We develop a method to estimate and correct atmospheric model errors
that can be trained using operational 6 hour analysis increments (a
proxy for the forecast errors). We separate the errors into bias
(including annual and diurnal cycles), and state-dependent errors, that
are estimated using SVD of the covariance between state anomalies and
forecast errors. Tests with the primitive equation SPEEDY model show
that 1) the method works well; 2) when integrating the model with the
corrections derived from the 6 hr yields significantly better results
than a statistical correction performed a posteriori; 3) the state
dependent correction based on SVD are many orders of magnitude better
than Leith (1978) and DelSole and Hou (1999) linear regression
approach, and are twice as effective in improving the forecast. The
method can be easily implemented in an operational setup.
Chris Danforth, Eugenia
Kalnay and Takemasa Miyoshi