Recently, Personalized Federated learning (PFL) has attracted much attention because it can effectively solve data heterogeneity and realize personalized models. Mainstream methods generate personalized models by weighted aggregation based on weights calculated from inter-client losses or parameter differences. However, these methods have been improved in performance, but there are still shortcomings in the processing of data heterogeneity and model generalization. In this paper, we propose a novel PFL method, pFLAWC. Firstly, pFLAWC utilizes Residual Attention Synergy Module(RASM) to deal with heterogeneity and better learn personalized feature information of clients. Secondly, the distance between the user’s personalized information and the global information was reflected by the guidance model. Finally, the proposed model uses Adaptive Top-K Aggregation mechanism to dynamically aggregate the model parameters collected from the client to obtain a personalized model. We conduct extensive experiments using three real-world datasets and show that pFLAWC outperforms 14 state-of-the-art methods in terms of effectiveness, scalability, and stability. Furthermore, pFLAWC alleviates the generalization problem and achieves up to 4.08% relative improvement in accuracy.