To identify physical parameters of a large structural system, the computational challenges in dealing with a large number of unknowns are formidable. A divide-and-conquer approach is often required to partition the structural system into many substructures, each with much lesser unknowns for more accurate and efficient identification. Furthermore, in view of the ill-conditioned nature of inverse analysis, it is highly beneficial to adopt nongradient-based search methods such as genetic algorithm (GA). To this end, this paper presents a GA-based substructural identification strategy for large structural systems. As compared with some recent work on substructural identification, the proposed strategy presents two significant improvements: (i) the use of acceleration measurement to directly account for interaction between substructures without approximation of interface force; and (ii) the use of an improved identification method based on multi-feature GA. In numerical simulations, the mass, damping, and stiffness parameters of a 100-storey shear building, involving 202 unknowns, are identified with very good accuracy (mean error of less than 3%) based on incomplete acceleration measurements with 10% noise. In addition, an experimental study on a 10-storey small-scale steel frame further validates the superior performance of the proposed strategy over complete structural identification.