analog forecasting errors from a dynamical systems point ... · imt-a & riken feb. 10, 2020 paul...

21
Analog forecasting errors from a dynamical systems point of view Paul Platzer (feb. 10, 1993) IMT-A & RIKEN workshop : Statistical Modeling and Machine Learning in Meteorology and Oceanography, feb. 10, 2020

Upload: others

Post on 19-Oct-2020

8 views

Category:

Documents


0 download

TRANSCRIPT

  • Analog forecasting errors from a dynamical systems point of view

    Paul Platzer (feb. 10, 1993)IMT-A & RIKEN workshop :

    Statistical Modeling and Machine Learningin Meteorology and Oceanography, feb. 10, 2020

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer2/20

    Analog forecasting errors from a dynamical systems point of view

    Paul Platzer (feb. 10, 1993)IMT-A & RIKEN workshop :

    Statistical Modeling and Machine Learningin Meteorology and Oceanography, feb. 10, 2020

    Marc Schoenauer,Naonori Ueda,Said Ouala,Maha Mdini…→ bridge the gap betweendata-driven andprocess-driven→ « 5th science »

    Pierre Tandeo,Yicun ZhenAnDA

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer3/20

    Two analog maps correspond to close points in phase space(Courtesy of D. Farandaand P. Yiou)

    What is an analog ?

    → Introduced by Lorenzin 1969 for atmospheric predictability

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer4/20

    ● ak0 = analogs of x0 : « close » to x0

    ● akt successors ( time t ) → estimate xt

    Analog forecasting : the idea

    (Lguensat et al. 2017)

    0 t

    a10

    aK-10aK0

    a1t

    aK-1taKt

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer5/20

    Analog forecasting : cheap ML

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer6/20

    Analog forecasting : in practice● Large database called « catalog » (reanalysis, model output) ● Select K nearest neighbors of x0 (fast thanks to ML libraries)● Assign weights to the analogs (discard bad analogs)● Apply a given analog forecasting operator

    • • •

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer7/20

    Analog forecasting operators

    0

    0

    0

    0 0

    0 t

    t

    t

    t

    t

    t

    x0a10

    a20

    x0a10

    a20

    x0 a10

    a20

    x0

    a20

    a10

    x0

    a20

    a10

    x0

    a20

    a10

    µLCa1t

    a2t

    µLIa1t

    a2tµLLa1t

    a2t

    µLCa2t

    a1t

    µLI

    a2t

    a1t

    µLL

    a2t

    a1t

    (figure from Lguensat et al. 2017)

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer8/20

    Dynamical systems● Hypotheses :

    ● Distance between successor and future state :

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer9/20

    Dynamical systems● Here → looking at local error

    → similar to Nicolis et al. (2009)● Zhao and Giannakis (2016)

    → looking at global dynamical properties

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer10/20

    Dynamical systems : mean error● Hypotheses :

    ● Mean error of analog forecasting operators :

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer11/20

    Dynamical systems : mean error● Hypotheses :

    ● Mean error of analog forecasting :

    → experiments on L96 with δ=0

    (figure from Lguensat et al. 2017)

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer12/20

    Dynamical systems : mean error● Hypotheses :

    ● Mean error of analog forecasting :→experiments on L63 with δ=0

    – LC empirical error-- LC error estimation– LI empirical error-- LI error estimation

    (figure from Platzer et al. 2019)

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer13/20

    Dynamical systems : mean error● Hypotheses :

    ● Mean error of analog forecasting :→experiments on L63 with δ=0

    – LC -- LI-- LI+mean correction

    -·-· LI+local correction(figure from Platzer et al. 2019)

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer14/20

    Relation between operators● Locally-linear :

    residuals

    ● Locally-constant, locally-incremental :

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer15/20

    The locally-linear and the Jacobian

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer16/20

    The locally-linear and the Jacobian

    Mean error of LL forecast

    Quality of Jacobian estimation using LL

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer17/20

    What about the covariance ?

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer18/20

    What about the variance ?

    LC (Lguensat estimation)LC (empirical from several catalogs)LI (Lguensait estimation)LI (empirical from several catalogs)

    → Experiments on L63

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer19/20

    What about the variance ?(logscale)

    occu

    ren c

    es

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer20/20

    Conclusion● Analog forecasting = empirical, simple ML● Using dynamical systems helps

    → interpretation→ error estimation→ data-based + model-based hybrid method ?→ finer tuning ?

    ● Combine analogs and NN ? (analogs=first guess)

  • IMT-A & RIKENFeb. 10, 2020

    Paul Platzer21/20

    Thank you !Bibliography :• Lguensat, R., Tandeo, P., Ailliot, P., Pulido, M., & Fablet, R. (2017). The analog data assimilation. Monthly Weather Review, 145(10), 4093-4107.• Lorenz, E. N. (1969). Atmospheric predictability as revealed by naturally occurring analogues. Journal of the Atmospheric sciences, 26(4), 636-646.• Nicolis, C., Perdigao, R. A., & Vannitsem, S. (2009). Dynamics of prediction errors under the combined effect of initial condition and model errors. Journal of the atmospheric sciences, 66(3), 766-778.• Platzer, P., Yiou, P., Tandeo, P., Naveau, P., & Filipot, J. F. (2019, October). Predicting Analog Forecasting Errors using Dynamical Systems.• Tippett, M. K., & DelSole, T. (2013). Constructed analogs and linear regression. Monthly Weather Review, 141(7), 2519-2525.• Zhao, Z., & Giannakis, D. (2016). Analog forecasting with dynamics-adapted kernels. Nonlinearity, 29(9), 2888.

    Diapo 1Diapo 2Diapo 3Diapo 4Diapo 5Diapo 6Diapo 7Diapo 8Diapo 9Diapo 10Diapo 11Diapo 12Diapo 13Diapo 14Diapo 15Diapo 16Diapo 17Diapo 18Diapo 19Diapo 20Diapo 21