Skill of Nowcasting of Precipitation by NWP and by Lagrangian Persistence. (where we chronicle the bridging of the gap )

Transcription

1 Skill of Nowcasting of Precipitation by NWP and by Lagrangian Persistence (where we chronicle the bridging of the gap )

2 Skill of Nowcasting of Precipitation by NWP and by Lagrangian Persistence (where we chronicle the bridging of the gap )

3 There were various attempts at improving precipitation nowcasting through addition of NWP: Skill-weighted average of Lagrangian Persistence (LP) and NWP Correction of positional errors (and more) of NWP Selectively adding NWP-predicted growth and decay to LP Correction of phase errors of NWP (As a standalone or in combination with LP, with or without data assimilation; deterministic or ensemble NWP)

4 Basic fact of life: short scales are ephemeral 2 16 Lifetime [h] % Band-pass Lifetime 1% Linear Scale [km]

5 Basic fact of life: Precipitation patterns have characteristics of pink noise

6 Basic fact of life: Precipitation patterns have characteristics of pink noise

7 Model vs. Nowcast & Merging (for 6 days of rain) R>.1mm/h POD O-MAPLE+WRF+GEM O-MAPLE WRF GEM GEM+WRF MAPLE CSI R>1mm/h CSI forecast hour forecast hour weight 1 1 CSI(i,t) 1 a better balance for t<4h is obtained with weight [CSI(i,t)]

8 Nowcasting Skill of Model and of Lagrangian Persistence The nowcast is improved when NWP nowcast and Lagrangian persistence nowcast are merged by a skill-weighted average. However, there is no advantage in doing this adaptively: climatological skill is as good as the skill determined in a particular situation just prior to the nowcast. Question: Why? Possible answers: Either model skill is not sufficiently persistent in time (ex: effect of diurnal cycle) or the skill of model and of LP are correlated

9 Scatterplots of CSI June to August 25 CORRELATION Lead time [h]

10 Scatterplots of CSI Jan. to March 25 CORRELATION Lead time [h]

11 Scatterplots of CSI Jan. to March 25 CORRELATION Lead time [h] Should we be asking why so much scatter?

12 We acquired outputs of ensemble runs (OU, Ming Xue) to further experiment with NWP contributions to nowcasting. The ensemble is generated by varying initial conditions and model physics. Radar data are assimilated in all members except c ; cn is identical to c except that radar data were assimilated Ensemble mean is re-calibrated by probability matching, PM (making the pdf of intensity equal to the average pdf of members)

13 We acquired outputs of ensemble runs (OU, Ming Xue) to further experiment with NWP contributions to nowcasting. The ensemble is generated by varying initial conditions and model physics. Radar data are assimilated in all members except c ; cn is identical to c except that radar data were assimilated Ensemble mean is re-calibrated by probability matching, PM (making the pdf of intensity equal to the average pdf of members) Note: the POD of the ensemble mean (before PM) is smaller than one, indicating that the ensemble does not cover all observed precipitation)

14 NWP Ensembles (poor and best predictability cases)

15 Diurnal cycle in pdf of rain

16 Models fail to correctly reproduce the diurnal cycle 5 Summer precipitation over this domain: Mean (no assim) CN (assim) GEM WRF 18 UTC time [h] Radar Mean C CN.7-85 GEM WRF UTC time [h] longitude [deg] longitude [deg] longitude [deg] longitude [deg] longitude [deg] longitude [deg] Note the more consistent diurnal cycle in observations Latitude Radar C Longitude Coverage Intensity -8-7

17 Models fail to correctly reproduce the diurnal cycle 36 coverage of the 24 h cycle 36 intensity of the 24h cycle phase [deg] GEM WRF ens. mean OU model cn OU model c NSSL mosaic peak time UTC [h] phase [deg] GEM WRF ens. mean OU model cn OU model c NSSL mosaic peak time UTC [h] longitude [deg] longitude [deg]

18 Models fail to correctly reproduce the diurnal cycle 36 coverage of the 24 h cycle 36 intensity of the 24h cycle phase [deg] GEM WRF ens. mean OU model cn OU model c NSSL mosaic peak time UTC [h] phase [deg] GEM WRF ens. mean OU model cn OU model c NSSL mosaic peak time UTC [h] longitude [deg] -36 {Region of transport of diurnal cycle longitude [deg] {

19 The larger forecast errors of diurnal cycle happens where LP is longer!! 5 45 Latitude N Lifetime Longitude E miss h

20 The larger forecast errors of diurnal cycle happens where LP is longer!! 5 45 Latitude N Lifetime Longitude E miss h Is this one of the reasons why the predictability by LP and NWP is not better correlated?

