J2.1 Potentials of Using Neural Networks for Bias Corrections

Tuesday, 24 January 2017: 4:00 PM
Conference Center: Skagit 2 (Washington State Convention Center )
Vladimir M. Krasnopolsky, NOAA/NWS/NCEP, College Park, MD; and Y. Fan

Model Output Statistics (MOS) is a technique used to objectively interpret numerical model output and produce improved weather forecast guidance.  There are about 3,000 sites over ConUS and METAR, for which MOS produces guidance. Using MOS allows meteorologists objectively interpret numerical model forecast, quantify uncertainties, remove systematic bias, and improve forecast.  MOS also allows meteorologists produce a site specific forecast, that is, it is a kind of downscaling technique.

MOS relates observed weather elements (predictands) to appropriate model variables (predictors) via a statistical approach (multiple linear regressions).  The current MOS operational system consists of about 9 million linear regression equations.  The reasons for such a huge number of equations are: (1) MOS has separate equations for different sites, (2) MOS has different equations for different variables (single site/single variable setup), (3) MOS has different equations for different projection times, and (4) MOS has different equations for different weather regimes.  Most of these reasons are rooted in the fact that the functional relationship between predictors and predictands is nonlinear; however, in MOS each of these nonlinear relationships is approximated by many linear regression equations.

In this pilot study we investigate if the MOS system can benefit from using nonlinear statistical techniques like neural networks (NN) instead of linear regressions.  We have found that if NN is used in single site/single variable setup (like linear regression is used), it does not give significant improvement upon the linear regression.  We showed that NN can be used in more beneficial setups: (1) a single site/several variables, (2) several sites/several variables, (3) and even all 3,000 ConUS sites/several variables.  In all these cases NN significantly reduces the number of equations and in cases (2) and (3) significantly improves the accuracy of forecast.  In the third case, NN trained on a discrete number of sites actually provides a gridded field over entire ConUS.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner