Linear and nonlinear postprocessing of ensemble forecasts

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 19 January 2010: 2:15 PM
B204 (GWCC)
Ranran Wang, University of Washington, Seattle, WA; and C. Marzban

Presentation PDF (335.3 kB)

It is well-known that forecasts from Numerical Weather Prediction (NWP) models suffer from errors which can be corrected a posteriori. Model Output Statistics (MOS) is the classic example of such postprocessing. Here, a similar approach is developed for ensemble forecasts. Specifically, linear and nonlinear statistical models are developed to map ensemble outputs to observations. This is done for 90 stations across the US. The data set is a 20-year forecast dataset generated retrospectively (i.e., reforecast) by using a regional fine-scale 10-member ensemble forecasting system based on the Weather Research and Forecast model (WRF-ARW Version 3.0.1). A resampling scheme is employed to assess the prediction error, and to identify the onset of overfitting. It is found that even the simplest nonlinear models tend to overfit the data. As such, focus is placed on the linear models, and the spatial distribution of their performance is examined.