Global Ensemble Predictions of 2009’s Tropical Cyclones Initialized with an Ensemble Kalman Filter

Abstract

Verification was performed on ensemble forecasts of 2009 Northern Hemisphere summer tropical cyclones (TCs) from two experimental global numerical weather prediction ensemble prediction systems (EPSs). The first model was a high-resolution version (T382L64) of the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS). The second model was a 30-km version of the experimental NOAA/ Earth System Research Laboratory's Flow-following finite-volume Icosahedral Model (FIM). Both models were initialized with the first 20 members of a 60-member ensemble Kalman filter (EnKF) using the T382L64 GFS. The GFS-EnKF assimilated the full observational data stream that was normally assimilated into the NCEP operational Global Statistical Interpolation (GSI) data assimilation, plus human-synthesized ''observations'' of tropical cyclone central pressure and position produced at the National Hurricane Center and the Joint Typhoon Warning Center. The forecasts from the two experimental ensembles were compared against four operational EPSs from the European Centre for Medium-Range Weather Forecasts (ECMWF), NCEP, the Canadian Meteorological Centre (CMC), and the Met Office (UKMO).The errors of GFS-EnKF ensemble track forecasts were competitive with those from the ECMWF ensemble system, and the overall spread of the ensemble tracks was consistent in magnitude with the track error. Both experimental EPSs had much lower errors than the operational NCEP, UKMO, and CMC EPSs, but the FIM-EnKF tracks were somewhat less accurate than the GFS-EnKF. The ensemble forecasts were often stretched in particular directions, and not necessarily along or across track. The better-performing EPSs provided useful information on potential track error anisotropy. While the GFS-EnKF initialized relatively deep vortices by assimilating the TC central pressure estimate, the model storms filled during the subsequent 24 h. Other forecast models also systematically underestimated TC intensity (e.g., maximum forecast surface wind speed). The higher-resolution models generally had less bias.Analyses were conducted to try to understand whether the additional central pressure observation, the EnKF, or the extra resolution was most responsible for the decrease in track error of the experimental Global Ensemble Forecast System (GEFS)-EnKF over the operational NCEP. The assimilation of the additional TC observations produced only a small change in deterministic track forecasts initialized with the GSI. The T382L64 GFS-EnKF ensemble was used to initialize a T126L28 ensemble forecast to facilitate a comparison with the operational NCEP system. The T126L28 GFS-EnKF EPS track forecasts were dramatically better than the NCEP operational, suggesting the positive impact of the EnKF, perhaps through improved steering flow.

Impact of perturbation methods in the ECMWF ensemble prediction system on tropical cyclone forecasts

“…The central pressure of the TCs is systematically weaker in both analyses and forecasts than in the best-track data. This is consistent with the results of Hamill et al (2010), who concluded that global models with their relatively coarse resolution are still not able to provide realistic TC intensity forecasts.…”

“…We follow the approach of Hamill et al (2010) to calculate the mean spread of the experiments. We only use a forecast if it is possible to identify the TC at least in 40% of the ensemble members at the respective forecast time.…”

“…Hence the number of forecasts used to calculate the ensemble spread of the experiments varies over time. The spread of the ensemble for a single TC ensemble forecast at a given forecast time is defined as (Hamill et al, 2010)…”

“…To test for statistical significance, we calculate 95% confidence intervals for the hypothesis that the mean spread values of two experiments differ from each other. To compute the confidence intervals, we use the paired block bootstrap algorithm described by Hamill (1999) and Hamill et al (2010). The different ensemble forecasts for one TC are not treated as independent within the computation but rather shuffled as a block to account for error correlations between consecutive initialization dates.…”

“…The differences between the two means are not statistically significant within 5 days forecast time. To measure how well the ALL50 ensemble is able to describe the anisotropy in the position uncertainty of the TCs, we follow the approach of Hamill et al (2010). We calculate the average…”