Ensemble Data Assimilation at DWD System and Selection of Research Projects Andreas Rhodin, Ana Fernandez, Roland Potthast, Christoph Schraff, Hendrik Reich, Harald Anlauf, Anne Walter, Alex Cress, u.v.m DWD, Germany & University of Reading, UK Bonn Sept 2016
64
Embed
Ensemble Data Assimilation at DWD - University of Bonn...Ensemble Data Assimilation at DWD System and Selection of Research Projects Andreas Rhodin, Ana Fernandez, Roland Potthast,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Ensemble Data Assimilation at DWD
System and Selection of Research Projects
Andreas Rhodin,
Ana Fernandez,
Roland Potthast,
Christoph Schraff,
Hendrik Reich,
Harald Anlauf,
Anne Walter,
Alex Cress,
u.v.m
DWD, Germany & University of Reading, UK
Bonn Sept 2016
Global NWP Modelling
0)(
0)(
vv
vpdh
vpdn
tn
vt
vt
gz
cz
wwwv
t
w
nc
z
vw
n
Kvf
t
v
ICON Model 13km + Nest over Europe (6.5km)
3
4. ICON Ensemble Datenassimilation
We are running ICON EDA in our
Routine since Jan 2016
• 40 Members each with 40km global
resolution and 20km NEST over Europe
• 1 deterministic 13km member
• EPS forecasts 40 Members 7 Days + 1
Deterministic
• Output for convective-scale EDA/EPS
• Hybrid System
Grafics by ICON EDA Head
Dr. Andreas Rhodin, FE12
Operational since January 2016 : Rhodin, Fernandez, Cress, Anlauf, etc.
Roland Potthast
ICON EnVar
Roland Potthast
ICON EnVar
Hybrid Methods: EnVAR Scores
ICON EDA
Hybrid Methods: EnVAR Scores
ICON EDA
Particle Filter
• Localized version of Particle Filter Classical Particle Filter PF and Localized Markov Chain Particle Filter LMCPF
(See book of Nakamura and Potthast)
• Hybrid Ensemble Var Particle Filter Particle filter coupled with Variational Method (3D-VAR)
Global NWP with ICON Model 40 Particles 40km global resolution, Deterministic run 13km
You get a prior distribution p(x) by some prior ensemble
Measurements define a data distribution p(y|x)
Bayes theorem defines a posterior distribution by
p(x|y) = c p(x) p(y|x)
The core game is how to get an analysis ensemble from p(x|y).
Particle Filter
PRIOR
DATA
Posterior
Analysis
Ensemble
BAYES Data Assimilation
Following the LETKF philosophy
Replacing the LETKF square root filter by a particle selection which works for non-Gaussian distributions
Localizing the EDA part in Observation space
Localizing the coupled variational part in state space.
Using standard tools for spread control from EnKF, i.e. multiplicative and additive covariance inflation, relaxation towards prior perturbations, …adaptively.
Preventing particle filter collapse by a pseudo-random draw in each analysis step around the particles with non-zero weight.
Particle Filter Details
Roland Potthast 2016
EnKF T on level 85
Roland Potthast 2016
PF T on level 85
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
TEMP T 3h, 5 days
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
Roland Potthast 2016
Summary
• Implemented a localized particle filter for the ICON EDA global assimilation
• Implemented a hybrid EnVar Particle Filter for the deterministic run • Testing the system in a case study • In principle we see that the system is functioning • The behaviour of the forecasts in the case study was useful • The 3h o-f scores of the PF were worse than for the LETKF • The 3h o-f scores of the Hybrid PF were better than for the EnVAR • The forecasts scores were comparable between EnVar-LETKF and
EnVar-PF, the comparison is not yet significant • We need some adaptive spread control in our particle filter, this is
ongoing work. • Further studies and investigation of many details are ongoing.
