Presentation is loading. Please wait.

Presentation is loading. Please wait.

CONSENS Priority Project Status report COSMO year 2008/2009 Involved scientists: Chiara Marsigli, Andrea Montani, Tiziana Paccagnella, Tommaso Diomede.

Similar presentations


Presentation on theme: "CONSENS Priority Project Status report COSMO year 2008/2009 Involved scientists: Chiara Marsigli, Andrea Montani, Tiziana Paccagnella, Tommaso Diomede."— Presentation transcript:

1 CONSENS Priority Project Status report COSMO year 2008/2009 Involved scientists: Chiara Marsigli, Andrea Montani, Tiziana Paccagnella, Tommaso Diomede (ARPA-SIMC) Flora Gofa, Petroula Louka (HNMS) Felix Fundel (MeteoSwiss)

2 Overview  Task 1: Running of the COSMO-SREPS suite  suite maintenance  implementation of the back-up suite  Task 2: Model perturbations  perturbation of physics parameters  perturbation of soil fields  Task 3: Ensemble merging  Multi-clustering  Task 4: Calibration

3 The COSMO-SREPS ensemble  COSMO-SREPS has been developed within the SREPS PP, aiming at the development of a Short-Range Ensemble Prediction System  3 days forecast range, 10 km of horizontal resolution  COSMO-SREPS provides boundary conditions for COSMO-DE-EPS, the 2.8 km ensemble system under development at DWD  Application: test of the use of COSMO-SREPS to estimate a flow-dependent B matrix in a 1D-Var DA of satellite data

4 COSMO-SREPS COSMO at 25 km on IFS IFS – ECMWF global by AEMET Spain COSMO at 25 km on GME GME – DWD global COSMO at 25 km on UM UM – UKMO global COSMO at 25 km on GFS GFS – NCEP global P1: control P2: physics pert p2 P3: physics pert p3 P4: physics pert p4 … COSMO (v 4.7) 00 UTC and12 UTC 10 km 40 levels 16 members 72 h

5 1. Running of the COSMO-SREPS suite ARPA-SIMC  Maintenance of the COSMO-SREPS suite at ECMWF  Adaptation of the data output for COSMO-DE-EPS  Implementation of a 12 UTC run (beside the 00 UTC one)  Implementation of the back-up suite: delayed (9 months)  The work involves also DWD, even if implicitly!  AEMET has provided the int2lm code adapted for the NCEP and UKMO models  An agreement with UKMO has been signed, in order to receive regularly the boundary conditions from the UM

6 Suite availability

7 2.1 Model perturbations: parameters CSPERT test suite ARPA-SIMC - HNMS  In order to study new parameter perturbations, a test suite (CSPERT) was already implemented at ECMWF, by ARPA-SIMC, during the SREPS PP. Results for SON 2007 can be found in the SREPS final report  According to the outcome of the SREPS PP, it was decided to analyse the impact of these perturbations on a dry season as well  New runs of the CSPERT suite were performed in autumn 2008, for the JJA 2008 period  Analysis of the results completed in May 2009

8 The CSPERT suite 16 LM runs at 10 km P1: control (ope) P2: conv. scheme (KF) P3: parameter 1 P4: parameter 2 P5: … IFS – ECMWF global SON 07 + JJA 08

9 JJA 2008 – IT BIAS MAE T2m Td2m

10 JJA 2008 – GR BIAS RMSE T2m Td2m

11 | T BIAS | T RMSE | Td BIAS | Td RMSE | U BIAS | U RMSE KF ==~ == tur_len=150 === >== > (day) tur_len=1000 ~ ==<=== pat_len=10000 < =(day) > < (day)> = (day)~ = rat_sea=1 >>>= >~ = rat_sea=60 <<>= >~ = crsmin=50 >>>>= > (day) crsmin=200 << =<=== c_lnd=1 = < (day)> (day)>=== c_lnd=10 ><<>> (night)= < (day)= rlam_heat=0.1 > < (day)= > (day)>>== rlam_heat=10 (day)= < (day)>>> <~ = ---+++= Summary of the perturbation impact

12 Remarks from the CSPERT suite  The effect of perturbing each physics parameter on improving or worsening the statistical values of the results in comparison to the corresponding control was investigated  Based on these results, the next step was to explore the importance and the effect of selected physical perturbations further  It seems that the particular parameter perturbations do not influence greatly the mean horizontal wind apart from a few exceptions. Possibly looking at the vertical wind component would make the effects more apparent for some parameters

13 Remarks (cont)  Looking separately at each parameter perturbation compared to the control run:  scaling factors related to the laminar layer (rlam_heat, rat_sea), turbulent length scale (tur_len) and evapotranspiration (crsmin), all associated with the development of the turbulent surface layer, are the physical parameters on which the main focus is given

