Upgrade to Pro — share decks privately, control downloads, hide ads and more …

時系列データ予測手法の宇宙天気予報への応用

 時系列データ予測手法の宇宙天気予報への応用

TECH in Kyoto #1
TECH in Kyoto 第1回目の発表資料です。時系列データ予測手法の宇宙天気予報への応用についてご紹介します。

227382dbd5e033db211c159edf32853c?s=128

Hacarus Inc.

April 28, 2022
Tweet

More Decks by Hacarus Inc.

Other Decks in Science

Transcript

  1. ࣌ܥྻσʔλ༧ଌख๏ͷ Ӊ஦ఱؾ༧ใ΁ͷԠ༻ ಺໺޺ढ़!5&$)JO,ZPUP ▪ Qiita https://qiita.com/ufield/items/9acc9f6e69b6617f8ef1 ▪ Github https://github.com/ufield/swpy

  2. ✓ ಺໺ ޺ढ़ (͏ͪͷ ͻΖͱ͠) ✓ 2017೥ ཧֶ (ത࢜) -

    ஍ٿ࣓ؾݍͷ෺ཧաఔͷݚڀ ✓ 2021೥ HACARUS ʹத్ೖࣾ - ΞϓϦέʔγϣϯΤϯδχΞ ✓ ݸਓతڵຯ - ػցֶशͷࣗવՊֶ΁ͷԠ༻ - WebγεςϜ։ൃ 2 ࣗݾ঺հ -20℃ͷத࣓ྗܭΛຒΊ͍ͯΔ @Χϓεέʔγϯά (Χφμ)
  3. ✓ ʰΨ΢εաఔͱػցֶशʱ - Ψ΢εաఔΛཧղ͢ΔͨΊͷྑॻ ✓ Ψ΢εաఔͷԠ༻ྫ - ϕΠζ࠷దԽ (࣮ݧܭըɾϋΠύϥαʔν) -

    ༧ଌͷ৴པ۠ؒͷ෇༩ ✓ Ψ΢εաఔͷ࣮σʔλ΁ͷԠ༻ྫ͕গͳ͍ 3 ࠓճͷൃදͰಘΒΕΔ͜ͱ ࣮ࡍͷ࣌ܥྻσʔλ΁ͷԠ༻ํ๏͕Θ͔Δʂ ࠓճͷൃද
  4. 4 ൃද֓ཁ 1. ଠཅ஍ٿ؀ڥܥͱӉ஦ఱؾ༧ใ 2. LSTM Λ༻͍ͨ Dstࢦ਺ ༧ଌ 3.

    Ψ΢εաఔճؼʹ͍ͭͯ 4. Ψ΢εաఔճؼΛಋೖͨ͠ Dst ࢦ਺ ༧ଌ ✓ ଠཅ෩ͱ஍ٿ࣓ؾݍ ✓ Ӊ஦ఱؾͱ͸ ✓ ࣓ؾཛྷ ͱ Dstࢦ਺ ✓ σʔληοτ ✓ ೖྗύϥϝʔλ ✓ LSTMʹΑΔ Dstࢦ਺ ༧ଌ ✓ Ψ΢εաఔճؼͷಋೖ ✓ ༧ଌ݁Ռߟ࡯ ✓ Future Work ✓ ·ͱΊ ✓ ઢܗճؼͱΨ΢εաఔ ✓ Χʔωϧ ✓ Ψ΢εաఔճؼ
  5. 5 1. ଠཅ஍ٿ؀ڥܥͱӉ஦ఱؾ༧ใ ✓ ଠཅ෩ͱ஍ٿ࣓ؾݍ ✓ Ӊ஦ఱؾͱ͸ ✓ ࣓ؾཛྷ ͱ

