In today’s data-driven era, concerns about data privacy and security are increasingly prominent. With the tightening of data protection regulations, restrictions on data sharing have led traditional centralized machine learning to face the challenge of data silos. Federated Learning (FL), as an advanced framework, effectively addresses issues related to data silos and privacy. However, in real-world scenarios, the heterogeneity of data can lead to inefficient models, degraded performance, and stagnation in development. To address these challenges, we propose an enhanced FL approach called Federated Bat algorithm (FedBat). This method integrates the echolocation characteristics of bat algorithms to dynamically balance global and local searches, thereby optimizing the model weight update process. To tackle the decline in the generalization performance of the global model due to client model discrepancies, we extend FedBat with Jensen-Shannon (JS) divergence. Clients evaluate the differences between their local models and the global model to decide whether to upload their models. This approach not only enhances the generalization capability of the global model but also effectively reduces communication costs. In our theoretical analysis, we derive the convergence rate formula for FedBat and explore the impact of various parameters on its performance. Extensive experimental results demonstrate that FedBat significantly improves model convergence speed and test accuracy in practical applications: compared to other methods, FedBat achieves up to a 5-fold increase in convergence speed, over a 40 \(\%\) improvement in accuracy, and approximately 20 \(\%\) reduction in communication overhead.