Comparison Of Multi-model Ensemble Forecasts At Surface And 500hPa Based On The Kalman Filter Method

This paper applied Kalman filter (KF) to multi-model ensemble forecast, and made some improvements according to different forecast variables. Based on ensemble forecasts taken from the THORPEX Interactive Grand Global Ensemble project, we used the KF to construct multi-model forecasts to explore their performances in predicting a constant blocking event and related variables such as geopotential height, temperature and winds at 500 hPa, sea level pressure (SLP), as well as surface air temperature (T2m), during the summer of 2010. The widely used super-ensemble and linear equation based super-ensemble were used as comparison as well.For near-surface forecast, the root mean squared error (RMSE) of SLP and T2m of member models became lager with longer leads time. The multi-model forecasts showed a more skillful performance than any single model in near-surface prediction. The improvement was spatially and temporally consistent. In areas with low-resolution observations or complex topography (e.g. Tibet Plateau and Sahara Desert), the multi-models exhibited at least a 20% reduction in RMSE at short lead times. Multi-model forecast removed the systematic errors of member model to different extent. Say, the RMSE of KF was reduced more than 30% in comparison with that of ECMWF in 1-3 day forecast, but only about 20% existed in SEs.For mid-troposphere forecast, compared to the forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF, the best single model in this study), the multivariable forecasts of multi-model combination enhanced the forecast skill by at least 15% (measured as RMSE) at mid-troposphere for short to medium lead times. Similar to near-surface forecast, more than 20% reduction in RMSE was gained in areas with low-resolution observations or complex topography, bur no significant skill difference between the three multi-model techniques was found.Skill gap between forecasts at 500 hPa and near surface was analyzed. Near-surface forecasts are more dependent on physical parameterization, but upper-level forecast are more dynamically controlled. The physical parameterization made for a less consistent forecast among member models that lead to a large error cancellation in multi-model combination.Finally, we explored the tracks and intensities (central mean sea level pressure) of tropical cyclones (TCs) in the northwestern Pacific basin in 2010 and 2011. The KF reduced the TC mean absolute track forecast error by approximately 50,80 and 100 km in the 24-,48-and 72-h forecasts, respectively, compared with the best individual model (ECMWF). Also, the intensity forecasts were improved by the KF to some extent in terms of average intensity deviation (AID) and correlation coefficients with reanalysis intensity data. Overall, the Kalman Filter technique performed better compared to multi-models, the ensemble mean, and the super-ensemble in 3-day forecasts.