14 2.1 Model perturbations: parameters the new COSMO-SREPS configuration ARPA-SIMC - HNMS  On the basis of the analysis of these results, a new configuration of the COSMO-SREPS suite has been implemented in May 2009  An analysis of its performance over summer 2009 (JJA) has been carried out:  in terms of 2m temperature only over the Alpine area  In term of the continuous parameters (T, U and Td) over Greece  Precipitation has not been considered up to now mainly due to the summer season

15 COSMO-SREPS new configuration (from the 5th of May 2009) convection scheme: 0 Tiedtke 1 Kain-Fritsch maximal turbulent length scale length scale of thermal surface patterns scaling factor of the laminar layer depth ratio of laminar scaling factors for heat over sea minimal stomata resistance

16 IFS GME NCEP UM Tiedtke Kain-Fritsch tur_len < tur_len > rlam_heat < crsmin > pat_len > rlam_heat < crsmin > rat_sea < tur_len >tur_len < rlam_heat > rat_sea > crsmin <

17 Nearest grid point Relationship between error and spread JJA09 Small sample, 30 days only SYNOP over the MAP D-PHASE domain SYNOP over the whole domain t2m

18 Relationship between error and spread JJA09 SYNOP over the whole domain - Nearest grid point 00 UTC 12 UTC t2m

19 Nearest grid point 2m T – deterministic scores global model JJA09 SYNOP over the MAP D-PHASE domain ecmwf gme ncep ukmo MAEBIAS -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0

20 Nearest grid point 2m T – deterministic scores convection scheme JJA09 SYNOP over the MAP D-PHASE domain Tiedtke Kain-Fritsch -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0 MAEBIAS

21 Nearest grid point 2m T – deterministic scores tur_len JJA09 SYNOP over the MAP D-PHASE domain tur_len=150 – ecmwf T tur_len=1000 – ecmwf KF tur_len=1000 – ncep T tur_len=150 – ncep KF -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0 MAEBIAS

22 tur_len: maximal turbulent length scale

23 Nearest grid point 2m T – deterministic scores pat_len JJA09 SYNOP over the MAP D-PHASE domain pat_len=10000 – ecmwf KF pat_len=10000 – gme T MAEBIAS -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0

24 Nearest grid point 2m T – deterministic scores rlam_heat JJA09 SYNOP over the MAP D-PHASE domain rlam_heat=0.1 crsmin=200 – ecmwf T rlam_heat=0.1 – gme KF rlam_heat=10 – ncep T rlam_heat=10 – ncep KF MAEBIAS -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0

25 rlam_heat: scaling factor of the laminar layer depth

26 rat_sea: ratio of laminar scaling factor for heat over sea

27 Nearest grid point 2m T – deterministic scores crsmin JJA09 SYNOP over the MAP D-PHASE domain rlam_heat=1 crsmin=200 – ecmwf T rat_sea=1 crsmin=200 – gme T crsmin=50 – ukmo T crsmin=50 – ukmo KF MAEBIAS -0.5 0.0 0.5 1.0 1.5 2.0 1.8 2.0 2.2 2.4 2.6 2.8 3.0

28 crsmin: minimal stomata resistance

29 Remarks  Some of the perturbations produce common effects on both regions (e.g. rlam_heat, crsmin, tur_len)  However, the impact of some of the physical perturbations (e.g. rat_sea) depends on the geographical characteristics of the region  Large values of rlam_heat produce an increase in the error, implying that, theoretically, a deeper laminar layer suppresses the vertical fluxes  The value of pat_len will be decreased in the new implementation to be more consistent  A paper about the SREPS outcomes is in preparation!

30 Test of new parameter perturbations (new CSPERT suite) memberconvpat_lenrlam_heatrat_seacrsmincloudmu_raingscp 1T5000.1202005.00e+080.54 2KF5000.1202005.00e+080.54 3T500112005.00e+080.54 4KF500112005.00e+080.54 5T5001201505.00e+070.54 6KF5001201505.00e+070.54 7T5001201505.00e+0804 8KF5001201505.00e+0804 9T5001201505.00e+080.53 (no gra) 10KF5001201505.00e+080.53 (no gra) 11T100001201505.00e+070.54 12KF100001201505.00e+070.54 13T5001201505.00e+0704 14KF5001201505.00e+0704 15T5001201505.00e+080.54 16KF5001201505.00e+080.54 15: ctrl T 16: ctrl KF Nov 08 - MAMJ 09

31 2.2 Model perturbations: Developing perturbations for the lower boundary HNMS Aim Implement a technique for perturbing soil moisture conditions and explore its impacts Reasoning The lack of spread is typically worse near the surface rather than higher in the troposphere. Also, soil moisture is of primary importance in determining the partition of energy between surface heat fluxes, thus affecting surface temperature forecasts