High-resolution data needed, indirect measurements,
sparse data not resolving all processes
Strong non-linearities in the processes
Convective Scale EDA
Goal is the prediction of convection and subsequent precipitation,
here model grid (left) and upscaled probability (right)
(Courtesy: FE15)
Upscaling/downscaling of statistics is non-trivial! Roland Potthast - September 2016
Convective Scale EDA
Design of a convective scale EDA System
(Image: A. Rhodin and C. Schraff)
Roland Potthast - September 2016
Convective Scale EDA+EPS Design of a convective scale EPS System 1
LBC + IC + Physics
ICON, IFS, GFS, GSM
perturb.
COSMO-DE EPS
Construction of atmospheric Probability Distribution by very different Perturbation Techniques
Roland Potthast - September 2016
Convective Scale EDA+EPS
Design of a convective scale EPS System 2
LBC + IC + Physics
ICON EPS
perturb.
COSMO-DE EPS
Construction of atmospheric Probability Distribution by very different Perturbation Techniques
Roland Potthast - September 2016
Convective Scale EDA
Snow Analysis Deterministic analysis The snow analysis for COSMO-DE deterministic runs every 6 hours using observations from snow depth, precipitation combined with 2m temperature, and weather observations ww to analyse snow depth. Background field is the previous analysis. Ensemble system For the ensemble system no explicit snow depth perturbations are applied, differences result from free running snow variables for each member. The ensemble is adjusted after each deterministic analysis to ensure the ensemble mean matches the deterministic analysis. Collocation Method with radial basis functions = Cressman Method, Successive Correction
NOAA snow depth analysis previous day Roland Potthast - September 2016
Convective Scale EDA
Sea Surface Temperature (SST) Deterministic System SST analysis for COSMO-DE deterministic runs daily at 0:00 UTC using background fields from ICON which are based on NCEP input data. Sea ice is updated using the BSH ice mask.
Ensemble System The SST analysis for the ensemble system is based on the analysis from COSMO-DE deterministic. Perturbations are generated by a stochastic method with random perturbations and a localization based on Gaspari Cohn functions.
Roland Potthast - September 2016
Convective Scale EDA Sea Surface Temperature (SST) and Soil Moisture (w_SO) Perturbations Random algorithm with two scales Surface temperature differences from soil moisture perturbations and model
dynamics
Difference Member 3 – Mean (left) or Member 1 – Mean (right) of T_SO Roland Potthast - September 2016
Convective Scale EDA Sea Surface Temperature (SST) and Soil Moisture (w_SO) Perturbations Random algorithm with two scales Surface temperature differences from soil moisture perturbations and model
dynamics
Difference Member 3 – Mean (left) or Member 1 – Mean (right) of W_SO Roland Potthast - September 2016
Hourly Analysis of Atmospheric Fields No Soil-Moisture Analysis, but hourly soil-
moisture perturbations (with spread control) and relaxation of soil moisture towards the deterministic run
Snow Analysis every 6 hours at 0, 6, 12, 18 UTC
SST once per day at 0 UTC
EDA Component Schedule
Roland Potthast - September 2016
Distributions EPS Members
Histogram T50 Full temperature Distribution Of COSMO Model, 1 time slice
Histogram ΔT50 With subtraction of mean for each point
Roland Potthast - September 2016
• Talagrand Rank Histogram
• Checks the distribution of observations compared with the distribution of the ensemble
Distributions EPS Members
T2m is underdispersive
ensemble
obs
Verification Scores Survey
Upper Air Verification
Surface Verification
Precipitation Verification
Satellite Data Verification
Scores Metrics Bias Field Properties Spectral Distributions
Roland Potthast - September 2016
High Impact Weather Verification
Verification Scores Survey 1
bias RMSE
Nudging + LHN vs. LETKF + LHN
T [K] RH wind [m/s]
bias RMSE RMSE RMSE
Verification of 6-h forecasts against radiosondes , 28 days (18.05. – 15.06. 2014)