Abstract: Due to lack of valid observational data over the tropical oceans, TC (tropical c yclone) initialization for numerical models is one of major difficulties in accurate TC track forecasting. Since TC numerical prediction model at Natio nal Meteorological Center of China (NMC/China) was put into operation in 1996, it has always been applied an empirical and statistical method which imposes a scale-selective sophisticated filtering to remove shallow vortex from the initial field and inserts bogus vortex into it, constructed by a few observed parameters-such as central pressure, position, and maximum sustained wind. In fact, TC str ucture varies from one form to another for different TCs and even the same TC in different time of life, which is not directly related to these observed paramet ers. The spatial and temporal variability of TC structure are not represented by bogus vortex with a few observed parameters. Recently, a new TC initialization scheme based on NMC global analysisforecast cycle has been developed successfu lly. The new scheme is composed of three procedures: vortex formation, vortex relocation, and vortex modification. Specifically, an appropriate vortex (generated by global model forecast) is inserted into background field when TC occurs at first time (it is called vortex formation). In the next forecast time, there must be a little weak and misplaced vortex in the background (6-h forecast at first time), what all we need to do is to move the vortex to correct position and mo dify it close to observed parameters by relative dynamic formula (it is called v ortex relocation and modification). It can be seen that initial vortex of the ne w scheme is mostly formed with the integration of the global model and is compat ible with the dynamics and physics of the global model. Using the new scheme of TC initialization, a series of experimental forecasts were performed for TC Kaemi (0605) that occurred in the western North Pacific in 2006. From study it is fo und that the new scheme can produce the reasonable and realistic vortex in the initial field and produced much more accurate TC track prediction compared with operational bogus one. It indicates that mean track forecast biases can be reduce d effectively by new TC initialization scheme.