Kalman filtered analogs to improve Numerical Weather Predictions
KF is a linear, adaptive, recursive and optimal (in a least-square sense) method to estimate NWP biases. After a weather regime change, when the distribution of bias often changes, the KF requires few adjustment cycles before the new bias is accurately estimated. This drawback can be largely mitigated by rearranging past predictions based on their similarity to the current prediction (i.e., by searching for "analogs"). Least similar forecasts are placed first, and most similar are placed last in the KF sequence of cycles. The resulting series is then Kalman filtered. By Kalman filtering in analog space instead of time space the adjustment period present in the original KF method can be bypassed.
KFAN is tested on wind speed predictions of the fifth-generation Pennsylvania Sate University-NCAR Mesoscale Model (MM5), with data from 22 stations over New Mexico and a period of a year (starting 1 January 2005). KFAN largely improves the raw forecast skill, significantly outperforming also KF, persistence, and a simple 7-day bias correction.