21 Summary: Model-LP comparison of precipitation nowcasting 1..8 GEM15 C CN N2 PM mean MAPLE a) 1. b).8 CSI.6.4 correlation d) EPS mean is re-calibrated by PM Note the diurnal cycle in the RMS error RMSE [mm] UTC [h]

22 NWP Ensembles (the best case) OU & 4 km resolution Scores at 15 dbz threshold CSI Radar coverage MEAN CN C Fractional coverage Forecast time [h]. No single member is better Effect of data assimilation is short-lived

23 Position distance between model and radar WRF Model Corrected model Radar CSI MODEL, POSITION CORRECTED by VET MODEL Time [h] CRMSE Time [h] CORRELATION Time [h]

24 Nowcasting by correcting the model-radar distance WRF Model Corrected model Radar CSI MODEL MODEL, POSITION CORRECTED by VET MAPLE 15 dbz threshold Lead time [h] CRMSE Lead time [h] CORRELATION Lead time [h]

25 Nowcasting by correcting the model-radar distance WRF Model Corrected model Radar CSI MODEL MODEL, POSITION CORRECTED (VET) MAPLE 15 dbz Lead time [h] CRMSE Lead time [h] CORRELATION Lead time [h] 5 6

26 Blending NWP with Lagrangian persistence of a 3 radars network (Catalunya).8 LE radar NWP+Assim Blending.7 LAPS with 3D VAR; position corrected; blended by averaging; re-calibrated by PMM..6 CSI Lead-time [h] 5 6

27 Morphing model into radar by phase correction (one wavelength at a time)

28 Morphing model into radar by phase correction (one wavelength at a time)

29 Phase distance between model and radar RMS error (dbz) Phase & power corrected WRF (km) hour Phase contribution (%) (km) Power contribution (%) (km)

30 Phase distance between model and radar D(kx, ky) /( dbz threshold (km)

31 Phase distance between model and radar D(kx, ky) /( dbz threshold Dashed: Random within threshold (km)

32 Defining growth and decay in radar (a) UTC (b) 1 UTC ï11 25 ï11 ï15 ï1 ï15 ï1 ï95 ï9 ï (c) ï95 ï9 ï85 Growth and decay Maximize cross-correlation between (a) and (b) by a solid translation and rotation; 4 The difference defines growth and decay (c) ï11 ï15 ï1 ï5 ï95 ï9 ï85

33 Lifetime of growth and decay 4 8/21 8/13 8/1 7/3 7/24 7/22 7/18 7/13 7/9 6/6 5/24 5/6 4/18 35 Lifetime (h) Precipitation lifetime Scale (cutoff wavelength) [km] 2 Growth & Decay 8/21 8/13 8/1 7/3 7/24 7/22 7/18 7/13 7/9 6/6 5/24 5/6 4/18 3 Lifetime (h) Growth & decay lifetime Scale (cutoff wavelength) [km] 2 25

34 Effect of model error due to resolution

35 Effect of model error due to resolution Rel RMS Diff = i i ( X i Y ) 2 i ( X i + Y ) i Relative RMS Diff WRF3Km } WRF1Km vs. radar WRF333m WRF333m-1Km dashed WRF3Km-1Km dotted 1 1 Cutoff Scale [km]

36 Effect of model error & data assimilation (average of 24 cases) c during spinup; clear effect of assimilation on cn rapid loss of assimilation effect at the small scales rapid loss of assimilation effect at the all scales 1% difference between c and cn at smal scales Relative RMS Diff C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar t=1h t=4h Cutoff Scale [km] C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar t=2h t=5h Cutoff Scale [km] C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar C-CN CN-Radar C-Radar t=3h t=6h t=13h t=14h t=15h t=16h t=17h t=18h Cutoff Scale [km]

37 Ensembles (EnKF) to the rescue?

38 Models have qualitative errors

39 Effect of model erros on assimilation Simulation using data assimilation (model as strong constraint) into a simple model of freely falling rain-shaft with a 2-parameter DSD representation. Note that 3 parameters are needed to correctly describe the DSDs of falling drops. Observations Note a second shaft due to model error

40 Effect of model erros on assimilation Simulation using data assimilation (model as strong constraint) into a simple model of freely falling rain-shaft with a 2-parameter DSD representation. Note that 3 parameters are needed to correctly describe the DSDs of falling drops. Observations Note a second shaft due to model error

41 Conclusions The lifetime of scales below 1 km is SHORT

42 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar)

43 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE)

44 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE) Present data assimilation does seem to lead to nowcasts better than MAPLE

45 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE) Present data assimilation does seem to lead to nowcasts better than MAPLE Uncertainties in LP nowcast (not discussed here) are handled by pure statistical ensembling; some physics is in order

46 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE) Present data assimilation does seem to lead to nowcasts better than MAPLE Uncertainties in LP nowcast (not discussed here) are handled by pure statistical ensembling; some physics is in order Letʼs shutdown the supercomputers for a decade so there is time to study model errors

47 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE) Present data assimilation does seem to lead to nowcasts better than MAPLE Uncertainties in LP nowcast (not discussed here) are handled by pure statistical ensembling; some physics is in order Letʼs shutdown the supercomputers for a decade so there is time to study model errors

48 Conclusions The lifetime of scales below 1 km is SHORT At scales below 1 km NWP has no skill (when compared to radar) All tried corrections to forecast errors did not lead to nowcast better than LP (MAPLE) Present data assimilation does seem to lead to nowcasts better than MAPLE Uncertainties in LP nowcast (not discussed here) are handled by pure statistical ensembling; some physics is in order Letʼs shutdown the supercomputers for a decade so there is time to study model errors