Presentation is loading. Please wait.

Presentation is loading. Please wait.

Verification Summit AMB verification: rapid feedback to guide model development decisions Patrick Hofmann, Bill Moninger, Steve Weygandt, Curtis Alexander,

Similar presentations


Presentation on theme: "Verification Summit AMB verification: rapid feedback to guide model development decisions Patrick Hofmann, Bill Moninger, Steve Weygandt, Curtis Alexander,"— Presentation transcript:

1 Verification Summit AMB verification: rapid feedback to guide model development decisions Patrick Hofmann, Bill Moninger, Steve Weygandt, Curtis Alexander, Susan Sahm

2 Motivation There is a critical need for both rapid and comprehensive statistical and graphical verification of model forecasts from various AMB experimental models: RUC, RR, and HRRR - Real-time parallel cycles as well as retrospective runs - Two primary types: - Station verification : Upper-air, surface and clouds - Gridded verification: Precipitation, radar reflectivity, convective probabilities - Illuminate model biases and patterns to errors - Essential for evaluating model/assimilation configuration changes Rapid verification feedback enables timely improvement in forecast skill

3 Design Goals Fast computation and display of verification results (real- time for real-time cycles, day or two for retros) Simple procedures, but with sufficient options to elucidate key aspects (quantify visual impressions) Built-in capabilities to allow quick stratification by key parameters (metric, threshold, scale, valid time, initial time, forecast length, region) Easily accessible web-based presentation of verification results  ability to quickly examine aggregate statistics AND single-case plots in complementary manner Verification design driven by needs of forecast system developers

4 Design Details Use modified NCEP IPOLATES routines for interpolation and upscaling of input fields to multiple common grids. Calculate contingency table fields (YY, YN, NY, NN) for multiple scales, domains, and thresholds: -- database storage for statistical aggregation -- graphics for each event for detailed evaluation Web-based interface for aggregate statistics and event graphics Apply to multiple gridded fields (reflectivity, precipitation, probabilities) and multiple model runs (several version each of RUC, RR, HRRR as well as RCPF, HCPF, etc.)

5 Statistics Webpages Composite Reflectivity Time Series: http://ruc.noaa.gov/stats/radar/beta/timeserieshttp://ruc.noaa.gov/stats/radar/beta/timeseries Valid Times: http://ruc.noaa.gov/stats/radar/beta/validtimeshttp://ruc.noaa.gov/stats/radar/beta/validtimes Lead Times: http://ruc.noaa.gov/stats/radar/beta/leadtimeshttp://ruc.noaa.gov/stats/radar/beta/leadtimes 24 Hour Precipitation Time Series: http://ruc.noaa.gov/stats/precip/beta/timeserieshttp://ruc.noaa.gov/stats/precip/beta/timeseries Thresholds: http://ruc.noaa.gov/stats/precip/beta/thresholdshttp://ruc.noaa.gov/stats/precip/beta/thresholds Convective Probabilities Time Series: http://ruc.noaa.gov/stats/prob/beta/timeserieshttp://ruc.noaa.gov/stats/prob/beta/timeseries CSI vs Bias: http://ruc.noaa.gov/stats/prob/beta/csibiashttp://ruc.noaa.gov/stats/prob/beta/csibias Reliability Diagrams: http://ruc.noaa.gov/stats/radar/prob/reliabilitydiagrams http://ruc.noaa.gov/stats/radar/prob/reliabilitydiagrams ROC Curves: http://ruc.noaa.gov/stats/prob/beta/rochttp://ruc.noaa.gov/stats/prob/beta/roc

6 Sample “time-series” stats interface Model Region Scale Averaging period Metric Threshold Forecast Length Valid time Date Range Many R/T runs and retros

7 RR-dev w/ Pseudo-obs HRRR-dev HRRR HRRR-dev better HRRR better Reflectivity (> 25 dBZ) CSI Eastern US on 40 km grid (3-day avg) Models Thresh “Time series” mode Metric Region Scale Sample application of “time-series” stats Difference

8 HRRR-dev HRRR RR-dev w/ Pseudo-obs HRRR-dev better HRRR better Reflectivity (> 25 dBZ) CSI Eastern US on 40 km grid (3-day avg) Models Thresh “Time series” mode Metric Region Scale Sample application of “time-series” stats Difference

