Walter C. Kolczynski, Jr. -- Ph.D. Thesis Defense

(Penn State, Department of Meteorology)

"Evaluation of Linear Variance Calibration for Use in Atmospheric Transport and Dispersion Forecasting"

What PhD Defense Homepage GR
When Jun 22, 2011
from 01:00 pm to 02:30 pm
Where 529 Walker Bldg.
Contact Name Walter Kolczynski
Contact email
Contact Phone 814-771-6501
Add event to calendar vCal
iCal
Meteorological conditions are an important component in many important decisions, and an accurate estimate of meteorological uncertainty is required to make proper risk assessments. Ensembles of numerical weather prediction models are often used to quantify the meteorological uncertainty, but ensembles are subject to sampling error and model deficiencies that can make the measure of uncertainty, ensemble variance, are not representative of the actual uncertainty, as quantified by the error variance. Thus, there is a need to calibrate ensemble variances to more accurately represent the actual uncertainty in the forecast. This dissertation examines and further develops the Linear Variance Calibration (LVC) method as a way to calibrate meteorological wind variances for use in atmospheric transport and dispersion predictions. LVC is investigated in three ways. First, an idealized stochastic model is developed and used to determine fundamental properties and identify the role of sampling error. Second, LVC is applied to real short-term (0-4 day) meteorological ensemble data and evaluated for a number of different conditions over a year-long period. Finally, calibrations are applied to meteorological forecasts and used for atmospheric transport and dispersion forecasts to assess the impact of calibration on these forecasts. The stochastic model shows that calibration is necessary for even “perfect” ensembles due to the sampling error created by using a finite-sized ensemble. The stochastic model also provides insight on the behavior of LVC parameters (slope and intercept). Application of LVC to real meteorological ensemble data from the NCEP Short-Range Ensemble Forecast shows that it performs well for most forecasts of 10-m above ground level wind. Calibrations calculated from real data show sensitivity to forecast led time, season, time of day, and surface type (land vs. water). Results also show that calibrations calculated over a small geographic area may be suspect due to increased spatial correlation that lowers the effective sample size. Surface concentration and dosage forecasts that use calibrated meteorological wind variances from a meteorological ensemble generally have improved reliability and CRPS over forecasts using uncalibrated meteorological wind variances.