    Dstࢦ਺
  6. 6 ஍࣓ؾ ஍ٿ͸ݻ༗࣓৔(஍࣓ؾ)Λ࣋ͭ࿭੕

  7. ✓ ଠཅ෩ - ଠཅ͔Βৗ࣌ਧ͖ग़͢ϓϥζϚ - ࣓৔΋ଘࡏ ✓ ஍ٿ࣓ؾݍ - ଠཅ෩ͱ஍࣓ؾͷ૬ޓ࡞༻ʹΑΓੜ·ΕΔྖҬ

    - ஍࣓ؾ༝དྷͷ࣓৔͕ϓϥζϚͷӡಈΛίϯτϩ ʔϧ - ൃදऀͷֶੜ࣌୅ͷݚڀͷઐ໳ྖҬ 7 ଠཅ෩ͱ஍ٿ࣓ؾݍ ·ͱΊͯଠཅ஍ٿ؀ڥܥ(Solar-Terrestrial Environment)ͱݺͿ ଠཅ෩ ஍ٿ࣓ؾݍ
  8. 8 ଠཅ׆ಈͷ஍ٿ΁ͷӨڹ ଠཅ׆ಈΛىҼͱͨ͠Ұ࿈ͷࣗવݱ৅Λ·ͱΊͯӉ஦ఱؾͱݺͿ Ӊ஦ఱؾ༧ใ = ଠཅ஍ٿ؀ڥܥશମͷঢ়ଶΛ೺Ѳɾ༧ใ͢Δ͜ͱ

  9. ✓ ࣓ؾཛྷ - ஍࣓ؾ͕શ஍ٿతʹେ͖͘มಈ͢Δݱ৅ - େن໛ͳଠཅϑϨΞ͕ൃੜͨ͠ͱ͖ʹى ͜Γ΍͍͢ ✓ Dst ࢦ਺

    - ஍࣓ؾͷมಈΛද͢ࢦ਺ - -100 nT ҎԼͰ࣓ؾཛྷͱݴΘΕΔ (-50nTͱ͍͏ఆٛ΋) 9 ࣓ؾཛྷͱ Dst ࢦ਺ ࣓ؾཛྷͷݪཧ ࣓ؾཛྷ࣌ͷ Dst ࢦ਺ͷมಈ
  10. ✓ ࣓ؾཛྷ࣌ʹ஍্༠ಋిྲྀ͕ൃੜ - 1989೥3݄ͷେ࣓ؾཛྷ • ΧφμͰ600ສੈଳఀి • χϡʔδϟʔδʔभͷݪൃఀࢭ ✓ ࣓ؾཛྷͷن໛(=

    Dstࢦ਺)Λ༧ଌΛ͢Δ͜ͱ͸ࣾ ձతҙٛ΋େ͖͍ - ༧ଌʹج͍ͮͯ༧๷ࡦΛͱΕΔ 10 ࣓ؾཛྷͷࣾձతӨڹ https://www.jaxa.jp/article/interview/vol65/index_j.html 1989೥3݄ͷ࣓ؾཛྷ࣌Ͱނোͨ͠ χϡʔδϟʔδʔभͷૹి༻มѹث 1989೥10݄21೔ ๺ւಓްాଜ(౰࣌)ͰࡱӨ͞ΕͨΦʔϩϥ ▪ ࣓ؾཛྷͷਖ਼ͷ໘ ௿Ң౓஍Ҭ΋ΦʔϩϥΛݟΔ͜ͱ͕Ͱ͖Δ
  11. 11 2. LSTMΛ༻͍ͨ Dstࢦ਺ ༧ଌ ✓ σʔληοτ ✓ ೖྗύϥϝʔλ ✓

    LSTMʹΑΔ Dstࢦ਺ ༧ଌ ࢀߟ࿦จ Gruet, Marina A., et al. "Multiple-Hour-Ahead Forecast of the Dst Index Using a Combination of Long Short- Term Memory Neural Network and Gaussian Process." Space Weather 16.11 (2018): 1882-1896.
  12. ✓ Dst ࢦ਺σʔλ ✓ ଠཅ෩σʔλ - |B|: ଠཅ෩࣓৔ڧ౓ [nT] -

    Bz: ଠཅ෩ೆ޲͖੒෼ڧ౓ [nT] - V: ଠཅ෩଎౓ [km/s] - N: ଠཅ෩ϓϥζϚີ౓ [cm^-3] 12 ࣓ؾཛྷ࣌ͷσʔλ ࣓ؾཛྷճ෮૬ ࣓ؾཛྷओ૬ ࣓ؾཛྷͷൃੜن໛͸ɺଠཅ෩͔Β஍ٿ࣓ؾ ݍʹྲྀೖ͢ΔΤωϧΪʔʹґଘɻ্هͷଠ ཅ෩தͷ෺ཧύϥϝʔλ͕ॏཁ
  13. 13 ܇࿅σʔληοτ Start time End Time Min. Dst (nT) 1