32 Soil Perturbation method Based on the method proposed by Sutton and Hamill (2004) Select a period that provides variability in soil moisture e.g. spring Use of data from a land–surface model analysis for the defined period for a few years in order to create some “climatology” (DWD SMA) Implement the EOF (Empirical Orthogonal Function – Principal Component Analysis) to the data in order to generate random perturbations while retaining the spatial structure of the field Define the number of perturbations that will be initially used Test the impact of the perturbation within the COSMO- SREPS suite

33 3. Ensemble merging: development of the COSMO-LEPS clustering (A. Montani, A. Corigliano)  A dynamical downscaling where driving members for COSMO are taken from different global ensembles is under testing  The cluster analysis is applied on a large set of members coming from different global ensembles  Up to now, ECMWF EPS and UKMO MOGREPS have been considered initial conditions by EPS initial conditions by MOGREPS

34 Issues  Consider both ECMWF EPS and UKMO MOGREPS and study the properties of the cluster analysis on multi- ensemble:  How many times do the 2 ensembles mix?  Where do the RMs come from? How to they score depending on their “origin”?  Is there added value with respect to single-model ensemble:  BEFORE dowscaling  AFTER downscaling

35 Forecast and analysis datasets  data from TIGGE-PORTAL (everything in GRIB2)  90 days (MAM09) of ECMWF-EPS and UKMO-MOGREPS run at 00 and 12 UTC  use Z500 at fc+96h as clustering variable;  for verifying analysis (at 00 and 12 UTC), consider Z500:  “consensus analysis” (average of UKMO and ECMWF high-res analyses),  independent analysis (e.g. from NCEP);  generate the following global ensembles:  EPS (50+1): 51 members  MOGREPS (23+1): 24 members  MINI-MIX (EPS24 + MOGREPS24): 48 members  MEGA-MIX (EPS51 +MOGREPS24): 75 members

36 Strategy  perform cluster analysis with 16 clusters and select RMs (like in operations);  generate 16-member global ensembles (EPS_REDU, MOGREPS_REDU, MINI_REDU, MEGA_REDU).  How do “REDUs” ensembles rank with respect to EPS, MOGREPS, MINI-MIX, MEGA-MIX?  Where do the best (and the worst) elements of REDU ensembles come from?  How do they score depending on their “origin”?  BEFORE DOWNSCALING: is there added value with respect to single-model ensemble?

37 Future plans finish by March 2010! Future future plans  Implement dynamical downscaling: nest COSMO model in the selected RMs and generate “hybrid” COSMO-LEPS using boundaries from members of different global ensembles.  For a number of case, compare operational COSMO-LEPS and “hybrid” COSMO-LEPS.

38 Summary results  The availability of the COSMO-SREPS suite has been around 90% during this year, but the system is complete only about 50-60% of the times -> back-up suite!  The analysis of the parameter perturbations introduced in the SREPS PP has been completed in Spring, and new selected perturbations have been introduced in the COSMO-SREPS suite in May  There is a good impact of the new perturbations on the spread of the system  A new set of perturbations, also for the microphysics scheme, is currently under testing  A methodology for soil moisture perturbation has been selected and is being implemented at HNMS  The work on multi-clustering has started, using the GRIB2 fields of the TIGGE-PORTAL

39 4. Calibration ARPA-SIMC - MeteoSwiss  At MeteoSwiss (F. Fundel, Sep 08-Feb 09):  Sensitivity tests  Documentation/paper  At ARPA-SIMC (T. Diomede):  Data collection: observations MeteoSwiss reforecast COSMO-LEPS forecasts  Choice of the methods  Code implementation  Evaluation Preparatory step: visit of Tom Hamill and Felix Fundel at ARPA-SIMC, June 2008

40 Calibration Method x Return Period x Reforecasts ObservationsReforecasts Return Period 30 years COSMO-LEPS reforecasts (1971-2000)Observations (stations, gridded fields) CDF (for one grid point)

41 Verification Results I raw forecasts are overconfident calibrated forecasts nearly perfect reliable strong improvements during winter summer forecast already are reliable, only little improvement possible

42 Sensitivity Study (precip) current setup (18% improvement) best, cheap setup (15% improvement) Cost for 1 member is ~equal to 2 reforecasts 15% improvement (over 16 member CLEPS DMO) using 11-12 members and calibrate with 8-10 years reforecasts Depending on season: - more improvement during winter - less improvement during summer rel. improvement in RPS over 16 Member CLEPS DMO

43  Observations  Emilia-Romagna Region  24-h precipitation (08-08 UTC), 1970-2007  COSMO-LEPS reforecasts (done by MeteoSwiss)  30 years: 1971-2000  1 member, nested on ERA40, COSMO v4.0  1 run every third day (+90h)  COSMO-LEPS QPFs operational  5 years: 2003-2007 [m] Emilia-Romagna Region (22000 km 2 ) 281 COSMO-LEPS grid points 158 raingauges Calibration – data collection