9 HRRR-dev HRRR RR-dev w/ Pseudo-obs HRRR-dev better HRRR better Reflectivity (> 25 dBZ) CSI Eastern US on 40 km grid (3-day avg) Models Thresh “Time series” mode Metric Region Scale Sample application of “time-series” stats Difference Implemented in RR-prim HRRR-dev Longer time-step

10 HRRR-dev HRRR RR-dev w/ Pseudo-obs HRRR-dev better HRRR better Reflectivity (> 25 dBZ) CSI Eastern US on 40 km grid (3-day avg) Models Thresh “Time series” mode Metric Region Scale Sample application of “time-series” stats Difference Implemented in RR-prim HRRR-dev Longer time-step RR-dev Added shorter vert. length-scales in RR-dev/GSI Imple- mented In HRRR

11 CSI 25 dBZ 40-km EUS +6h fcst 8-22 Aug RUC HRRR Better RR HRRR better Sample “time-series” stats to examine scatter in forecast differences August

12 Sample application of “lead-time” stats illustrating CSI and bias “die-off” for different strengths of radar heating CSI (X100) Bias (X100) Forecast Length (hours) 0 2 4 6 8 10

13 Upscaled verification (especially to 40km and 80km) reveals “neighborhood” skill in HRRR forecasts, especially around the time of convective initiation 20-km 80-km 40-km 3-km HRRR 25dBZ, 6-h fcst Valid Time (GMT) CSI (x 100) Sample application of “valid time” stats illustrating diurnal variation in scale-dependent skill Convective Initiation time 00z 04z 08z 12z 16z 20z 00z

14 Reflectivity Graphics Webpage http://ruc.noaa.gov/crefVerif/Welcome.cgi

15 12z + 6 hr 3-km 40-km Miss FA Hit Single case plots showing “neighborhood” skill Obs Refl. HRRR fcst

16 13-km CONUS Comparison 2 X 12 hr fcst vs. CPC 24-h analysis 1 – 31 Dec 2010 Matched RR vs. RUC Precipitation Verification RR RUC | | | | | | | | 0.01 0.10 0.25 0.50 1.00 1.50 2.00 3.00 in. | | | | | | | | 0.01 0.10 0.25 0.50 1.00 1.50 2.00 3.00 in. CSI (x 100) RUC RR 100 (1.0) bias (x 100) Sample application of “threshold” stats to show skill for range of precip amounts

17 Precipitation Graphics Webpages http://ruc.noaa.gov/precipVerif

18 CPC 24-h precip RUC Thrs CSI Bias 1.00.45 1.22 2.00.29 1.95 observed Thrs CSI Bias 1.00.31 0.69 2.00.21 0.58 2 x 12h fcst interpolated to 20-km grid RR vs. RUC 24-h precip. verif Single case plots showing forecast skill for precip. RR

19 RUC Thrs CSI Bias 1.00.45 1.22 2.00.29 1.95 Thrs CSI Bias 1.00.31 0.69 2.00.21 0.58 1” threshold Miss FA Hit CPC 24-h precip observed 2 x 12h fcst interpolated to 20-km grid RR vs. RUC 24-h precip. verif Single case plots showing forecast skill for precip.

20 2-h fcst 4-h fcst 6-h fcst ROC curve CSI vs. bias Sample display of probability verification statistics Work in progress, have display for CCFP and CoSPA probabilities Plan to add HCPF, RCPF, expand to probabilities of other hazards (fog, high echo-tops, etc.)

21 2-h fcst 4-h fcst 6-h fcst Sample Reliability Diagram All plots can zoom

22 Conclusion The verification system, including both the statistical and graphical webpages, greatly aids evaluation of model performance within AMB and facilitates rapid assessment of experimental configurations and improvements in real-time. We are also able to verify retrospective cases of scientific interest in very quick succession for use in presentations and publications for outreach endeavors.


Download ppt "Verification Summit AMB verification: rapid feedback to guide model development decisions Patrick Hofmann, Bill Moninger, Steve Weygandt, Curtis Alexander,"

Similar presentations


Ads by Google