 2025, the tool doubled the low average skill score of ECMWF's AI forecasting system. The tool raised skill for extreme heat in week three by 51 percent over ECMWF's own corrected forecasts. For rainfall during floods, skill was 76 to 92 percent higher than in ECMWF's uncorrected forecasts. In fall 2025, a system built on the tool came first in an ECMWF contest for every weather type in both weeks.

5The researchers said the tool only targets errors that can be predicted and does not replace better models. Skill at three to four weeks is still modest, and a single forecast can be wrong. Anyone can use the team's computer code for free.

Words 単語

steadilyB1副着実にslowly but without stopping
sharplyB2副急に、ぐっとsuddenly and by a large amount
biases外名かたより、くせways of always leaning to one side
chaoticB2形カオスの(予測しにくい)very hard to predict because small changes grow
observedB1形観測されたwatched and measured carefully
subtractsB2動引く、引き算するtakes one number away from another
extremeB1形極端な、ひどいmuch stronger than usual
rainfallB1名雨量、降水the amount of rain that falls
modestB2形控えめな、それほど大きくないnot very large
replaceA2動~のかわりになるto take the place of something

単語のレベル:CEFR-J(『CEFR-J Wordlist Version 1.6』 東京外国語大学投野由紀夫研究室. (URL: https://www.cefr-j.org/download.html より 2026年8月ダウンロード))

1. Reading Quiz 本文の内容に合うものを選ぼう

  1. Which of these is NOT given as a reason forecasts get worse after two weeks?

    A. Small errors build up at each step.B. Computer models have biases.C. The atmosphere is chaotic.D. Weather centers stop collecting data.

  2. How does the new method work?

    A. It runs on a huge supercomputer once a year.B. It compares past forecasts with observed weather and subtracts repeated errors.C. It replaces all the computer models at ECMWF.D. It sends weather balloons into the atmosphere.

  3. In the study, what did "weeks three and four" mean?

    A. Days 19 to 32B. Days 1 to 14C. Days 3 to 4D. Days 29 to 42

  4. By how much did the tool raise skill for extreme heat in week three, compared with ECMWF's own corrected forecasts?

    A. By 20 percentB. By 92 percentC. By 51 percentD. By 40 percent

  5. What did the researchers say about the tool's limits?

    A. It can predict every storm perfectly.B. It only works for rainfall.C. It is too expensive to use every day.D. It only targets errors that can be predicted.

2. True or False 正しければ T、違っていれば F

  1. (     ) The new method makes forecasts on its own without any other model.

  2. (     ) The tool can be retrained every day because it needs little computing power.

  3. (     ) Before the tool was used, ECMWF's AI system already had a high skill score at three to four weeks.

  4. (     ) A system built on the tool came first in an ECMWF contest in fall 2025.

3. Fill in the Blanks 下の語から選んで空らんをうめよう(形を変えずに)

steadilyobservedbiasessubtractschaoticextreme

  1. Small errors build up at each step, and computer models have _____, which are errors that appear again and again.

  2. It then _____ the errors that keep coming back.

  3. The tool raised skill for _____ heat in week three by 51 percent over ECMWF's own corrected forecasts.

  4. The atmosphere is also _____.

  5. Forecasts for the next two weeks have improved _____, but their skill falls sharply after that.

  6. Instead, part of it compares about 20 years of past forecasts with the _____ weather, place by place and for each time of year.

4. Word Order 語を並べかえて正しい文にしよう

  1. ( chaotic / is / atmosphere / The / also ) .

  2. ( in / study / been / published / the / Nature / The / has / journal ) .

  3. ( the / use / computer / free / code / Anyone / for / can / team's ) .

5. Think and Write 英語で答えてみよう

  1. Should people trust AI weather forecasts for important plans such as school trips or sports days? Why or why not?(修学旅行や運動会のような大事な予定に、AIの天気予報を信じてもいいと思いますか?それはなぜ?)

  2. The tool improves by studying repeated mistakes. Do you think this is also a good way for students to study? Explain your opinion.(この道具は、くり返すまちがいを調べてよくなります。これは生徒の勉強法としてもよい方法だと思いますか?意見を説明しよう。)

Answers 解答

1. (1) D (2) B (3) A (4) C (5) D

2. (1) F (2) T (3) F (4) T

3. (1) biases (2) subtracts (3) extreme (4) chaotic (5) steadily (6) observed

4.

  1. The atmosphere is also chaotic.
  2. The study has been published in the journal Nature.
  3. Anyone can use the team's computer code for free.

全文訳

1昔の予報のまちがいから学ぶAIの道具が、3〜4週間先の天気予測をよくしたことが、研究でわかりました。研究チームには、マイクロソフト・リサーチ、マサチューセッツ工科大学(MIT)、ヨーロッパ中期予報センター(ECMWF)などの科学者が入っています。この研究は科学誌「ネイチャー」にのりました。

2この先2週間の予報は着実によくなってきましたが、その先になると、予報の力がぐっと落ちます。小さな誤差は予報の1歩ごとに積み重なります。また、コンピューターのモデルには「くせ」、つまり何度もくり返し出るまちがいがあります。そのうえ、大気もカオス(ほんの小さなちがいで先がまったく変わること)なのです。「2週間が限界」という考えは、50年以上前のMITの科学者エドワード・ローレンツの研究にさかのぼります。

3「確率的バイアス補正」と呼ばれる新しい方法は、それだけで予報をするものではありません。そのかわりに、道具の一部が、約20年分の昔の予報を、観測された実際の天気と、場所ごと・時期ごとにくらべます。そして、くり返し出るまちがいを引き算します。この道具は計算の力をあまり使わないので、毎日学び直すことができます。

4チームは、気温・気圧・雨の予報の「3週目と4週目」でためしました。この研究では、19日目から32日目のことです。ここでの「予報の力」とは、その季節のいつもの天気をそのまま使うより、どれだけよく当たるかということです。2025年を使ったテストで、この道具は、ECMWFのAI予報のしくみの、もともと低かった平均の点数を2倍にしました。この道具は、3週目の猛暑の予報の力を、ECMWFが自分で補正した予報より51%上げました。洪水のときの雨の予報の力は、ECMWFの補正していない予報より76〜92%高くなりました。2025年の秋には、この道具をもとにしたしくみが、ECMWFのコンテストで、どの気象の項目でも、3週目・4週目の両方で1位になりました。

5研究者たちは、この道具は予測できるまちがいだけをねらうもので、よりよいモデルのかわりにはならないと述べました。3〜4週間先の予報の力はまだそれほど高くなく、1回1回の予報ははずれることもあります。チームのプログラムは、だれでも無料で使えます。

文法ポイント:関係代名詞 which の非制限用法(, which ~)

コンマのあとの which が、前の語句にあとから説明を足す。本文の computer models have biases, which are errors that appear again and again(モデルには「くせ」がある。それは何度もくり返し出るまちがいのことだ)や weeks three and four, which in the study meant days 19 to 32 がその例。

出典:arXiv(論文の原稿。Nature 掲載論文と同じ研究)、ECMWF(ヨーロッパ中期予報センター)、Earth.com、GitHub(研究チームのプログラム公開ページ) ほか(記事は原田英語の時事英語ジュニアのオリジナル)|https://www.haradaeigo.com/jijieigo-junior/news/2026-11-12-ai-weather-forecast-two-weeks/