    2000-02-11 23:00 2000-02-14 11:00 -133 2 2000-05-23 16:00 2000-05-27 07:00 -147 3 2000-07-15 06:00 2000-07-19 06:00 -301 4 2000-09-17 11:00 2000-09-21 07:00 -201 5 2000-10-12 15:00 2000-10-16 06:00 -107 6 2000-10-28 13:00 2000-10-31 16:00 -127 7 2000-11-26 13:00 2000-12-01 10:00 -119 8 2001-03-30 19:00 2001-04-04 17:00 -387 9 2001-04-11 07:00 2001-04-16 01:00 -271 10 2001-04-17 17:00 2001-04-20 19:00 -114 11 2001-08-17 01:00 2001-08-19 04:00 -105 12 2001-09-25 11:00 2001-09-27 11:00 -102 13 2001-09-30 13:00 2001-10-06 01:00 -166 14 2001-10-21 08:00 2001-10-25 20:00 -187 15 2001-10-27 18:00 2001-11-01 01:00 -157 16 2001-10-31 10:00 2001-11-03 13:00 -106 17 2002-03-23 06:00 2002-03-26 05:00 -100 Start time End Time Min. Dst (nT) 18 2002-04-17 02:00 2002-04-23 16:00 -149 19 2002-05-11 04:00 2002-05-14 01:00 -110 20 2002-08-01 01:00 2002-08-04 05:00 -102 21 2002-11-20 07:00 2002-11-24 01:00 -128 22 2003-05-29 07:00 2003-06-01 22:00 -144 23 2003-06-17 17:00 2003-06-20 20:00 -141 24 2003-08-17 09:00 2003-08-21 01:00 -148 25 2003-11-19 23:00 2003-11-24 14:00 -422 26 2004-01-21 21:00 2004-01-24 15:00 -130 27 2004-04-03 06:00 2004-04-05 02:00 -117 28 2004-07-22 11:00 2004-07-31 09:00 -170 29 2004-08-29 20:00 2004-09-02 12:00 -129 30 2005-05-07 13:00 2005-05-11 13:00 -110 31 2005-05-29 13:00 2005-06-02 01:00 -111 32 2005-06-12 08:00 2005-06-15 05:00 -103 33 2005-08-31 03:00 2005-09-02 14:00 -122 34 2005-09-10 05:00 2005-09-15 20:00 -139 Start time End Time Min. Dst (nT) 35 2006-12-14 06:00 2006-12-19 04:00 -159 36 2011-10-24 12:00 2011-10-28 13:00 -147 37 2012-04-23 08:00 2012-04-28 02:00 -120 38 2012-07-14 16:00 2012-07-19 23:00 -139 39 2012-09-30 05:00 2012-10-03 16:00 -122 40 2012-10-07 20:00 2012-10-12 11:00 -109 41 2013-03-16 21:00 2013-03-20 23:00 -132 42 2013-05-31 16:00 2013-06-04 20:00 -124 43 2014-02-18 06:00 2014-02-23 16:00 -119 44 2015-03-16 21:00 2015-03-22 14:00 -222 45 2015-06-22 03:00 2015-06-29 22:00 -204 46 2015-10-06 17:00 2015-10-11 20:00 -124 47 2015-12-31 02:00 2016-01-02 17:00 -110 48 2017-05-27 13:00 2017-05-29 19:00 -125 49 2018-08-25 09:00 2018-08-29 09:00 -174 શ49Πϕϯτ
  14. 14 ςετσʔληοτ Start time End Time Min. Dst (nT) 1

