The modelling of large dynamic systems very often leads to linearized models of the form X =AX+BU with the state-vector XE Rn, the input-vector U s RP and mat-rices. A and B of appropriate dimensions. However, the large-ness of a technical process often results in a high order n of the model . Therefore these models may be very difficult to employ for simulation or control design. To circumvent the drawback of high order, many authors suggest the application of order reduction techniques. The previous work in the area of order reduction suffers from the following main drawbacks:         1. Most of the methods rec.uires a priori knowledage of the be-haviour of each state variable, which is not always known; 2. The level of parcentage error associated with order reduction is not indicated; 3. The effect of the imaginary part of sigenvalues has been neglected. The present paper introduces a new method for simplification of linear dynamic systems. The level of percentage error is calculated and the effect of the imaginary part of the sigen-values is considered. The determination of significant and less-significant state va:iables does not require a priori knowledege of the nature of the state variables. The method is tested through its apE.ication to a dynamic model of a synchonous machine .