Interactive Cluster Tool

1
Interactive Cluster Tool

• For HPC Medium Range Forecast Operations
• Keith Brill and Mike Schichtel

2
What is this Interactive Cluster Tool (ICT)?

• The ICT for HPC MEDR is a procedure that finds a
  subset of an ensemble matching either 500-mb Z or
  PMSL contour fragments drawn by a forecaster and
  uses that subset to generate a full suite of
  means or probabilities for all required forecast
  products.

3
Motivation for Development

• Provide an objective way to resolve forecast
  uncertainties
• Allow increased productivity
• Allow more extensive use of ensembles in the
  forecasting process
• Introduce a different approach to forecast
  product creation
• Motivate future development

4
Forecast Desk Requirements

• Must be easy to use
• Must run fast
• Must provide integrated consistent set of
  guidance products at each forecast hour
• PMSL contours
• Max/min temperatures
• 12-h probabilities of precipitation (POP)

5
Technical Requirements

• Use existing HPC forecaster desktop tools for the
  user interface (NMAP2)
• Include ECMWF and operational GFS as well as
  global ensemble
• Use GEMPAK diagnostic functions
• Use as little stand-alone compiled code as
  possible
• Be capable of producing all fields for the HPC
  NDFD output (e.g., day 8 max/min temps, 6-h dew
  points, 6-h winds, 12-h POP, 6-h cloud fraction,
  6-h weather type)

6
Forecaster Workflow Using ICT
Forecaster Workflow Without ICT
7
Summary of What Is Next

• Presentation of technical details of how the ICT
  works
• Demonstration of an ICT session and presentation
  of an actual ICT forecast case
• Verification of ICT output

8
ICT Data Sources

• Uses lagged ensemble composed of last three runs
  of Global Ensemble
• Uses corresponding cycles of GFS
• Includes ECMWF high-resolution grids (00 and/or
  12 UTC) if available

In the following, each cycle is referred to as a
subcluster.
9
Overview of ICT Execution

• Extract control data from the user-created VG
  file.
• HGHT or PMSL contour fragments
• Loop over three subclusters determining a minimum
  error value (tolerance) for each that allows
  inclusion of at least 15 of members. (Either
  exceed maximum allowed tolerance and exclude
  subcluster or accept at least 15 of members.)
• Use tolerance values to compute weight for each
  subcluster.
• Compute cluster means and probabilities as
  weighted average over subclusters.
• Apply MOS and PRISM adjustments to temperatures.
• Make output VG files for forecasters.

10
Extracting Control Data From the VG File

• Use vgftoascii to convert control VG file to text
• Apply gdctlgd (GEMPAK-like program) to populate
  the control grid
• Is the only compiled code (FORTRAN) developed for
  the ICT
• Works somewhat like the GEMPAK grphgd program
• Assign missing values to all grid points
• Assign the value associated with the contour
  fragment to the single grid point nearest each
  vertex along the contour fragment
• Repeat the value assignment above for labelled
  contour fragments

The control grid consists of a mix of missing
values and values from contour fragments.
11
Determine Tolerance Value for Each Subcluster

• Set the tolerance value low (e.g., 90 m for HGHT)
• Use GEMPAK ensemble probability function to
  compute the fraction of ensemble members for
  which the magnitude of the difference between the
  control grid values and the ensemble members is
  less than the tolerance at all non-missing grid
  points.
• ens_prob( sgmn(ge(ctol,miss(abs(sub(valcntl
  2gmpkdtm,cfunc)),0))) ), where
  ctoltolerance, cntlh5 or sp,
  gmpkdtmdate-time, cfuncHGHT or PMSL
• sgmn() evaluates to 1 or 0 everywhere on the
  grid and serves a selector function effectively
  turning individual members on or off. It is used
  as a multiplier for fields for which ensemble
  means or probabilities are required. Each mean
  or probability so obtained must be divided by the
  fraction computed above by ens_prob().
• If the fraction of the ensemble member count
  exceeds the target (.15) then the tolerance value
  is established otherwise, increment the
  tolerance value and return to step 1 above.

12
Compute Cluster Means Using Subcluster Weights

• Use the tolerance values for each subcluster
• Compute the weight for each subcluster as the
  difference between the sum of all the tolerance
  values and subcluster tolerance value divided by
  the sum of all such differences
• Compute cluster means as weighted averages of
  subcluster values

13
MOS Adjustment

• 00 UTC Ensemble MOS at points is considered as
  verifying data for the corresponding gridded
  Global Ensemble member.
• Max/Min Temperature and dew point values at MOS
  points are analyzed to a subset of the global
  grid using GEMPAK Barnes objective analysis.
• Additional smoothing is applied on the grid.
• The MOS correction is the average over all
  members of the difference (MOS mbr)
• The MOS correction is added to max/min
  temperatures and dew points.

