Using Bayesian Statistical Postprocessing Methods To Improve Local Wind Forecasts
Abstract
This thesis explores the use of Bayesian statistical postprocessing to rapidly train a highly accurate forecast from a 1 km resolution gridded WRF model forecast over a 100 km by 100 km area. These methods leverage three modeled forecast variables--10 m winds, sea-level pressure, and terrain elevation--in conjunction with downstream observations and prior model runs to identify model inaccuracies. Using only three days of data, a Bayesian corrected forecast is produced and analyzed for accuracy and improvement over the original model run relative to real-world observations. Over 90 of the resulting forecasts saw improvement over the raw model forecasts in root mean squared error, and over 87 of the forecasts saw improvement in mean error over the raw model forecasts. Extreme circumstances saw improvements in accuracy of over 9 knots while overall improvements were reliably seen both in accuracy and precision among Bayesian corrected forecasts. These findings are significant as they suggest that Bayesian statistical postprocessing methods work and should be both employable at rapid rates, and result in more accurate forecasts.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2021
- Accession Number
- AD1150687
Entities
People
- Darby J. Maier
Organizations
- Naval Postgraduate School