44 Calibration – choice of the methods  choice of methodologies which enable a calibration of the quantitative precipitation forecasts, not only of the probabilities of exceeding a threshold  aim:  improve COSMO-LEPS output (QPF)  hydrological applications  chosen methods up to now:  Cumulative Distribution Function (CDF) based  Linear regression  Analogues, based on the similarity of forecast fields: precipitation geopotential height

45 CDF-based corrections Ref: Zhu and Toth, 2005 AMS Annual Conf., and many others For each model grid point: blue line  CDF of COSMO-LEPS reforecasts red line  CDF of historical observations “raw forecast”  each member of the operational COSMO-LEPS Calibration methodologies

46 Linear Regression Ref: any applied statistics textbook For each model grid point: x-axis: COSMO-LEPS reforecasts y-axis: historical observations Calibration methodologies

47 1 analog date for the whole Emilia-Romagna Region and for each 24-h forecast period For each ensemble member’s forecast and 24-h forecast period (+ 20-44h, 44-68h, 68- 92h, 92-116h) : - the analog search is performed in terms of 24-h rainfall pattern over the Emilia-Romagna Region - the root-mean-square (rms) difference between the current forecast and each reforecast is computed, over all the grid points of the Emilia-Romagna Region - the historical date with the smallest rms difference is chosen as the date of the analog, then the past raingauge recordings are used as the calibrated forecast Calibration methodologies Analogues

48 Calibration – analogues domain used for the analogue search example on the methodology used for the analogue search in terms of geopotential at 700 hPa

49 Method comparison autumn threshold: 5 mm/24 hthreshold: 20 mm/24 h +20-44h

50 Method comparison threshold: 5 mm/24 hthreshold: 20 mm/24 h +68-92h autumn

51 Method comparison threshold: 5 mm/24 hthreshold: 20 mm/24 h +20-44h spring

52 Different sub-areas mountain plain +20-44h 20 mm/24 h

53 raw calibrated (CDF) Spatial variability of the Mean Error Mean Error of the ensemble mean autumn mm/24h underestimation overestimation +20-44h

54 Spatial variability of the Mean Error raw calibrated CDF mm/24h underestimation overestimation Mean Error of the ensemble mean spring +20-44h

55 Spatial variability of the Mean Error raw calibrated CDF Mean Error of the ensemble mean summer +20-44h mm/24h underestimation overestimation

56 raw calibrated CDF Spatial variability of the Mean Error Mean Error of the ensemble mean winter +20-44h mm/24h underestimation overestimation

57 Flow direction mountain plain +20-44h 20 mm/24 h

58 upper mountainous macro-areas overestimationunderestimation Wind S-SW-W Linear Regression autumn [m2/s2]

59 Remarks  The lack of improvement can be ascribed to the lack of a strong relationship between forecast and observed data.  It is necessary to generate correction functions which are weather-type specific. The training sample should be divided into sub-samples which have similar characteristic with respect to the meteorological situation. Hence, a model error which is systematic with respect to the meteorological situation could be identified and reduced by a specific correction function.

60 Next developments  improve the analogue search, by using upper air fields (geopotential and specific humidity) at different levels and daytimes and testing the size of the domain used for the analogue search  apply LR and CDF on a limited sample of analogues  verify results by the coupling with hydrologic models  extend the calibration over other areas, if observed data will be made available (data over Switzerland are available)  reduce the size of the reforecast dataset (in order to use more recent hourly data to calibrate precipitation forecasts accumulated over 12 or 6 h and enable more detailed hydrological applications)

61 Final remarks

62 Problems encountered  The implementation of the back-up suite has just started (with delay, not critical for the moment)  Difficulty in objectively evaluating COSMO-SREPS since the ensemble is often incomplete; problems in verifying precipitation in the summer season  Calibration:  The performance of the calibration methodology is dependent on the precipitation threshold and on the considered area  Difficulty in “catching the bias” of precipitation over Emilia-Romagna, dependent on weather type  Are good data over other areas available?

63 Decisions needed  In order to calibrate the ensemble over the whole domain, very good (dense and covering a log period, i.e. years) observations should be made available by other regions/countries within the Consortium. And in principle also outside!

64 Lessons learned  The development of a methodology always introduces new questions, not foreseen, which need to be answered within the project => increase of the amount of work needed (e.g. assess the effect of parameter perturbations in a robust manner, calibration)  Some re-shuffling of the timing of the tasks has been applied, but without influencing the project development, since the tasks are independent


Download ppt "CONSENS Priority Project Status report COSMO year 2008/2009 Involved scientists: Chiara Marsigli, Andrea Montani, Tiziana Paccagnella, Tommaso Diomede."

Similar presentations


Ads by Google