Kalman filter and analog schemes to postprocess numerical weather predictions

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Monday, 24 January 2011: 1:45 PM
Kalman filter and analog schemes to postprocess numerical weather predictions
615-617 (Washington State Convention Center)
Luca Delle Monache, NCAR, Boulder, CO; and T. Nipen, Y. Liu, G. Roux, R. B. Stull, T. T. Warner, and P. Childs

Two new postprocessing methods are proposed to reduce numerical weather predictions systematic and random errors. The first method (ANKF) consists in running a bias-correction algorithm based on the Kalman filter (KF) in analog space rather than in time. The analog of a forecast for a given location and time is defined as a past prediction that matches selected features of the current forecast. The second method is based on the analog concept (AN): it is the weighted average of the observations that verified when the 10 best analogs were issued.

ANKF and AN are tested with 10-m wind speed forecasts from the Weather Research and Forecasting (WRF) modeling system with a 24-hour lead time and initialized at 0200 UTC. Hourly predictions are compared to 500 surface wind observations for a 6-month period (May-October, 2009) and over a domain centered on western Colorado, USA. The analog-based methods (i.e., ANKF and AN) performance is compared to the skill of the raw forecast (Raw), a 7-day running mean bias correction (7-Day), and KF. Both AN and ANKF are able to predict drastic changes in forecast error (e.g., associated with rapid weather regime changes), a feature lacking in KF, 7-Day, and standard model output statistics algorithms. AN almost eliminates the bias of the raw prediction, while ANKF drastically reduces it with values slightly worse than KF. Both analog-based methods are also able to reduce random errors, therefore improving the predictive skill of Raw. AN is consistently the best, with average improvements of 10%, 20%, 25%, and 35% with respect to ANKF, KF, 7-Day, and Raw, as measured by centered-root-mean-square-error, and of 5%, 20%, 25%, and 40%, as measured by rank correlation.