    2000-08-10 11:00 2000-08-15 08:00 -235 2 2000-10-02 06:00 2000-10-08 07:00 -182 3 2000-11-06 04:00 2000-11-08 21:00 -159 4 2001-03-19 04:00 2001-03-22 21:00 -149 5 2001-04-21 17:00 2001-04-25 09:00 -102 6 2001-11-23 21:00 2001-11-27 19:00 -221 7 2002-09-03 21:00 2002-09-06 13:00 -181 8 2002-09-30 19:00 2002-10-03 20:00 -176 9 2003-07-10 16:00 2003-07-13 17:00 -105 10 2004-11-07 11:00 2004-11-16 14:00 -374 11 2005-05-14 18:00 2005-05-20 15:00 -247 12 2005-08-23 23:00 2005-08-28 18:00 -184 13 2011-09-26 06:00 2011-10-01 07:00 -118 14 2012-03-08 16:00 2012-03-12 20:00 -145 15 2012-11-13 08:00 2012-11-17 01:00 -108 16 2013-06-28 01:00 2013-07-02 09:00 -102 17 2015-12-19 19:00 2015-12-23 17:00 -155 18 2016-10-12 21:00 2016-10-15 07:00 -103 શ18Πϕϯτ ✓ Πϕϯτநग़ʹ͍ͭͯ - -100nTҎԼʹୡͨ͠Πϕϯτ - ࣓ؾཛྷओ૬ɾճ෮૬ΛؚΉظؒ
  15. 15 LSTMʹΑΔ Dst ࢦ਺༧ଌϞσϧ LSTM Cell Fully Connected Layer 1

    with Dropout & ReLu Dst T+p Dst T−5 h ˜ Bz T−5 h ˜ N T−5 h ˜ V T−5 h ˜ |B| T−5 h Dst T−4 h ˜ Bz T−4 h ˜ N T−4 h ˜ V T−4 h ˜ |B| T−4 h Dst T−3 h ˜ Bz T−3 h ˜ N T−3 h ˜ V T−3 h ˜ |B| T−3 h Dst T−2 h ˜ Bz T−2 h ˜ N T−2 h ˜ V T−2 h ˜ |B| T−2 h Dst T−1 h ˜ Bz T−1 h ˜ N T−1 h ˜ V T−1 h ˜ |B| T−1 h Dst T ˜ Bz T ˜ N T ˜ V T ˜ |B| T : ଠཅ෩࣓৔ೆ޲͖੒෼ : ଠཅ෩࣓৔ڧ౓ : ଠཅ෩ϓϥζϚີ౓ : ଠཅ෩଎౓ Bz |B| N V ※ ͸෺ཧྔ ͷ࣌ࠁ ͷฏۉ ˜ X T X T − 1 ∼ T ✓ p = 4 h ✓ Output dim : 125 (LSTM), 25(FC) ✓ Dropout ratio : 0.5 ✓ Batch size: 32 ✓ Epochs: 100 LSTM Cell LSTM Cell LSTM Cell LSTM Cell LSTM Cell Fully Connected Layer 2 with Dropout ݱࡏ࣌ࠁ ʹରͯ͠ɺ ࣌ؒޙͷ Dstࢦ਺ Λ༧ଌ T p ࣮ݧύϥϝʔλ
  16. ✓ ࣌ؒޙͷ Dstࢦ਺ (੺ઢ) ✓ ෆे෼ͳ఺͋Γ (ಈ࡞ʹ஫ྗ) - ςετσʔλʹର͢ΔධՁ -

    ύϥϝʔλνϡʔχϯά p = 4 16 LSTMʹΑΔ Dst ࢦ਺༧ଌ݁Ռྫ ͜͜·Ͱ͸Α͋͘ΔLSTMΛ༻͍ͨ࣌ܥྻ༧ଌ
  17. 17 ༧ଌσʔλͷ࣮ࣾձʹ͓͚Δ׆༻ ✓ ୯७ͳ༧ଌ஋͚ͩͰͷରࡦ͸े෼͔ʁ - ޡࠩ΋ؚΊͯͲͷఔ౓ͷରࡦ͕ඞཁ͔Λ൑அ͍ͨ͠ ޡ͕ࠩΘ͔Ε͹࠷ѱέʔε ͷରࡦΛͱΕΔ ༧ଌ஋͸ͲΕ͙Β͍ͷޡࠩʁ ޡࠩͷ࠷ѱέʔεΛߟ͑ͨͱ͖ʹ

    Ͳͷఔ౓ͷରࡦ͕ඞཁʁ ୯७ͳ༧ଌ https://www.bousai.go.jp/kohou/kouhoubousai/h29/88/news_02.html
  18. 18 3. Ψ΢εաఔճؼʹ͍ͭͯ ✓ ઢܗճؼͱΨ΢εաఔ ✓ Χʔωϧ ✓ Ψ΢εաఔճؼ ࢀߟॻ੶ͷ

