As a portion of the model deficiencies are due to physics processes in numerical weather prediction (NWP), perturbing the parameterization schemes of the physics processes will lead to improve ensemble forecasting (EF). In this paper, a method for producing appropriate ensemble members from a series ensemble of combinations of physics process parameterization schemes is proposed and examined. The stepwise similarity analysis (SSA) procedure aims to reduce the weighting of similar cluster populations to improve the EF performance. Fourteen tropical cyclones (TCs) were selected in the Western North Pacific (WNP) basin for 2000-2014, and the ensembles were examined out to 5 days.
It was observed that the ensemble of 13 members, to which the screening method was applied, has outstanding advantages in terms of either accuracy or consistency compared with ensembles consisting of other members at 24, 72 and 120 h. The verification of the performance of the optimized ensemble forecasting system (OEFS) shows that the average absolute track error of the OEFS mean at 120 h is reduced by 13.5% compared with the control forecasts and by 15.9% compared with the ensemble with initially 27 members. At long forecast period, forecasts show ahead-and on the right side-of-track biases and underdispersion in the WNP basin, whereas the OEFS develops both a reduced bias and a larger dispersion to better reflect the typhoon trajectories. A variety of physical parameters and a larger combinations size need to be examined for further accurate conclusions.