Improvements in Deterministic and Probabilistic Tropical Cyclone Wind Predictions: A Joint Hurricane Testbed Project Update Mark De. Maria and Ray Zehr NOAA/NESDIS/ORA, Fort Collins, CO John Knaff and Kimberly Mueller CIRA/CSU, Fort Collins, CO Presented at The Interdepartmental Hurricane Conference March 2005 Jacksonville, FL
Outline • Deterministic Intensity Prediction – GOES and Recon Intensity Prediction (GRIP) model • Predictors from aircraft recon and IR radial structure combined with SHIPS forecasts – Evaluate neural network techniques • Probabilistic Intensity Prediction – Monte Carlo wind probability model • Results from 2004 • 2005 Plans • Are Intensity Forecasts Improving?
The 2004 SHIPS Model • • • Statistical-dynamical intensity model (12 -120 hr) Developed from 1982 -2003 sample Empirical decay for portion of track over land Track from adjusted 6 -hour old NHC forecast Version with satellite input operational for 2004 SHIPS Input – – Climatological: Julian Day Atmospheric Environment: Shear, T 200, 850 Oceanic Environment: SST, Ocean Heat Content Storm Properties: Vm, d. Vm/dt, motion, PSL, lat, GOES Cold Pixel Count, GOES TB Std Dev • Most storm property inputs are indirect measurements
SHIPS Forecast Skill 2004 Atlantic Sample
Aircraft Data in the GRIP Model • USAF Reserve and NOAA aircraft data – Highly utilized for intensity estimation – Under utilized for intensity prediction • Real time automated analysis system – Real time aircraft database set up on NCEP IBM (C. Sisko) – Move data to storm-relative coordinates – Automated quality control • Test for data coverage • Gross error check • Check deviations from pre-analysis – Variational objective analysis in cylindrical coordinates • Greater azimuthal than radial smoothing
Sample Analysis for Hurricane Jeanne 2004 Input Data Wind Analysis Isotachs 358 Dependent Cases (1995 -2003, 12 hour intervals) 124 Independent Cases (2004, 6 hour intervals) Input to GRIP Model: Azimuthally Averaged Tangential Wind
GOES Data in the GRIP Model • SHIPS already includes cold pixel count and Tb standard deviation (area averages) • Examine radial structure of GOES data for predictive signal Azimuthal Average
GRIP Model Statistical Development • GRIP Predictors – EOF Version • SHIPS Forecast • Amplitudes of first four EOFs of GOES and Recon profiles (principal components) – Physical Version • SHIPS Forecast • 10 physical parameters from GOES and recon profiles • Final GRIP Model – EOF Version • SHIPS forecast, 2 recon PCs, 1 GOES PC – Physical Version • SHIPS forecast, 3 recon variables, 1 GOES variable • Both versions tested on 124 cases from 2004 Atlantic season
GRIP Model Results 2004 Independent Cases
2005 GRIP Model • Add 2004 cases and re-derive the coefficients – ~20% increase in sample size • Consider combined EOF and physical variable version • Run in real time during 2005 season for further evaluation
Neural Network Model (Short Version: It didn’t work) • NN Model Development – SHIPS dependent dataset used for training • Non-satellite version – Development by Prof. Chuck Anderson, CSU computer science department – 5 to 10% reduction in mean absolute error in dependent sample (12 -120 hr) • Independent tests – 2 -5% degradation – NN Method appears to over-fit training data – One final try with more stringent fitting requirements • Restrict input to only those predictors selected by SHIPS
Monte Carlo Wind Probability Model • Provides 5 day surface wind probabilities – 34, 50 and 64 kt • Historical NHC track, intensity and radii-CLIPER error distributions – Includes forecast interval time continuity and bias corrections • Run in real time on NCEP IBM during 2004 • Results displayed on password-protected CIRA web site – Atlantic, east, central and western N. Pacific sectors
Sample 34 kt Wind Probabilities
2005 Monte Carlo Model • Move web page to TPC w N-AWIPS graphics • Add t=0 hour probabilities • Include radii adjustment – Convert max in quadrant to average in quadrant • Ratios based upon H*Wind analyses • • Provide TPC with distribution calculation code Text product under development Training being developed Verification system still needed – Verification system could be used for all TC probabilistic forecasts (ensemble based, etc)
Are Intensity Forecasts Improving? • 20 Year Atlantic sample (1985 -2004) • Verification with consistent set of rules – All cases except extra tropical – Official, Persistence, SHIFOR, SHIPS and GFDL • Consider only 48 hour forecasts
48 Hour Intensity Errors 1985 -2004
48 Hour Intensity Forecast Errors Normalized by Persistence Errors
Summary • GRIP Model to be tested in real time during 2005 season – 2004 results are encouraging • Last chance for neural network model • Monte Carlo probability model development continuing in 2005 • Intensity forecasts are improving Ref: Further improvements to SHIPS, Weather and Forecasting, in press.
Скачать презентацию Improvements in Deterministic and Probabilistic Tropical Cyclone Wind
4c06a80ee6c61edf22cb23fae0fc6ec5.ppt