    3.1~3.5 ষ
  19. 19 ઢܗճؼ جຊతͳ1࣍ݩͷೖྗʹର͢ΔઢܗճؼΛߟ͑Δ ˒ ઢܗճؼͷදݱྗΛߴΊ͍ͨ 㱺 جఈؔ਺Λ૿΍ͤ͹ྑ͍ y = w

    0 + w 1 x + w 2 x2 + w 3 x3 = wTϕ(x) w = (w 0 , w 1 , w 2 , w 3 )T, ϕ(x) = (1, x, x2, x3)T جఈؔ਺ ϕ h (x) = exp ( − (x − μ h )2 σ2 ) y = H ∑ h=−H w h exp ( − (x − μ h )2 σ2 ) ͷͱ͖جఈؔ਺ͷݸ਺͸ 21 ݸɻ ͷ࣍ݩΛ 10࣍ݩʹ͢Δͱ? ɹˠ ݸ (࣍ݩͷढ͍) H = 10 x 2110 = 16,679,880,978,201 Ψ΢εաఔʹΑΔղܾʂ
  20. 20 Ψ΢εաఔ ࣍ݩͷढ͍ͷղܾํ๏ɿ ʹ͍ͭͯظ଴஋ΛͱΔ ɹɹˠ Ϟσϧ͔Β Λੵ෼ফڈ͢Δ w w Nݸͷೖग़ྗϖΞ

    ʹରͯ͠ɺ ઢܗճؼϞσϧ͸ (y 1 , x1 ), . . . , (y N , xN ) y 1 y 2 ⋮ y N = ϕ 0 (x1 ) ϕ 1 (x1 ) ⋯ ϕ H (x1 ) ϕ 0 (x2 ) ϕ 1 (x2 ) ⋯ ϕ H (x2 ) ⋮ ⋮ ϕ 0 (xN ) ϕ 1 (xN ) ⋯ ϕ H (xN ) w 0 w 1 ⋮ w H y = Φw : ܭըߦྻ Φ ॏΈ ͕ɺฏۉ Ͱ෼ࢄ ͷΨ΢ ε෼෍͔Βੜ੒ w 0 λ2I w ∼ (0, λ2I) ͷ෼෍΋Ψ΢ε෼෍ʹै͏ y = Φw ͷظ଴஋ͱڞ෼ࢄߦྻ y = Φw [y] = Φ [w] = 0 Σ = [yyT] − [y] [y]T = [(Φw)(Φw)T] = Φ [wwT]ΦT = λ2ΦΦT y ∼ (0, λ2ΦΦT) ೚ҙͷೖྗू߹ ʹ͍ͭͯɺରԠ͢Δग़ྗ ͷಉ࣌෼෍ ͕ଟมྔΨ΢ε෼෍ʹै͏ͱ͖ɺ ͱ ͷؔ܎͸Ψ΢εաఔ (x1 , x2 , …, xN ) (y 1 , y 2 , …, y N ) x y Ծఆ ఆٛ
  21. 21 Ψ΢εաఔͷҙຯ ڞ෼ࢄߦྻΛ ͱ͓͘ͱɺڞ෼ࢄߦྻͷ֤ཁૉ͸ K = λ2ΦΦT K n, n′

    = λ2ϕ(xn )Tϕ(xn′ ) ͱ༩͑ΒΕΔɻڞ෼ࢄߦྻͷ֤ཁૉ͸ɺೖྗ ͱ ͷಛ௃ϕΫτϧ ͱ ͷ ಺ੵͷఆ਺ഒͰ͋ΔɻଟมྔΨ΢ε෼෍ʹ͓͍ͯɺڞ෼ࢄߦྻ͕େ͖͍ 㱺 ࣅͨ஋Λऔ Γ΍͍͢ɻ Ψ΢εաఔͷੑ࣭ ɹɹࣅͨಛ௃Λࣔ͢ೖྗͷ૊Έ߹ΘͤͰ͋Ε͹ɺग़ྗͷ ͱ ΋ࣅͨ஋ͱͳΔ xn xn′ ϕ(xn ) ϕ(xn′ ) y n y′ n y ∼ (0, K) ฏۉ ͷΨ΢εաఔ 0
  22. 22 Χʔωϧ K n, n′ = ϕ(xn )Tϕ(xn′ ) લϖʔδͷٞ࿦ʹΑΓɺ

