Virtual power plant (VPP) serves as an effective solution for maintaining internal power balance and participating in external peak shaving auxiliary services within grid-connected microgrid involved in multi-type flexible resources (FRs). However, with increasing prominence of the feature heterogeneity in response behaviors of diverse FRs and their coupling in peak shaving poses challenges in the accurate decomposition of VPP scheduling commands. This paper proposes a de-aggregation strategy, utilizing discrete choice model and feature matching methods, to dynamically sequence FRs responses while optimizing VPP's peak shaving capability. Initially, heterogeneous features are refined and modeled to characterize the response capability of multi-type FRs in meeting the scheduled demand of grid-connected microgrid (SDGM). Subsequently, a feature difference quantification model and matching priority criterion are formulated to describe the feature mapping relationship and guide dynamic decision-making process. On this basis, the multi-type FRs are co-scheduled in the considered VPP to form a dynamic response sequence achieving peak shaving objectives. Case studies based on real data from a region-connected microgrid demonstrate the proposed strategy's performance in improving return on investment by 6.1 %, reducing peak shaving deviation and power exchange with main grid by 70 % and 13.1 %, respectively, and effectively improve the ability of grid-connected microgrid to balance the power and participate in peaking auxiliary services.