Abstract: Locally refined elements in FEA enable the computational cost to be reduced with the accuracy ensured simultaneously. The multilevel refined method (MRM) is therefore proposed on the basis of the h method and existing element transition method, with its implementation procedure given. To alleviate the effect of local over-rigidity on the refined element boundary, an improving scheme is given. A new concept, refined scale, is described and how the computational accuracy varies with the scale is studied to establish the functional relationship between them. Thus, the computational accuracy of finite element model can be forecast and, what's more, a reference is provided for determining the element refining extent so as to balance the computational accuracy and cost. The method proposed is highly applicable to technological processes.