Strategy for Using CPC Precipitation and Temperature Forecasts to Create Ensemble Forcing for NWS Ensemble Streamflow Prediction (ESP)

Transcription

1 Strategy for Using CPC Precipitation and Temperature Forecasts to Create Ensemble Forcing for NWS Ensemble Streamflow Prediction (ESP) John Schaake (Acknowlements: D.J. Seo, Limin Wu, Julie Demargne, Rob Hartman, Satish Regonda) CPPA PI Workshop October 1, 28

2 Discussion Topics CPC Forecasts: Why is an Ensemble Preprocessor (EPP) needed? ESP Forcing Requirements Ensemble Preprocessor (EPP) Goals Science Issues for Hydrologic Application of CPC Forecasts Ensemble Preprocessor (EPP) Science Strategy Some EPP/ESP Results What to Expect

3 Why is EPP needed? Weather and Climate forecasts are biased Weather and Climate ensemble forecasts do not account for all of the uncertainty Spread is typically under-estimated Some events that might occur are not represented Space-time structure of members not fully representative: (i.e. space-time structure of future events is essentially unknown) Ensemble spread is not properly related to differences between ensemble mean and observed values Atmospheric ensemble members are not equally likely they are more like sensitivity analyses to estimated uncertainty in initial conditions Major research is needed to adjust raw ensemble members for hydrologic application Need to combine information from many forecast sources

4 Hydrology Ensemble Forcing Requirements Skillful and reliable precipitation and temperature ensemble forcing time-series for each subbasin/segment (1-day to ~ 1-year) at ESP time-step Ensemble forcing must preserve space-time variability of precipitation and temperature forcing. Uncertainty in future forcing is time and space scale dependent and this must be represented in ensemble input to ESP Ensemble members must be consistent in space and time (over entire forecast space-time domain). Required to assure streamflow members can properly be routed downstream Ensemble members must be equally likely Need to make best possible use of all weather and climate forecast information

5 Ensemble Preprocessor (EPP) Goals Create ensemble future forcing for ESP from single-value atmospheric forecasts Use all available forecast information from NCEP (CPC,EMC,HMC) and RFC HAS forecaster Produce future MAP/MAT values that are consistent in space and time over the entire forecast domain (so that hydrologic forecast members can be routed downstream) Maintain skill of raw forecasts Represent future uncertainty reliably Include conditional uncertainty if possible Allow for multi-model forecast input Keep it simple as possible Make it good not perfect

6 Science Issues for Hydrologic Application of CPC Forecasts How to construct an ensemble of CFS forecasts from individual CFS forecasts (created during the past month) How to downscale/rescale CFS ensembles to create ESP forcing How to use empirical CPC forecast products in addition to CFS to construct ensemble forcing for hydrological models

7 Ensemble Preprocessor (EPP) Science Strategy - General Historical forecasts and observations from same historical period are used to calibrate EPP Historical forecasts from same system as used in operations Climatology of observations should be same as used for hydrologic model calibration Climatology of forecasts must be known

8 Ensemble Preprocessor (EPP) Science Strategy Input Forecasts RFC single-value precipitation forecasts (archive required) MOS temperature forecasts (days 1-7) GFS Ensemble Mean from Fixed Version of GFS for now (days 1-14). CFS forecasts beyond 2wks (singlevalue/ensemble-mean) CPC empirical forecast products SHREF ensemble mean products Other forecasts (in a multi-model context)

9 Ensemble Preprocessor (EPP) Science Strategy - Procedures Forecast probability distributions for many overlapping future events at different space/time scales Use historical space-time patterns to create ensemble members with prescribed forecast probability distributions over RFC forecast domain ( Schaake Shuffle ) 1-6hr time steps 1-day to 1-year forecast periods forecast basin segments entire forecast area Ensemble members preserve all forecast probability distributions for all futue events

10 Forecast Valid Time (6-hr periods) moving 6hr window average from t= average from t=4 average from t=8 average from t=16 Ensemble Preprocessor (EPP) Two Step Process Science Strategy 1. Define events [location (space-time); scale (space-time)] and derive a forecast probability distribution for each event 55 Forecast Window Width (6-hr periods) week 2 weeks Correlation Coefficient GFS Precipitation Forecast Correlation vs Lead Time NFDC1HUF - January 15 Forecast Creation Time Increasing lead time to end of forecast period Forecast Lead Time to Beginning of Forecast Window (6-hr periods) 2. Create ensemble members that preserve all forecast probability distributions (using Schaake Shuffle )

11 CPC Events Temporal Scale Weeks 1 4 (individually and cumulatively) Months 1 8 (individually and cumulatively) 3-month Seasons Spatial Scale Gridded fields ¼ degree to ~1 degrees? Derive ESP forcing from finest grid

12 Discussion Topics Hydrology Ensemble Forcing Requirements NCEP Ensemble Forecasts: Why is EPP needed? Ensemble Preprocessor (EPP) Goals Science Issues for Hydrologic Application of CPC Forecasts Ensemble Preprocessor (EPP) Strategy Some EPP Results What to Expect

13 Thank You Water Predictions for Life Decisions

14 a 4b c 4d space (HRAP bins) 6 space (HRAP bins) Return precip. depth (mm) precip. depth (mm) precip. depth (mm) precip. depth (mm) space (HRAP bins) space (HRAP bins)

15 Joint Distribution of Forecasts and Observations 25 Observation (mm) Forecast (mm)

16 Observed vs Global Ensemble Climatologies (July, Lat = 35., Lon = 82.5) 35 Forecast vs Observed Precipitation - July Probability Observed (mm/day) Forecast (mm/day) Observations Global Ensemble Daily Precipitation (mm) Return

17 Temporal Scale Dependency of Precipitation Forecasts GFS Precipitation Forecast Correlation vs Lead Time NFDC1HUF - January 15 Forecast Creation Time Correlation Coefficient Forecast Valid Time (6-hr periods) moving 6hr window average from t= average from t=4 average from t=8 average from t=16

18 GFS Precipitation Forecast Correlation Coefficient: Temporal Scale-Dependency NFDC1HUP January 15 Forecasts Notes: Forecast Window Width (6-hr periods) week 2 weeks Correlation Coefficient GFS Precipitation Forecast Correlation vs Lead Time NFDC1HUF - January 15 Forecast Creation Time moving 6hr window.6 average from t=.5 average from t=4.4 average from t=8.3 average from t= Forecast Valid Time (6-hr periods) Increasing lead time to end of forecast period This diagram would be different for a different time of year This diagram would be different for a different geographical location. This diagram is for correlation at one spatial scale. This diagram would be different for a different spatial scale Forecast Lead Time to Beginning of Forecast Window (6-hr periods)

19 Atmospheric Pre-Processing Requirements Mesh ensemble forcing from short, medium, and long range techniques. Maintain spatial and temporal relationships across very large areas. time mesoscale wx models downscaling medium range wx models downscaling long range global circulation models downscaling rainy + cold clear + warm variable downscaling snowing cloudy + hot forecaster skill climate forecasts and indexes Irrational outcomes

20 Smith River at Crescent City, CA Forecasts and Simulations CREC1: December Forecasts and Simulations CREC1: March Forecasts and Simulation 1 1 Correlation with Observations calibration rg14c3 rgc3 Correlation with Observations calibration rg14c3 rgc Forecast Period (Days) Forecast Period (days) Return

21 CFS Hindcasts were made: -Once per day -15 days per month -55 +/- 12-hr time-steps -Beginning 12/9/198 CFS Hindcasts Day of Month