Ensemble forecastingdata assimilation and model error estimation algorithm
Ensemble forecasting/data assimilation and model error estimation algorithm Prepared by Dusanka Zupanski and Milija Zupanski CIRA/CSU Denning group meeting December 29, 2004 References Zupanski, M. , 2005: The Maximum Likelihood Ensemble Filter. Theoretical aspects. Accepted in Mon. Wea. Rev. [Available at ftp: //ftp. cira. colostate. edu/milija/papers/MLEF_MWR. pdf] Zupanski, D. , and M. Zupanski, 2005: Model error estimation employing ensemble data assimilation approach. Submitted to Mon. Wea. Rev. [Available at ftp: //ftp. cira. colostate. edu/Zupanski/manuscripts/MLEF_model_err. revised 2. pdf] Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
Maximum Likelihood Ensemble Filter (MLEF) Developed using ideas from: ØVariational data assimilation (3 DVAR, 4 DVAR) ØIterated Kalman Filters ØEnsemble Transform Kalman Filter (ETKF, Bishop et al. 2001) What the MLEF can do? ØCalculate optimal estimates of: - model state variables (e. g. , carbon fluxes, sources, sinks) - empirical parameters (e. g. , light response, allocation, drought stress) - model error (bias) - boundary conditions error (lateral, top, bottom boundaries) ØCalculate uncertainty of all estimates ØCalculate information content of observations (observability in ensemble subspace) Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
What the MLEF can do (continued)? ØCalculate sensitivity, defined in calculus of variations - find the most likely sources/sinks of carbon ØDefine targeted observations strategies, based on the forecast uncertainty - the system knows where, when and which observations are needed in order to reduce forecast uncertainty ØAcquire new knowledge about atmospheric and carbon processes - the system learns from the past about the state variables, model errors, empirical parameters, etc. - the system is adaptive (updates error covariance matrices in each data assimilation cycle) Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
DATA ASSIMILATION (ESTIMATION THEORY) Discrete stochastic-dynamic model Gurney et al. (2003, Tellus) (? ) x k-1 – model state w k-1 – model error (stochastic forcing) M – non-linear dynamic (NWP) model G – operator reflecting the state dependence of model error Discrete stochastic observation model Gurney et al. (2003, Tellus) e k – measurement + representativeness error H – non-linear observation operator (M D ) Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF APPROACH Minimize cost function J Change of variable - model state vector of dim Nstate >>Nens - control vector in ensemble space of dim Nens Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF APPROACH (continued) Analysis error covariance In Gurney et al. (2003) Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF APPROACH (continued) Forecast error covariance In standard Kalman filter Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF application to calculate information content RAMS model example Observation categories within the same cycle Multiple cycles Initial cycles carry more information. Model has a capability to learn from the observations in later cycles. Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
Ensemble Data Assimilation results with CSU global shallow-water model Observation categories within the same cycle Multiple cycles Impact (contribution) from each observation type can be quantified ! Milija Zupanski, CIRA/CSU Zupanski. M@CIRA. colostate. edu
MLEF Algorithm prep_ensda. sh WARM start: Copy files from previously completed cycle COLD start: Run randomly-perturbed ensemble forecasts to initialize fcst err cov cycle_ensda. sh icycle < N_cycles_max fcsterr_cov. sh - Prepare first-guess (background) vector - Prepare forecast error covariance (from ensembles) prep_obs. sh Given ‘OBSTYPE’ and ‘delobs’, select and copy available obs files assimilation. sh Iterative minimization of cost function, save current cycle output Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF Algorithm assimilation. sh script assimilation. sh iter < ioutmax forward. sh: Transformation from model space to observation space Hessian preconditioning (only for iter=1) Gradient calculation (ensembles) Cost function calculation (diagnostic) Minimization (ensemble subspace) Step-length (line-search) Control variable update (transformation from ensemble subspace to model (physical) space) - Analysis error covariance calculation - Save current cycle output files - Post-processing (chi-square, RMS, etc. ) Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
MLEF Algorithm Options Models: Kd. VB, GEOS, SWM, RAMS, … Observations: Synthetic or include various real observations Obs. operators: Include various forward operators Solution Type: Mode (max likelihood-MLEF) or Mean (ensemble mean ETKF) Estimator: Filter or Smoother Control variable: Initial conditions, Model bias, Model parameters Covariances: Localized, or Non-localized forecast error covariance Minimization: Minimization algorithm (C-G, L-BFGS) MPI: Parallel MPI run or a Single processor run Verifications: Innovation statistics (chi-square test, K-S test), RMS-errors Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
Tasks Ø Decide which MODELs will be used (LPDM, Si. B-CASA-RAMS, PCTM) Ø In initial experiments use MODEL as in Gurney et al. (2003) Ø Install LAPACK on your computer, unless it already exists. Compile with 32 bit object code option (LAPACK can be found at http: //www. cs. colorado. edu/~lapack) Ø Install MLEF algorithm on your computer Ø Include MODEL into MLEF - develop an interface between MODEL and MLEF (a template from SWM, RAMS is available) - prepare all input files for the model Ø Prepare observations - synthetic observations of atmospheric and carbon variables - separate observations into groups (if needed) - define data assimilation interval Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
Tasks (continued) ØPerform initial data assimilation experiments with synthetic observations - perfect model assumption experiments - parameter estimation experiments - model error estimation experiments Ø Develop observation operators for: - atmospheric observations (u, v, T, p, q, etc. ) - carbon observations ØData assimilation and ensemble forecasting experiments with real observations - LPDM - Si. B-CASA-RAMS - PCTM Major development assignments Ø Model related tasks (LPDM, Si. B-CASA-RAMS, PCTM) ØObservation operators related tasks Dusanka Zupanski, CIRA/CSU Zupanski@CIRA. colostate. edu
- Slides: 14