With the increasing penetration of high proportion of renewable energy into power grids, the frequency regulation pressure on power systems has intensified, making flexibility resources an important supplementary means for frequency regulation. However, the characteristics of wide distribution and heterogeneous response parameters of flexibility resources make it difficult for centralized control to adapt to their dynamic disparities, leading to low coordination efficiency. To address these challenges, this paper proposes a consensus modeling method for multi-type flexibility resources and a cooperative frequency regulation strategy with dynamically allocated weights. By extracting common parameters such as power and state of charge, a consensus model (CM) based on the first-order lag element is constructed to uniformly characterize the dynamic response characteristics of resources. On this basis, a collaborative architecture with “centralized training-decentralized execution” is designed. The upper layer coordinates global frequency deviation and frequency regulation cost optimization, and the lower layer proposes a differential reward-driven state interaction weight allocation mechanism, which dynamically screens the high-value interaction information by quantifying the contribution of agent actions to the frequency control, realizing the coordinated control with “higher contribution weights leading to stronger strategy influence”. Finally, the effectiveness of the proposed method is verified through a modified IEEE 39-bus system.