Key question. ì Model uncertainty a priori (insert uncertainty where it occurs) or a posteriori ( holis6c approaches: SKEBS, SPPT)

Transcription

1

2

3 Key question Model uncertainty a priori (insert uncertainty where it occurs) or a posteriori ( holis6c approaches: SKEBS, SPPT)

4 Outline Weather applica6on: Improve reliability and reduce ensemble error in analysis and ensemble forecasts Climate applica6on: Reducing systema6c error in mean and variability, improvement of physical processes Seasonal applica6on: Somewhere in between

5 Outline Weather applica6on: Improve reliability and reduce ensemble error in analysis and ensemble forecasts - increase spread Climate applica6on: Reducing systema6c error in mean and variability, improvement of physical processes - oien entails a reduc6on of reduc6on of variability, e.g. along a wave train Seasonal applica6on: Somewhere in between

6 A priori vs a posteriori Model uncertainty added a posteriori: Process uncertainty added a priori during model development: Model Stochas6city Forecast uncertainty

7 A priori vs a posteriori Model uncertainty added a posteriori: Process uncertainty added a priori during model development: Model Stochas6city Forecast uncertainty

8

9 Stochastic parameter perturbations (SPP) Grell convec6on scheme 1 Timeseries; <= PDF Autocorrelation function (in h); Decorrelation time = =6h MYNN PBL Timeseries; <= PDF -1 1 Autocorrelation function (in h); Decorrelation time = =6h With Isidora Jankov, Jeff Beck, George Grell, Joe B, Tanya Smirnova

10 s

11 Experiment setup v v v v v v Limited area model: Con6guous United States (CONUS)/NARRE Weather Research and Forecast Model WRFV km horizontal resolu6on 1-member ensemble, integrated for 4h 1 dates in June of 15, Z and 1Z Boundary and ini6al condi6ons are taken from GEFS

12 Multi-physics ensemble

13 RMSE Spread CRPSS T85 U1 CNTL SPP SPPT SKEBS SPPT+SKEBS+SPP Jankov et al., submiaed

14 Brierscore skill score near.7 the surface.4.6. Resolution CNTL PARAM SKEBS PHYS1 PHYS1_SKEBS PHYS3_SKEBS_PARAM Reliability Resolution Spread;Error Brier Score Brier Skill Score eliability.8 e)..6 i) a) g) Zonal Wind U at 1m c) Forecast Lead time e) Brier skill measures probabilis6c skill in regard to a reference (here CNTL). Verified event: μ<x<μ+σ i) Forecast Lead time..4 h) f) j) b) h) CNTL PARAM SKEBS PHYS1 PHYS1_SKEBS PHYS3_SKEBS_PAR Temperature at m d) Forecast Lead time.1 j) f) Forecast Lead time.1. CNTL PARAM SKEBS PHYS1 PHYS1_SKEBS PHYS3_SKEBS_PAR Berner et al., et al 15

15 C Relative skill improvement Debiased Change of Model Raw Version 1 Brier Skill Score Reliability Resolution Calibrated & Debiased Average over all forecast lead 6mes Variables are U7, T7, U1, T Reference is CNTL Brier Skill Score CNTL PARAM SKEBS PHYS1 PHYS1_SKEBS PHYS3_SKEBS_PARAM U7 T7 U1 T Berner et al., et al 15 U7 T7 U1 T U7 T7 U1 T

16 C Brier Skill Score Reliability Debiased Calibrated & Debiased Change of Model Raw Version Average over all forecast lead 6mes Variables are U7, T7, U1, T Reference is CNTL Brier Skill Score NB: Role of bias Brier Skill Score Reliability Resolution U7 T7 U1 U7 T T7 U7 U1 T7 T U1 U7TT7 U7 U1 T7 T U1 U7TT7 U1 T & Raw Calibrated Debiased CNTL PARAM SKEBS PHYS1 PHYS1_SKEBS PHYS3_SKEBS_PARAM