    ͷ෼෍͸ڞ෼ࢄߦྻ ͷཁૉ y K Ͱఆ·Δɻ͔͠͠ɺ Λ௚઀ॻ͖Լͯ͠ܭࢉ͢Δͷ͸ߴ࣍ݩͷ৔߹ࠔ೉ͱͳΔɻ 㱺 ͱ ͷ಺ੵͷܗঢ়(2ͭͷಛ௃͕ͲͷΑ͏ʹྨࣅ͢Δ͔)Λ௚઀Ծఆ͢Δ ϕ(x) ϕ(x) ϕ(x′ ) ΧʔωϧτϦοΫ ΧʔωϧτϦοΫʹΑΓɺ൥ࡶͳܭࢉΛආ͚ͯΧʔωϧؔ਺ Λ௚઀ܭࢉͯ͠ڞ෼ࢄߦྻΛಘΔ͜ͱ͕Ͱ͖Δ k(xn , xn′ ) = ϕ(xn )Tϕ(xn′ ) ▪ Χʔωϧؔ਺ͷྫ k(x, x′ ) = θ 1 exp ( − |x − x′ |2 θ2 ) k(x, x′ ) = θ 1 exp θ 2 cos ( |x − x′ | θ3 ) RBF Χʔωϧ पظ Χʔωϧ
  23. 23 Ψ΢εաఔճؼϞσϧ ط஌ͷσʔληοτ ͕༩͑ΒΕ͍ͯΔͱ͖ɺ ೖྗ ʹର͢Δग़ྗ Λ༧ଌ͍ͨ͠ = {(x1 ,

    y 1 ), (x2 , y 2 ), …, (xN , y N )} x* y* ΋Ψ΢ε෼෍ʹै͏͜ͱΛར༻͢Δ y* ( y y*) ∼ ( 0, ( K k* kT * k ** )) k* = (k(x*, x1 ), k(x*, x2 ), …, k(x*, xN ))T k ** = k(x*, x*) ʹ͍ͭͯղ͘ͱɺΨ΢εաఔͷ༧ଌ෼෍͸ y* p(y*|x*, ) = (kT * K−1y, k ** − kT * K−1k* ) ɾɾɾ(※)
  24. 24 ϋΠύʔύϥϝʔλਪఆ Χʔωϧؔ਺ͷϋΠύʔύϥϝʔλਪఆ 㱺 ط஌ͷσʔλͷֶशʹΑΔ࠷దԽ ϋΠύʔύϥϝʔλΛ·ͱΊͯ ͱ͓͘ͱֶशσʔλͷ֬཰෼෍͸ θ = (θ

    1 , …, θ J ) p(y|x, θ) = (y|0, K θ ) = 1 (2π)N/2 1 |Kθ |1/2 exp ( − 1 2 yTK−1 θ y ) ର਺ΛͱΔ L = log p(y|x, θ) ∝ − log|K θ | + yTK−1 θ y + const . ্هͷର਺໬౓ؔ਺Λ࠷େԽ͢ΔΑ͏ʹ࠷దԽ (ޯ഑๏ͳͲ)
  25. 25 4. Ψ΢εաఔճؼΛಋೖͨ͠ Dst ࢦ਺ ༧ଌ ✓ Ψ΢εաఔճؼͷಋೖ ✓ ༧ଌ݁Ռߟ࡯

