Adaptation for global application of calibration and downscaling methods of medium range ensemble weather forecasts Nathalie Voisin Hydrology Group Seminar UW 11/18/2009
Objective Develop a medium range global flood prediction system using: Global Circulation Model ensemble weather forecasts High spatial resolution remote sensing precipitation to downscale and calibrate the precipitation forecasts Using a semi distributed hydrology model Daily time steps, up to 2 weeks lead time For application in all areas with no flood warning system and sparse in situ observations ( radars, gauge stations, etc).
Forecast Schematic Several years back Medium range forecasts (2 weeks) Daily TRMM Precip. (TMPA v6) 2002-2008 0.25 degree VIC Hydrology Model Hydrologic model spinup 0.25 degree DETERMINISTIC Daily ECMWF Analysis Tmin, Tmax, Wind 2002-2008 interpolation to 0.25 degree Initial State ECMWF EPS 50 ensemble members 2002-2008 Forecast calibration and downscaling, 0.25 degree Hydrologic fcst (soil moist., SWE, runoff ) Atmospheric inputs Flow fcst ENSEMBLE FORECASTS
Outline Calibration and downscaling methods: 1. The meteorologist approach versus the hydrologist approach 2. Adaptation of calibration and downscaling methods for global application 3. Ensemble Forecast Verification Focus on medium range ensemble precipitation forecasts
1. Meteorologist vs. hydrologist approaches ( to calibrate and downscale ensemble precip fcst) Different Goals Met: precipitation is the end product Hydro: precipitation is used to derive flood risks ( input to hydrology models, rainfall-runoff regression, etc) Different domains Met: Large spatial domain : entire U.S. or larger Hydro: River Basins specific Different Forecast Types Met: Probability that thresholds will be reached (generally) Hydro: requires precipitation quantities
1. Meteorologist vs. hydrologist approaches ( to calibrate and downscale ensemble precip fcst) Meteorologist approaches: Model Output Statistics (MOS, Glahn and Lowry 1972): regression performed between retrospective forecasts and observation. It is stationdependent. Bayesian Model Averaging for probabilistic quantitative precipitation forecasts (Sloughter et al. 2007): It is station-dependent. Probabilistic forecast that any threshold can be reached. Work is ongoing to add spatial correlation. Need to consider the correlation with other weather variables. Analog Methods (Hamill and Whitaker 2006): over a spatial domain, choose a retrospective forecast (the analog) similar to the forecast, assign the higher spatial resolution observed field corresponding to the date of the analog. Probabilistic forecast that a threshold will be reached.
1. Meteorologist vs. hydrologist approaches ( to calibrate and downscale ensemble precip fcst) Hydrologist approaches: Bias Correction and statistical disaggregation for seasonal flow prediction (Wood and Lettenmaier 2006): monthly bias correction performed at each coarse grid cell, spatial and temporal disaggregation based on scaling precipitation anomalies. Basin dependent limits on the size of the basin for daily application. Precipitation field estimation from probabilistic quantitative precipitation forecasts (Seo et al. 2000) produces daily ensembles of basin mean areal precipitation forecasts. It needs long obs-retrofcst overlap period, computationally intensive. National Weather Service Ensemble Precipitation Processor ( Schaake et al. 2007): a single value forecast is used to derive parameters for conditional distributions of the forecast and observation pairs at each sub basin location. Marginal distributions are derived for both daily events and n-day aggregates. Transformation of both the forecast and the observation to new variables using the normal quantile transform. Joint probability distribution and the Schaake Shuffle (Clark et al. 2004) are used to create ensemble forecasts (mean areal precipitation) for each sub basin.
