2010 Dec 26 Blizzard – Using ensemble sensitivity analyses to examine forecast jump between Dec 24 and Dec 25 Sensitivity calculations performed by Minghua Zheng (SBU student) and Edmund Chang
Forecasts from 24 Dec 2010 Courtesy: Dan Petersen Short-Range Forecast Issues
Forecasts from 25 Dec 2010 Courtesy: Dan Petersen
Data analyzed –From TIGGE –ECMWF ensemble forecast from Dec 24 00Z and Dec 25 00Z 50 members MSLP and Z500 –Also examined combined NCEP+CMC+UKMO ensemble from same initial times members Results are similar and will not be shown
ECMWF Difference between MSLP ens mean f/c at 2010Dec2712Z Initial time 2010Dec2500Z (60hr) – 2010Dec2400Z (84hr)
Ensemble sensitivity calculations: Recall: - J is any forecast metric at the final forecast time - x i is any variable within the model state vector Here, J is the projection of the “shift” pattern onto each ensemble member at final forecast time (Dec27 12Z) Based on ECMWF 84-hr forecast ensemble initialized at Dec 24 00Z Note: In all plots, -0h is the forecast valid time (Dec27 12Z) DEC Z DEC Z Projection of a pattern: Pattern: p i Ensemble member anomaly: x i Projection of Pattern onto Ensemble member: p i x i Basically value of projection is large when the anomaly of the ensemble member resembles the pattern
ECMWF (50 members): evolution of forecast difference between DEC25Z00 and DEC24Z00 MSLP (hPa)Z500 (dm) DEC27 12Z DEC25 00Z
Results from sensitivity analyses suggest that: –Cyclone E-W location at -0 hr (Dec 27 12Z) is sensitive to conditions near the Gulf coast and over northern Canada at -48 to -60 hr –Forecast errors from Dec 24 00Z forecast (comparing to Dec 25 00Z analysis and forecast) also show errors over those sensitive regions –Can we assess the impacts of those short- range (24-hr) forecast errors?
Here, we introduce an alternative way of ensemble sensitivity analysis to examine the “forward” impact of forecast errors at early time on forecasts at later time We believe that this way of analysis complements the usual way of ensemble sensitivity analysis in which we compute sensitivity backwards based on forecast anomalies at forecast validation time
ECMWF (50 members): evolution of forecast difference between DEC25Z00 and DEC24Z00 MSLP (hPa)Z500 (dm) DEC27 12Z DEC25 00Z How do the forecast errors at -60 hr over these regions impact the forecast at -0 hr?
Region near gulf coast at -60h Z500 difference (m) between Dec2500Z analysis and Dec2400Z forecast
Note: Regression based on projection of the pattern on each ensemble member can be considered to be weighted mean of ensemble members that have anomalies resembling the chosen pattern Regression based on Z500 projection in black boxed region using forecast ensemble from Dec 24 00Z Difference between Dec 25 and Dec 24 forecasts Z500 (unit dm)
Difference between Dec 25 and Dec 24 forecasts Regression based on Z500 projection in black boxed region using forecast ensemble from Dec 24 00Z MSLP (unit mb)
Region over North Canada at -60h Z500 difference (m) between Dec2500Z analysis and Dec2400Z forecast
Regression based on Z500 projection in black boxed region using forecast ensemble from Dec 24 00Z Difference between Dec 25 and Dec 24 forecasts Z500 (unit dm)
Difference between Dec 25 and Dec 24 forecasts Regression based on Z500 projection in black boxed region using forecast ensemble from Dec 24 00Z MSLP (unit mb)
Regression based on Z500 projection in black boxed region using forecast ensemble from Dec 24 00Z MSLP (unit mb) Comparing results from the 2 regions
Summary Results from sensitivity analyses suggest that: –Cyclone location at Dec 27 12Z sensitive to prior conditions (at -48hr to -60hr) near the gulf coast as well as over northern Canada west of Hudson Bay –Short range (24-hr) forecast errors over those locations from forecasts based on Dec 24 00Z apparently led to cyclone being forecast too far east –Hypothesis: Forecasts based on Dec 25 00Z are better because of observations over those regions? Need to confirm with obs denial studies –Results based on ECMWF ensemble (50 members) and (NCEP+CMC+UKMO) ensemble (63 members) are qualitatively consistent
Outlook for CSTAR Can something similar be done in near real time for CSTAR? –Can access GEFS data (20 members) –Access to Canadian ensemble data? –No access to ECMWF and UKMO (2-day delay for TIGGE) –Perhaps combine 2 GEFS ensemble (40 members) with 6-hr difference initial time? Automation difficult since reference regions have to be selected subjectively However, once data are available and regions have been selected, computation of sensitivity and regression are rather straightforward