Formulating Model Output Statistics Using Support Vector Regression

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 4 February 2014: 3:45 PM
Room C204 (The Georgia World Congress Center )
Andrew E. Mercer, Mississippi State Univ., Mississippi State, MS; and J. L. Dyer

Model output statistics (hereafter MOS) are commonly utilized by forecasters when preparing local forecasts for major cities across the United States. However, MOS is primarily based on multivariate linear regression techniques that are unique for each region being addressed. This linearity limitation is a likely contributor to error in MOS-based forecasts. As such, a non-linear support vector regression (SVR) MOS product for ten major eastern United States cities is presented for different WRF-simulated weather events. The SVR-MOS product provides forecasts of high and low temperature, as well as daily peak wind gust. A SVR is trained for each of the 10 cities individually, similar to the statistical models currently utilized in MOS products. Output from SVR-MOS is verified against operational MOS products from the NAM, which uses the same basic simulation core as the WRF. Results from this study will help test the viability of artificial intelligence techniques such as SVR for improving current MOS forecasts from operational models.