1. Meteorologist vs. hydrologist approaches ( to calibrate and downscale ensemble precip fcst) Develop a calibration and downscaling approach that would be: - NOT basin dependent ( global approach - Met) - distributed : not mean areal precipitation (Met) - preserve the spatial and temporal rank structure of each weather forecast variables and the correlation between them for hydrology simulation (Hydro) - an ensemble of quantitative forecasts ( vs. probability that an event will be reached ) (Hydro)
2. Adaptation of calibration and downscaling methods for global application 2 methods were adapted for the daily time scale, global application: Bias correction and statistical downscaling for seasonal flow prediction (Wood et al. 2006) Analog methods (Hamill and Whitaker 2006) To be compared to a simple interpolation method
2. Adaptation of calibration and downscaling methods for global application Adaptation of the BCSD (Wood and Lettenmaier 2006): Precip Bias correction performed on a daily time scale (vs. monthly), with adaptation to handle intermittency of precipitation. ( each ensemble, each lead time, each grid cell independently) No temporal disaggregation Spatial disaggregation ( scaling ) made independently to each 1 degree grid cell ( not basin dependent) TRMM 3B42 V6 daily 0.25 degree precipitation (Huffman et al. 2007) is used as the high spatial precipitation observed dataset for the spatial disaggregation. biascorrected fcst,.25degree Precip biascorrected fcst,1deg ree Precip Pr ecip TRMM,.25degree TRMM,1deg ree
5 degree 2. Adaptation of calibration and downscaling methods for global application Adaptation of the Analog method ( Hamill and Whitaker 2006) Retrosp. FCST dataset, +/- 45 days around day n 1 degree FCST D DAY OBS D DAY FCST day n 1 degree 5 degree FCST D DAY FCST D DAY FCST D DAY FCST D DAY FCST D DAY FCST D DAY FCST D DAY FCST D DAY FCST D FCST DAY FCST D nday +/- 45 days Year-1 Corresp. Observation (TRMM) 0.25 degree OBS D DAY OBS D DAY OBS D DAY OBS D DAY OBS D DAY OBS D DAY OBS D DAY OBS D DAY OBS D OBS DAY OBS D nday +/- 45 days Year-1 3 methods for choosing the analog: -Closest in terms of RMSD, for each ensemble -15 closest in terms of RMSD, to the ensemble mean fcst -Closest in terms of rank, for each ensemble Downscaled FCST day n 0.25 degree
2. Adaptation of calibration and downscaling methods for global application Adaptation of the Analog method ( Hamill and Whitaker 2006) Choose an analog for the entire domain (Maurer et al. 2008): entire US, or the globe Ensure spatial rank structure Need a long dataset of retrofcst-observation. Moving spatial window (Hamill and Whitaker 2006): 5x5 degree window (25 grid points) Choose analog based on ΣRMSD, or Σ(Δrank) Date of analog is assigned to the center grid point
Ens. Fcst Mean 20050713 Fcst, 20050713 Analog technique 4 closest analogs in the retrospective forecast dataset Corresponding 0.25 degree TRMM for the analogs, Downscaled ensemble forecast members Downscaled ens. mean forecast TRMM (obs) ( adapted from Hamill and Whitaker 2006)
2. Adaptation of calibration and downscaling methods for global application Adaptation of the Analog method ( Hamill and Whitaker 2006) Modified Schaake Shuffle (Clark et al. 2004) to ensure spatial and temporal rank structure for each ensemble and between weather forecast variables ( temperatures, wind, precipitation). select N dates in the +/-45 days centered around day n in the observed climatology ( N is the number of ensembles) Rank the dates from 1 to N, at each grid cell Rank each ensemble forecast member from 1 to N, each grid cell, lead time one day Reorder the ensembles: New ensemble 1 has the spatial rank structure of date 1. Use consecutive days to the N dates for the other lead times: gives the temporal structure Keep the same dates and do the same for temperature and wind: rank structures across weather forecast variables.
2. Adaptation of calibration and downscaling methods for global application
3. Ensemble Forecast Verification Compare the adapted BCSD and analogs methods with a simple interpolation method For different forecast categories ( conditioned on the forecasts vs. for specific observed events) For different skills: Mean errors : bias, RMSE Reliability: Talagrand ( rank) diagrams,continuous Rank Probability Skill Score Predictability: correlation Over the Ohio River Basin, 2002-2006 period.
Spatial coherence 2005, Jul 13 th 75 th Percentile basin daily acc., 2002-2006 TRMM
Conclusions We intend to pursue the analog RMSDmean method for subsequent applications to flood forecasting because of the following conclusions: Downscaled precipitation patterns for the analog methods were more realistic than the interpolated forecasts in general, while the BCSD was consistently spotty and unrealistic. Accuracy of the analog methods was either maintained or improved upon a simple interpolation in all forecast categories, and was better than BCSD for wet forecasts. The analog rank was the most accurate of the analog methods but the analog RMSDmean was competitive. Reliability of the analog methods was considerably higher than for the interpolated forecasts, and the BCSD. Predictability was not improved with respect to interpolated forecasts for any of the methods. The analog RMSDmean was the best at maintaining reliability at shorter lead times when using TRMM and was improving it when using a gridded station dataset and for 5 day accumulation periods.
3. Ensemble Forecast Verification Reliability: Reliability plot: Choose an event = event specific Each time the event was forecasted with a specific probability ( 20%, 40%, etc), how many times did it happen ( observation >= chosen event). It requires a sharpness diagram to give the confidence in each point. It should be on a 1:1 line. Talagrand diagram ( rank): Give a rank to the observation with respect to the ensemble forecast ( 0 if obs below all ensemble members, Nmember + 1 if obs larger ) Is uniform if ensemble spread is reliable, (inverse) U-shaped if ensemble is too small (large), asymetric is systematic bias.