Improving Excessive Rainfall Forecasts at HPC by using the “Neighborhood - Spatial Density“ Approach to High Res Models Michael Eckert, David Novak, and Keith Brill NOAA/NWS/NCEP Hydrometeorological Prediction Center
HPC Excessive Rainfall Categorical probability of exceeding flash flood guidance (Out to Day 3) Rainfall RFC 4km FFG SLGT = 5-10% MOD = 10-15% HI = >15%
Motivation Excessive Rainfall Forecasts have always been very subjective and relied on human input/experience & low resolution Flash Flood Guidance (FFG) FFG has improved w/inclusion of several important land/physical characteristics & increased resolution to 4km Lower resolution models don’t forecast extreme rainfall (unless parameterized grid-scale feedback occurs – correct ~ 50% of the time) Higher resolution (<5km) models shows promise forecasting extreme rainfall (convection allowing) Kain et al – 2005, Weisman et al Need to objectively quantify extreme rainfall
Integration of Hi-Resolution WRF guidance High-resolution models are not accurate on the scale of individual grid points However, high-resolution models can capture realistic amplitude of events Use neighborhood approach (e.g., Schwartz et al. 2009) to give credit for the correct event/phenomenon, even if the placement is not perfect Also called “Spatial Density “ “Close only counts in Horseshoes“
Model 1 h QPF1” Binary1”/h Density a) Create binary field where threshold exceeded Neighborhood / Spatial Density Approach Schwartz et al. (2009) g) Can be calibrated to provide a probability of exceedance c) Smooth the resulting binary (1 or 0) distribution (using a Gaussian Smoother)
Background Information Jack Kain (NSSL) visits HPC in June 2009 to discuss High Resolution ARW (Advanced Research WRF) QPF Discussions led to NSSL internally produced “experimental” Neighborhood/Spatial Density Plots of rainfall rates Exceeding 1” per hour in a 6 hour period 00-06z, 06-12z, 12-18z & 18-00z) Exceeding 3” per 24 hours Exceeding 5” per 24 hours Derived from summing hourly rainfall
Background Information Led to manual use of NSSL ARW rainfall density output combined with NWS 1, 3 & 6 hour 4km FFG Forecaster “eyeballed” the two parameters FFG on N-AWIPS Density Plots of Rainfall Rate on PC Very inefficient process and not very precise Better way of doing this was required
4 km resolution No convective parameterization NMM (Non-Hydrostatic Mesoscale Model) Domain covers CONUS Initialized with NAM Run at 00 & 12 UTC each day out to 36 hours Data available by 05 & 17 UTC The NCEP WRF Model Run
High-res model QPF Define your forecast period (one, three or six hours) & pull rainfall out of model. HPC Density Approach
Specified rainfall period is overlaid with corresponding FFG 1 hourly rainfall with 1 hour FFG 3 hourly rainfall with 3 hour FFG 6 hour rainfall with 6 hour FFG 1 hour rainfall 4km 1 hour FFG
HPC Density Approach Grid point where rainfall rate exceeds FFG = EVENT GRID-POINT EVENT
GRID-POINT EVENT If Grid box rainfall > FFG assigned value = 1 Otherwise Grid box is assigned a value = 0 Define a “GRID-POINT EVENT“ as: An event occurring within a certain distance (20 or 40 km) of a point Similar to SPC tornado & severe thunderstorm probabilities X HPC Density Approach
Raw data (1s & 0s) are run data through a Gaussian Weighted Filter to create an index of values from Produce both 20km and 40km indices Future calibration will give us probabilities & best spread (20 or 40km) Creating the Exceeding FFG Density Plot
Not a true probability, as values have not been calibrated. For now we call it an “INDEX” with values ranging from 0 to 100 Higher INDEX values means a greater likelihood of rainfall exceeding the FFG However, the guidance is just from one model run which may be in error. The approach applies a spatial smoother to the amplitude information from 1 run, so if the model has the wrong amplitude or placement (say beyond the chosen radius) there will be error. Interpretation
Excessive rainfall forecasts cover either 21 or 30 hours Generation of Density plot of rainfall exceeding FFG: 1.Model rainfall accumulation at 1, 3 & 6 hour periods compared with FFG 2.21 hour forecast has: 1.21 chances of exceeding the 1 hour FFG 2.19 chances of exceeding the 3 hour FFG 3.16 chances of exceeding the 6 hour FFG 3.30 hour forecast has: 1.30 chances of exceeding the 1 hour FFG 2.28 chances of exceeding the 3 hour FFG 3.25 chances of exceeding the 6 hour FFG Application to HPC Excessive Rainfall
Examples
November 25, h Flash Flood Guidance FFG is 1-2” per 3 hours over eastern Iowa
NCEP NMM-WRF 3-h QPF NCEP NMM-WRF forecasts isolated 2.5” amounts during the UTC period. November 25, 2009
20 km Spatial Density Model QPF exceeds FFG, so there is a probability of exceeding FFG November 25, 2009
40 km Spatial Density If we use the 40 km radius, we find the probabilities have been reduced and the area expanded November 25, 2009
Observations Isolated rainfall exceeding FFG occurred 1-1.5” November 25, 2009
5-10” Rainfall 3 hr – 40km density plots 6 hr – 40km density plots 20-30% 12z 12/7/09 – 12z 12/8/09 Observed NMM- WRF Forecast
5-10” Rainfall 3 hr – 40km density plots 6 hr – 40km density plots 90-95% 80-90% 00z 12/24/09 – 00z 12/25/09 Observed NMM – WRF Forecast
00z 12/24/09 – 00z 12/25/ ” Observed Rainfall HPC “High” Risk of Rainfall Exceeding FFG
Conclusions 1.Initial results looks promising 2.Verification & Calibration underway Spatial Density of 20% ≠ Observed 20% frequency 3.Only one model, so treating it as a “perfect prog” If model is wrong (placement/amplitude) then there will be error 4.Several member ensemble (NCEP, NSSL, NCAR WRF) would help build uncertainty into the forecasts 5.Not appropriate for west of Rockies Spatial smoothing does not make sense when pcpn tied to terrain 6.Technique has given forecasters better situational awareness of potential excessive rainfall events, that they otherwise would of not considered
Questions ???