    ✓ Future Work ✓ ·ͱΊ
  26. 26 Dst ࢦ਺༧ଌ΁ͷΨ΢εաఔͷಋೖ p(y*|x*, ) = (m(x*) + kT *

    K−1(y − m(x)), k ** − kT * K−1k* ) (※)ΛฏۉΛՃຯͨ͠΋ͷʹมܗ ʹ LSTM + FC ૚ ͷग़ྗΛ࢖༻͢Δ m(x), m(x*) k(x, x′ ) = θ 1 exp ( − |x − x′ |2 θ2 ) Χʔωϧ͸ RBF ΧʔωϧΛ࢖༻ (࿦จͰ͸NNΧʔωϧ) Ψ΢εաఔͷܭࢉ͸ GpyTorch[https://gpytorch.ai/]Λ࢖༻
  27. 1. ܇࿅σʔλʹରͯ͠ LSTM Λར༻ͨ͠ϞσϧʹΑΔֶशΛ࣮ࢪ 2. ܇࿅σʔλʹରͯ͠ RBF ΧʔωϧͷϋΠύʔύϥϝʔλͷ࠷దԽΛ࣮ࢪ 3. Ψ΢εաఔճؼΛ༻͍ͯςετσʔλʹରͯ͠༧ଌ஋Λࢉग़

    27 Dst ༧ଌͷશମखॱ
  28. ✓ Ψ΢εաఔճؼΛؚΜͩ༧ଌ݁ Ռྫ - 2σͷ৴པ۠ؒ෇༩ - ৴པ۠ؒʹ؍ଌ஋͕΄΅ೖΔ - ҆ఆྖҬͰ৴པ۠ؒͷ෯͕খ ͘͞ͳΔ

    ✓ վળ఺ - Χʔωϧͷબఆ 28 ༧ଌ݁Ռྫ
  29. ✓ ςετσʔλʹର͢Δ༧ଌͷධՁ - Dst ࢦ਺มಈ͕େ͖͍෦෼ͷ༧ଌਫ਼౓Λॏཁࢹ͢ΔΑ͏ͳࢦඪʁ ✓ ࣌ؒޙҎ֎Ͱͷ༧ଌ - ࠓճͷଠཅ෩ͷ؍ଌ఺͔Β਺࣌ؒͰ஍ٿ΁Өڹ͕ੜ͡Δͱ͍ΘΕΔ㱺 ͨͱ

    ͑͹ɺ ͳͲ͸ҙຯ͕ͳ͞ͳ͘ͳΔͱߟ͑ΒΕΔ ✓ ϦΞϧλΠϜσʔλରԠ - ֶशσʔλͷࣗಈߋ৽ɻ༧ଌ΁ͷ൓ө p = 4 p = 10 29 Dst ࢦ਺༧ଌશମͷFuture Work
  30. ✓ ଠཅ஍ٿ؀ڥܥͱӉ஦ఱؾͷղઆ - ࣓ؾཛྷͱDstࢦ਺ - Dstࢦ਺Λ༧ଌ͢Δҙٛ ✓ LSTM + Ψ΢εաఔճؼΛ༻͍ͯ

    Dst ࢦ਺ΛޡࠩΛؚΊͯ༧ଌ - Ψ΢εաఔճؼͷཧ࿦Λղઆ - ࣮σʔλ΁ͷద༻ํ๏Λ঺հɺ࣮ફ 30 ·ͱΊ αϯϓϧσʔλͰຬ଍ͤͣɺ࣮σʔλͱઓͬͯΈΑ͏ʂ
  31. ✓ Ӊ஦ఱؾ༧ใ[https://swc.nict.go.jp/knowledge/] by ࠃཱݚڀ։ൃ๏ਓ৘ใ௨৴ݚڀػߏ - ը૾Ҿ༻౳ ✓ ࣋ڮ େ஍ɾେӋ ੒੐

    (2019)ʰΨ΢εաఔͱػցֶशʱɹߨஊࣾ ✓ Gruet, Marina A., et al. "Multiple-Hour-Ahead Forecast of the Dst Index Using a Combination of Long Short-Term Memory Neural Network and Gaussian Process." Space Weather 16.11 (2018): 1882-1896. 31 ࢀߟࢿྉ σʔλऔಘݩ ✓ Dst ࢦ਺ - ژ౎େֶେֶӃཧֶݚڀՊ෇ଐ஍࣓ؾੈքࢿྉղੳηϯλʔ [http://wdc.kugi.kyoto-u.ac.jp/wdc/Sec3-j.html] ✓ ଠཅ෩σʔλ - High resolution OMNI (5-min) ͷଠཅ෩σʔλ [https://omniweb.gsfc.nasa.gov/ow_min.html] - Pyspedas [https://github.com/spedas/pyspedas] Λར༻ͯ͠μ΢ϯϩʔυ