model description

since the method and model building procedures have been discussed in some detail previously (see publications: fraedrich and rueckert 1998, sievers et al. 2000, and fraedrich et al. 2003) a short outline will suffice. the self-adapting analog scheme used for the cyclone track predictions is an extension of the previously used analog methods (hurran) by adapting the state space metric to predict cyclone tracks in an error minimizing fashion. model building proceeds as follows. the first two steps are basic for analog forecasting:

1) introduction of an error measure and, after state space reconstruction,

2) model building, which adapts the metric weights and ensemble weights at minimum forecast error.

3) optimization of the procedure.


here, the so-called recursive Local Averaging Optimization (LAO) 1st introduced by McNames 2002 (see publications) is used in a modified way. the algorithm applies both metric weights for building the ensemble members and compounded weights for each considered ensemble member. instead of considering also past forecasts in the error measure, only the latest time window of observed phase space values is used here as the embedding dimension for searching the analogs. the method has been modified by Langmack, H., F. Sielmann, and K. Fraedrich, 2007 (see publications).

two different forecast models based on that scheme are applied here:

a) latitude effect, using the latitude as additional phase space member, or alternatively

b) coriolis effect, using the positional differences of coriolis and persistence prediction

in contrast to the optimized LAO forecast, an additional cluster analysis of the best 100 LAO ensemble members using 2 clusters is done. euclidean distances of all ensemble tracks are built by summing up at each lead time as base for the cluster analysis. the resulting forecast track for each cluster as the unweighted mean of all ensemble tracks belonging to that cluster is shown with the corresponding ellipses enveloping the 88 % confidence area at each lead time. a normal distribution of ensemble track distances is assumed.