Statistical Post-Processing of Dynamical Tropical Cyclone Model Track Forecasts

Russell L. Elsberry Department of Meteorology, Naval Postgraduate School, Monterey, CA 93940

Search for other papers by Russell L. Elsberry in
Current site
Google Scholar
PubMed
Close
and
Dennis R. Frill Department of Meteorology, Naval Postgraduate School, Monterey, CA 93940

Search for other papers by Dennis R. Frill in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

The Fleet Numerical Oceanography Center Tropical Cyclone Model (TCM) exhibits a systematic bias in track forecasts. Based on a 1975–78 sample of western North Pacific tropical cyclones, the mean zonal and meridional 72 h errors are about 280 and 350 km, respectively. A series of statistical regression equations is derived to adjust the TCM forecasts. Predictors include zonal and meridional displacements and speeds over various forecast intervals to 72 h. The dynamical model is also integrated backward in time for 36 h to derive additional predictors. Inclusion of predictors derived from comparing the backward-integrated positions with the known warning positions at −12, −24 and −36 h resulted in smaller forecast errors than the case which involved only forward integration predictors. Use of the forward-backward statistical post-processing with an independent set of TCM forecasts resulted in an error reduction of ∼200 km at 72 h.

Abstract

The Fleet Numerical Oceanography Center Tropical Cyclone Model (TCM) exhibits a systematic bias in track forecasts. Based on a 1975–78 sample of western North Pacific tropical cyclones, the mean zonal and meridional 72 h errors are about 280 and 350 km, respectively. A series of statistical regression equations is derived to adjust the TCM forecasts. Predictors include zonal and meridional displacements and speeds over various forecast intervals to 72 h. The dynamical model is also integrated backward in time for 36 h to derive additional predictors. Inclusion of predictors derived from comparing the backward-integrated positions with the known warning positions at −12, −24 and −36 h resulted in smaller forecast errors than the case which involved only forward integration predictors. Use of the forward-backward statistical post-processing with an independent set of TCM forecasts resulted in an error reduction of ∼200 km at 72 h.

Save