Federated learning faces a key challenge of performance degradation in heterogeneous data environments. In order to resolve this issue, We introduce an innovative personalized federated learning framework leveraging Particle Swarm Optimization (PSO), referred to as pFLPSO. Our approach dynamically adjusts the fusion ratio between global and local models using PSO, combined with element-wise weight allocation, enabling better adaptation to data distribution associated with each client. To assess the performance of pFLPSO, we carried out comprehensive evaluations across five widely-used benchmark datasets spanning both computer vision (CV) and natural language processing (NLP) tasks. Experimental findings indicate that pFLPSO achieves improvements of up to 2.85% over eight advanced FL approaches across diverse levels of data heterogeneity and different numbers of participating clients. Notably, the advantages of pFLPSO are more pronounced on larger datasets and more complex models.