Use of Mesoscale Ensemble Weather Predictions to Improve Short-Term
Precipitation and Hydrological ForecastsMichael Erickson1, Brian A. Colle1, Jeffrey Tongue2, Alan Cope3, and Joseph
Ostrowski4
1 School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY2 National Weather Service, Upton, NY
3 National Weather Service, Mt. Holly, NJ4 Mid-Atlantic River Forecast Center, State College, PA
Motivations and Goals- Determine whether the probability of river flood forecasts can be improved using a
multi-model ensemble (NCEP SREF, Stony Brook ensemble, GFS, and NAM).
- Since ensembles will be run using convective parametrizations for several years (dx
> 4-km grid spacing), it is important to understand the precipitation errors fed into
the hydrological models.
- Using the SBU ensemble, verify QPF using a high resolution precipitation dataset
(stage IV) to determine whether model performance varies spatially and temporally.
- Determine if certain members outperform
others in order to allow for unequal
weighting in the ensemble streamflow
forecasts.
Hydrologic Flowchart
SBU EnsembleSBU Ensemble
Upton, NY and Mount Holly Upton, NY and Mount Holly
WFO: Ingested into AWIPSWFO: Ingested into AWIPS
MARFC: MARFC:
Downscaling and Downscaling and
Basin AveragingBasin Averaging
6hr accumulated QPF 6hr accumulated QPF
ingested into Ensemble ingested into Ensemble
Streamflow PredictionStreamflow Prediction
Data ingested into Data ingested into
Site Specific for Site Specific for
the Passaic Basinthe Passaic Basin
0000 UTC 13-Member MM5/WRF Ensemble
7 MM5 Members:
-**WRF-NMM (Grell, MRF, Sice)
-WRF-NMM (Grell, M-Y, Reis2)
-GFS (Betts-Miller, M-Y, Sice)
-GFS (KF2, MRF, Reis2)
-NOGAPS (Grell, Blackadar, Sice)
-CMC (KF2, M-T, Sice)
-18 Z GFS + FDDA (Grell, Blackadar,
Sice)
6 WRF-ARW Members:
-**WRF-NMM (KF2, YSU, Ferrier)
-WRF-NMM (Betts-Miller, M-Y, WSM3)
-GFS (Grell, YSU, Ferrier)
-GFS (KF2, M-Y, Ferrier)
-NOGAPS (Betts-Miller, YSU, WSM3)
-CMC (KF2, M-Y, WSM3)
All runs integrated down to 12-km grid spacing to hour 48**
MM5/WRF members are run down to 4-km grid spacing
Ensemble MM5 36- and 12-km Domains
Hanna Case – 9/6/08 00z Run Select WRF Members Select MM5 Members
Ensemble Mean Stage IV Data
Ensemble Streamflow Prediction: Hanna Case9/6/08 00z Run: Saddle River: Lodi, NY
QPF from Ensemble River Response from Ensemble
River Response: Mean and Spread Forecast and Observed River Height
Stage IV and model details
Stage IV data consists of radar estimates and rain gauge data that were
blended with some additional manual quality control. Accumulated rainfall
between hours 18 and 42 of the model run were considered.
The stage IV rain data was interpolated
to the 12 km MM5/WRF model grid.
Regions sufficiently offshore were masked.
The 2007 and 2008 cold seasons (12/1-
3/31) and 2006 to 2008 warm seasons (5/1-
8/31) were analysed.
Total Stage IV Warm Season Precip.
Model Error – Warm Season 2006 - 200818-42 Hr Acc. Precip
Mean Absolute Error
Average Member Ranking
Model Bias
MM
5W
RF
Spatial Bias Plots – Warm Season 2006 – 200818-42 Hr Acc. Precip
BMMY-ccm2.NEUS.avn GRMRF.NEUS.eta
Ensemble Mean Variance in Bias across Members
Member Bias Variability: Passaic, NYC and LI - Warm Season 2008
7 Day % of Obs. ME – 2008 Season WRF7 Day % of Obs. ME – 2008 Season MM5
Biases vary greatly in time and are negatively correlated (-0.35 to -0.50) to
Stage IV rain data for the MM5 models.
Brier Score Plots – Warm Season 2006 – 200818-42 Hr Acc. Precip
Brier Score: Threshold > 0.1” Brier Score: Threshold > 0.5”
Ensemble Mean Bias: Threshold > 0.1” Ensemble Mean Bias: Threshold > 0.5”
Rank Histograms: Warm and Cool Seasons18-42 Hr Acc. Precip
Rank Histogram: Warm Season
Rank Histograms are consistently underdispersed and show a general wet bias.
Rank Histogram: Cool Season
Model Error – Cool Season 2007 - 200818-42 Hr Acc. Precip
Mean Absolute Error
Average Member Ranking
Model Bias
MM
5W
RF
Spatial Bias Plots – Cool Season 2007 – 200818-42 Hr Acc. Precip
GFS.MYJ.KFE.WSM3 GRMRF.NEUS.eta
Ensemble Mean Variance in Bias across Members
Member Bias Variability: Passaic, NYC and LI - Cool Season
7 Day % of Obs. ME – 2008 Season WRF7 Day % of Obs. ME – 2008 Season MM5
Biases not as sensitive to low stage IV rain days, although there is still a slight
negative correlation.
Brier Score Plots – Cool Season 2007 – 200818-42 Hr Acc. Precip
Brier Score: Threshold > 0.1” Brier Score: Threshold > 0.5”
Ensemble Mean Bias: Threshold > 0.1” Ensemble Mean Bias: Threshold > 0.5”
Conclusions
Most ensemble members tend to have a overprediction bias for precipitation
during both warm and cool seasons. The overprediction variability among members is largest during the warm
season and areas that experience more convection (MD and DE area, major
valleys, etc...). This suggests large sensitivities to the convective
parametrization (and other physics). Overprediction is not as large for many members during the cool season, but
the raw ensemble is still positively biased and underdispersed. Some ensemble members perform better than others, with WRF members
better than MM5 during the cool season. SBU Ensemble data is now being used by the Ensemble Streamflow
Prediction (ESP) system at MARFC and Site Specific at Upton/Mt. Holly.
Ensemble will have to be bias corrected and weighted given the errors noted
above.