14
PRISM Adjustment

• PRISM adjustment is done after MOS correction.
• PRISM climatology consists of very high
  resolution gridded (NDFD) monthly means of
  max/min tempertures, dew points, and
  precipitation.
• For temperatures and dew points, the PRISM
  adjustments are values to be added to
  low-resolution data and are obtained as follows
• Map PRISM data to global grid preserving area
  averages
• Map this data back to the NDFD grid using
  bilinear interpolation
• Subtract the coarse data from the original data
  on NDFD grid
• For precipitation, the PRISM adjustments are
  values to multiply the low-resolution data and
  are obtained as above, except in step 3 the ratio
  of the original data to the coarse data is
  computed and used as the correction factor.

15
Example ICT Session

• Forecasters may draw a few contour fragments of
  the 500-mb HGHT or the PMSL to use as a control
  field for the ICT.
• For the MEDR forecast example, the forecaster
  used a blender tool to generate a 500-mb HGHT
  field over the Pacific and western US.

16
Day 6 FRCST from 20060522 valid 12 UTC 20060528
17
Selected Text Output From ICT

• Control field is 500 MB HGHT.
• ENSEMBLE cycle 1 gefs060522/00,gefc060522/0
  0,gfs060522/00,ecm_2200.grd,ecm_2200.grd,ecm_2200
  .grd
• Sub-cluster of 3 members is 16 of ensemble cycle
  1.
• Sub-cluster members are within 120 M of contour
  fragments.
• ENSEMBLE cycle 2 gefs060522/06,gefc060522/0
  6,gfs060522/06
• Sub-cluster of 4 members is 25 of ensemble cycle
  2.
• Sub-cluster members are within 120 M of contour
  fragments.
• ENSEMBLE cycle 3 gefs060522/12,gefc060522/1
  2,gfs060522/12
• Sub-cluster of 7 members is 44 of ensemble cycle
  3.
• Sub-cluster members are within 150 M of contour
  fragments.
• Weighted cluster is 35ict_1_fd6.grd,35ict_2_fd6
  .grd,30ict_3_fd6.grd.

Full ensemble specification
Final cluster
Cluster tolerance value
18
HPC MEDIUM RANGE FORECASTER MIKE SCHICHTEL
PRESENTS A DAY 5 ICT FORECAST CASE VALID 8 MAY
2006
Although developed as an interactive tool, the
ICT is run automatically on a Pacific- Western US
500-mb height forecast field created by the HPC
Fronts/Pressures Forecaster using model blending
tools.
19
24 HOUR QPF VALID 00 UTC 9 MAY 2006
NCEP ENSEMBLE MEAN
GFS
20
DAY 5 500 MB FCSTS
NCEP ENSEMBLE MEAN
GFS
UKMET
ECMWF
21
Although developed as an interactive tool, the
ICT is also run automatically on a
Pacific-Western US 500-mb height forecast field
created by HPC Fronts/Pressure Forecasters
using model blending tools.
22
Forecaster Input Red 500-mb HGHT contours
23
ENSEMBLE MEAN PMSL VERSUS ANALYSIS
ICT PMSL VERSUS ANALYSIS
24
Additional ICT Output(not shown)

• Maximum Temperatures (from 6-h 2-m T)
• Minimum Temperatures (from 6-h 2-m T)
• 12-h Probability of Precipitation (POP)
  (univariate frequency of QPFgt.03 in cluster)
• These are displayed as values plotted at
  forecast station locations (points).

25
ICT Verification Procedure

• Run the ICT automatically using the HPC final
  Pacific/western US 500-mb HGHT field issued
  around 1400 local time as the control field
• Includes the 12 UTC model runs, but not the 18
  UTC runs
• Verify PMSL against HPC analysis to compute
  standardized anomaly correlations
• Verify max/min temperatures at points against the
  MDL verifying station data (the data used to
  verify MOS) to compute Root Mean Squared Error
  (RMSe)
• Verify POP at points against MDL station data
  using Brier Score

26
(No Transcript)
27
(No Transcript)
28
(No Transcript)
29
(No Transcript)
30
Summary Future Work

• ICT is a good tool for creating the PMSL forecast
• ICT does not outperform MOS for max/min temps and
  POPs
• FUTURE work ..
• Use bias corrected ensemble forecasts
• Include all the NAEFS members
• Add MOS correction for POPS
• Investigate other methods of applying MOS
  corrections