17 NB

18 WRF-DART:Verification of surface analysis against independent observations V-1m Tm CNTL SKEBS PHYS Ø Including a model-error representa6on reduces the RMS error of the surface analysis (also prior) in 1m wind and Temperature at m Ha et al. 15

19 Difference in SST Mean With Hannah Christensen and Jus6n Small See Christensen et al. 16, submiaed

20 Difference in SST standard deviation

21 Nino3.4 Power spectra

22 Difference in U85 standard deviation

23 Double-well potential with weak additive white noise Double-well potential with strong additive white noise a)) c)) Stochas6c parameteriza6ons can change the mean and variance of a PDF Impacts variability Impacts mean bias b)) d))

24 Northern Annular Mode (MAM) 1 st EOF of sea level pressure over Northern Hemispheric Extratropics CAM4 AMIP simula6ons (prescribed SSTs), 194 NCEP CNTL 1% 5% Stochas6c parameteriza6on improves paaern and reduces explained variance SKEBS SPPT 35% 37% Degenerate response: SKEBS and SPPT have same effect

25 Northern Annular Mode (MAM) 1 st EOF of sea level pressure over Northern Hemispheric Extratropics CAM4 AMIP simula6ons (prescribed SSTs), 194 NCEP CNTL 1% 5% Stochas6c parameteriza6on improves paaern and reduces explained variance SKEBS SPPT 35% 37% Degenerate response: SKEBS and SPPT have same effect

26 Conclusions Parameter perturba6ons (a priori) do not generate enough uncertainty in ensemble systems. Currently, a posteriori methods (SKEBS, SPPT, residual methods) are necessary for reliable ensemble system The same holds for other a priori methods (Plant- Greg scheme, PBL varia6on scheme) In NWP model-error schemes are mainly used to generate spread Role of bias needs to be revisited In climate models, stochas6c methods can reduce coupling and reduce variability

27 Future work Perturb dynamical tendencies land surface models divergence in skebs stochas6c perturba6on in GF convec6on scheme as func6on of large-scale forcing Will impact of stochas6c parameteriza6ons change at convec6on-resolving/allowing scales, where largest uncertainty moves to micro-physics? Is climate sensi6vity influenced by subgrid-scale variability (Seiffert and von Storch 9) SKEBS will be implemented into MPAS and used to study predictability in the presence of kink in spectrum

28

29 Berner, J, K. Fossell, S.-Y. Ha, J. P. Hacker, C. Snyder 15: Increasing the skill of probabilis6c forecasts: Understanding performance improvements from model-error representa6ons, Mon. Wea. Rev., 143, Berner, J., S.-Y. Ha, J. P. Hacker, A. Fournier, C. Snyder, 11: Model uncertainty in a mesoscale ensemble predic6on system: Stochas6c versus mul6-physics representa6ons, Mon. Wea. Rev, 139, Romine, G. S., C. S. Schwartz, J. Berner, K. R. Smith, C. Snyder, J. L. Anderson, and M. L. Weisman, 14: Represen6ng forecast error in a convec6on-permipng ensemble system, Mon. Wea. Rev, 14, 1, Ha, S.-Y., J. Berner, C. Snyder, 15: Model-error representa6on in mesoscale WRF-DART cycling, under review at Mon. Wea. Rev.

30 LOWRES Mean systematic error of 5 hpa Z5 Difference eto4-er4 ( ) 6b) geopotential height fields 6 HIGHRES a) c) -6 SKEBS Z5 Difference eut3-er4 ( ) b) Z5 Diffe d) Z5 Diffe d) Z5 Difference f8o8-er4 ( ) Reduc6on of z5 bias in all simula6ons with model-refinement Berner et al., 1

31 Impact of SKEBS on precipitation bias Coupled control run shows significant bias due to split inter-tropical convergence zone SKEBS reduced bias in precipita6on

32