1 / 20

Ensemble Methods and tools

Ensemble Methods and tools. Tara L. Jensen , Barbara Brown, John Halley Gotway , Tressa Fowler and Randy Bullock National Center for Atmospheric Research, Boulder Colorado, USA and Developmental Testbed Center, Boulder Colorado, USA.

meris
Download Presentation

Ensemble Methods and tools

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Ensemble Methods and tools Tara L. Jensen, Barbara Brown, John Halley Gotway, Tressa Fowler and Randy Bullock National Center for Atmospheric Research, Boulder Colorado, USA andDevelopmental Testbed Center, Boulder Colorado, USA 6th NCEP Ensemble Users Workshop, 25-27 March 2014 °NCWCP, Maryland

  2. Overview • Tools • Methods • New Research

  3. Tools and Methods

  4. R Statistics • R ~ the dominant language in the statistical research community. • R is Open Source and free. • Runs on most operating systems • Nearly 2,400 packages contributed. Good for point probability forecasts Struggles with large gridded files • SpatialVx • Rgis, spatstat (GIS) • ncdf ( support for netcdf files ) • rgdal (support for grib1 files) • rNOMADS (support for grib2 files archived by NCEP) • Rcolorbrewer • randomForests Some useful packages • verification • fields (spatial stats) • radiosondes • extRemes • BMA(Bayesian Model Averaging) • BMAensemble • circular • Rsqlite

  5. Model EVALUATION TOOLS (MET) Input Reformat Statistics Analysis Gridded GRIB Input: Observation Analyses Model Forecasts Gen Poly Mask NetCDF Mask MODE ASCII NetCDF PS MODE Analysis ASCII PCP Combine Gridded NetCDF Wavelet Stat STAT ASCII NetCDF PS Grid Stat STAT ASCII NetCDF ASCII Point Obs ASCII2NC Stat Analysis ASCII Ensemble Stat STAT ASCII NetCDF PrepBufr Point Obs PB2NC NetCDF Point Obs Point Stat STAT MADIS Point Obs MADIS2NC = optional Developed by the Developmental Testbed Center, Boulder Colorado, USA

  6. MET capabilities Ensemble Characteristics (Ensemble Stat) Probability Measures (Grid and Point stat) Brier Score + Decomposition Brier Skill Score ROC and Area Under ROC Reliability • Rank Histogram • PIT • CRPS • Ignorance Score • Spread-Skill

  7. New in MET: Series Analysis toolgeographic representation of scores

  8. Object Oriented Method: MODEHow it works • Matched Object 1 • Matched Object 2 • Unmatched Object ENS FCST OBS Radius=5 ObjectThresh>6.35 mm MergingThresh>5.7 mm Radius=5 ObjectThresh>6.35 mm MergingThresh> 5.7 mm Merging Merging Matching No false alarms Misses

  9. Using MODE on Probability Fields

  10. QPE_06 >12.7 mm vs. 50% Prob(APCP_06>12.7 mm) Good Forecastwith DisplacementError? TraditionalMetrics Brier: 0.07 Area UnderROC: 0.62 Spatial Metrics Centroid Distance: Obj1) 200 km Obj2) 88km Area Ratio: Obj1) 0.69 Obj2) 0.65 1 Obj PODY: 0.72 Obj FAR: 0.32 2

  11. MODE for Different Probabilities – May 11, 2013 NWS PoP - Percent chance that rain will occur at any given point in the area. Intersection Area Forecast Area 22,600 0.61 0.58 21,500 SymmetricDifference (non-intersecting area) 0.36 1,050 6,500 0.11 22,200 0.75 19,400 0.41 23,006 0.35 Prob>2.54 mm >50% Prob>2.54 mm >25% Observation NWS PoP= C x A where "C" = the confidence that precipitation will occur somewhere in the forecast area "A" = the percent of the area that will receive measureable precipitation. 16,600 0.72 Forecast Prob>2.54 mm >75%

  12. ENSEMBLE MODE

  13. As probabilities: Areas do not have “shape” of precipitation areas; may “spread” the area As mean: Area is not equivalent to any of the underlying ensemble members Applying spatial methods to ensembles

  14. ExampleMay 11, 2013 Matched Observed Object No Ensemble Mean Matched Ensemble Mean Matched Forecast Object Unmatched Observed Object Unmatched Forecast Object

  15. Spreadincreases With Time

  16. IndividualMATCHED OBSERVED OBJECTs 2 1 3 4 5

  17. median Fcst Matched Forecast Area (grid squares) Obs Object Areas Some observed objects do not have ensemble means matched – this may be good information to have 4 1 3 2 5

  18. May 2013: 27 Days of Matched Observed/Forecast Pairs FORECAST AREA Dominated by 1 member Matched Forecast Area (grid squares) Mean appears to be middle of attribute distribution High bias on forecast area. Sometimes ensemble mean is in middle of attribute distribution and sometimes is it dominated by 1 member

  19. What Next? • Now that we have ability to keep track of all member objects paired with observed objects • Rank Histograms of individual attributes (i.e. area, centroid latitude, longitude, complexity) • Distribution of Attributes • Individual attributes may be used prognostically • Paired Forecast-Observed attributes used diagnostically • Summary with Inner Quartile Range • Spread-Skill diagrams • Diagnostic Analysis • How many ensemble means are not matched even though members are indicating the event may occur • Evaluation of probability fields using MODE objects in new ways

  20. Thank You DTC: http://www.dtcenter.org MET: http://www.dtcenter.org/met/users MET HELP: met_help@ucar.edu Email: jensen@ucar.edu

More Related