Parameter estimation using EnKF

World Weather Open Science Conference.
Montreal, Canada, 16 to 21 August 2014
How do model errors and localization
approaches affects model parameter
estimation
Juan Ruiz, Takemasa Miyoshi and Masaru Kunii
[email protected]
Centro de Investigaciones del Mar y la AtmósferaCONICET
University of Buenos Aires
Advanced Institute for Computational Science RIKEN
Several works showed that surface exchange parameters
have a large impact upon model performance
These parameters might be estimated using data
assimilation based parameter estimation (Ito et al. 2010,
Kang et al. 2012, Green and Zhang 2014).
In this work we evaluate a simple approach for data
assimilation based parameter estimation using the
LETKF-WRF system (Miyoshi and Kunii 2012).
Experiments goes from ideal to real observations tests
Fq  F

WRF
q
Simple parameter estimation approach:
a multiplicative correction factor is
introduced and is estimated using the
LETKF-WRF system.
Model sensitivity to surface fluxes:
TC Sinlaku (2008)
Less sensitive (heat exchange)
More sensitive (latent heat exchange)
Given the stronger impact of latent heat fluxes we test the
methodology focusing on these fluxes.
Ruiz , Miyoshi and Kunii (2014, in preparation)
OSSE experiments:
Realistic observation distribution quasi perfect
model and boundary conditions.
Estimated parameter is identifiable.
Observation network seems to be adequate for the estimation of the
parameter.
-> OSSE experiments are successful
OSSE experiments:
Realistic observation distribution, prefect BC but
imperfect model
Estimated parameter is seems to converge to a different value
Error reduction is not as large as in the perfect model scenario but
improvements can be found in all variables.
Real world experiments:
Estimated model parameters as a function of time
Estimated parameters are below one indicating that surface
moisture flux is reduced in the parameter estimation experiment.
Horizontal distribution is quite homogeneous particularly over the
tropical ocean where the model sensitivity to the parameter is
stronger.
Real world experiments:
Impact upon the analysis (compared with GDAS)
BIAS
RMSE
relative improvement
Low level biases are removed in almost all variables.
Upper level biases are increased.
RMSE improved for wind. Moisture and temperature
shows mixed behaviour
Real world experiments:
Impact upon the forecast (compared with GDAS)
40 member ensemble forecast
IMPROVEMENT
DEGRADATION
PS
Wind improved at almost all levels
Temperature and moisture improved at low levels but degraded at
middle and upper levels.
Real world experiments:
Precipitation forecast (compared with CMORPH)
72 hr
BIAS
ETS
24 hr
48 hr
Precipitation forecast improved ETS. Precipitation frequency
decreases .
Real world experiments:
Impact upon TC forecast
Forecast improved
Forecast degraded
Some cases shows a consistent improvement while
others shows a consistent degradation...
Real world experiments:
Impact upon TC forecast
The mean track error is slightly better for the parameter estimation
experiment.
The sample is too small to have robust results.
Sensitivity to localization strategy:
Large biases near the surface might significantly affect the
estimated parameter values
Sensitivity to the parameters not necessarily confined to low levels
2D estimation
With vertical
localization
Without vertical
localization
0D estimation
Without vertical
localization
Three experiments have been conducted to explore the sensitivity
of the estimated parameters to the localization strategy.
Sensitivity to localization strategy:
Impact upon the estimated parameters
Estimated parameters
as a function of time.
All strategies estimate parameter values that are below the default.
0D estimation produces noisier results.
Experiment with vertical localization produce lower parameter
values.
Sensitivity to localization strategy:
Impact upon the estimated parameters
Vertical profile of
RMSE improvement
0D strategy seems to provide the best results for wind and
temperature (although larger degradation is introduce in the
moisture field)
Similar results are obtained for the forecast
Conclusions:
Parameters are successfully estimated using the LETKF-WRF
system. In all the experiments parameters indicate that moisture
surface fluxes are too strong and are possible responsible for the
moist biases at low levels.
Parameter estimation impact upon the forecast is positive in some
variables including precipitation.
Impact upon the TC forecast is still unclear although results
suggest that estimated parameters can potentially improve TC
forecasting.
Localization has an impact upon the estimated parameters. Best
results has been obtained with 0D parameters (maybe because of
small spatial variability of the estimated parameter).
Thank you!!