The empirical Gramian balance reduction method is gradually applied to the nonlinear control and design of power system. How to obtain effective empirical controllability and observability Gramian matrices is the core of this method. In order to further improve the reduced effect, this paper forms nonlinear dynamic model of power system and analyzes the relationship of among research problem, the choice of control variables and reduction object. Given the variation characteristic of state and control variables, this paper proposes a scheme which is used to determine the disturbed values of state and control variables to form empirical controllability and observability Gramian matrices which contains rich system dynamic behaviors information. It takes advantage of some relative power system to verify the effectiveness of the proposed scheme. The simulation results show that this scheme can further decrease the reduced order and the effect of reduced model are improved greatly on the premise of guarantee the upper bound of error for order reduction. This work is supported by National Natural Science Foundation of China (No. 51077053).