This paper aims to propose a novel large-signal order reduction (LSOR) approach for microgrids (MG) by embedding a stability and accuracy assessment theorem. Different from the existing order reduction methods, the proposed approach prevails mainly in two aspects. Firstly, the dynamic stability of full-order MG models can be assessed by only leveraging their derived reduced-order models and boundary layer models with our method. Specially, when the reduced-order system is input-to-state stable and the boundary layer system is uniformly globally asymptotically stable, the original MGs system can be proved to be stable under several common growth conditions. Secondly, a set of accuracy assessment criterion is developed and embedded into a tailored feedback mechanism to guarantee the accuracy of derived reduced model. It is proved that the errors between solutions of reduced and original models are bounded and convergent under such conditions. Strict mathematical proof for the proposed stability and accuracy assessment theorem is provided. The proposed LSOR method is generic and can be applied to arbitrary dynamic systems. Multiple case studies are conducted on MG systems to show the effectiveness of proposed approach.